docs: restore the Acknowledgements section in the README

- Add docs/shannon-xbow-aikido-benchmark.md with methodology, per-model
  cost/coverage tables, and links to each model's report and SARIF
- Link the writeup from the README "Shannon in Action" section

.dockerignore

@@ -18,6 +18,7 @@ xben-benchmark-results/
 # Development files
 *.md
 !CLAUDE.md
+!THIRD_PARTY_NOTICES.md
 .DS_Store
 Thumbs.db
 
@@ -69,4 +70,5 @@ coverage/
 docs/
 README.md
 LICENSE
+!LICENSE
 CHANGELOG.md

.env.example

@@ -1,66 +1,60 @@
-# Shannon Environment Configuration
-# Copy this file to .env and fill in your credentials
+# Copy to .env and uncomment one provider block.
+# SHANNON_AI_MODEL is <provider>:<model-id>, split on the first colon.
+# Defaults to anthropic:claude-sonnet-4-6.
 
-# Recommended output token configuration for larger tool outputs
-CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000
-
-# Adaptive thinking is enabled automatically on Opus 4.6/4.7. Set to false to disable.
-# CLAUDE_ADAPTIVE_THINKING=false
-
-# Shannon forwards your machine's /etc/hosts entries into the worker container. Set to false to disable.
-# SHANNON_FORWARD_HOSTS=false
-
-# =============================================================================
-# OPTION 1: Direct Anthropic
-# =============================================================================
-ANTHROPIC_API_KEY=your-api-key-here
-
-# OR use OAuth token instead
+# --- Anthropic ---------------------------------------------------------------
+SHANNON_AI_API_KEY=your-api-key-here
+SHANNON_AI_MODEL=anthropic:claude-sonnet-4-6
 # CLAUDE_CODE_OAUTH_TOKEN=your-oauth-token-here
 
-# =============================================================================
-# OPTION 2: Custom Base URL (compatible proxies, gateways, etc.)
-# =============================================================================
-# Point the SDK at an alternative Anthropic-compatible endpoint.
-# ANTHROPIC_BASE_URL=https://your-proxy.example.com
-# ANTHROPIC_AUTH_TOKEN=your-auth-token   # Auth token for the custom endpoint
+# --- OpenAI ------------------------------------------------------------------
+# SHANNON_AI_API_KEY=your-api-key-here
+# SHANNON_AI_MODEL=openai:gpt-5.5
 
-# =============================================================================
-# Model Tier Overrides (Anthropic API / OAuth / Custom Base URL / Bedrock)
-# =============================================================================
-# Override which model is used for each tier. Defaults are used if not set.
-# Optional for direct Anthropic and custom base URL modes. Required for Bedrock/Vertex.
-# ANTHROPIC_SMALL_MODEL=...     # Small tier  (default: claude-haiku-4-5-20251001)
-# ANTHROPIC_MEDIUM_MODEL=...    # Medium tier (default: claude-sonnet-4-6)
-# ANTHROPIC_LARGE_MODEL=...     # Large tier  (default: claude-opus-4-7)
+# --- xAI ---------------------------------------------------------------------
+# SHANNON_AI_API_KEY=your-api-key-here
+# SHANNON_AI_MODEL=xai:grok-4.5
 
-# =============================================================================
-# OPTION 3: AWS Bedrock
-# =============================================================================
-# https://aws.amazon.com/blogs/machine-learning/accelerate-ai-development-with-amazon-bedrock-api-keys/
-# Requires the model tier overrides above to be set with Bedrock-specific model IDs.
-# Example Bedrock model IDs for us-east-1:
-#   ANTHROPIC_SMALL_MODEL=us.anthropic.claude-haiku-4-5-20251001-v1:0
-#   ANTHROPIC_MEDIUM_MODEL=us.anthropic.claude-sonnet-4-6
-#   ANTHROPIC_LARGE_MODEL=us.anthropic.claude-opus-4-7
-
-# CLAUDE_CODE_USE_BEDROCK=1
+# --- AWS Bedrock -------------------------------------------------------------
+# Bearer token only; model must be enabled in your region.
 # AWS_REGION=us-east-1
 # AWS_BEARER_TOKEN_BEDROCK=your-bearer-token
+# SHANNON_AI_MODEL=amazon-bedrock:us.anthropic.claude-opus-4-8
 
-# =============================================================================
-# OPTION 4: Google Vertex AI
-# =============================================================================
-# https://cloud.google.com/vertex-ai/generative-ai/docs/partner-models/use-partner-models
-# Requires a GCP service account with roles/aiplatform.user.
-# Download the SA key JSON from GCP Console (IAM > Service Accounts > Keys).
-# Requires the model tier overrides above to be set with Vertex AI model IDs.
-# Example Vertex AI model IDs:
-#   ANTHROPIC_SMALL_MODEL=claude-haiku-4-5@20251001
-#   ANTHROPIC_MEDIUM_MODEL=claude-sonnet-4-6
-#   ANTHROPIC_LARGE_MODEL=claude-opus-4-7
+# --- Custom Base URL ---------------------------------------------------------
+# Route through a proxy or gateway (LiteLLM, an internal endpoint).
+# Pick the block matching the API dialect your gateway speaks, and uncomment all
+# three lines. The provider prefix picks the dialect; the model id is whatever
+# name your gateway serves it under.
 
-# CLAUDE_CODE_USE_VERTEX=1
-# CLOUD_ML_REGION=us-east5
-# ANTHROPIC_VERTEX_PROJECT_ID=your-gcp-project-id
-# GOOGLE_APPLICATION_CREDENTIALS=./credentials/google-sa-key.json
+# Anthropic compatible - Anthropic Messages:
+# SHANNON_AI_API_KEY=your-gateway-key-here
+# SHANNON_AI_BASE_URL=https://llm-gateway.example.com
+# SHANNON_AI_MODEL=anthropic:claude-sonnet-4-6
+
+# OpenAI compatible - Chat Completions (default) or Responses:
+# SHANNON_AI_API_KEY=your-gateway-key-here
+# SHANNON_AI_BASE_URL=https://llm-gateway.example.com/v1
+# SHANNON_AI_MODEL=openai:gpt-5.5
+# SHANNON_AI_OPENAI_FORMAT=responses
+
+# --- Other provider ----------------------------------------------------------
+# Any other provider the Pi harness supports. Name it in SHANNON_AI_MODEL and
+# supply the key via the generic SHANNON_AI_API_KEY. Pi validates the provider
+# and model at preflight.
+# SHANNON_AI_MODEL=openrouter:moonshotai/kimi-k3
+# SHANNON_AI_API_KEY=your-api-key-here
+
+# --- Misc --------------------------------------------------------------------
+# Forward /etc/hosts entries into the worker container.
+# SHANNON_FORWARD_HOSTS=false
+
+# See the guide below to use an OpenAI subscription
+# https://github.com/KeygraphHQ/shannon/blob/main/docs/ai-providers.md#openai-codex-chatgpt-pluspro-subscription
+# SHANNON_USE_PI_AUTH=1
+# SHANNON_AI_MODEL=openai-codex:gpt-5.5
+
+# Or the guide below to use an xAI subscription
+# https://github.com/KeygraphHQ/shannon/blob/main/docs/ai-providers.md#xai-grok-subscription
+# SHANNON_USE_PI_AUTH=1
+# SHANNON_AI_MODEL=xai:grok-4.6

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -69,9 +69,9 @@ body:
 
         Issues without this information may be difficult to triage.
 
-        - Check the workflow log:
-            - **npx mode:** `~/.shannon/workspaces/<workspace>/workflow.log`
-            - **Local mode:** `./workspaces/<workspace>/workflow.log`
+        - Check the scan log:
+            - **npx mode:** `~/.shannon/workspaces/<workspace>/.shannon/workflow.log`
+            - **Local mode:** `./workspaces/<workspace>/.shannon/workflow.log`
           Use `grep` or search to identify errors.
           Paste the relevant error output below.
         - Temporal:
@@ -117,9 +117,12 @@ body:
       options:
         - "Anthropic (API key)"
         - "Anthropic (OAuth token)"
-        - "Custom base URL (proxy/gateway)"
+        - "OpenAI"
+        - "xAI"
         - "AWS Bedrock"
-        - "Google Vertex AI"
+        - "Custom base URL - Anthropic Messages"
+        - "Custom base URL - OpenAI Chat Completions"
+        - "Custom base URL - OpenAI Responses"
     validations:
       required: true
 

.github/workflows/release-beta.yml

@@ -30,15 +30,17 @@ jobs:
         run: |
           set -euo pipefail
 
+          BASE="3.0.0"
           LATEST=$(npm view "@keygraph/shannon" dist-tags.beta 2>/dev/null || echo "")
 
-          if [[ -z "$LATEST" ]]; then
-            echo "version=1.0.0-beta.1" >> "$GITHUB_OUTPUT"
-          else
-            # Extract N from 1.0.0-beta.N and increment
+          if [[ "$LATEST" == "$BASE-beta."* ]]; then
+            # Same base version — increment the beta counter (e.g. 3.0.0-beta.2 -> 3.0.0-beta.3)
             N=$(echo "$LATEST" | grep -oE 'beta\.([0-9]+)' | grep -oE '[0-9]+')
             NEXT=$((N + 1))
-            echo "version=1.0.0-beta.$NEXT" >> "$GITHUB_OUTPUT"
+            echo "version=$BASE-beta.$NEXT" >> "$GITHUB_OUTPUT"
+          else
+            # No prior beta, or a different base (e.g. last beta was 2.0.0-beta.N) — start over.
+            echo "version=$BASE-beta.1" >> "$GITHUB_OUTPUT"
           fi
 
       - name: Print version
@@ -187,8 +189,6 @@ jobs:
 
       - name: Publish npm package
         working-directory: apps/cli
-        env:
-          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
         run: |
           if npm view "@keygraph/shannon@${{ needs.preflight.outputs.version }}" version 2>/dev/null; then
             echo "Version already published, skipping"

.github/workflows/release.yml

@@ -201,8 +201,6 @@ jobs:
 
       - name: Publish npm package
         working-directory: apps/cli
-        env:
-          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
         run: |
           if npm view "@keygraph/shannon@${{ needs.preflight.outputs.version }}" version 2>/dev/null; then
             echo "Version already published, skipping"

.github/workflows/rollback-beta.yml

@@ -4,7 +4,7 @@ on:
   workflow_dispatch:
     inputs:
       version:
-        description: "Beta version to roll back to (example: 1.0.0-beta.2)"
+        description: "Beta version to roll back to (example: 3.0.0-beta.2)"
         required: true
         type: string
 
@@ -31,7 +31,7 @@ jobs:
           VERSION="${RAW_VERSION#v}"
 
           if ! [[ "$VERSION" =~ ^[0-9]+\.[0-9]+\.[0-9]+-beta\.[0-9]+$ ]]; then
-            echo "Version must be in format X.Y.Z-beta.N (e.g. 1.0.0-beta.2)"
+            echo "Version must be in format X.Y.Z-beta.N (e.g. 3.0.0-beta.2)"
             exit 1
           fi
 

CLAUDE.md

@@ -4,7 +4,7 @@ AI-powered penetration testing agent for defensive security analysis. Automates
 
 ## Commands
 
-**Prerequisites:** Docker, AI provider credentials (`.env` for local, `shn setup` or env vars for npx)
+**Prerequisites:** Docker, AI provider credentials (`.env` for local, `npx @keygraph/shannon setup` or env vars for npx)
 
 ### Dual CLI
 
@@ -15,8 +15,8 @@ Shannon supports two CLI modes, auto-detected based on the current working direc
 | **Install** | Zero-install via npm | Clone the repo |
 | **Image** | Pulled from Docker Hub (`keygraph/shannon:latest`) | Built locally (`shannon-worker`) |
 | **State** | `~/.shannon/` | Project directory |
-| **Credentials** | `~/.shannon/config.toml` (via `shn setup`) or env vars | `./.env` |
-| **Config** | `~/.shannon/config.toml` (via `shn setup`) | N/A |
+| **Credentials** | `~/.shannon/config.toml` (via `npx @keygraph/shannon setup`) or env vars | `./.env` |
+| **Config** | `~/.shannon/config.toml` (via `npx @keygraph/shannon setup`) | N/A |
 | **Prompts** | Bundled in Docker image | Mounted from `./apps/worker/prompts/` (live-editable) |
 
 Mode auto-detection: local mode activates when env var `SHANNON_LOCAL=1` is set by the `./shannon` entry point (`apps/cli/src/mode.ts`). Otherwise npx mode.
@@ -44,8 +44,8 @@ echo "ANTHROPIC_API_KEY=your-key" > .env
 ./shannon build
 
 # Run
-./shannon start -u <url> -r my-repo
-./shannon start -u <url> -r my-repo -c ./apps/worker/configs/my-config.yaml
+./shannon start -u <url> -r ./my-repo
+./shannon start -u <url> -r ./my-repo -c ./apps/worker/configs/my-config.yaml
 ./shannon start -u <url> -r /any/path/to/repo
 ```
 
@@ -56,22 +56,27 @@ echo "ANTHROPIC_API_KEY=your-key" > .env
 npx @keygraph/shannon setup
 
 # Workspaces & Resume
-./shannon start -u <url> -r my-repo -w my-audit    # New named workspace
-./shannon start -u <url> -r my-repo -w my-audit    # Resume (same command)
-./shannon workspaces                                 # List all workspaces
+./shannon start -u <url> -r ./my-repo -w my-audit    # New named workspace
+./shannon start -u <url> -r ./my-repo -w my-audit    # Resume (same command)
 
 # Monitor
-./shannon logs <workspace>            # Tail workflow log
-./shannon status                      # Show running workers
-# Temporal Web UI: http://localhost:8233
+./shannon scans                       # List running and completed scans, with each report's path
+./shannon logs [<workspace>]          # Show a scan's live log (default: the single running scan, else the most recent)
+./shannon logs [<workspace>] --agent <name>   # Tail one agent's own log (from .shannon/agents/)
+./shannon logs [<workspace>] --list-agents    # List the agents that have their own log
+./shannon status [<workspace>]        # Live phase/agent progress of one scan, read from Temporal (redraws, then exits; same default target)
+# Dashboard: http://localhost:8233
 
 # Stop
-./shannon stop                        # Preserves workflow data
-./shannon stop --clean                # Full cleanup including volumes (confirms first)
+./shannon stop [<workspace>]          # Stop one scan (default: the single running scan; confirms first; --yes/-y to skip)
+./shannon stop --all                  # Stop all running scans (Temporal stays up; confirms first)
+./shannon reset                       # Stop everything and wipe all Temporal data + volumes (type 'confirm' to proceed; cannot be skipped)
+
+# Version
+./shannon version                     # npx: package version; local: git SHA
 
 # Image management
 ./shannon build [--no-cache]          # Local mode: build worker image
-npx @keygraph/shannon uninstall             # npx mode: remove ~/.shannon/ (confirms first)
 
 # Build TypeScript (development)
 pnpm run build                       # Build all packages via Turborepo
@@ -82,7 +87,7 @@ pnpm biome:fix                       # Auto-fix lint, format, and import sorting
 
 **Monorepo tooling:** pnpm workspaces, Turborepo for task orchestration, Biome for linting/formatting. TypeScript compiler options shared via `tsconfig.base.json` at the root. All packages extend it, overriding only `rootDir` and `outDir`. Shared devDependencies (`typescript`, `@types/node`, `turbo`, `@biomejs/biome`) are hoisted to the root workspace.
 
-**Options:** `-c <file>` (YAML config), `-o <path>` (output directory), `-w <name>` (named workspace; auto-resumes if exists), `--pipeline-testing` (minimal prompts, 10s retries), `--debug` (preserve worker container after exit for log inspection)
+**Options:** `-c <file>` (YAML config), `-o <path>` (output directory), `-w <name>` (named workspace; auto-resumes if exists), `--pipeline-testing` (minimal prompts, 10s retries), `--keep-container` (preserve worker container after exit for log inspection), `--yes`/`-y` (skip the confirmation prompt on `stop`; required for non-interactive use; `reset` requires a typed `confirm` and cannot be skipped)
 
 ## Architecture
 
@@ -94,24 +99,29 @@ apps/worker/     — @shannon/worker (private, Temporal worker + pipeline logic)
 ```
 
 ### CLI Package (`apps/cli/`)
-Published as `@keygraph/shannon` on npm. Contains only Docker orchestration logic — no Temporal SDK, business logic, or prompts. Bundled with tsdown for single-file ESM output.
+Published as `@keygraph/shannon` on npm. Contains Docker orchestration and a direct `@temporalio/client` integration for read-only status plus bounded workflow lifecycle operations; no worker/pipeline business logic or prompts. Bundled with tsdown for single-file ESM output (deps stay external).
 
-- `apps/cli/src/index.ts` — CLI dispatcher (`setup`, `start`, `stop`, `logs`, `workspaces`, `status`, `build`, `uninstall`, `info`)
+- `apps/cli/src/index.ts` — CLI dispatcher (`setup`, `start`, `stop`, `reset`, `logs`, `status`, `scans`, `build`, `version`)
+- `apps/cli/src/temporal-client.ts` — `@temporalio/client` integration: connects to the frontend on `127.0.0.1:7233` (published by compose), provides `describeScan` (status + `pendingActivities` → running agents), `queryProgress` (live `getProgress` query → `PipelineState`), `getTerminalOutcome` (workflow `result()`), and bounded lifecycle RPCs for `stop`. `stop` requests cancellation first, waits up to 10 seconds, then requests termination only when necessary and verifies closure within a bounded window. No worker of its own; scans are visible within Temporal's retention window, which `ensureInfra` (`apps/cli/src/docker.ts`) converges to `168h` (7 days) on every successful `shannon start`; override with `SHANNON_TEMPORAL_RETENTION` (a positive whole-hour value like `72h`)
+- `apps/cli/src/scan/` — `status` rendering: `pipeline.ts` (static phase/agent plan + `run*Agent` activity-type→agent map + mirrored `PipelineState`/`AgentMetrics` types; keep in sync with the worker), `derive.ts` (pure phase/agent state derivation shared by the tree and `--json`), `render.ts` (one renderer for both the live query state and the terminal result). The tree shows model work only: every row is an agent, an Agentic SAST stage, or a report step that is currently running or failed. Reconciliation is model work owned by a class, so its wall time renders as a trailing `+ duration` on that class's exploitation row (its analysis row when `exploit: false`) rather than as a row of its own; deterministic bookkeeping stages (`report:*` renumber/assemble/finalize/surface) never appear once they complete. `DerivedPhase.children` (renders sub-rows) and `DerivedPhase.meta` (`duration` vs a `k/N done` tally) are independent — Agentic SAST lists stages under a duration, exploitation lists classes under a tally
 - `apps/cli/src/mode.ts` — Auto-detection: local mode if `SHANNON_LOCAL=1` env var is set
-- `apps/cli/src/docker.ts` — Compose lifecycle, image pull/build, ephemeral `docker run` worker spawning
+- `apps/cli/src/docker.ts` — Compose lifecycle, image pull/build, and ephemeral `docker run` worker spawning. Each worker carries workspace, task-queue, and preselected workflow-ID labels so stop can correlate the local worker with its Temporal execution before `session.json` exists. Before `docker run`, start fsyncs that exact candidate under the workspace's hidden internals and clears it only when `session.json` registers the same ID; stop reconciles any candidate left by an interrupted launch. Start also checks the image's workflow-ID protocol label and refuses a stale worker that would ignore the preselected ID
 - `apps/cli/src/home.ts` — State directory management (`~/.shannon/` for npx, `./` for local)
-- `apps/cli/src/env.ts` — `.env` loading, TOML fallback (npx only) via `apps/cli/src/config/resolver.ts`, credential validation, env flag building
+- `apps/cli/src/env.ts` — `.env` loading, TOML fallback (npx only) via `apps/cli/src/config/resolver.ts`, credential validation, provider-scoped env flag building
+- `apps/cli/src/model-spec.ts` — `SHANNON_AI_MODEL` (`<provider>:<model-id>`) parsing; mirrors `apps/worker/src/ai/models.ts`
 - `apps/cli/src/config/resolver.ts` — Cascading config (npx only): env vars → `~/.shannon/config.toml` (parsed with `smol-toml`)
 - `apps/cli/src/config/writer.ts` — TOML serialization and secure file persistence (0o600)
 - `apps/cli/src/commands/setup.ts` — Interactive TUI wizard (`@clack/prompts`) for provider credential setup (npx only)
-- `apps/cli/src/paths.ts` — Repo/config path resolution (bare name → `./repos/<name>`, or any absolute/relative path)
+- `apps/cli/src/paths.ts` — Repo/config path resolution (any absolute or relative path)
+- `apps/cli/src/version.ts` — Version reporting (npx: `package.json` version; local: `git-<sha>`)
+- `apps/cli/src/tty.ts` — Terminal capability detection: `requireInteractive` guard (fails fast off-TTY instead of hanging on a prompt), `supportsColor` color gating (`NO_COLOR`/`FORCE_COLOR`), and `stdoutIsTerminal` for spinner/cursor output
 - `apps/cli/src/commands/` — Command handlers
 - `apps/cli/infra/compose.yml` — Bundled Temporal compose file for npx mode
 - `apps/cli/tsdown.config.ts` — tsdown bundler config
 - `shannon` — Node.js entry point (`#!/usr/bin/env node`) that delegates to `apps/cli/dist/index.mjs`
 
 ### Docker Architecture
-Infra (Temporal) runs via `docker-compose.yml`. Workers are ephemeral `docker run --rm` containers, one per scan, each with a unique task queue and isolated volume mounts.
+Infra (Temporal) runs via `docker-compose.yml`. Workers are ephemeral `docker run --rm` containers, one per scan, each with a unique task queue, preselected workflow ID, matching identity labels, and isolated volume mounts. `shannon stop --all` takes the union of labeled running workers and Temporal-running workflows, so an orphaned workflow is still stopped after its worker has disappeared.
 
 - `docker-compose.yml` — Infra only: `shannon-temporal` (port 7233/8233). Network: `shannon-net`
 - `Dockerfile` — 2-stage build (builder + Chainguard Wolfi runtime). Uses pnpm. Entrypoint: `CMD ["node", "apps/worker/dist/temporal/worker.js"]`
@@ -122,7 +132,7 @@ Infra (Temporal) runs via `docker-compose.yml`. Workers are ephemeral `docker ru
 - `apps/worker/src/paths.ts` — Centralized path constants (`PROMPTS_DIR`, `CONFIGS_DIR`, `WORKSPACES_DIR`)
 - `apps/worker/src/session-manager.ts` — Agent definitions (`AGENTS` record). Agent types in `apps/worker/src/types/agents.ts`
 - `apps/worker/src/config-parser.ts` — YAML config parsing with JSON Schema validation
-- `apps/worker/src/ai/claude-executor.ts` — Claude Agent SDK integration with retry logic
+- `apps/worker/src/ai/pi/pi-executor.ts` — pi harness integration (agent-level retry disabled so Temporal owns restarts; provider-level retry on, see `apps/worker/src/ai/pi/retry-settings.ts`)
 - `apps/worker/src/services/` — Business logic layer (Temporal-agnostic). Activities delegate here. Key: `agent-execution.ts`, `error-handling.ts`, `container.ts`
 - `apps/worker/src/types/` — Consolidated types: `Result<T,E>`, `ErrorCode`, `AgentName`, `ActivityLogger`, etc.
 - `apps/worker/src/utils/` — Shared utilities (file I/O, formatting, concurrency)
@@ -144,13 +154,22 @@ Durable workflow orchestration with crash recovery, queryable progress, intellig
 4. **Exploitation** (5 parallel agents, conditional) — Exploits confirmed vulnerabilities
 5. **Reporting** (`report`) — Executive-level security report
 
+Around those phases:
+
+- Optional agentic static analysis runs before the pentest when `agentic_sast.enabled` is `"true"`, as a child workflow.
+- After each class's analysis, reconciliation groups its findings into exploitation tasks.
+- Findings outside the five classes form an internal `miscellaneous` class with its own exploitation agent (`miscellaneous-exploit`).
+- A scan can finish `completed`, `partial`, `failed`, or `cancelled`; `partial` carries an ordered set of reasons.
+
 ### Supporting Systems
-- **Configuration** — YAML configs in `apps/worker/configs/` with JSON Schema validation (`config-schema.json`). Supports auth settings (MFA/TOTP), URL/code rule scoping (`rules.avoid`/`rules.focus`), run-scope steering (`vuln_classes`, `exploit`), free-form `rules_of_engagement`, and post-hoc `report` filters (`min_severity`, `min_confidence`, `guidance`). `code_path` avoid rules are written into `~/.claude/settings.json` `permissions.deny` (`Read`/`Edit`) once per workflow by `apps/worker/src/temporal/activities.ts:syncCodePathDenyRules` so the SDK enforces them at the tool layer even in `bypassPermissions` mode. `vuln_classes`/`exploit` scope is locked into `session.json` on first run; resumes with a different scope fail fast (`persistOrValidateRunScope`). Credential resolution — local mode: env vars → `./.env`; npx mode: env vars → `~/.shannon/config.toml` (via `shn setup`)
+- **Configuration** — YAML configs in `apps/worker/configs/` use the closed JSON Schema in `config-schema.json`. Every fresh scan runs the fixed five analysis classes; there is no public class selector. `agentic_sast.enabled` is the only public agentic-SAST setting. Finding reconciliation runs on every scan and has no public setting of its own. Config also supports authentication (MFA/TOTP), URL/code rule scoping (`rules.avoid`/`rules.focus`), `exploit`, free-form `rules_of_engagement`, and post-hoc `report` options (`min_severity`, `min_confidence`, `guidance`, and exploit-only `sarif` output via `apps/worker/src/services/sarif-renderer.ts`, on by default for exploit runs and opt out with `report.sarif: "false"`). `code_path` avoid rules are enforced via the `@gotgenes/pi-permission-system` extension: `apps/worker/src/temporal/activities.ts:syncCodePathDenyRules` writes a global `path` deny config once per workflow (`apps/worker/src/ai/pi/permission-system.ts:syncPermissionSystemConfig`), and the executor loads the extension when that config is present (`apps/worker/src/ai/pi/pi-executor.ts`), so denies fire across every tool and child `task` session. Credential resolution — local mode: env vars → `./.env`; npx mode: env vars → `~/.shannon/config.toml` (via `npx @keygraph/shannon setup`)
+- **Agentic SAST progress** — Capella runs as a child workflow, so its activities are absent from the parent's `pendingActivities` and invisible to the CLI. The child signals each stage boundary up via `capellaStageProgress` (`apps/worker/src/temporal/shared.ts`); the parent's handler validates the payload and writes the child-supplied `startedAt` and `durationMs` directly to `operationalStages['agentic-sast:<stage>']`, so both the live `getProgress` query and the terminal result carry per-stage rows. Signalling is best-effort and every failure is swallowed — a closed or unreachable parent must never fail a SAST run. `CAPELLA_STAGE_LABELS` in `apps/worker/src/ai/sast/types.ts` is the one label table, shared by the scan log and the status tree; `CAPELLA_PROGRESS_STAGES` omits `export`, which runs no model and so never becomes a row. Scans predating the signal keep the aggregate `agentic-sast` span and render as a bare phase line
 - **Prompts** — Per-phase templates in `apps/worker/prompts/` with variable substitution (`{{TARGET_URL}}`, `{{CONFIG_CONTEXT}}`). Shared partials in `apps/worker/prompts/shared/` via `apps/worker/src/services/prompt-manager.ts`, including `_code-path-rules.txt` (focus/avoid `[FILE]`/`[GLOB]` routing) and `_rules-of-engagement.txt` (free-text engagement rules). When `exploit: false`, `apps/worker/src/services/findings-renderer.ts` deterministically converts each `*_exploitation_queue.json` into a `*_findings.md` for report assembly — no LLM in the loop
-- **SDK Integration** — Uses `@anthropic-ai/claude-agent-sdk` with `maxTurns: 10_000` and `bypassPermissions` mode. Adaptive thinking is enabled by default on Opus 4.6/4.7 (`supportsAdaptiveThinking` in `apps/worker/src/ai/models.ts`); disable per-scan via `CLAUDE_ADAPTIVE_THINKING=false` (env) or `core.adaptive_thinking = false` (npx TOML). Browser automation via `playwright-cli` with session isolation (`-s=<session>`). TOTP generation via `generate-totp` CLI tool. Login flow template at `apps/worker/prompts/shared/login-instructions.txt` supports form, SSO, API, and basic auth. On authenticated whitebox scans, the `validate-authentication` preflight performs the single real login and saves the browser session to `auth-state.json` in the per-session audit directory (path from `authStateFile()` in `apps/worker/src/audit/utils.ts`, derived from `generateAuditPath()`). The validation activity (`apps/worker/src/services/validate-authentication.ts`) removes any stale file from a prior run before the agent runs and verifies the file parses and contains cookies or storage before the preflight is marked complete; `logWorkflowComplete` deletes it when the workflow ends so authenticated cookies don't sit on disk between scans. Agent prompts opt in to session reuse by `@include(shared/_shared-session.txt)` before their `<login_instructions>` block — the partial restores the session and falls through to the full login flow if verification fails. `vuln-auth`/`exploit-auth` omit the include and own their own login
-- **Audit System** — Crash-safe append-only logging in `workspaces/{hostname}_{sessionId}/`. Tracks session metrics, per-agent logs, prompts, and deliverables. WorkflowLogger (`apps/worker/src/audit/workflow-logger.ts`) provides unified human-readable per-workflow logs, backed by LogStream (`apps/worker/src/audit/log-stream.ts`) shared stream primitive
-- **Deliverables** — Saved to `deliverables/` in the target repo via the `save-deliverable` CLI script (`apps/worker/src/scripts/save-deliverable.ts`)
-- **Workspaces & Resume** — Named workspaces via `-w <name>` or auto-named from URL+timestamp. Resume detects completed agents via `session.json`. `loadResumeState()` in `apps/worker/src/temporal/activities.ts` validates deliverable existence, restores git checkpoints, and cleans up incomplete deliverables. Workspace listing via `apps/worker/src/temporal/workspaces.ts`
+- **Agent Harness (pi)** — Uses the **pi harness** (`@earendil-works/pi-coding-agent`, requires Node ≥ 22.19) via `apps/worker/src/ai/pi/pi-executor.ts` (`runPiPrompt` → `createAgentSession`). Retry is split in `apps/worker/src/ai/pi/retry-settings.ts`: pi's agent-level loop is off so Temporal owns agent restarts, while `provider.maxRetries` stays on — pi reads the `provider` block independently of the `enabled` flag — so transport faults are absorbed in-session rather than costing a full agent re-run. `maxRetryDelayMs` is left at pi's 60s default. One model runs every phase, named by `SHANNON_AI_MODEL=<provider>:<model-id>` (default `anthropic:claude-sonnet-4-6`). `apps/worker/src/ai/models.ts` parses the spec — splitting on the **first** colon only, so Bedrock IDs keep theirs — and resolves it through pi's `ModelRuntime`. pi ships the `CredentialStore` interface but no in-memory implementation (its own reads `auth.json` from disk), so `RuntimeCredentialStore` in that file supplies one: credentials arrive as env vars in an ephemeral container and must never touch disk. `createModelRuntime(providerId, apiKey)` builds the runtime; `allowModelNetwork` stays at its default `false` so a scan never blocks on a catalog refresh. `resolveModelSelection()` is **async** because `ModelRuntime.create()` is. Any pi-ai provider id is accepted — `parseModelSpec` no longer rejects against a hardcoded list, so pi's registry is the authority (an unknown provider/model surfaces as a clear "not found in pi registry" error at preflight, which points to the browsable catalogue at `pi.dev/models` — `PI_CATALOG_URL` in `apps/worker/src/ai/models.ts`, appended to the not-found errors and shown in the setup wizard's "Other provider" hint). Four providers are **curated** (`CURATED_PROVIDERS`: `anthropic`, `openai`, `xai`, `amazon-bedrock`) with their own credential variables, config sections, and setup flows; each provider's API key env var is declared once in `PROVIDER_API_KEY_ENV` — Shannon uses each vendor's own variable name (`OPENAI_API_KEY`, `XAI_API_KEY`, …), never an invented one; Bedrock's entry is `AWS_BEARER_TOKEN_BEDROCK`, paired with `AWS_REGION`, which preflight requires separately as provider config rather than a credential. Any other provider uses the **generic** credential path: `SHANNON_AI_API_KEY` (`GENERIC_API_KEY_ENV`) supplies the key for any provider whose credential is a plain API key. Curated providers' own variables take precedence over it, and it also works as a fallback for them — Bedrock is the sole exception (it authenticates through its AWS_ variables, so the generic key never stands in for it). The CLI forwards `SHANNON_AI_API_KEY` in `COMMON_FORWARD_VARS` (it is provider-neutral, binding to whatever `SHANNON_AI_MODEL` names, so the "only one provider configured" guard counts only named credentials), and stores it under a generic `[provider]` config.toml section (`provider.api_key`). `npx @keygraph/shannon setup` exposes this as the "Other provider" option: free-text provider id + model id + key (a curated provider id is rejected there, since it has its own option). `SHANNON_AI_BASE_URL` overrides the endpoint for any provider (proxies/gateways); the credential is unchanged. `pointAtGateway` (`apps/worker/src/ai/models.ts`) applies the one dialect change: behind a base URL, `openai` follows `SHANNON_AI_OPENAI_FORMAT` (`chat-completions` default, or `responses`). On `chat-completions` it switches the API to `openai-completions` and drops the catalogue's Responses-shaped `compat` block so pi's `detectCompat` derives completions settings; on `responses` the descriptor is unchanged but for the endpoint. `resolveGatewayFormat` rejects the variable when the provider is not `openai` or no base URL is set, since it cannot take effect there. All other providers keep their API. The CLI mirrors the accepted values in `apps/cli/src/model-spec.ts`, forwards the variable in `COMMON_FORWARD_VARS`, and maps it to `openai.format` in config.toml. `buildEnvFlags` forwards only the selected provider's credential into the worker container. The CLI mirrors the parse rule and the provider/credential tables in `apps/cli/src/model-spec.ts` (it cannot import from the worker package); the two must stay in sync. pi ships no JSON-schema output or `Task`/`TodoWrite` built-ins, so structured queues are captured via a `submit_exploitation_queue` custom tool (`apps/worker/src/ai/queue-schemas.ts`), and `task` (child sessions scoped to `read`, `grep`, `find`, `ls`, `write`, and `bash` — no nested `task` or collector tools; `CHILD_TOOLS` in `apps/worker/src/ai/pi/task-tool.ts`) + `todo_write` (`apps/worker/src/ai/pi/session-tools.ts`) are provided as custom tools; the per-phase collectors are pi custom tools (TypeBox `defineTool` in `apps/worker/src/collectors/`). Shannon sets no thinking configuration at all — no `thinkingLevel` is passed to any `createAgentSession` call, so pi's own default applies. There Line truncated
+- **Pi Credential Reuse** — `SHANNON_USE_PI_AUTH=1` opts into reusing the host's Pi login, including an `openai-codex` ChatGPT Plus/Pro subscription (`SHANNON_AI_MODEL=openai-codex:<model-id>`) or an `xai` Grok subscription (`SHANNON_AI_MODEL=xai:<model-id>`); the mechanism is provider-agnostic and works for any Pi login. `apps/cli/src/env.ts` requires `~/.pi/agent/auth.json`; `start.ts` passes its path to `spawnWorker`, which mounts only that file read-write at `/tmp/.pi/agent/auth.json`. The flag itself is not forwarded: the worker detects the file with `piAuthPresent()` and passes its path to `ModelRuntime.create`. CLI and worker API-key presence checks are skipped on this path, but the normal preflight model probe still validates the credential. The image and UID-remapping entrypoint keep `/tmp/.pi/agent` owned by `pentest` so adjacent Pi/Shannon configuration remains writable. Refreshed OAuth state is persisted to the host for subsequent scans.
+- **Audit System** — Crash-safe append-only logging in `workspaces/{hostname}_{sessionId}/`. The run directory's top level holds the human-facing report in both formats (`Security-Assessment-Report.pdf` and `Security-Assessment-Report.md`, `FINAL_REPORT_PDF_FILENAME`/`FINAL_REPORT_MD_FILENAME` in `apps/worker/src/paths.ts`); everything else — deliverables, per-agent logs, prompts, `session.json`, `workflow.log`, and browser artifacts — is nested under a hidden `.shannon/` internals dir (`INTERNAL_DIR`) so a customer sees only the report. Audit path helpers route through `generateInternalPath` (`apps/worker/src/audit/utils.ts`); the CLI nests the overlay backing dirs under the same `.shannon/` (`apps/cli/src/docker.ts`, `start.ts`). `session.json`/`workflow.log` reads use dual-read resolvers (`resolveSessionJsonPath`, `resolveRunFile`) that prefer `.shannon/` and fall back to the legacy run-root layout, so pre-restructure workspaces stay listable (`scans`/`logs`) without migration. A pre-restructure workspace cannot be resumed: `classifyWorkspaceLaunch` (`apps/cli/src/commands/start.ts`) requires `.shannon/launch.json`, and its absence fails the launch as "created by an earlier version of Shannon" before anything on disk is touched. There is no in-place migration — the workspace's files and report are left untouched, and the operator starts a new scan under a different `-w` name. The report agent writes structured findings to `report.json`, from which `report-renderer.ts` renders the assembled markdown and `report-json-adapter.ts` produces the Typst-shaped JSON that `pdf-renderer.ts` compiles into `comprehensive_security_assessment_report.pdf` using the bundled `apps/worker/templates/typst/report.typ` template (the `typst` binary is installed in the worker image). `copyReportToRunRoot` (`apps/worker/src/services/reporting.ts`) surfaces both the PDF and the markdown to the run root as `Security-Assessment-Report.pdf` and `Security-Assessment-Report.md`; the deliverables-dir copies remain as the git-checkpointed sources. PDF compilation is best-effort — a failure is logged and the run still completes. WorkflowLogger (`apps/worker/src/audit/workflow-logger.ts`) provides unified human-readable per-workflow logs, backed by LogStream (`apps/worker/src/audit/log-stream.ts`) shared stream primitive. Every combined-log line is also projected into a per-agent file under `.shannon/agents/<slug>.log` (one per pipeline agent, one per Capella stage; subagents fold into the parent's file, and a stage's concurrent sessions share its file with an inline session label). The projection boundary is `apps/worker/src/audit/actor-projection.ts` (`projectActor` maps a `TraceActor` to its combined prefix and owning file slug — slugs come only from closed fields); fan-out is best-effort and never blocks the canonical combined log. A lifecycle owner holds a `LogStream` lease per agent file (the pipeline agent's `logAgent` span, or a Capella stage activity's `try/finally`) so per-line writes ride the reference count; `CapellaStageTrace.drain()` flushes a stage's trace queue before its activity returns. The CLI tails one file with `shannon logs --agent <name>` (`--list-agents` to enumerate); the default `shannon logs` path is unchanged
+- **Deliverables** — Saved to `.shannon/deliverables/` in the target repo via the `save-deliverable` CLI script (`apps/worker/src/scripts/save-deliverable.ts`)
+- **Workspaces & Resume** — Named workspaces via `-w <name>` or auto-named from URL+timestamp. Resume detects completed agents via `session.json`. `loadResumeState()` in `apps/worker/src/temporal/activities.ts` validates deliverable existence, restores git checkpoints, and cleans up incomplete deliverables
 
 ## Development Notes
 
@@ -168,7 +187,7 @@ Durable workflow orchestration with crash recovery, queryable progress, intellig
 ### Key Design Patterns
 - **Configuration-Driven** — YAML configs with JSON Schema validation
 - **Progressive Analysis** — Each phase builds on previous results
-- **SDK-First** — Claude Agent SDK handles autonomous analysis
+- **Harness-First** — the pi harness (`@earendil-works/pi-coding-agent`) handles autonomous analysis
 - **Modular Error Handling** — `ErrorCode` enum, `Result<T,E>` for explicit error propagation, automatic retry (3 attempts per agent)
 - **Services Boundary** — Activities are thin Temporal wrappers; `apps/worker/src/services/` owns business logic, accepts `ActivityLogger`, returns `Result<T,E>`. No Temporal imports in services
 - **DI Container** — Per-workflow in `apps/worker/src/services/container.ts`. `AuditSession` excluded (parallel safety)
@@ -228,7 +247,7 @@ Comments must be **timeless** — no references to this conversation, refactorin
 
 **Entry Points:** `apps/worker/src/temporal/workflows.ts`, `apps/worker/src/temporal/activities.ts`, `apps/worker/src/temporal/worker.ts`
 
-**Core Logic:** `apps/worker/src/session-manager.ts`, `apps/worker/src/ai/claude-executor.ts`, `apps/worker/src/ai/settings-writer.ts` (writes `code_path` deny rules to `~/.claude/settings.json`), `apps/worker/src/config-parser.ts`, `apps/worker/src/services/` (incl. `preflight.ts`, `findings-renderer.ts`, `reporting.ts`), `apps/worker/src/audit/`
+**Core Logic:** `apps/worker/src/session-manager.ts`, `apps/worker/src/ai/pi/pi-executor.ts`, `apps/worker/src/ai/pi/permission-system.ts` (writes `code_path` deny rules to the `@gotgenes/pi-permission-system` global config), `apps/worker/src/config-parser.ts`, `apps/worker/src/services/` (incl. `preflight.ts`, `findings-renderer.ts`, `reporting.ts`), `apps/worker/src/audit/`
 
 **Config:** `docker-compose.yml`, `apps/cli/infra/compose.yml`, `apps/worker/configs/`, `apps/worker/prompts/`, `tsconfig.base.json` (shared compiler options), `turbo.json`, `biome.json`
 
@@ -240,9 +259,9 @@ Package managers are configured with a minimum release age (7 days). Requires pn
 
 ## Troubleshooting
 
-- **"Repository not found"** — Pass a bare name (`-r my-repo`) for `./repos/my-repo`, or a path (`-r /path/to/repo`) for any directory
+- **"Repository not found"** — Pass a path to the target repo (`-r /path/to/repo` or `-r ./my-repo`)
 - **"Temporal not ready"** — Wait for health check or `docker compose logs temporal`
 - **Worker not processing** — Check `docker ps --filter "name=shannon-worker-"`
-- **Reset state** — `./shannon stop --clean`
+- **Reset state** — `./shannon reset`
 - **Local apps unreachable** — Use `host.docker.internal` instead of `localhost`
 - **Container permissions** — On Linux, may need `sudo` for docker commands

Dockerfile

@@ -43,6 +43,9 @@ RUN rm -rf node_modules apps/*/node_modules && pnpm install --frozen-lockfile --
 # Runtime stage - Minimal production image
 FROM cgr.dev/chainguard/wolfi-base:latest AS runtime
 
+# Lifecycle protocol consumed by the CLI before it trusts a container workflow-id label.
+LABEL shannon.worker-protocol="workflow-id-v1"
+
 # Install only runtime dependencies
 USER root
 RUN apk update && apk add --no-cache \
@@ -52,6 +55,8 @@ RUN apk update && apk add --no-cache \
     curl \
     ca-certificates \
     shadow \
+    # Typst tarball decompression
+    xz \
     # Language runtimes (minimal)
     nodejs-22 \
     npm \
@@ -73,6 +78,22 @@ RUN apk update && apk add --no-cache \
     # Font rendering
     fontconfig
 
+# Install Typst (report PDF compilation)
+ARG TYPST_VERSION=0.14.2
+RUN case "$(uname -m)" in \
+      x86_64) TYPST_ARCH=x86_64-unknown-linux-musl ;; \
+      aarch64) TYPST_ARCH=aarch64-unknown-linux-musl ;; \
+      *) echo "unsupported arch $(uname -m)" && exit 1 ;; \
+    esac && \
+    mkdir -p /tmp/typst-dl /usr/local/bin && cd /tmp/typst-dl && \
+    curl -fsSL "https://github.com/typst/typst/releases/download/v${TYPST_VERSION}/typst-${TYPST_ARCH}.tar.xz" -o typst.tar.xz && \
+    xz -d typst.tar.xz && \
+    tar -xf typst.tar && \
+    mv "typst-${TYPST_ARCH}/typst" /usr/local/bin/typst && \
+    chmod +x /usr/local/bin/typst && \
+    cd / && rm -rf /tmp/typst-dl && \
+    typst --version
+
 # Create non-root user
 RUN addgroup -g 1001 pentest && \
     adduser -u 1001 -G pentest -s /bin/bash -D pentest
@@ -91,7 +112,11 @@ COPY --from=builder /app/node_modules /app/node_modules
 COPY --from=builder /app/apps/worker /app/apps/worker
 COPY --from=builder /app/apps/cli/package.json /app/apps/cli/package.json
 
-RUN npm install -g --ignore-scripts @anthropic-ai/claude-code@2.1.84 @playwright/cli@0.1.1
+# Third-party license and notice material travels with the distributed image
+COPY LICENSE THIRD_PARTY_NOTICES.md /usr/share/licenses/shannon/
+COPY LICENSES/ /usr/share/licenses/shannon/LICENSES/
+
+RUN npm install -g --ignore-scripts @playwright/cli@0.1.1
 RUN mkdir -p /tmp/.claude/skills && \
     playwright-cli install --skills && \
     cp -r .claude/skills/playwright-cli /tmp/.claude/skills/ && \
@@ -101,16 +126,18 @@ RUN mkdir -p /tmp/.claude/skills && \
 RUN ln -s /app/apps/worker/dist/scripts/save-deliverable.js /usr/local/bin/save-deliverable && \
     chmod +x /app/apps/worker/dist/scripts/save-deliverable.js && \
     ln -s /app/apps/worker/dist/scripts/generate-totp.js /usr/local/bin/generate-totp && \
-    chmod +x /app/apps/worker/dist/scripts/generate-totp.js
+    chmod +x /app/apps/worker/dist/scripts/generate-totp.js && \
+    ln -s /app/apps/worker/dist/scripts/set-report-meta.js /usr/local/bin/set-report-meta && \
+    chmod +x /app/apps/worker/dist/scripts/set-report-meta.js
 
 # Create directories for session data and ensure proper permissions
 RUN mkdir -p /app/sessions /app/repos /app/workspaces && \
-    mkdir -p /tmp/.cache /tmp/.config /tmp/.npm && \
+    mkdir -p /tmp/.cache /tmp/.config /tmp/.npm /tmp/.pi/agent && \
     chmod 777 /app && \
     chmod 777 /tmp/.cache && \
     chmod 777 /tmp/.config && \
     chmod 777 /tmp/.npm && \
-    chown -R pentest:pentest /app /tmp/.claude
+    chown -R pentest:pentest /app /tmp/.claude /tmp/.pi
 
 COPY entrypoint.sh /app/entrypoint.sh
 RUN chmod +x /app/entrypoint.sh

LICENSES/Apache-2.0.txt

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

LICENSES/MIT-Pi.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Mario Zechner
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -1,188 +1,278 @@
->[!NOTE]
-> **[Better Steerability, Authentication Improvements, and the Migration to the Pi Harness](https://github.com/KeygraphHQ/shannon/discussions/348)**
+> [!NOTE]
+> **Shannon 3.0 is live:** deeper security code analysis, a rebuilt terminal experience, native CI/CD workflows, professional PDF reports, and SARIF—still fully open source, self-hosted, and bring-your-own-model.
 
 <div align="center">
 
-<img src="./assets/github-banner.png" alt="Shannon - AI Pentester for Web Applications and APIs" width="100%">
-
-# Shannon - AI Pentester by Keygraph
+<picture>
+<source media="(prefers-color-scheme: dark)" srcset="./assets/github-banner-dark.png">
+<source media="(prefers-color-scheme: light)" srcset="./assets/github-banner-light.png">
+<img src="./assets/github-banner-light.png" alt="Shannon, AI Pentester for Web Apps and APIs, by Keygraph" width="100%">
+</picture>
 
 <a href="https://trendshift.io/repositories/15604" target="_blank"><img src="https://trendshift.io/api/badge/repositories/15604" alt="KeygraphHQ%2Fshannon | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
 
-Shannon is an autonomous, white-box AI pentester for web applications and APIs. <br />
-It analyzes your source code, identifies attack paths, and executes real exploits to prove vulnerabilities before they reach production.
+### Shannon is an autonomous, AI pentester for web applications and APIs.
+
+It analyzes your source code, identifies attack paths, and executes real exploits to prove vulnerabilities before they reach production. **No exploit, no report.**
+
+**This repository is Shannon Open Source: the full agent, run locally from your command line.**
 
 ---
 
-<a href="https://discord.gg/9ZqQPuhJB7"><img src="./assets/discord.png" height="40" alt="Join Discord"></a>
-<a href="https://keygraph.io/"><img src="./assets/Keygraph_Button.png" height="40" alt="Visit Keygraph.io"></a>
+<a href="https://discord.gg/9ZqQPuhJB7"><picture><source media="(prefers-color-scheme: dark)" srcset="./assets/discord_button_dark.png"><source media="(prefers-color-scheme: light)" srcset="./assets/discord_button_light.png"><img src="./assets/discord_button_light.png" height="40" alt="Join Discord"></picture></a>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;<a href="https://keygraph.io/"><picture><source media="(prefers-color-scheme: dark)" srcset="./assets/keygraph_button_dark.png"><source media="(prefers-color-scheme: light)" srcset="./assets/keygraph_button_light.png"><img src="./assets/keygraph_button_light.png" height="40" alt="Visit Keygraph.io"></picture></a>
 
 ---
+
 </div>
 
 > [!TIP]
 > **AI agents and LLMs:** start with [llms.txt](llms.txt) for a concise map of this repository, or use [llms-full.txt](llms-full.txt) for the README and docs combined into one file.
 
+
+
 ## Table of Contents
 
+- [Table of Contents](#table-of-contents)
 - [What is Shannon?](#what-is-shannon)
-- [Product Line](#product-line)
-- [Shannon Lite in Action](#shannon-lite-in-action)
+  - [Why Shannon Exists](#why-shannon-exists)
+  - [Why "Shannon"?](#why-shannon)
+  - [Not a replacement for human pentesters](#not-a-replacement-for-human-pentesters)
+- [Shannon in Action](#shannon-in-action)
 - [Quick Start](#quick-start)
+  - [Prerequisites](#prerequisites)
+  - [Run Shannon](#run-shannon)
 - [Key Capabilities](#key-capabilities)
-- [Shannon Lite and Shannon Pro](#shannon-lite-and-shannon-pro)
+- [CI/CD Integrations](#cicd-integrations)
+  - [GitHub Actions](#github-actions)
+- [Editions](#editions)
 - [Architecture](#architecture)
 - [Documentation](#documentation)
 - [Safety, Scope, and Limitations](#safety-scope-and-limitations)
-- [License and Enterprise Licensing](#license-and-enterprise-licensing)
+- [License](#license)
+- [Acknowledgements](#acknowledgements)
+- [About Keygraph](#about-keygraph)
 - [Community and Support](#community-and-support)
+- [Common Questions](#common-questions)
+  - [Can I self-host Shannon?](#can-i-self-host-shannon)
+  - [Does Shannon support bring your own key (BYOK)?](#does-shannon-support-bring-your-own-key-byok)
+  - [Does Shannon output SARIF?](#does-shannon-output-sarif)
+  - [Which AI providers does Shannon support?](#which-ai-providers-does-shannon-support)
+  - [Can I run Shannon on a local or self-hosted model?](#can-i-run-shannon-on-a-local-or-self-hosted-model)
+  - [Does Shannon actually exploit vulnerabilities, or just scan?](#does-shannon-actually-exploit-vulnerabilities-or-just-scan)
+
+
 
 ## What is Shannon?
 
-Shannon is an AI pentester developed by [Keygraph](https://keygraph.io). It performs white-box security testing of web applications and their underlying APIs by combining source-code analysis with live exploitation.
+Shannon is an autonomous AI pentester developed by [Keygraph](https://keygraph.io). It performs security testing of web applications and their underlying APIs by combining source-code analysis with live exploitation.
 
 Shannon analyzes your web application's source code to identify potential attack vectors, then uses browser automation and command-line tools to execute real exploits against the running application and its APIs. Only vulnerabilities with a working proof-of-concept are included in the final report.
 
+Shannon is the agent. This repository is Shannon Open Source, the standalone pentester you run yourself. The same Shannon also powers the [Keygraph platform](https://keygraph.io), Keygraph's commercial pentesting product. See [Editions](#editions) for how the two compare.
+
 ### Why Shannon Exists
 
 Thanks to tools like Claude Code and Cursor, your team ships code non-stop. But your penetration test? That happens once a year. This creates a massive security gap. For the other 364 days, you could be unknowingly shipping vulnerabilities to production.
 
 Shannon closes that gap by providing on-demand, automated penetration testing that can run against every build or release.
 
-## Product Line
+### Why "Shannon"?
 
-Shannon is developed by [Keygraph](https://keygraph.io) and available in two editions:
+It's named after Claude Shannon, the father of information theory. At its core, pentesting is an information problem: every probe reduces uncertainty about a system's state. The best tools maximize the signal gained from every request, turning those bits of knowledge into an exploit path.
 
-| Edition | License | Best For |
-| --- | --- | --- |
-| **Shannon Lite** | AGPL-3.0 | Local, strictly white-box testing of applications you own or are authorized to test. |
-| **Shannon Pro** | Commercial | Organizations needing a continuous pentesting and AppSec platform with black-box and white-box pentesting, parsed-code SAST, CI/CD gating, verified remediation, SLA tracking, and enterprise deployment. |
+Also, we wanted you to be able to say, "Hey Claude, run Shannon" to find all the security flaws in your vibe-coded app.
 
-## Shannon Lite in Action
+### Not a replacement for human pentesters
+
+Shannon is built to work alongside expert pentesters and red teamers, not replace them. Great pentesters understand the business, chain attacks in ways nobody anticipated, and bring years of judgment that current models can't match.
+
+Shannon solves a different problem: there is far more software to test than security teams have time to cover. Critical systems get periodic expert assessments, while the long tail of internal apps, APIs, and fast-moving services rarely gets tested at all.
+
+Shannon shifts pentesting left into the software development lifecycle (SDLC). Use it to run exploitation-backed tests against staging environments and releases at the cadence they actually ship, and save expert human time for the risks that need someone who knows the organization.
+
+## Shannon in Action
+
+![Shannon running an autonomous pentest](assets/Shannon3GIF.gif)
+
+Penetration test reports from Shannon Open Source scanning Photoview 2.4.0. Read the [announcement][announcement] and the full [benchmark writeup][benchmark] for methodology, cost, and the comparison against Aikido and XBOW.
+
+
+| Model             | Report                                                                       | SARIF                                                            |
+| ----------------- | ---------------------------------------------------------------------------- | --------------------------------------------------------------- |
+| DeepSeek v4 Flash | [View report](benchmark/photoview-deepseek-v4-flash.pdf) | [SARIF](benchmark/photoview-deepseek-v4-flash.sarif) |
+| Grok 4.6          | [View report](benchmark/photoview-grok-4-6.pdf)          | [SARIF](benchmark/photoview-grok-4-6.sarif)          |
+| Claude Opus 5     | [View report](benchmark/photoview-opus-5.pdf)            | [SARIF](benchmark/photoview-opus-5.sarif)            |
+
+[announcement]: https://github.com/KeygraphHQ/shannon/discussions/439
+[benchmark]: docs/shannon-xbow-aikido-benchmark.md
 
-<p align="center">
-  <img src="assets/shannon-action.gif" alt="Shannon Lite running an autonomous pentest" width="100%">
-</p>
 
-Sample Shannon Lite penetration test reports from intentionally vulnerable applications:
 
-| Target | Summary | Report |
-| --- | --- | --- |
-| OWASP Juice Shop | 20+ vulnerabilities, including authentication bypass, SQL injection, IDOR, and SSRF. | [View report](sample-reports/shannon-report-juice-shop.md) |
-| c{api}tal API | Approximately 15 critical and high-severity API findings, including command injection, auth bypass, and mass assignment. | [View report](sample-reports/shannon-report-capital-api.md) |
-| OWASP crAPI | 15+ critical and high-severity findings across JWT, injection, SSRF, and API authorization paths. | [View report](sample-reports/shannon-report-crapi.md) |
 
 ## Quick Start
 
+
+
 ### Prerequisites
 
-- **Docker** - required for the worker container.
-- **Node.js 18+** - required for the recommended `npx` workflow.
-- **AI provider credentials** - Anthropic is recommended; AWS Bedrock, Google Vertex AI, and compatible proxy setups are documented separately.
+- **Docker**: required for the worker container.
+- **Node.js 18+**: required for the recommended `npx` workflow.
+- **AI provider credentials**: Shannon runs on Anthropic, OpenAI, xAI, AWS Bedrock, [any other provider](docs/ai-providers.md#any-other-provider) in the harness catalogue, and any endpoint that speaks the Anthropic Messages API or the OpenAI Chat Completions or Responses API through a [custom base URL](docs/ai-providers.md#custom-base-url). You bring your own key, and Keygraph never proxies your model traffic. Shannon is provider-agnostic. See [AI providers](docs/ai-providers.md#suggested-models) for suggested model IDs.
+- **Cyber safeguards cleared with your provider**: Anthropic and OpenAI apply real-time safeguards to cyber-security workloads, which can interrupt a scan mid-run. Complete their guidance for legitimate security testers before your first run - see [AI providers](docs/ai-providers.md#cyber-safeguards-do-this-before-your-first-scan).
 
-### Run Shannon Lite
+
+
+### Run Shannon
 
 > [!WARNING]
-> Shannon Lite actively executes exploits. Run it only against applications and environments you own or have explicit written authorization to test. Do not run Shannon Lite against production systems.
+> Shannon actively executes exploits. Run it only against applications and environments you own or have explicit written authorization to test. Do not run Shannon against production systems.
 
 ```bash
 # Configure credentials with the interactive wizard.
-npx @keygraph/shannon setup
+npx @keygraph/shannon@latest setup
 
 # Run a pentest against a source-available target.
-npx @keygraph/shannon start -u https://your-app.com -r /path/to/your-repo
+npx @keygraph/shannon@latest start \
+  -u https://your-app.com \
+  -r /path/to/your/repo
 ```
 
-Shannon Lite pulls the worker image from Docker Hub, starts the required local infrastructure, mounts the target repository read-only inside an ephemeral worker container, and writes results to a local workspace.
+Shannon pulls the worker image from Docker Hub, starts the required local infrastructure, mounts the target repository read-only inside an ephemeral worker container, and writes results to a local workspace.
 
 For source builds, authenticated scans, provider-specific setup, and platform notes, see [Documentation](#documentation).
 
+> [!TIP]
+> **Prefer to use a subscription instead of API credits?**
+>
+> - **OpenAI Codex:** The latest version of Shannon supports ChatGPT Plus and Pro subscriptions. Follow the [OpenAI Codex subscription setup guide](docs/ai-providers.md#openai-codex-chatgpt-pluspro-subscription) to get started.
+> - **xAI (Grok):** The latest version of Shannon supports xAI subscriptions. Follow the [xAI subscription setup guide](docs/ai-providers.md#xai-grok-subscription) to get started.
+> - **Claude Code:** The latest version of Shannon does not support Claude Code subscriptions. Follow the [Claude Code subscription setup guide](docs/ai-providers.md#claude-code-subscription) to use version `1.9.0`, which is the final release built on the Claude Agent SDK.
+
+
+
 ## Key Capabilities
 
-- **Proof-by-exploitation reports**: Shannon Lite reports validated findings with reproducible proof-of-concept steps instead of speculative warnings.
-- **White-box attack planning**: Shannon Lite uses source-code analysis to guide dynamic testing and focus on realistic attack paths.
-- **Autonomous execution**: Shannon Lite launches reconnaissance, vulnerability analysis, exploitation, and report generation from a single command.
-- **Authenticated testing**: Shannon Lite configuration files can describe login flows, test credentials, TOTP, email-based login flows, focus areas, and rules of engagement.
-- **OWASP-focused coverage**: Shannon Lite targets exploitable Injection, XSS, SSRF, Broken Authentication, and Broken Authorization issues.
-- **Resumable workspaces**: Shannon Lite can resume interrupted runs without re-running completed agents.
+- **No exploit, no report**: Shannon includes a vulnerability only after validating it with a working, reproducible proof of concept—eliminating the speculative warnings typical of scanners.
+- **Advanced security code analysis**: Before it sends a single payload, Shannon reads the codebase and builds a picture of the application: architecture, trust boundaries, exposed interfaces, data flows, and the assets worth attacking. From there it opens targeted investigations and filters the candidates they turn up. What survives goes to the live pentesting agents.
+- **Autonomous execution**: Shannon launches reconnaissance, vulnerability analysis, exploitation, and report generation from a single command.
+- **Live terminal experience**: A rebuilt CLI makes scans easy to configure and shows agent progress and clean results without requiring operators to inspect the underlying orchestration logs.
+- **Authenticated testing**: configuration files can describe login flows, test credentials, TOTP, email-based login flows, focus areas, and rules of engagement.
+- **OWASP-focused coverage**: Shannon targets exploitable Injection, XSS, SSRF, Broken Authentication, and Broken Authorization issues.
+- **Resumable workspaces**: Shannon can resume interrupted runs without re-running completed agents.
+- **Native CI/CD integrations**: Run Shannon through the official GitHub Action or reusable GitLab CI/CD component. Preserve reports, SARIF, and logs as pipeline artifacts; publish findings into native security workflows; and gate releases only on vulnerabilities Shannon actually demonstrates.
+- **Professional and machine-readable reports**: Shannon generates evidence-rich PDF and Markdown reports plus structured JSON and SARIF 2.1.0. SARIF is enabled by default on exploit-mode scans and can be disabled with `report.sarif: "false"`.
+- **Bring your own key, provider-agnostic**: Shannon runs on Anthropic, OpenAI, xAI, AWS Bedrock, and any endpoint speaking the Anthropic Messages API or the OpenAI Chat Completions or Responses API, including self-hosted models served through Ollama, vLLM, or LM Studio and gateways such as OpenRouter and LiteLLM. You supply the credentials and choose exactly where model traffic goes. Local and self-hosted models are supported.
+- **Private by design**: Shannon runs inside your infrastructure and writes results to a local workspace. Model requests go straight to the provider or endpoint you configure, and they carry source and application context with them, so choose that endpoint deliberately. Point Shannon at a local model endpoint and nothing leaves your environment.
 
-## Shannon Lite and Shannon Pro
 
-This repository contains **Shannon Lite**, the AGPL-3.0 open-source CLI for strictly white-box, proof-by-exploitation testing of web applications and APIs you own or are authorized to test. Shannon Lite requires access to the target application's source code and repository layout.
 
-**Shannon Pro** is Keygraph's commercial continuous pentesting and AppSec platform for teams running security across many repositories, services, and environments. While Shannon Lite is a local white-box pentesting CLI, Shannon Pro is a full platform: it combines parsed-code SAST, source-to-sink analysis, black-box and white-box agentic pentesting, verified remediation, CI/CD gating, SLA tracking, and reporting for security and compliance teams.
+## CI/CD Integrations
 
-Shannon Pro supports both **white-box and black-box agentic pentesting**: use source-aware testing when code is available, or run autonomous black-box testing against deployed applications and APIs when source access is unavailable or unnecessary.
+Shannon can run continuously against deployed staging and development environments through official integrations for [GitHub Actions](https://github.com/KeygraphHQ/shannon-action) and [GitLab CI/CD](https://gitlab.com/KeygraphHQ/shannon-ci).
 
-Shannon Pro covers the full vulnerability lifecycle: finding exploitable issues, deduplicating and prioritizing them, syncing work into developer workflows, generating verified remediations, re-testing fixes, tracking SLAs, and producing dashboards for security reporting and compliance.
+Both integrations:
 
-For enterprise deployments, Shannon Pro supports self-hosted and air-gapped environments, strict bring-your-own-key model access, and customer-controlled LLM gateway patterns. Deployments can be designed so source code, scan results, prompts, completions, and model traffic remain inside your security perimeter.
+- analyze the checked-out source repository while attacking a running target;
+- preserve PDF, Markdown, and SARIF reports as pipeline artifacts;
+- preserve scan and agent logs for debugging, including incomplete runs;
+- support pull-request, release, and scheduled pentests;
+- distinguish an incomplete assessment from a completed scan with no findings; and
+- optionally fail the pipeline when Shannon exploits a vulnerability at or above a configured severity threshold.
 
-Shannon Lite is a strong fit for local and project-level white-box testing. Shannon Pro is intended for organizations that need continuous AppSec coverage, black-box and white-box pentesting, centralized triage, verified remediation workflows, compliance-ready reporting, enterprise integrations, and commercial support.
+A code-analysis hypothesis does not fail the pipeline. Severity gates count only findings with `status: exploited`.
 
-| Need | Shannon Lite | Shannon Pro |
-| --- | --- | --- |
-| License | AGPL-3.0 | Commercial |
-| White-box pentesting | Yes; source code required | Yes; source-aware testing with platform workflows |
-| Black-box pentesting | No | Yes; autonomous testing without source-code access |
-| Code analysis / SAST | Prompting and source pass-through to guide pentesting | Actual code parsing, Code Property Graph analysis, source-to-sink path analysis, and agentic SAST |
-| AppSec coverage | OWASP-focused agentic pentesting | Agentic pentesting, SAST, SCA, secrets, IaC, containers, and business logic testing |
-| CI/CD and gating | Manual/local CLI runs | Headless commercial CLI for CI/CD gating across enterprise CI/CD platforms |
-| Finding lifecycle | Local Markdown reports | Canonical findings, deduplication, ownership, status, SLA tracking, workflow sync, and reporting dashboards |
-| Remediation | Manual | User-initiated remediation with verification before delivery |
-| Fix verification | None; manual reruns only | Targeted verification without rerunning the entire scan, completing the remediation lifecycle |
-| Enterprise deployment | Local CLI and Docker worker | Self-hosted, air-gapped, BYOK, and customer-controlled LLM gateway options |
-| Support | Community | Commercial support |
+### GitHub Actions
 
-Learn more on the [Keygraph website](https://keygraph.io), read the [Shannon Pro technical overview](docs/shannon-pro.md), start a free trial or book a [Shannon Pro demo](https://cal.com/team/keygraph/shannon-pro), or contact [shannon@keygraph.io](mailto:shannon@keygraph.io).
+```yaml
+name: Shannon Pentest
+
+on:
+  workflow_dispatch:
+
+permissions:
+  security-events: write
+
+jobs:
+  pentest:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Run Shannon
+        uses: KeygraphHQ/shannon-action@v1
+        with:
+          url: https://staging.example.com
+          api-key: ${{ secrets.SHANNON_AI_API_KEY }}
+          fail-on-severity: high
+          upload-sarif: true
+```
+
+The Action defaults `repo` to the checked-out GitHub workspace. It uploads one artifact containing the security assessment reports and SARIF, plus a separate run artifact containing scan and agent logs. Enabling `upload-sarif` publishes supported findings to GitHub code scanning.
+
+Requirements:
+
+- a private repository;
+- a runner with Docker and Docker Compose v2;
+- access to the running staging or development target; and
+- a model-provider credential stored as a GitHub Actions secret.
+
+See the [Shannon GitHub Action documentation](https://github.com/KeygraphHQ/shannon-action) and [GitHub Marketplace listing](https://github.com/marketplace/actions/shannon-ai-pentester).
+
+## Editions
+
+**Shannon Open Source** is the complete autonomous pentester for developers and security teams. It is optimized for fast local and CI/CD runs: understand the application, execute real attacks, and report only proven vulnerabilities.
+
+**Keygraph Enterprise Platform** turns Shannon's proof engine into an organization-wide AppSec program, adding exhaustive analysis, centralized vulnerability management, automated remediation, enterprise governance, and continuous operation at scale.
+
+
+|                     | Shannon Open Source                                                                                                                                                    | Keygraph Enterprise Platform                                                                                                                                                                              |
+| ------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| Best for            | Local and CI/CD pentesting                                                                                                                                             | Continuous AppSec across teams and repositories                                                                                                                                                           |
+| Security analysis   | Multi-stage agentic review models architecture, trust boundaries, and data flows, filters candidate vulnerabilities, and hands the survivors to live pentesting agents | Exhaustive parsed-code agentic SAST: persistent Code Property Graphs, interprocedural source-to-sink and sanitizer modeling, cross-repository context, exploit-chain analysis, and business-logic testing |
+| Additional coverage | Not included                                                                                                                                                           | SCA with reachability, secrets scanning, and business-logic testing                                                                                                                                       |
+| AppSec operations   | N/A — standalone CLI                                                                                                                                                   | Canonical findings, deduplication, SLAs, analytics, automated remediation, and targeted verification                                                                                                      |
+| Governance          | N/A — local, single-operator CLI                                                                                                                                       | SSO, SCIM, granular access control, APIs, and full audit logging                                                                                                                                          |
+| Deployment          | Self-hosted, air-gapped, BYOM, AGPL-3.0                                                                                                                                | On-premises or air-gapped, granular model routing, commercial support                                                                                                                                     |
+
+
+Shannon Open Source is not a trial edition. Choose Keygraph Enterprise when you need deeper analysis and a governed, closed-loop AppSec program.
+
+[Explore the Keygraph Enterprise Platform →](docs/keygraph-platform.md)
 
 ## Architecture
 
-Shannon Lite uses a multi-agent workflow that combines source-code analysis with live exploitation:
+Shannon combines multi-stage security code analysis with live reconnaissance and exploitation:
 
-```text
-        ┌──────────────────────┐
-        │   Pre-Reconnaissance │
-        │   (source code scan) │
-        └──────────┬───────────┘
-                   │
-                   ▼
-        ┌──────────────────────┐
-        │   Reconnaissance     │
-        │  (attack surface     │
-        │   mapping)           │
-        └──────────┬───────────┘
-                   │
-                   ▼
-        ┌──────────┴───────────┐
-        │          │           │
-        ▼          ▼           ▼
-  ┌───────────┐ ┌───────────┐ ┌───────────┐
-  │ Vuln      │ │ Vuln      │ │   ...     │
-  │(Injection)│ │  (XSS)    │ │           │
-  └─────┬─────┘ └─────┬─────┘ └─────┬─────┘
-        │              │             │
-        ▼              ▼             ▼
-  ┌───────────┐ ┌───────────┐ ┌───────────┐
-  │ Exploit   │ │ Exploit   │ │   ...     │
-  │(Injection)│ │  (XSS)    │ │           │
-  └─────┬─────┘ └─────┬─────┘ └─────┬─────┘
-        │              │             │
-        └──────┬───────┴─────────────┘
-               │
-               ▼
-        ┌──────────────────────┐
-        │      Reporting       │
-        └──────────────────────┘
+```mermaid
+flowchart TD
+    S["Source code"] --> EXISTING["Recon + vulnerability analysis"]
+    S --> SAST["Agentic security code analysis"]
+
+    EXISTING -- "Pentest candidates" --> REC["Finding reconciliation<br/>(merge + deduplicate)"]
+    SAST -- "SAST candidates" --> REC
+
+    REC -- "Reconciled exploitation queue" --> EXP["Exploitation agents"]
+    APP["Running application"] --> EXP
+
+    EXP -- "Exploit demonstrated" --> REPORT["Reporting<br/>PDF · Markdown · SARIF"]
+    EXP -- "No exploit demonstrated" --> DROP["Discard"]
+
+    REPORT --> CICD["CI/CD gate"]
 ```
 
-At a high level:
 
-- **Pre-reconnaissance** identifies frameworks, entry points, data flows, and likely attack surfaces from the repository.
-- **Reconnaissance** explores the live application and correlates runtime behavior with code-level context.
-- **Vulnerability analysis** runs specialized agents for Injection, XSS, SSRF, Authentication, and Authorization.
-- **Exploitation** attempts real proof-of-concept attacks and discards hypotheses that cannot be proven.
-- **Reporting** compiles validated findings, evidence, and remediation guidance into a final Markdown report.
+
+Stage by stage:
+
+1. **Recon and vulnerability analysis** explores the running application, ties runtime behavior back to the source, and runs specialized agents across Injection, XSS, SSRF, Authentication, and Authorization.
+2. **Agentic security code analysis** maps the application's architecture, trust boundaries, exposed interfaces, dependencies, data flows, and high-risk assets, then opens targeted investigations against them.
+3. **Finding reconciliation** merges both streams of candidates, deduplicates the overlap, and groups what remains into an exploitation queue.
+4. **Exploitation agents** attempt real proof-of-concept attacks against the running application.
+5. **Validation** throws out every candidate Shannon can't demonstrate.
+6. **Reporting** produces PDF and Markdown reports with the evidence attached, plus structured JSON and SARIF for downstream systems.
+
+Only live-validated vulnerabilities become Shannon pentest findings or count toward CI/CD severity gates.
 
 Each scan runs in an ephemeral Docker container with an isolated workspace and per-invocation orchestration.
 
@@ -190,41 +280,57 @@ Each scan runs in an ephemeral Docker container with an isolated workspace and p
 
 Use these guides for operational detail:
 
-| Guide | Use it for |
-| --- | --- |
-| [Source build and CLI commands](docs/development.md) | Cloning, building, common commands, output paths, and local development. |
-| [Configuration](docs/configuration.md) | Authenticated testing, login flows, rules of engagement, report filters, and rate-limit settings. |
-| [AI providers](docs/ai-providers.md) | Anthropic, AWS Bedrock, Google Vertex AI, and custom Anthropic-compatible endpoints. |
-| [Platforms and networking](docs/platforms.md) | Windows/WSL2, Linux, macOS, Docker networking, local apps, and custom hostnames. |
-| [Workspaces and resuming](docs/workspaces.md) | Naming workspaces, resuming interrupted scans, and workspace storage. |
-| [Safety and limitations](docs/safety.md) | Authorized-use requirements, non-production guidance, mutative effects, cost, and model caveats. |
-| [Coverage and roadmap](docs/coverage-roadmap.md) | Current vulnerability coverage and planned work. |
-| [Shannon Pro](docs/shannon-pro.md) | Commercial platform, black-box and white-box pentesting, full lifecycle workflows, and enterprise deployment. |
+
+| Guide                                                     | Use it for                                                                                                                                                                |
+| --------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| [Source build and CLI commands](docs/development.md)      | Cloning, building, common commands, output paths, and local development.                                                                                                  |
+| [Configuration](docs/configuration.md)                    | Authenticated testing, login flows, rules of engagement, and report filters.                                                                                              |
+| [AI providers](docs/ai-providers.md)                      | Selecting the model, the supported providers (Anthropic, OpenAI, xAI, AWS Bedrock, and any other Pi-supported provider), and custom gateways.                             |
+| [Platforms and networking](docs/platforms.md)             | Windows/WSL2, Linux, macOS, Docker networking, local apps, and custom hostnames.                                                                                          |
+| [Workspaces and resuming](docs/workspaces.md)             | Naming workspaces, resuming interrupted scans, and workspace storage.                                                                                                     |
+| [Safety and limitations](docs/safety.md)                  | Authorized-use requirements, non-production guidance, mutative effects, cost, and model caveats.                                                                          |
+| [Coverage and roadmap](docs/coverage-roadmap.md)          | Current vulnerability coverage and planned work.                                                                                                                          |
+| [Keygraph Enterprise Platform](docs/keygraph-platform.md) | Exhaustive agentic SAST, continuous pentesting, full-lifecycle finding management, remediation, targeted verification, enterprise governance, and on-premises deployment. |
+
+
+
 
 ## Safety, Scope, and Limitations
 
-Shannon Lite is not a passive scanner. Its exploitation agents can create users, submit forms, mutate application state, trigger outbound requests, and otherwise affect the target system. Use sandboxed, staging, or local development environments with disposable data.
+Shannon is not a passive scanner. Its exploitation agents can create users, submit forms, mutate application state, trigger outbound requests, and otherwise affect the target system. Use sandboxed, staging, or local development environments with disposable data.
 
-You are responsible for using Shannon Lite legally and ethically. Do not point Shannon Lite at systems, repositories, or applications you do not own or do not have explicit authorization to test.
+You are responsible for using Shannon legally and ethically. Do not point Shannon at systems, repositories, or applications you do not own or do not have explicit authorization to test.
 
 Important limitations:
 
-- Shannon Lite focuses on actively exploitable issues such as Injection, XSS, SSRF, Broken Authentication, and Broken Authorization. Broader static-analysis findings, including vulnerable dependencies and insecure configurations, are a core focus of Shannon Pro.
+- Shannon Open Source is tuned for fast, code-informed pentesting in everyday development and CI/CD. Exhaustive agentic SAST, broader scanner coverage, centralized governance, and full-lifecycle vulnerability management are delivered through the Keygraph Enterprise Platform.
 - Findings still require human review. LLM-generated reports can contain weakly supported or incorrect details.
-- Shannon Lite is officially supported with Claude models. Smaller, alternative, or proxied non-Claude models may be incomplete or unstable.
+- Anthropic, OpenAI, xAI, and AWS Bedrock are built-in providers, and any Anthropic Messages API or OpenAI Chat Completions or Responses API endpoint works through a custom base URL. Model capability varies, and a model that does not follow Shannon's instructions or tool-use constraints reliably will produce weaker results.
 - A full run can take roughly 1 to 1.5 hours and may incur LLM API costs depending on model pricing and application complexity.
-- Do not scan untrusted or adversarial codebases; AI-powered tools that read source code can be exposed to prompt injection.
+- Do not scan untrusted or adversarial codebases. AI-powered tools that read source code can be exposed to prompt injection.
 
-Read the full [Safety and limitations](docs/safety.md) guide before running Shannon Lite in a new environment.
+Read the full [Safety and limitations](docs/safety.md) guide before running Shannon in a new environment.
 
-## License and Enterprise Licensing
+## License
 
-Shannon Lite is licensed under the [GNU Affero General Public License v3.0](LICENSE).
+Shannon Open Source is licensed under the [GNU Affero General Public License v3.0](LICENSE).
 
-Commercial and enterprise licensing is available for organizations that need different license terms, commercial support, private redistribution, managed-service use, or broader deployment options.
+Commercial and enterprise licensing is available for organizations that need different license terms, commercial support, private redistribution, managed-service use, or broader deployment options, including the Keygraph platform.
 
 For commercial licensing, contact [shannon@keygraph.io](mailto:shannon@keygraph.io).
 
+## Acknowledgements
+
+Thanks to [Pi](https://github.com/earendil-works/pi),
+[Playwright CLI](https://github.com/microsoft/playwright-cli),
+and [Mantis](https://github.com/google/mantis).
+
+See [THIRD_PARTY_NOTICES.md](./THIRD_PARTY_NOTICES.md) for licensing and attribution details.
+
+## About Keygraph
+
+**Keygraph** is the company behind Shannon. It also builds the **Keygraph platform**, the commercial agentic pentesting product that closes the full AppSec lifecycle and runs an enhanced build of Shannon as its pentesting engine.
+
 ## Community and Support
 
 **Community office hours** are available for hands-on help with bugs, deployments, and configuration questions.
@@ -233,7 +339,7 @@ For commercial licensing, contact [shannon@keygraph.io](mailto:shannon@keygraph.
 - Asia: Thursday, 2:00 PM IST
 - [Book a slot](https://cal.com/george-flores-keygraph/shannon-community-office-hours)
 
-[Join Discord](https://discord.gg/cmctpMBXwE) to ask questions, share feedback, and connect with other Shannon Lite users.
+[Join Discord](https://discord.gg/cmctpMBXwE) to ask questions, share feedback, and connect with other Shannon users.
 
 At this time, Keygraph is not accepting external code contributions. Issues are welcome for bug reports and feature requests:
 
@@ -246,6 +352,36 @@ Stay connected:
 - [Twitter/X: @KeygraphHQ](https://twitter.com/KeygraphHQ)
 - [LinkedIn: Keygraph](https://linkedin.com/company/keygraph)
 
-<p align="center">
-  <b>Built by <a href="https://keygraph.io">Keygraph</a></b>
-</p>
+
+
+## Common Questions
+
+
+
+### Can I self-host Shannon?
+
+Yes. Shannon Open Source runs inside your infrastructure in an ephemeral worker container. It mounts the repository read-only and writes results to a local workspace.
+
+Keygraph never receives your source code and never proxies your model traffic. Your model requests go straight to the provider or endpoint you configure, and they carry source and application context with them. Point Shannon at a locally hosted endpoint and that traffic stays inside your environment too.
+
+### Does Shannon support bring your own key (BYOK)?
+
+Yes, always. You provide the LLM credentials Shannon uses to run a pentest, in every deployment, open source and commercial. Keygraph never proxies your model traffic.
+
+### Does Shannon output SARIF?
+
+Yes. Shannon emits SARIF 2.1.0, the OASIS standard format for static analysis results, alongside structured JSON. Any SARIF consumer reads it: code scanning services, vulnerability management platforms, security dashboards, and CI/CD pipelines. It is written by default on exploit-mode scans; set `report.sarif` to `"false"` in your configuration file to opt out.
+
+### Which AI providers does Shannon support?
+
+Anthropic, OpenAI, xAI, and AWS Bedrock are built in and configured directly by provider ID. Beyond those, Shannon runs on any endpoint that implements the Anthropic Messages API or the OpenAI Chat Completions or Responses API, reached through a custom base URL. The rule is the API format, not the vendor. Shannon uses a single unified model setting throughout a pentest.
+
+### Can I run Shannon on a local or self-hosted model?
+
+Shannon works with local models served through Ollama, vLLM, or LM Studio, which expose an OpenAI-compatible endpoint, as well as routers such as OpenRouter and gateways such as LiteLLM. Point Shannon at the endpoint with a custom base URL. Capability varies, and a model that does not follow Shannon's instructions or tool-use constraints reliably will produce weaker pentests than a frontier model, so take this path only if you know how your chosen model behaves. See [AI providers](docs/ai-providers.md#custom-base-url).
+
+### Does Shannon actually exploit vulnerabilities, or just scan?
+
+Shannon executes real exploits. It reports a finding only when it has produced a working proof-of-concept, and discards hypotheses it cannot prove. It is a pentester, not a passive scanner.
+
+**Built by [Keygraph](https://keygraph.io)**

THIRD_PARTY_NOTICES.md

@@ -0,0 +1,46 @@
+# Third-Party Notices
+
+Shannon incorporates and adapts material from third-party open-source projects.
+
+Shannon as a whole is distributed under the GNU Affero General Public License,
+version 3.0 (see LICENSE). Third-party material incorporated into Shannon
+remains subject to the attribution and notice requirements of its own license.
+
+## Pi
+
+Shannon uses Pi as part of its agent framework.
+
+Project: https://github.com/earendil-works/pi
+License: MIT
+
+Copyright (c) 2025 Mario Zechner
+
+The applicable license is reproduced at `LICENSES/MIT-Pi.txt`.
+
+## Mantis
+
+Portions of Shannon's Capella agentic SAST implementation, specifically the
+agent prompts, are derived from the Mantis project.
+
+- Project:         https://github.com/google/mantis
+- Upstream commit: 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52
+- Retrieved:       2026-08-25
+- License:         Apache License, Version 2.0
+
+The Apache License, Version 2.0 is reproduced at LICENSES/Apache-2.0.txt.
+
+The Mantis-derived files are individually marked with a provenance header and
+reside under:
+
+- apps/worker/prompts/partials/      (capella-*.hbs prompt partials)
+- apps/worker/prompts/sast/capella/  (prompt templates)
+
+The Mantis-derived material has been substantially modified by Keygraph
+for use within Shannon, including adaptation to Shannon's agent
+architecture and the Pi agent framework.
+
+Copyright and attribution notices from the original Mantis material
+remain the property of their respective copyright holders.
+
+Modifications:
+Copyright © 2026 Keygraph, Inc.

apps/cli/README.md

@@ -1,22 +1,60 @@
 <div align="center">
 
-<img src="https://raw.githubusercontent.com/KeygraphHQ/shannon/main/assets/github-banner.png" alt="Shannon — AI Pentester for Web Applications and APIs" width="100%">
+<img src="https://raw.githubusercontent.com/KeygraphHQ/shannon/main/assets/github-banner-light.png" alt="Shannon, AI Pentester for Web Apps and APIs, by Keygraph" width="100%">
 
-# Shannon — AI Pentester by Keygraph
+### Shannon is an autonomous, AI pentester for web applications and APIs.
 
-Shannon is an autonomous, white-box AI pentester for web applications and APIs. <br />
-It analyzes your source code, identifies attack vectors, and executes real exploits to prove vulnerabilities before they reach production.
+It analyzes your source code, identifies attack paths, and executes real exploits to prove vulnerabilities before they reach production.
+
+**This package is Shannon Open Source: the full agent, run locally from your command line.**
 
 ---
 
-<a href="https://github.com/KeygraphHQ/shannon/discussions/categories/announcements"><img src="https://raw.githubusercontent.com/KeygraphHQ/shannon/main/assets/announcements.png" height="40" alt="Announcements"></a>
-<a href="https://discord.gg/9ZqQPuhJB7"><img src="https://raw.githubusercontent.com/KeygraphHQ/shannon/main/assets/discord.png" height="40" alt="Join Discord"></a>
-<a href="https://keygraph.io/"><img src="https://raw.githubusercontent.com/KeygraphHQ/shannon/main/assets/Keygraph_Button.png" height="40" alt="Visit Keygraph.io"></a>
-<a href="https://www.linkedin.com/company/keygraph/"><img src="https://raw.githubusercontent.com/KeygraphHQ/shannon/main/assets/linkedin.png" height="40" alt="Follow Us on Linkedin"></a>
+<a href="https://discord.gg/9ZqQPuhJB7"><img src="https://raw.githubusercontent.com/KeygraphHQ/shannon/main/assets/discord_button_light.png" height="40" alt="Join Discord"></a>&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;<a href="https://keygraph.io/"><img src="https://raw.githubusercontent.com/KeygraphHQ/shannon/main/assets/keygraph_button_light.png" height="40" alt="Visit Keygraph.io"></a>
 
 ---
 
-**Full README and usage guide**  
-[https://github.com/KeygraphHQ/shannon#readme](https://github.com/KeygraphHQ/shannon#readme)
-
 </div>
+
+## Quick Start
+
+### Prerequisites
+
+- **Docker**: required for the worker container.
+- **Node.js 18+**: required for the recommended `npx` workflow.
+- **AI provider credentials**: Shannon runs on Anthropic, OpenAI, xAI, AWS Bedrock, any other provider in the harness catalogue, and any endpoint that speaks the Anthropic Messages API or the OpenAI Chat Completions or Responses API through a custom base URL. You bring your own key, and Keygraph never proxies your model traffic. Shannon is provider-agnostic.
+- **Cyber safeguards cleared with your provider**: Anthropic and OpenAI apply real-time safeguards to cyber-security workloads, which can interrupt a scan mid-run. Complete their guidance for legitimate security testers before your first run.
+
+### Run Shannon
+
+> **Warning:** Shannon actively executes exploits. Run it only against applications and environments you own or have explicit written authorization to test. Do not run Shannon against production systems.
+
+```bash
+# Configure credentials with the interactive wizard.
+npx @keygraph/shannon setup
+
+# Run a pentest against a source-available target.
+npx @keygraph/shannon start -u https://your-app.com -r /path/to/your-repo
+```
+
+Shannon pulls the worker image from Docker Hub, starts the required local infrastructure, mounts the target repository read-only inside an ephemeral worker container, and writes results to a local workspace.
+
+## Editions
+
+Shannon ships in two ways. **Shannon Open Source** is this package: the standalone pentester you run yourself, on demand, and complete in that lane. The **Keygraph platform** is the commercial product that runs an enhanced build of Shannon continuously and closes the full AppSec lifecycle around it - code analysis, finding management, automated remediation, verification, and enterprise deployment.
+
+## Documentation
+
+**Full README, guides, and usage documentation:** [github.com/KeygraphHQ/shannon](https://github.com/KeygraphHQ/shannon#readme)
+
+## License
+
+Shannon Open Source is licensed under the [GNU Affero General Public License v3.0](https://github.com/KeygraphHQ/shannon/blob/main/LICENSE).
+
+Commercial and enterprise licensing is available for organizations that need different license terms, commercial support, private redistribution, managed-service use, or broader deployment options, including the Keygraph platform.
+
+For commercial licensing, contact [shannon@keygraph.io](mailto:shannon@keygraph.io).
+
+<p align="center">
+  <b>Built by <a href="https://keygraph.io">Keygraph</a></b>
+</p>

apps/cli/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@keygraph/shannon",
   "version": "0.0.0",
-  "description": "Shannon - Autonomous white-box AI pentester for web applications and APIs by Keygraph",
+  "description": "Shannon is an autonomous white-box AI pentester for web applications and APIs, by Keygraph.",
   "type": "module",
   "main": "dist/index.mjs",
   "bin": {
@@ -18,6 +18,7 @@
   },
   "dependencies": {
     "@clack/prompts": "^1.1.0",
+    "@temporalio/client": "1.15.0",
     "chokidar": "^5.0.0",
     "dotenv": "^17.3.1",
     "smol-toml": "^1.6.1"
@@ -34,8 +35,12 @@
     "appsec",
     "keygraph"
   ],
-  "author": "",
+  "author": "Keygraph, Inc.",
   "license": "AGPL-3.0-only",
+  "bugs": {
+    "url": "https://github.com/KeygraphHQ/shannon/issues"
+  },
+  "homepage": "https://github.com/KeygraphHQ/shannon#readme",
   "repository": {
     "type": "git",
     "url": "git+https://github.com/KeygraphHQ/shannon.git",

apps/cli/src/args.ts

@@ -0,0 +1,106 @@
+/**
+ * Shared argument parsing for CLI commands.
+ *
+ * Every command declares which boolean flags, value options, and positionals it
+ * accepts; `parseArgs` resolves aliases, rejects anything unrecognized, and hands
+ * back a typed result. This centralizes the common flags (notably `--yes`/`-y`) so
+ * each command no longer re-hardcodes `args.includes('--yes')`, and it makes
+ * unknown flags and stray arguments fail loudly instead of being silently ignored.
+ */
+
+import { closestMatch } from './suggest.js';
+
+/** Thrown when argv does not match a command's schema. The dispatcher formats it. */
+export class ArgError extends Error {}
+
+/** Tokens that set the "skip confirmation" flag, declared once for every command. */
+export const YES_FLAGS = ['--yes', '-y'] as const;
+
+export interface ArgSchema {
+  /** Boolean flags: result key -> accepted tokens (canonical plus any aliases). */
+  readonly booleans?: Record<string, readonly string[]>;
+  /** Value-taking options: result key -> accepted tokens. */
+  readonly values?: Record<string, readonly string[]>;
+  /** Maximum positional arguments allowed. Defaults to 0. */
+  readonly maxPositionals?: number;
+  /** Extra guidance appended to the error when too many positionals are given. */
+  readonly positionalHint?: string;
+}
+
+export interface ParsedArgs {
+  readonly flags: Record<string, boolean>;
+  readonly values: Record<string, string>;
+  readonly positionals: readonly string[];
+}
+
+/** Build a token -> result-key lookup from a schema section. */
+function indexTokens(section: Record<string, readonly string[]>): Map<string, string> {
+  const byToken = new Map<string, string>();
+  for (const [key, tokens] of Object.entries(section)) {
+    for (const token of tokens) {
+      byToken.set(token, key);
+    }
+  }
+  return byToken;
+}
+
+export function parseArgs(argv: readonly string[], schema: ArgSchema): ParsedArgs {
+  const booleanByToken = indexTokens(schema.booleans ?? {});
+  const valueByToken = indexTokens(schema.values ?? {});
+  const maxPositionals = schema.maxPositionals ?? 0;
+
+  const flags: Record<string, boolean> = {};
+  const values: Record<string, string> = {};
+  const positionals: string[] = [];
+
+  for (let i = 0; i < argv.length; i++) {
+    const arg = argv[i];
+    if (arg === undefined) {
+      continue;
+    }
+
+    const equalsIndex = arg.startsWith('--') ? arg.indexOf('=') : -1;
+    const token = equalsIndex === -1 ? arg : arg.slice(0, equalsIndex);
+    const inlineValue = equalsIndex === -1 ? undefined : arg.slice(equalsIndex + 1);
+
+    const booleanKey = booleanByToken.get(token);
+    if (booleanKey !== undefined) {
+      if (inlineValue !== undefined) {
+        throw new ArgError(`Flag ${token} does not take a value`);
+      }
+      flags[booleanKey] = true;
+      continue;
+    }
+
+    const valueKey = valueByToken.get(token);
+    if (valueKey !== undefined) {
+      if (inlineValue !== undefined) {
+        values[valueKey] = inlineValue;
+        continue;
+      }
+      const next = argv[i + 1];
+      if (next === undefined || next.startsWith('-')) {
+        throw new ArgError(`Option ${token} requires a value`);
+      }
+      values[valueKey] = next;
+      i++;
+      continue;
+    }
+
+    if (arg.startsWith('-')) {
+      const suggestion = closestMatch(token, [...booleanByToken.keys(), ...valueByToken.keys()]);
+      const hint = suggestion ? `\nDid you mean '${suggestion}'?` : '';
+      throw new ArgError(`Unknown option: ${token}${hint}`);
+    }
+
+    positionals.push(arg);
+  }
+
+  if (positionals.length > maxPositionals) {
+    const extra = positionals[maxPositionals];
+    const hint = schema.positionalHint ? `\n${schema.positionalHint}` : '';
+    throw new ArgError(`Unexpected argument: ${extra}${hint}`);
+  }
+
+  return { flags, values, positionals };
+}

apps/cli/src/colors.ts

@@ -0,0 +1,34 @@
+/**
+ * ANSI color and style escapes — the single source for the CLI's palette.
+ *
+ * Codes are plain constants; callers decide whether to emit them via `paint`
+ * (wrap-and-reset) or `gate` (prefix-or-empty), gating on `supportsColor()` from
+ * `tty.ts`. Cursor-control escapes live with their sole consumer, not here — this
+ * module is color only.
+ */
+
+export const RESET = '\x1b[0m';
+
+/** Shannon brand gold — the running/completed accent, shared with the splash logo. */
+export const GOLD = '\x1b[38;2;244;197;66m';
+
+export const BOLD = '\x1b[1m';
+export const RED = '\x1b[31m';
+export const YELLOW = '\x1b[33m';
+export const DIM = '\x1b[90m';
+
+// The splash logo uses bolder variants of cyan/white/yellow than the progress tree.
+export const CYAN = '\x1b[36;1m';
+export const WHITE = '\x1b[1;37m';
+export const GRAY = '\x1b[0;37m';
+export const BOLD_YELLOW = '\x1b[1;33m';
+
+/** Wrap `text` in `code` and reset, or return it unchanged when color is off. */
+export function paint(text: string, code: string, enabled: boolean): string {
+  return enabled ? `${code}${text}${RESET}` : text;
+}
+
+/** A style code when color is on, or an empty string when off — for templates that interleave prefixes directly. */
+export function gate(code: string, enabled: boolean): string {
+  return enabled ? code : '';
+}

apps/cli/src/commands/build.ts

@@ -1,19 +1,20 @@
 /**
- * `shannon build` command — build the worker Docker image locally.
- * Only available in local mode (running from cloned repository).
+ * `shannon build` command — build the worker Docker image from the repository.
+ * Requires a clone (Dockerfile in the working directory).
  */
 
-import { buildImage } from '../docker.js';
-import { isLocal } from '../mode.js';
+import { buildImage, canBuildImage, ensureDocker } from '../docker.js';
+import { fail } from '../errors.js';
 
-export function build(noCache: boolean): void {
-  if (!isLocal()) {
-    console.error('ERROR: Build is only available when running from the Shannon repository');
-    console.error('  (Dockerfile not found in current directory)');
-    console.error('');
-    console.error('For npx usage, run: shannon update');
-    process.exit(1);
+export function build(noCache: boolean, version: string): void {
+  ensureDocker();
+
+  if (!canBuildImage()) {
+    fail(
+      'Build is only available when running from the Shannon repository',
+      '  (Dockerfile not found in current directory)',
+    );
   }
 
-  buildImage(noCache);
+  buildImage(noCache, version);
 }

apps/cli/src/commands/logs.ts

@@ -1,20 +1,89 @@
 /**
- * `shannon logs` command — tail a workspace's workflow log.
+ * `shannon logs` command — tail a scan's live log.
  *
- * Uses chokidar for reliable cross-platform file watching and
- * bounded synchronous reads to prevent duplicate output.
+ * The log file is streamed for its content and ends the tail on its own terminal marker
+ * (`Scan COMPLETED/PARTIAL/FAILED/CANCELLED`) or Ctrl-C. Temporal's workflow status is a backstop
+ * that also closes the tail when a worker dies without writing a marker — for interactive `logs` a
+ * Temporal outage is never fatal (it keeps tailing); only `start --follow` (CI) treats a sustained
+ * outage as a failure. Uses chokidar for reliable cross-platform file watching and bounded
+ * synchronous reads to prevent duplicate output.
  */
 
 import fs from 'node:fs';
 import path from 'node:path';
+import { StringDecoder } from 'node:string_decoder';
+import { setTimeout as sleep } from 'node:timers/promises';
 import { watch } from 'chokidar';
+import { fail } from '../errors.js';
 import { getWorkspacesDir } from '../home.js';
+import { resolveRunFile } from '../paths.js';
+import { resolveWorkflowId } from '../session.js';
+import { waitForWorkflowClose } from '../temporal-client.js';
+import { stdoutIsTerminal } from '../tty.js';
 
-// Match the exact line the worker writes — anchored to prevent false positives from agent output
-const COMPLETION_PATTERN = /^Workflow (COMPLETED|FAILED)$/m;
+const TERMINAL_HEADINGS = new Set(['Scan COMPLETED', 'Scan PARTIAL', 'Scan FAILED', 'Scan CANCELLED']);
 
-/** Read a byte range from a file and return it as a UTF-8 string. */
-function readRange(filePath: string, start: number, end: number): string {
+// The combined log resets completion on the bare `RESUMED` heading; a per-agent file carries the
+// distinct `--- RESUMED (<workflow id>) ---` boundary that WorkflowLogger.logResumeBoundary writes
+// (kept distinct per resume so it stays idempotent per file). Both mean a new execution began, so a
+// `--agent` tail must clear a stale terminal marker on either, matching the combined tail.
+const AGENT_RESUME_BOUNDARY = /^--- RESUMED \(.+\) ---$/u;
+
+function isResumeBoundary(line: string): boolean {
+  return line === 'RESUMED' || AGENT_RESUME_BOUNDARY.test(line);
+}
+
+/** Tracks only complete structural lines while output remains byte-for-byte unchanged. */
+export class LogCompletionState {
+  private pendingLine = '';
+  private terminalIsLastMarker = false;
+  private failureIsLastMarker = false;
+
+  ingest(chunk: string): void {
+    const lines = `${this.pendingLine}${chunk}`.split('\n');
+    this.pendingLine = lines.pop() ?? '';
+    for (const line of lines) {
+      if (isResumeBoundary(line)) {
+        this.terminalIsLastMarker = false;
+        this.failureIsLastMarker = false;
+      } else if (TERMINAL_HEADINGS.has(line)) {
+        this.terminalIsLastMarker = true;
+        this.failureIsLastMarker = line === 'Scan FAILED';
+      }
+    }
+  }
+
+  isComplete(): boolean {
+    return this.terminalIsLastMarker;
+  }
+
+  hasFailureMarker(): boolean {
+    return this.failureIsLastMarker;
+  }
+}
+
+/** Append the forced-stop marker after the worker has exited, unless this execution already ended. */
+export function appendCancellationFallback(logFile: string): void {
+  fs.mkdirSync(path.dirname(logFile), { recursive: true });
+  const state = new LogCompletionState();
+  try {
+    state.ingest(fs.readFileSync(logFile, 'utf8'));
+  } catch {
+    // A pre-registration stop may not have created the file yet.
+  }
+  if (state.isComplete()) return;
+
+  const descriptor = fs.openSync(logFile, 'a', 0o600);
+  try {
+    fs.writeSync(descriptor, '\nScan CANCELLED\n');
+    fs.fsyncSync(descriptor);
+  } finally {
+    fs.closeSync(descriptor);
+  }
+}
+
+/** Read a byte range without decoding across an arbitrary live-write boundary. */
+function readRange(filePath: string, start: number, end: number): Buffer {
   const length = end - start;
   const buffer = Buffer.alloc(length);
   const fd = fs.openSync(filePath, 'r');
@@ -23,84 +92,247 @@ function readRange(filePath: string, start: number, end: number): string {
   } finally {
     fs.closeSync(fd);
   }
-  return buffer.toString('utf-8');
+  return buffer;
 }
 
 /** Resolve a workspace ID to its workflow.log path, or exit with an error. */
-function resolveLogFile(workspaceId: string): string {
+export function resolveLogFile(workspaceId: string): string {
   const workspacesDir = getWorkspacesDir();
 
   // 1. Direct match
-  const directPath = path.join(workspacesDir, workspaceId, 'workflow.log');
+  const directPath = resolveRunFile(path.join(workspacesDir, workspaceId), 'workflow.log');
   if (fs.existsSync(directPath)) return directPath;
 
   // 2. Resume workflow ID (e.g. workspace_resume_123)
   const resumeBase = workspaceId.replace(/_resume_\d+$/, '');
   if (resumeBase !== workspaceId) {
-    const resumePath = path.join(workspacesDir, resumeBase, 'workflow.log');
+    const resumePath = resolveRunFile(path.join(workspacesDir, resumeBase), 'workflow.log');
     if (fs.existsSync(resumePath)) return resumePath;
   }
 
   // 3. Named workspace ID (e.g. workspace_shannon-123)
   const namedBase = workspaceId.replace(/_shannon-\d+$/, '');
   if (namedBase !== workspaceId) {
-    const namedPath = path.join(workspacesDir, namedBase, 'workflow.log');
+    const namedPath = resolveRunFile(path.join(workspacesDir, namedBase), 'workflow.log');
     if (fs.existsSync(namedPath)) return namedPath;
   }
 
-  console.error(`ERROR: Workflow log not found for: ${workspaceId}`);
-  console.error('');
-  console.error('Possible causes:');
-  console.error("  - Workflow hasn't started yet");
-  console.error('  - Workspace ID is incorrect');
-  console.error('');
-  console.error('Check the Temporal Web UI at http://localhost:8233 for workflow details');
-  process.exit(1);
+  fail(
+    `No scan found named: ${workspaceId}`,
+    '',
+    'Possible causes:',
+    "  - The scan hasn't started yet",
+    '  - The workspace name is incorrect',
+    '',
+    'Check the dashboard at http://localhost:8233 for scan details',
+  );
 }
 
-export function logs(workspaceId: string): void {
-  const logFile = resolveLogFile(workspaceId);
-  let position = 0;
-
+export interface TailOptions {
+  /** Workflow whose Temporal status can end the tail (alongside the file's own terminal marker). */
+  readonly workflowId?: string;
+  /** Called if the tail ends because Temporal became unreachable, with the captured error. */
+  readonly onUnreachable?: (lastError: string) => void;
   /**
-   * Output any new content appended since the last read.
-   * Returns true when the workflow completion marker is detected.
+   * Consecutive Temporal-outage polls before the watch gives up. Interactive `logs` passes
+   * Infinity so a blip never ends the tail (the file marker or Ctrl-C do); `start --follow` (CI)
+   * leaves it bounded so a genuinely dead Temporal fails the run instead of hanging.
    */
-  function flush(): boolean {
-    try {
-      const { size } = fs.statSync(logFile);
-      if (size <= position) return false;
+  readonly maxConnectFailures?: number;
+}
 
-      const data = readRange(logFile, position, size);
-      process.stdout.write(data);
-      position = size;
+/** Outcome of a tail: whether the streamed log already contained the worker's `Scan FAILED` block. */
+export interface TailResult {
+  readonly sawFailure: boolean;
+}
 
-      return COMPLETION_PATTERN.test(data);
-    } catch {
-      // File deleted or unreadable — treat as done
-      return true;
+/**
+ * Stream a scan's log to the terminal until the file shows a terminal marker, the workflow closes,
+ * or Ctrl-C. A Temporal outage is warned about; if it reaches `maxConnectFailures` the tail ends
+ * with a diagnostic (bounded for `start --follow`), but interactive `logs` sets that to Infinity so
+ * an outage keeps tailing. Never exits the process itself. Reports whether the log already showed
+ * the failure, so a caller need not print it a second time.
+ */
+export function tailUntilComplete(logFile: string, opts: TailOptions = {}): Promise<TailResult> {
+  return new Promise((resolve) => {
+    let position = 0;
+    const completion = new LogCompletionState();
+    const completionDecoder = new StringDecoder('utf8');
+    let done = false;
+    const controller = new AbortController();
+    let watcher: ReturnType<typeof watch> | undefined;
+
+    /** Output any new content appended since the last read. */
+    function flush(): boolean {
+      try {
+        const { size } = fs.statSync(logFile);
+        if (size <= position) return completion.isComplete();
+        const data = readRange(logFile, position, size);
+        process.stdout.write(data);
+        position = size;
+        completion.ingest(completionDecoder.write(data));
+        return completion.isComplete();
+      } catch {
+        // File not present yet or transiently unreadable — nothing to flush this round.
+        return false;
+      }
     }
+
+    function finish(): void {
+      if (done) return;
+      done = true;
+      controller.abort();
+      process.off('SIGINT', finish);
+      const result = { sawFailure: completion.hasFailureMarker() };
+      if (watcher) {
+        watcher.close().finally(() => resolve(result));
+        // Safety net — resolve anyway if watcher.close() stalls.
+        setTimeout(() => resolve(result), 1000).unref();
+      } else {
+        resolve(result);
+      }
+    }
+
+    // 1. Output existing content, then stream anything appended. A per-agent file can be created
+    //    after the watcher starts, so `add` is handled too and streams it from its first line.
+    //    The file's own `Scan COMPLETED/PARTIAL/FAILED/CANCELLED` marker ends the tail on its own —
+    //    a Temporal round-trip is a backstop for a worker that dies without writing one, not the
+    //    only way to stop.
+    watcher = watch(logFile, { persistent: true });
+    const onFsEvent = (): void => {
+      if (flush()) finish();
+    };
+    watcher.on('change', onFsEvent);
+    watcher.on('add', onFsEvent);
+    if (flush()) {
+      finish();
+      return;
+    }
+
+    // 2. Ctrl-C stops watching.
+    process.once('SIGINT', finish);
+
+    // 3. Temporal backstops completion for a worker that dies without a marker. Without a workflow
+    //    id, the tail relies on the file marker or Ctrl-C alone.
+    if (opts.workflowId) {
+      waitForWorkflowClose(opts.workflowId, {
+        signal: controller.signal,
+        ...(opts.maxConnectFailures !== undefined ? { maxConnectFailures: opts.maxConnectFailures } : {}),
+        onConnectionTrouble: (lastError) => {
+          if (!done) console.error(`\n⚠ Lost contact with Temporal, retrying… (${lastError})`);
+        },
+        onReconnected: () => {
+          if (!done) console.error('  Reconnected to Temporal.');
+        },
+      })
+        .then(async (end) => {
+          if (done) return;
+          // Flush, let a just-written final summary land, then flush the tail once more.
+          flush();
+          await sleep(750).catch(() => {});
+          flush();
+          if (end.reason === 'unreachable') {
+            console.error('\nScan watch aborted: lost contact with Temporal.');
+            console.error(`  Last error: ${end.lastError}`);
+            console.error('  Temporal may have crashed — check `docker compose logs temporal`.');
+            opts.onUnreachable?.(end.lastError);
+          }
+          finish();
+        })
+        .catch(() => {
+          // waitForWorkflowClose never rejects; guard only against an aborted race.
+        });
+    }
+  });
+}
+
+/** The `.shannon/agents/` directory that sits beside a scan's combined workflow.log. */
+function agentsDirFor(logFile: string): string {
+  return path.join(path.dirname(logFile), 'agents');
+}
+
+/** List the per-agent log names available for a scan (filename stems, sorted), or an empty list. */
+export function listAgentLogNames(logFile: string): string[] {
+  try {
+    return fs
+      .readdirSync(agentsDirFor(logFile))
+      .filter((entry) => entry.endsWith('.log'))
+      .map((entry) => entry.slice(0, -'.log'.length))
+      .sort();
+  } catch {
+    return [];
   }
+}
 
-  console.log(`Tailing workflow log: ${logFile}`);
+/**
+ * Resolve an agent name to its per-agent log path. The name must be a closed-charset basename, and
+ * the resolved file must stay inside the agents directory: traversal and symlink escapes are
+ * rejected. Returns undefined when the name is structurally invalid or escapes the directory.
+ */
+export function resolveAgentLogFile(logFile: string, agentName: string): string | undefined {
+  if (!/^[a-z0-9][a-z0-9-]{0,63}$/u.test(agentName)) return undefined;
+  const agentsDir = agentsDirFor(logFile);
+  const target = path.join(agentsDir, `${agentName}.log`);
+  try {
+    const realDir = fs.realpathSync(agentsDir);
+    const realTarget = fs.realpathSync(target);
+    if (realTarget !== path.join(realDir, `${agentName}.log`)) return undefined;
+  } catch {
+    // The file does not exist yet (scan still starting); the closed-charset check already proved
+    // the path cannot traverse out of the agents directory, so it is safe to watch for creation.
+  }
+  return target;
+}
 
-  // 1. Output existing content
-  if (flush()) {
+export interface LogsOptions {
+  readonly agent?: string;
+  readonly listAgents?: boolean;
+}
+
+function tailFileToExit(logFile: string, workflowId: string | undefined, label: string): void {
+  console.error(stdoutIsTerminal() ? `${label}: ${logFile}` : label);
+  tailUntilComplete(logFile, {
+    ...(workflowId ? { workflowId } : {}),
+    // Interactive tail: a Temporal outage must never end the session. The file's terminal marker or
+    // Ctrl-C stop it; Temporal stays a soft backstop that reconnects and closes the tail on a
+    // silent worker death, but its unreachability is never fatal here.
+    maxConnectFailures: Number.POSITIVE_INFINITY,
+  }).finally(() => process.exit(0));
+}
+
+export function logs(workspaceId: string, options: LogsOptions = {}): void {
+  const logFile = resolveLogFile(workspaceId);
+
+  if (options.listAgents) {
+    const names = listAgentLogNames(logFile);
+    if (names.length === 0) {
+      console.error('No per-agent logs for this scan yet.');
+      process.exit(0);
+    }
+    for (const name of names) console.log(name);
     process.exit(0);
   }
 
-  // 2. Watch for appended content via chokidar
-  const watcher = watch(logFile, { persistent: true });
+  const workflowId = resolveWorkflowId(workspaceId);
 
-  const shutdown = (): void => {
-    watcher.close().finally(() => process.exit(0));
-    // Safety net — force exit if watcher.close() stalls
-    setTimeout(() => process.exit(0), 1000).unref();
-  };
+  if (options.agent !== undefined) {
+    const agentFile = resolveAgentLogFile(logFile, options.agent);
+    if (agentFile === undefined) {
+      fail(`No agent log named: ${options.agent}`, '', 'Available agents:', ...withBullets(listAgentLogNames(logFile)));
+    }
+    const known = listAgentLogNames(logFile);
+    // If the directory already lists agents, a name not among them is a typo, not a not-yet-created
+    // file; fail loudly rather than tailing a path that will never appear.
+    if (known.length > 0 && !known.includes(options.agent)) {
+      fail(`No agent log named: ${options.agent}`, '', 'Available agents:', ...withBullets(known));
+    }
+    tailFileToExit(agentFile, workflowId, `Tailing ${options.agent} log`);
+    return;
+  }
 
-  watcher.on('change', () => {
-    if (flush()) shutdown();
-  });
+  tailFileToExit(logFile, workflowId, 'Tailing scan log');
+}
 
-  process.on('SIGINT', shutdown);
+function withBullets(names: readonly string[]): string[] {
+  return names.length === 0 ? ['  (none yet)'] : names.map((name) => `  - ${name}`);
 }

apps/cli/src/commands/reset.ts

@@ -0,0 +1,26 @@
+/**
+ * `shannon reset` command — stop everything and wipe all Temporal data and volumes,
+ * returning the machine to a clean slate. The destructive counterpart to `stop`.
+ */
+
+import * as p from '@clack/prompts';
+import { confirmByTyping } from '../confirm.js';
+import { ensureDocker, runningContainers, stopContainers, stopInfra, WORKER_FILTER } from '../docker.js';
+
+export async function reset(): Promise<void> {
+  ensureDocker();
+
+  console.log('This will stop all running scans and permanently remove all Temporal data and volumes.');
+  await confirmByTyping('reset', 'confirm');
+
+  const spinner = p.spinner();
+  spinner.start('Stopping scans');
+  const running = runningContainers(WORKER_FILTER);
+  await stopContainers(running);
+  spinner.stop(
+    running.length > 0 ? `Stopped ${running.length} scan${running.length === 1 ? '' : 's'}` : 'No scans running',
+  );
+
+  await stopInfra(true);
+  console.log('Reset complete.');
+}

apps/cli/src/commands/scans.ts

@@ -0,0 +1,229 @@
+/**
+ * `shannon scans` command — list scans, running and completed, and where each report lives.
+ *
+ * Running scans come from Docker: every worker container is stamped with the shannon.workspace
+ * label, so `runningScanWorkspaces()` is the authoritative live-scan list (shared with `stop`).
+ * A scan counts as completed once it has produced a report; the report can live in any of a few
+ * locations depending on the version that ran it, so `findReport` probes them in order and the
+ * first hit is both the completion signal and the link target behind the workspace name. Dates and
+ * durations come from each run's session.json (createdAt/completedAt), with the report file's mtime
+ * as the date fallback for runs that lack a recorded time; a running scan's duration is elapsed time
+ * so far (now − createdAt).
+ *
+ * Running scans are listed first, then completed newest-first. Human-readable by default; `--json`
+ * emits the same rows as raw machine values on stdout.
+ *
+ * The completed list is filesystem-only (local ./workspaces/ or npx ~/.shannon/workspaces/ via
+ * getWorkspacesDir); the running list needs Docker but degrades to empty when the daemon is down,
+ * which is the correct answer (no scan can be running then).
+ */
+
+import fs from 'node:fs';
+import path from 'node:path';
+import { pathToFileURL } from 'node:url';
+import { BOLD, CYAN, GOLD, paint } from '../colors.js';
+import { runningScanWorkspaces } from '../docker.js';
+import { getWorkspacesDir } from '../home.js';
+import { commandPrefix } from '../mode.js';
+import { FINAL_REPORT_PDF_FILENAME, INTERNAL_DIR, resolveRunFile } from '../paths.js';
+import { stdoutIsTerminal, supportsColor } from '../tty.js';
+
+/** Assembled report in the deliverables dir. Must match ASSEMBLED_REPORT_FILENAME in the worker package. */
+const ASSEMBLED_REPORT_FILENAME = 'comprehensive_security_assessment_report.md';
+
+/** Run-root markdown surfaced by older versions, before the PDF. Kept so those runs still list. */
+const FINAL_REPORT_MD_FILENAME = 'Security-Assessment-Report.md';
+
+const DELIVERABLES_SUBDIR = 'deliverables';
+
+/** One scan, running or completed; raw values so the table and --json render from one source. */
+interface ScanRow {
+  readonly workspace: string;
+  readonly state: 'running' | 'completed';
+  /** Completion time in ms — sort key and date source. Null while a scan is still running. */
+  readonly finishedMs: number | null;
+  /** Wall-clock duration in ms: elapsed-so-far for running, total for completed. Null when unknown. */
+  readonly durationMs: number | null;
+  /** Absolute path to the report file — the link target behind the workspace name. Null while running. */
+  readonly report: string | null;
+}
+
+/** The --json row shape: raw machine values, one per scan. */
+interface JsonRow {
+  readonly workspace: string;
+  readonly state: 'running' | 'completed';
+  readonly finishedAt: string | null;
+  readonly durationMs: number | null;
+  readonly reportPath: string | null;
+}
+
+/** Compact wall-clock duration from milliseconds: "47s", "1m 32s", "1h 47m". */
+function formatDuration(ms: number): string {
+  const totalSeconds = Math.round(ms / 1000);
+  if (totalSeconds < 60) {
+    return `${totalSeconds}s`;
+  }
+  const totalMinutes = Math.floor(totalSeconds / 60);
+  if (totalMinutes < 60) {
+    return `${totalMinutes}m ${totalSeconds % 60}s`;
+  }
+  return `${Math.floor(totalMinutes / 60)}h ${totalMinutes % 60}m`;
+}
+
+/**
+ * Wrap `text` in an OSC 8 hyperlink to `url` so a supporting terminal opens it on click,
+ * or return `text` unchanged. Terminals without OSC 8 simply show the text.
+ */
+function hyperlink(text: string, url: string): string {
+  return `\x1b]8;;${url}\x1b\\${text}\x1b]8;;\x1b\\`;
+}
+
+/** First existing report path for a run (newest-surfaced first), or null if it has none. */
+function findReport(runDir: string): string | null {
+  const candidates = [
+    path.join(runDir, FINAL_REPORT_PDF_FILENAME),
+    path.join(runDir, FINAL_REPORT_MD_FILENAME),
+    path.join(runDir, INTERNAL_DIR, DELIVERABLES_SUBDIR, ASSEMBLED_REPORT_FILENAME),
+    path.join(runDir, DELIVERABLES_SUBDIR, ASSEMBLED_REPORT_FILENAME),
+  ];
+
+  for (const candidate of candidates) {
+    if (fs.existsSync(candidate)) {
+      return candidate;
+    }
+  }
+  return null;
+}
+
+interface SessionData {
+  readonly session: { readonly createdAt?: string; readonly completedAt?: string };
+}
+
+/** Read a run's session.json (dual-read across layouts). Missing or unreadable → empty shape. */
+function readSession(runDir: string): SessionData {
+  try {
+    const parsed = JSON.parse(fs.readFileSync(resolveRunFile(runDir, 'session.json'), 'utf8'));
+    return { session: parsed?.session ?? {} };
+  } catch {
+    return { session: {} };
+  }
+}
+
+/** Gather every workspace that has a report, one row each. */
+function collectCompletedScans(workspacesDir: string): ScanRow[] {
+  let entries: fs.Dirent[];
+  try {
+    entries = fs.readdirSync(workspacesDir, { withFileTypes: true });
+  } catch {
+    // Workspaces directory does not exist yet — no scans have ever run.
+    return [];
+  }
+
+  const rows: ScanRow[] = [];
+  for (const entry of entries) {
+    if (!entry.isDirectory()) {
+      continue;
+    }
+
+    const runDir = path.join(workspacesDir, entry.name);
+    const reportPath = findReport(runDir);
+    if (!reportPath) {
+      continue;
+    }
+
+    const { session } = readSession(runDir);
+    const completedMs = Date.parse(session.completedAt ?? '');
+    const createdMs = Date.parse(session.createdAt ?? '');
+    const finishedMs = Number.isNaN(completedMs) ? fs.statSync(reportPath).mtimeMs : completedMs;
+    const durationMs = Number.isNaN(completedMs) || Number.isNaN(createdMs) ? null : completedMs - createdMs;
+
+    rows.push({ workspace: entry.name, state: 'completed', finishedMs, durationMs, report: reportPath });
+  }
+  return rows;
+}
+
+/** Gather every currently-running scan, one row each. Elapsed time is now − createdAt. */
+function collectRunningScans(workspacesDir: string, nowMs: number): ScanRow[] {
+  const rows: ScanRow[] = [];
+  for (const workspace of runningScanWorkspaces()) {
+    const { session } = readSession(path.join(workspacesDir, workspace));
+    const createdMs = Date.parse(session.createdAt ?? '');
+    const durationMs = Number.isNaN(createdMs) ? null : nowMs - createdMs;
+    rows.push({ workspace, state: 'running', finishedMs: null, durationMs, report: null });
+  }
+  return rows;
+}
+
+function toJsonRow(row: ScanRow): JsonRow {
+  return {
+    workspace: row.workspace,
+    state: row.state,
+    finishedAt: row.finishedMs === null ? null : new Date(row.finishedMs).toISOString(),
+    durationMs: row.durationMs,
+    reportPath: row.report,
+  };
+}
+
+/** Print the scans as an aligned table with each completed workspace name linked to its report. */
+function printTable(workspacesDir: string, rows: readonly ScanRow[]): void {
+  if (rows.length === 0) {
+    const prefix = commandPrefix();
+    console.log(`No scans yet. Run '${prefix} start -u <url> -r <path>' to begin.`);
+    return;
+  }
+
+  const color = supportsColor();
+  // On a terminal a completed workspace name is an OSC 8 hyperlink that opens its report; when
+  // piped, or for a running scan that has no report yet, it prints as plain text.
+  const linkable = stdoutIsTerminal();
+
+  const table = rows.map((row) => ({
+    state: row.state === 'running' ? 'RUNNING' : 'COMPLETED',
+    finished: row.finishedMs === null ? '—' : new Date(row.finishedMs).toISOString().slice(0, 10),
+    duration: row.durationMs === null ? '—' : formatDuration(row.durationMs),
+    workspace: row.workspace,
+    report: row.report,
+  }));
+
+  const stateWidth = Math.max('STATE'.length, ...table.map((row) => row.state.length));
+  const dateWidth = Math.max('FINISHED'.length, 'YYYY-MM-DD'.length);
+  const durationWidth = Math.max('DURATION'.length, ...table.map((row) => row.duration.length));
+
+  console.log(`\nScans in ${workspacesDir}:\n`);
+  const header = `${'STATE'.padEnd(stateWidth)}  ${'FINISHED'.padEnd(dateWidth)}  ${'DURATION'.padEnd(durationWidth)}  WORKSPACE`;
+  console.log(paint(header, BOLD, color));
+
+  for (const row of table) {
+    const stateText = row.state.padEnd(stateWidth);
+    const state = row.state === 'RUNNING' ? paint(stateText, CYAN, color) : stateText;
+    const finished = row.finished.padEnd(dateWidth);
+    const duration = row.duration.padEnd(durationWidth);
+    // A running scan has no report to open, so its name stays plain; completed names are linked.
+    const name = row.report ? paint(row.workspace, GOLD, color) : row.workspace;
+    const workspace = row.report && linkable ? hyperlink(name, pathToFileURL(row.report).href) : name;
+    console.log(`${state}  ${finished}  ${duration}  ${workspace}`);
+  }
+  console.log('');
+}
+
+export function scans(opts: { readonly json: boolean }): void {
+  const workspacesDir = getWorkspacesDir();
+  const nowMs = Date.now();
+
+  const running = collectRunningScans(workspacesDir, nowMs);
+  const runningNames = new Set(running.map((row) => row.workspace));
+  // A running scan has no final report, so it can't also be completed; guard anyway.
+  const completed = collectCompletedScans(workspacesDir).filter((row) => !runningNames.has(row.workspace));
+
+  // Running scans on top (most recently started first), then completed newest-first.
+  running.sort((a, b) => (a.durationMs ?? 0) - (b.durationMs ?? 0));
+  completed.sort((a, b) => (b.finishedMs ?? 0) - (a.finishedMs ?? 0));
+  const rows = [...running, ...completed];
+
+  if (opts.json) {
+    console.log(JSON.stringify(rows.map(toJsonRow), null, 2));
+    return;
+  }
+
+  printTable(workspacesDir, rows);
+}

apps/cli/src/commands/setup.ts

@@ -1,61 +1,167 @@
 /**
- * `shn setup` — interactive TUI wizard for one-time credential configuration.
+ * `npx @keygraph/shannon setup` — interactive TUI wizard for one-time credential configuration.
  *
- * Walks the user through selecting a provider and entering credentials,
- * then persists everything to ~/.shannon/config.toml with 0o600 permissions.
+ * Walks the user through selecting a provider, entering credentials, and naming
+ * the model that runs the whole scan, then persists everything to
+ * ~/.shannon/config.toml with 0o600 permissions.
  */
 
-import fs from 'node:fs';
 import os from 'node:os';
 import path from 'node:path';
 import * as p from '@clack/prompts';
 import { type ShannonConfig, saveConfig } from '../config/writer.js';
+import { CURATED_PROVIDERS, type CuratedProviderId, isCuratedProvider, type OpenAiFormat } from '../model-spec.js';
+import { displaySplash } from '../splash.js';
+import { requireInteractive } from '../tty.js';
+import { getVersion } from '../version.js';
 
 const SHANNON_HOME = path.join(os.homedir(), '.shannon');
 
-type Provider = 'anthropic' | 'custom_base_url' | 'bedrock' | 'vertex';
+const CUSTOM_MODEL = '__custom__';
+const CUSTOM_BASE_URL = '__custom_base_url__';
+const OTHER_PROVIDER = '__other_provider__';
+
+/**
+ * Wire formats reachable through the gateway route. The format picks the provider
+ * that supplies the credential, and for OpenAI it also picks which of the two
+ * OpenAI APIs Shannon calls.
+ */
+const GATEWAY_DIALECTS: readonly {
+  value: string;
+  label: string;
+  provider: 'anthropic' | 'openai';
+  format?: OpenAiFormat;
+}[] = [
+  { value: 'anthropic', label: 'Anthropic Messages', provider: 'anthropic' },
+  {
+    value: 'openai-chat-completions',
+    label: 'OpenAI Chat Completions',
+    provider: 'openai',
+    format: 'chat-completions',
+  },
+  { value: 'openai-responses', label: 'OpenAI Responses', provider: 'openai', format: 'responses' },
+];
+
+/** Suggested models per curated provider, best-first. Free-text entry accepts any model in the provider's catalogue. */
+const MODEL_SUGGESTIONS: Readonly<Record<CuratedProviderId, readonly string[]>> = {
+  anthropic: ['claude-sonnet-4-6', 'claude-opus-4-8', 'claude-opus-4-7', 'claude-haiku-4-5-20251001'],
+  openai: ['gpt-5.6-sol', 'gpt-5.5', 'gpt-5.4'],
+  xai: ['grok-4.5'],
+  'amazon-bedrock': ['us.anthropic.claude-sonnet-4-6', 'us.anthropic.claude-opus-4-8', 'us.anthropic.claude-opus-4-7'],
+};
+
+/** Placeholder shown in the free-text model ID prompt, per curated provider. */
+const MODEL_ID_PLACEHOLDER: Readonly<Record<CuratedProviderId, string>> = {
+  anthropic: 'claude-sonnet-4-6',
+  openai: 'gpt-5.6-sol',
+  xai: 'grok-4.5',
+  'amazon-bedrock': 'us.anthropic.claude-opus-4-8',
+};
+
+/** Model ID placeholder for a provider, absent when the provider is not curated. */
+function modelIdPlaceholder(provider: string): string | undefined {
+  return isCuratedProvider(provider) ? MODEL_ID_PLACEHOLDER[provider] : undefined;
+}
 
 export async function setup(): Promise<void> {
-  p.intro('Shannon Setup');
+  requireInteractive('setup', 'For non-interactive use, export credentials as env vars (e.g. ANTHROPIC_API_KEY).');
+  displaySplash(getVersion());
+  p.intro('Setup');
 
-  // 1. Select provider
-  const provider = await p.select({
+  // 1. Select provider. "Custom Base URL" is a route, not a provider — it asks
+  //    which API dialect the gateway speaks and configures that provider. "Other
+  //    provider" reaches any pi-supported provider Shannon does not curate.
+  const selected = await p.select({
     message: 'Select your AI provider',
     options: [
-      { value: 'anthropic' as const, label: 'Claude Direct', hint: 'recommended' },
-      { value: 'custom_base_url' as const, label: 'Custom Base URL', hint: 'proxies, gateways' },
-      { value: 'bedrock' as const, label: 'Claude via AWS Bedrock' },
-      { value: 'vertex' as const, label: 'Claude via Google Vertex AI' },
+      { value: 'anthropic' as const, label: 'Anthropic', hint: 'Claude models - recommended' },
+      { value: 'openai' as const, label: 'OpenAI', hint: 'GPT models' },
+      { value: 'xai' as const, label: 'xAI', hint: 'Grok models' },
+      { value: 'amazon-bedrock' as const, label: 'AWS Bedrock', hint: 'Claude models via AWS' },
+      { value: CUSTOM_BASE_URL as typeof CUSTOM_BASE_URL, label: 'Custom Base URL', hint: 'your own proxy or gateway' },
+      {
+        value: OTHER_PROVIDER as typeof OTHER_PROVIDER,
+        label: 'Other provider',
+        hint: 'any other Pi-supported provider',
+      },
     ],
   });
-  if (p.isCancel(provider)) return cancelAndExit();
+  if (p.isCancel(selected)) return cancelAndExit();
 
-  const config = await setupProvider(provider as Provider);
+  // 2. Credentials — and, on the gateway route, the endpoint and its dialect.
+  const { provider, config, gateway } = await setupSelection(selected);
 
-  // 2. Adaptive thinking
-  await maybePromptAdaptiveThinking(config);
+  // 3. The model that runs every phase.
+  const modelId = await promptModel(provider);
+  config.core = { ...config.core, model: `${provider}:${modelId}` };
+  if (gateway) config.core = { ...config.core, base_url: gateway.baseUrl };
 
-  // 3. Save config
   saveConfig(config);
 
   const configPath = path.join(SHANNON_HOME, 'config.toml');
+  const summary = [`Provider   ${provider}`, `Model      ${modelId}`];
+  if (gateway) summary.push(`Endpoint   ${gateway.baseUrl}`);
+  if (gateway?.format) summary.push(`API        ${gateway.format}`);
+
   p.log.success(`Configuration saved to ${configPath}`);
+  p.log.info(summary.join('\n'));
   p.outro('Run `npx @keygraph/shannon start` to begin a scan.');
 }
 
-async function setupProvider(provider: Provider): Promise<ShannonConfig> {
+interface Selection {
+  provider: string;
+  config: ShannonConfig;
+  gateway?: GatewaySetup;
+}
+
+/** Resolve the provider selection into a provider id and its credential config. */
+async function setupSelection(
+  selected: CuratedProviderId | typeof CUSTOM_BASE_URL | typeof OTHER_PROVIDER,
+): Promise<Selection> {
+  if (selected === CUSTOM_BASE_URL) {
+    const gateway = await setupGateway();
+    return { provider: gateway.provider, config: gateway.config, gateway };
+  }
+  if (selected === OTHER_PROVIDER) {
+    return setupOtherProvider();
+  }
+  return { provider: selected, config: await setupProvider(selected) };
+}
+
+async function setupProvider(provider: CuratedProviderId): Promise<ShannonConfig> {
   switch (provider) {
+    case 'amazon-bedrock':
+      return setupBedrock();
     case 'anthropic':
       return setupAnthropic();
-    case 'custom_base_url':
-      return setupCustomBaseUrl();
-    case 'bedrock':
-      return setupBedrock();
-    case 'vertex':
-      return setupVertex();
+    case 'openai':
+      return { openai: { api_key: await promptSecret('Enter your OpenAI API key') } };
+    case 'xai':
+      return { xai: { api_key: await promptSecret('Enter your xAI API key') } };
   }
 }
 
+/**
+ * Any pi provider Shannon does not curate. The id is free text — the worker's
+ * preflight validates it — and the key is stored generically as SHANNON_AI_API_KEY.
+ */
+async function setupOtherProvider(): Promise<Selection> {
+  p.log.info('Browse supported providers and models at https://pi.dev/models');
+  const provider = await p.text({
+    message: 'Provider ID',
+    validate: (value) => {
+      const id = value?.trim();
+      if (!id) return 'Provider ID is required';
+      if (isCuratedProvider(id)) return `${id} has its own option.`;
+      return undefined;
+    },
+  });
+  if (p.isCancel(provider)) return cancelAndExit();
+
+  const apiKey = await promptSecret('Enter the API key');
+  return { provider: provider.trim(), config: { provider: { api_key: apiKey } } };
+}
+
 // === Provider Setup Flows ===
 
 async function setupAnthropic(): Promise<ShannonConfig> {
@@ -68,112 +174,13 @@ async function setupAnthropic(): Promise<ShannonConfig> {
   });
   if (p.isCancel(authMethod)) return cancelAndExit();
 
-  const config: ShannonConfig = {};
-
   if (authMethod === 'oauth') {
     const token = await promptSecret('Enter your OAuth token');
-    config.anthropic = { oauth_token: token };
-  } else {
-    const apiKey = await promptSecret('Enter your Anthropic API key');
-    config.anthropic = { api_key: apiKey };
+    return { anthropic: { oauth_token: token } };
   }
 
-  const customizeModels = await p.confirm({
-    message:
-      'Do you want to change the default models?\n' +
-      '    Small  - claude-haiku-4-5-20251001\n' +
-      '    Medium - claude-sonnet-4-6\n' +
-      '    Large  - claude-opus-4-7',
-    initialValue: false,
-  });
-  if (p.isCancel(customizeModels)) return cancelAndExit();
-
-  if (customizeModels) {
-    const small = await p.text({
-      message: 'Small model ID',
-      initialValue: 'claude-haiku-4-5-20251001',
-      validate: required('Small model ID is required'),
-    });
-    if (p.isCancel(small)) return cancelAndExit();
-
-    const medium = await p.text({
-      message: 'Medium model ID',
-      initialValue: 'claude-sonnet-4-6',
-      validate: required('Medium model ID is required'),
-    });
-    if (p.isCancel(medium)) return cancelAndExit();
-
-    const large = await p.text({
-      message: 'Large model ID',
-      initialValue: 'claude-opus-4-7',
-      validate: required('Large model ID is required'),
-    });
-    if (p.isCancel(large)) return cancelAndExit();
-
-    config.models = { small, medium, large };
-  }
-
-  return config;
-}
-
-async function setupCustomBaseUrl(): Promise<ShannonConfig> {
-  const baseUrl = await p.text({
-    message: 'Endpoint URL',
-    placeholder: 'https://your-proxy.example.com',
-    validate: (value) => {
-      if (!value) return 'Endpoint URL is required';
-      try {
-        new URL(value);
-      } catch {
-        return 'Must be a valid URL';
-      }
-      return undefined;
-    },
-  });
-  if (p.isCancel(baseUrl)) return cancelAndExit();
-
-  const authToken = await promptSecret('Enter the auth token for the custom endpoint');
-
-  const config: ShannonConfig = {
-    custom_base_url: { base_url: baseUrl, auth_token: authToken },
-  };
-
-  const customizeModels = await p.confirm({
-    message:
-      'Do you want to change the default models?\n' +
-      '    Small  - claude-haiku-4-5-20251001\n' +
-      '    Medium - claude-sonnet-4-6\n' +
-      '    Large  - claude-opus-4-7',
-    initialValue: false,
-  });
-  if (p.isCancel(customizeModels)) return cancelAndExit();
-
-  if (customizeModels) {
-    const small = await p.text({
-      message: 'Small model ID',
-      initialValue: 'claude-haiku-4-5-20251001',
-      validate: required('Small model ID is required'),
-    });
-    if (p.isCancel(small)) return cancelAndExit();
-
-    const medium = await p.text({
-      message: 'Medium model ID',
-      initialValue: 'claude-sonnet-4-6',
-      validate: required('Medium model ID is required'),
-    });
-    if (p.isCancel(medium)) return cancelAndExit();
-
-    const large = await p.text({
-      message: 'Large model ID',
-      initialValue: 'claude-opus-4-7',
-      validate: required('Large model ID is required'),
-    });
-    if (p.isCancel(large)) return cancelAndExit();
-
-    config.models = { small, medium, large };
-  }
-
-  return config;
+  const apiKey = await promptSecret('Enter your Anthropic API key');
+  return { anthropic: { api_key: apiKey } };
 }
 
 async function setupBedrock(): Promise<ShannonConfig> {
@@ -186,118 +193,121 @@ async function setupBedrock(): Promise<ShannonConfig> {
 
   const token = await promptSecret('Enter your AWS Bearer Token');
 
-  const small = await p.text({
-    message: 'Small model ID',
-    placeholder: 'us.anthropic.claude-haiku-4-5-20251001-v1:0',
-    validate: required('Small model ID is required'),
-  });
-  if (p.isCancel(small)) return cancelAndExit();
-
-  const medium = await p.text({
-    message: 'Medium model ID',
-    placeholder: 'us.anthropic.claude-sonnet-4-6',
-    validate: required('Medium model ID is required'),
-  });
-  if (p.isCancel(medium)) return cancelAndExit();
-
-  const large = await p.text({
-    message: 'Large model ID',
-    placeholder: 'us.anthropic.claude-opus-4-7',
-    validate: required('Large model ID is required'),
-  });
-  if (p.isCancel(large)) return cancelAndExit();
-
-  return {
-    bedrock: { use: true, region, token },
-    models: { small, medium, large },
-  };
+  return { bedrock: { region, token } };
 }
 
-async function setupVertex(): Promise<ShannonConfig> {
-  // 1. Collect region and project ID
-  const region = await p.text({
-    message: 'Google Cloud region',
-    placeholder: 'us-east5',
-    validate: required('Region is required'),
-  });
-  if (p.isCancel(region)) return cancelAndExit();
+interface GatewaySetup {
+  provider: CuratedProviderId;
+  config: ShannonConfig;
+  baseUrl: string;
+  format?: OpenAiFormat;
+}
 
-  const projectId = await p.text({
-    message: 'GCP Project ID',
-    validate: required('Project ID is required'),
+/**
+ * Gateway route: the endpoint decides where requests go, but the format still
+ * picks a real provider, because that is what supplies the credential and the
+ * wire protocol.
+ */
+async function setupGateway(): Promise<GatewaySetup> {
+  const choice = await p.select({
+    message: 'API format',
+    options: GATEWAY_DIALECTS.map(({ value, label }) => ({ value, label })),
   });
-  if (p.isCancel(projectId)) return cancelAndExit();
+  if (p.isCancel(choice)) return cancelAndExit();
 
-  // 2. File picker for service account key
-  p.log.info('Select the path to your GCP Service Account JSON key file.');
-  const keySourcePath = await p.path({
-    message: 'Service Account JSON key file',
+  const dialect = GATEWAY_DIALECTS.find((entry) => entry.value === choice);
+  if (!dialect) return cancelAndExit();
+  const provider = dialect.provider;
+
+  const baseUrl = await p.text({
+    message: 'Endpoint URL',
+    placeholder: 'https://llm-gateway.example.com',
     validate: (value) => {
-      if (!value) return 'Path is required';
-      if (!fs.existsSync(value)) return 'File not found';
-      if (!value.endsWith('.json')) return 'Must be a .json file';
+      if (!value) return 'Endpoint URL is required';
+      try {
+        new URL(value);
+      } catch {
+        return 'Must be a valid URL';
+      }
       return undefined;
     },
   });
-  if (p.isCancel(keySourcePath)) return cancelAndExit();
+  if (p.isCancel(baseUrl)) return cancelAndExit();
 
-  // 3. Copy key to ~/.shannon/ and lock permissions
-  const destPath = path.join(SHANNON_HOME, 'google-sa-key.json');
-  fs.mkdirSync(SHANNON_HOME, { recursive: true });
-  fs.copyFileSync(keySourcePath, destPath);
-  fs.chmodSync(destPath, 0o600);
-  p.log.success(`Key copied to ${destPath} (permissions: 0600)`);
+  const authToken = await promptSecret('Enter the auth token for the endpoint');
+  const config: ShannonConfig =
+    provider === 'anthropic'
+      ? { anthropic: { api_key: authToken } }
+      : { openai: { api_key: authToken, ...(dialect.format && { format: dialect.format }) } };
 
-  // 4. Model tiers
-  const models = await p.group({
-    small: () =>
-      p.text({
-        message: 'Small model ID',
-        placeholder: 'claude-haiku-4-5@20251001',
-        validate: required('Small model ID is required'),
-      }),
-    medium: () =>
-      p.text({
-        message: 'Medium model ID',
-        placeholder: 'claude-sonnet-4-6',
-        validate: required('Medium model ID is required'),
-      }),
-    large: () =>
-      p.text({
-        message: 'Large model ID',
-        placeholder: 'claude-opus-4-7',
-        validate: required('Large model ID is required'),
-      }),
+  return { provider, config, baseUrl, ...(dialect.format && { format: dialect.format }) };
+}
+
+// === Model Selection ===
+
+/**
+ * Ask for the one model that runs every phase. Providers with suggestions offer a
+ * pick list with a free-text escape hatch; the rest go straight to free text.
+ */
+async function promptModel(provider: string): Promise<string> {
+  const suggestions = isCuratedProvider(provider) ? MODEL_SUGGESTIONS[provider] : [];
+
+  if (suggestions.length === 0) {
+    return promptModelId(provider, modelIdPlaceholder(provider));
+  }
+
+  const choice = await p.select({
+    message: 'Model',
+    options: [
+      ...suggestions.map((model) => ({ value: model, label: model })),
+      { value: CUSTOM_MODEL, label: 'Enter a model ID…' },
+    ],
   });
-  if (p.isCancel(models)) return cancelAndExit();
+  if (p.isCancel(choice)) return cancelAndExit();
 
-  return {
-    vertex: {
-      use: true,
-      region,
-      project_id: projectId,
-      key_path: destPath,
+  if (choice === CUSTOM_MODEL) {
+    return promptModelId(provider, modelIdPlaceholder(provider));
+  }
+  return choice as string;
+}
+
+/**
+ * A leading `<provider>:` naming a supported provider other than the selected
+ * one. Bedrock model IDs carry their own colons (`…-v1:0`), so only a genuine
+ * provider id counts as a prefix.
+ */
+function conflictingProviderPrefix(provider: string, value: string): string | undefined {
+  const separator = value.indexOf(':');
+  if (separator === -1) return undefined;
+
+  const head = value.slice(0, separator);
+  if (head === provider) return undefined;
+  return (CURATED_PROVIDERS as readonly string[]).includes(head) ? head : undefined;
+}
+
+/**
+ * Ask for a model ID. The provider is already chosen, so this takes the bare ID
+ * and the caller pairs it with the provider — pasting a full `<provider>:<model>`
+ * spec just has its redundant prefix dropped.
+ */
+async function promptModelId(provider: string, placeholder?: string): Promise<string> {
+  const modelId = await p.text({
+    message: 'Model ID',
+    ...(placeholder && { placeholder }),
+    validate: (value) => {
+      if (!value) return 'Model ID is required';
+      const conflicting = conflictingProviderPrefix(provider, value);
+      if (conflicting) return `That model ID is for ${conflicting}, but you selected ${provider}.`;
+      return undefined;
     },
-    models: { small: models.small, medium: models.medium, large: models.large },
-  };
+  });
+  if (p.isCancel(modelId)) return cancelAndExit();
+
+  return modelId.startsWith(`${provider}:`) ? modelId.slice(provider.length + 1) : modelId;
 }
 
 // === Helpers ===
 
-async function maybePromptAdaptiveThinking(config: ShannonConfig): Promise<void> {
-  const m = config.models;
-  const hasOpus47 = !m || [m.small, m.medium, m.large].some((v) => v && /opus-4-[67]/.test(v));
-  if (!hasOpus47) return;
-
-  const enable = await p.confirm({
-    message: 'Enable adaptive thinking on Opus 4.6/4.7? Claude decides when and how deeply to reason.',
-    initialValue: true,
-  });
-  if (p.isCancel(enable)) return cancelAndExit();
-
-  config.core = { ...config.core, adaptive_thinking: enable };
-}
-
 async function promptSecret(message: string): Promise<string> {
   const value = await p.password({
     message,

apps/cli/src/commands/start.ts

@@ -8,12 +8,30 @@
 import { execFileSync } from 'node:child_process';
 import fs from 'node:fs';
 import path from 'node:path';
-import { ensureImage, ensureInfra, randomSuffix, spawnWorker } from '../docker.js';
-import { buildEnvFlags, loadEnv, validateCredentials } from '../env.js';
-import { getCredentialsPath, getWorkspacesDir, initHome } from '../home.js';
-import { isLocal } from '../mode.js';
-import { resolveConfig, resolveRepo } from '../paths.js';
-import { displaySplash } from '../splash.js';
+import { setTimeout as sleep } from 'node:timers/promises';
+import * as p from '@clack/prompts';
+import { ensureDocker, ensureImage, ensureInfra, randomSuffix, spawnWorker } from '../docker.js';
+import { buildEnvFlags, loadEnv, resolveHostPiAuthPath, shouldUsePiAuth, validateCredentials } from '../env.js';
+import { fail, warn } from '../errors.js';
+import { getWorkspacesDir, initHome } from '../home.js';
+import { commandPrefix, isLocal } from '../mode.js';
+import { resolveModelSpec } from '../model-spec.js';
+import {
+  expandHome,
+  FINAL_REPORT_MD_FILENAME,
+  FINAL_REPORT_PDF_FILENAME,
+  INTERNAL_DIR,
+  resolveConfig,
+  resolveRepo,
+  resolveRunFile,
+} from '../paths.js';
+import { clearPendingWorkflowIdentity, writePendingWorkflowIdentity } from '../pending-workflow.js';
+import { indentFailureSegments } from '../scan/failure.js';
+import { resolveWorkflowId } from '../session.js';
+import { displayPlainBanner, displaySplash } from '../splash.js';
+import { getTerminalOutcome } from '../temporal-client.js';
+import { stdoutIsTerminal } from '../tty.js';
+import { tailUntilComplete } from './logs.js';
 
 export interface StartArgs {
   url: string;
@@ -22,193 +40,475 @@ export interface StartArgs {
   workspace?: string;
   output?: string;
   pipelineTesting: boolean;
-  debug: boolean;
+  keepContainer: boolean;
+  follow: boolean;
   version: string;
 }
 
-export async function start(args: StartArgs): Promise<void> {
-  // 1. Initialize state directories and load env
-  initHome();
-  loadEnv();
+const LAUNCH_STATE_SCHEMA_VERSION = 1 as const;
+const LAUNCH_STATE_FILENAME = 'launch.json';
+const FIXED_CLASSES = ['injection', 'xss', 'auth', 'authz', 'ssrf'] as const;
 
-  // 2. Validate credentials
-  const creds = validateCredentials();
-  if (!creds.valid) {
-    console.error(`ERROR: ${creds.error}`);
-    process.exit(1);
+/**
+ * CLI-owned launch record at INTERNAL_DIR/launch.json, written once when a workspace is
+ * created and never rewritten. It pins the customer output destination so a resume with a
+ * different -o cannot silently redirect the final report. The worker does not read it.
+ */
+interface LaunchState {
+  readonly schema_version: typeof LAUNCH_STATE_SCHEMA_VERSION;
+  readonly customer_output_path?: string;
+}
+
+export interface WorkspaceLaunchDecision {
+  readonly isResume: boolean;
+  readonly outputDir?: string;
+}
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+  return value !== null && typeof value === 'object' && !Array.isArray(value);
+}
+
+function arraysEqual(left: readonly unknown[], right: readonly unknown[]): boolean {
+  return left.length === right.length && left.every((value, index) => value === right[index]);
+}
+
+/**
+ * Hand-rolled twin of the worker's durable-state validator in
+ * apps/worker/src/types/run-state.ts, which owns the session.json.durableScanState shape.
+ * Each array check accepts two variants because the worker appends 'miscellaneous' and
+ * 'miscellaneous-exploit' only after the miscellaneous pipeline admits findings. If the worker's shape
+ * changes and this twin lags, resume fails fast as incompatible instead of launching a
+ * worker against state it would misread.
+ */
+function isCurrentDurableState(value: unknown): boolean {
+  if (!isRecord(value) || value.schema_version !== 1 || typeof value.exploit !== 'boolean') return false;
+  if (!Array.isArray(value.participating_classes) || !Array.isArray(value.expected_agents)) return false;
+
+  const participating = value.participating_classes;
+  const validParticipation =
+    arraysEqual(participating, FIXED_CLASSES) || arraysEqual(participating, [...FIXED_CLASSES, 'miscellaneous']);
+  if (!validParticipation) return false;
+
+  const baselineAgents = ['pre-recon', 'recon', ...FIXED_CLASSES.map((name) => `${name}-vuln`)];
+  if (value.exploit) baselineAgents.push(...FIXED_CLASSES.map((name) => `${name}-exploit`));
+  baselineAgents.push('report');
+  const expected = value.expected_agents;
+  return arraysEqual(expected, baselineAgents) || arraysEqual(expected, [...baselineAgents, 'miscellaneous-exploit']);
+}
+
+/** One refusal for damaged CLI-owned or worker-owned workspace records, whichever reads first. */
+const DAMAGED_RECORDS_MESSAGE =
+  "This workspace's internal records are damaged and it cannot be resumed. Its report files are untouched. Start a new scan with a different -w name.";
+
+const NEWER_RELEASE_MESSAGE =
+  'This workspace was created by a newer version of Shannon. Upgrade Shannon, or start a new scan with a different -w name.';
+
+function readJsonFile(filePath: string): unknown {
+  try {
+    return JSON.parse(fs.readFileSync(filePath, 'utf8'));
+  } catch {
+    fail(DAMAGED_RECORDS_MESSAGE);
+  }
+}
+
+function readLaunchState(filePath: string): LaunchState {
+  if (!fs.existsSync(filePath)) {
+    fail(
+      'This workspace was created by an earlier version of Shannon and cannot be resumed. Its files and report are untouched. Start a new scan with a different -w name.',
+    );
+  }
+  const value = readJsonFile(filePath);
+  if (!isRecord(value)) fail(NEWER_RELEASE_MESSAGE);
+  // Unknown keys mean a newer release wrote this workspace; refuse rather than half-read it.
+  const keys = Object.keys(value).sort();
+  const keysAreValid = keys.every((key) => key === 'customer_output_path' || key === 'schema_version');
+  const customerPath = value.customer_output_path;
+  const pathIsValid =
+    customerPath === undefined ||
+    (typeof customerPath === 'string' && path.isAbsolute(customerPath) && path.resolve(customerPath) === customerPath);
+  if (value.schema_version !== LAUNCH_STATE_SCHEMA_VERSION || !keysAreValid || !pathIsValid) {
+    fail(NEWER_RELEASE_MESSAGE);
+  }
+  return {
+    schema_version: LAUNCH_STATE_SCHEMA_VERSION,
+    ...(typeof customerPath === 'string' && { customer_output_path: customerPath }),
+  };
+}
+
+/**
+ * Decide fresh-versus-resume from on-disk state alone, before start() mutates anything.
+ * A fresh launch requires the workspace directory to be absent or empty; a resume requires
+ * current-release session state, a matching target URL, and a customer output path that
+ * agrees with the recorded one. Every other combination fails the launch, so a typo in
+ * -w or -o stops here instead of spawning a worker into the wrong workspace.
+ */
+export function classifyWorkspaceLaunch(
+  workspacePath: string,
+  expectedUrl: string,
+  requestedOutputDir: string | undefined,
+): WorkspaceLaunchDecision {
+  const sessionPath = resolveRunFile(workspacePath, 'session.json');
+  const sessionExists = fs.existsSync(sessionPath);
+  if (!sessionExists) {
+    if (fs.existsSync(workspacePath) && fs.readdirSync(workspacePath).length > 0) {
+      fail(
+        'This directory is not a Shannon workspace, or its scan state is missing. Start a new scan with a different -w name.',
+      );
+    }
+    return { isResume: false, ...(requestedOutputDir !== undefined && { outputDir: requestedOutputDir }) };
   }
 
-  // 3. Resolve paths
+  const launchPath = path.join(workspacePath, INTERNAL_DIR, LAUNCH_STATE_FILENAME);
+  const launch = readLaunchState(launchPath);
+  const session = readJsonFile(sessionPath);
+  if (!isRecord(session) || !isRecord(session.session) || session.session.webUrl !== expectedUrl) {
+    fail(
+      'This workspace was created for a different target URL, so it cannot be resumed against this one. Check -u, or start a new scan with a different -w name.',
+    );
+  }
+  if (!isCurrentDurableState(session.durableScanState)) {
+    fail(
+      "This workspace's scan state cannot be read by this version. Its files are untouched. Start a new scan with a different -w name.",
+    );
+  }
+
+  const storedOutputDir = launch.customer_output_path;
+  if (requestedOutputDir !== undefined && requestedOutputDir !== storedOutputDir) {
+    fail(
+      'This workspace already copies its report to a different location than the -o path you passed. Re-run without -o to keep the original location, or start a new scan with a different -w name.',
+    );
+  }
+  return { isResume: true, ...(storedOutputDir !== undefined && { outputDir: storedOutputDir }) };
+}
+
+/**
+ * Crash-safe single write: exclusive temp file (pid plus random suffix keeps concurrent
+ * starts apart), fsync, rename into place, then directory fsync so the entry survives a
+ * host crash. Callers invoke this only for a fresh workspace; an existing launch.json is
+ * the resume contract and must never be replaced.
+ */
+export function writeLaunchStateAtomically(internalPath: string, outputDir: string | undefined): void {
+  const finalPath = path.join(internalPath, LAUNCH_STATE_FILENAME);
+  const temporaryPath = path.join(internalPath, `${LAUNCH_STATE_FILENAME}.tmp-${process.pid}-${randomSuffix()}`);
+  const launchState: LaunchState = {
+    schema_version: LAUNCH_STATE_SCHEMA_VERSION,
+    ...(outputDir !== undefined && { customer_output_path: outputDir }),
+  };
+  const descriptor = fs.openSync(temporaryPath, 'wx', 0o600);
+  try {
+    fs.writeFileSync(descriptor, `${JSON.stringify(launchState, null, 2)}\n`, 'utf8');
+    fs.fsyncSync(descriptor);
+  } finally {
+    fs.closeSync(descriptor);
+  }
+  try {
+    fs.renameSync(temporaryPath, finalPath);
+    const directory = fs.openSync(internalPath, 'r');
+    try {
+      fs.fsyncSync(directory);
+    } finally {
+      fs.closeSync(directory);
+    }
+  } catch (error) {
+    fs.rmSync(temporaryPath, { force: true });
+    throw error;
+  }
+}
+
+/** Select the workflow ID before Docker starts so the container can carry it as immutable identity. */
+export function createWorkflowId(workspace: string, isResume: boolean, timestamp: number = Date.now()): string {
+  if (isResume) return `${workspace}_resume_${timestamp}`;
+  return /_shannon-\d+$/.test(workspace) ? workspace : `${workspace}_shannon-${timestamp}`;
+}
+
+export async function start(args: StartArgs): Promise<void> {
+  // 1. Resolve non-mutating inputs and classify the workspace before changing it.
+  initHome();
+  loadEnv();
+  const creds = validateCredentials();
+  if (!creds.valid) {
+    fail(creds.error ?? 'Invalid credentials');
+  }
   const repo = resolveRepo(args.repo);
   const config = args.config ? resolveConfig(args.config) : undefined;
-
-  // 4. Ensure workspaces dir is writable by container user (UID 1001)
   const workspacesDir = getWorkspacesDir();
+  const workspace =
+    args.workspace ?? `${new URL(args.url).hostname.replace(/[^a-zA-Z0-9-]/g, '-')}_shannon-${Date.now()}`;
+  const workspacePath = path.join(workspacesDir, workspace);
+  const requestedOutputDir = args.output ? path.resolve(expandHome(args.output)) : undefined;
+  const launchDecision = classifyWorkspaceLaunch(workspacePath, args.url, requestedOutputDir);
+
+  // 2. Inputs are valid; identify the run before initializing shared infrastructure.
+  const bannerVersion = isLocal() ? undefined : args.version;
+  if (stdoutIsTerminal()) {
+    displaySplash(bannerVersion);
+  } else {
+    displayPlainBanner(bannerVersion);
+  }
   fs.mkdirSync(workspacesDir, { recursive: true });
   fs.chmodSync(workspacesDir, 0o777);
-
-  // 5. Ensure image (auto-build in dev, pull in npx) and start infra
+  ensureDocker();
   ensureImage(args.version);
-  await ensureInfra();
+  const spinner = p.spinner();
+  spinner.start('Starting scan');
+  await ensureInfra(spinner);
 
-  // 6. Generate unique task queue and container name
+  // 3. Generate the invocation identity.
   const suffix = randomSuffix();
   const taskQueue = `shannon-${suffix}`;
   const containerName = `shannon-worker-${suffix}`;
+  const workflowId = createWorkflowId(workspace, launchDecision.isResume);
 
-  // 7. Generate workspace name if not provided
-  const workspace =
-    args.workspace ?? `${new URL(args.url).hostname.replace(/[^a-zA-Z0-9-]/g, '-')}_shannon-${Date.now()}`;
-
-  // 8. Create writable overlay directories (mounted over :ro repo paths inside container)
-  // Workspace dir must be 0o777 so the container user (UID 1001) can create audit subdirs
-  const workspacePath = path.join(workspacesDir, workspace);
+  // 4. Create writable overlay directories after resume validation has succeeded.
+  // The run dir and its INTERNAL_DIR must be 0o777 so the container user can create audit
+  // subdirs and the overlay backing dirs.
+  const internalPath = path.join(workspacePath, INTERNAL_DIR);
   fs.mkdirSync(workspacePath, { recursive: true });
   fs.chmodSync(workspacePath, 0o777);
+  fs.mkdirSync(internalPath, { recursive: true });
+  fs.chmodSync(internalPath, 0o777);
   for (const dir of ['deliverables', 'scratchpad', '.playwright-cli', '.playwright']) {
-    const dirPath = path.join(workspacePath, dir);
+    const dirPath = path.join(internalPath, dir);
     fs.mkdirSync(dirPath, { recursive: true });
     fs.chmodSync(dirPath, 0o777);
   }
+  if (!launchDecision.isResume) {
+    writeLaunchStateAtomically(internalPath, launchDecision.outputDir);
+  }
 
-  // 9. Pre-create overlay mount points (:ro mounts can't auto-create them)
+  // 5. Pre-create overlay mount points (:ro mounts cannot create them).
   const shannonDir = path.join(repo.hostPath, '.shannon');
   for (const dir of ['deliverables', 'scratchpad', '.playwright-cli']) {
     fs.mkdirSync(path.join(shannonDir, dir), { recursive: true });
   }
   fs.mkdirSync(path.join(repo.hostPath, '.playwright'), { recursive: true });
 
-  const credentialsPath = getCredentialsPath();
-  const hasCredentials = fs.existsSync(credentialsPath);
-
-  if (hasCredentials) {
-    process.env.GOOGLE_APPLICATION_CREDENTIALS = '/app/credentials/google-sa-key.json';
-  }
-
-  // 10. Resolve output directory
-  const outputDir = args.output ? path.resolve(args.output) : undefined;
+  // 6. Create the validated customer-copy destination, if configured.
+  const outputDir = launchDecision.outputDir;
   if (outputDir) {
     fs.mkdirSync(outputDir, { recursive: true });
   }
 
-  // 11. Resolve prompts directory (local mode only)
+  // 7. Resolve prompts and capture the pre-launch resume counter.
   const promptsDir = isLocal() ? path.resolve('apps/worker/prompts') : undefined;
+  const sessionJson = resolveRunFile(workspacePath, 'session.json');
+  const isResume = launchDecision.isResume;
+  let initialResumeCount = 0;
+  if (isResume) {
+    // Docker and Temporal startup sit between this read and the classification that validated the
+    // same file, so a file that changed in between is a workspace-state failure, not a CLI bug.
+    const session = readJsonFile(sessionJson);
+    const attempts = isRecord(session) && isRecord(session.session) ? session.session.resumeAttempts : undefined;
+    initialResumeCount = Array.isArray(attempts) ? attempts.length : 0;
+  }
 
-  // 12. Display splash screen
-  displaySplash(isLocal() ? undefined : args.version);
+  // 8. Persist the exact launch candidate before Docker can start the worker. Session
+  // registration later replaces this bridge as the durable workflow identity.
+  try {
+    writePendingWorkflowIdentity(workspacePath, workflowId, taskQueue);
+  } catch {
+    spinner.error('Could not record the scan workflow identity');
+    process.exit(1);
+  }
 
-  // 13. Spawn worker container
+  // 9. Spawn the worker container.
   const proc = spawnWorker({
     version: args.version,
     url: args.url,
     repo,
     workspacesDir,
     taskQueue,
+    workflowId,
     containerName,
     envFlags: buildEnvFlags(),
     ...(config && { config }),
-    ...(hasCredentials && { credentials: credentialsPath }),
     ...(promptsDir && { promptsDir }),
     ...(outputDir && { outputDir }),
     workspace,
     ...(args.pipelineTesting && { pipelineTesting: true }),
-    ...(args.debug && { debug: true }),
+    ...(args.keepContainer && { keepContainer: true }),
+    ...(shouldUsePiAuth() && { piAuthHostPath: resolveHostPiAuthPath() }),
   });
 
-  // 14. Bail if `docker run -d` itself fails (mount error, image missing, etc.)
+  // Bail if `docker run -d` itself fails (mount error, image missing, etc.)
   const dockerExitCode = await new Promise<number>((resolve) => {
     proc.once('exit', (code) => resolve(code ?? 1));
-    proc.once('error', (err) => {
-      console.error(`Failed to start worker: ${err.message}`);
-      resolve(1);
-    });
+    proc.once('error', () => resolve(1));
   });
 
   if (dockerExitCode !== 0) {
+    spinner.error('Could not start the scan');
     process.exit(1);
   }
 
-  // Detect whether this is a fresh workspace or a resume by checking session.json existence
-  const sessionJson = path.join(workspacesDir, workspace, 'session.json');
-  const isResume = fs.existsSync(sessionJson);
-  let initialResumeCount = 0;
-  if (isResume) {
-    try {
-      const session = JSON.parse(fs.readFileSync(sessionJson, 'utf-8'));
-      initialResumeCount = session.session?.resumeAttempts?.length ?? 0;
-    } catch {
-      // Corrupted file — worker will handle validation
-    }
-  }
-
-  // Poll for workflow to register in session.json
-  process.stdout.write('Waiting for workflow to start...');
-  let workflowId = '';
   let started = false;
-  let attempts = 0;
-  const pollInterval = setInterval(() => {
-    attempts++;
-    if (attempts > 60) {
-      clearInterval(pollInterval);
-      process.stdout.write('\n');
-      console.error('Timeout waiting for workflow to start');
-      process.exit(1);
-    }
 
-    try {
-      const session = JSON.parse(fs.readFileSync(sessionJson, 'utf-8'));
-      const resumeAttempts: { workflowId: string }[] = session.session?.resumeAttempts ?? [];
+  // Set when the startup poll times out but session.json already holds durable state this
+  // release understands: the workflow is executing, so the exit handler must not stop its
+  // worker. An operator abort is a different intent and still stops it.
+  let scanRunningUnconfirmed = false;
 
-      // Fresh: session.json appears with originalWorkflowId. Resume: new resumeAttempts entry.
-      const ready = isResume ? resumeAttempts.length > initialResumeCount : !!session.session?.originalWorkflowId;
-
-      if (ready) {
-        clearInterval(pollInterval);
-        started = true;
-
-        // Latest workflow ID: last resume attempt, or originalWorkflowId for fresh scans
-        workflowId = resumeAttempts.at(-1)?.workflowId ?? session.session?.originalWorkflowId ?? '';
-
-        // Clear waiting line and show info
-        process.stdout.write('\r\x1b[K');
-        printInfo(args, workspace, workflowId, repo.hostPath, workspacesDir);
-        return;
-      }
-    } catch {
-      // File doesn't exist yet
-    }
-    process.stdout.write('.');
-  }, 2000);
-
-  // Stop the worker container only if it hasn't started yet
+  // Stop the worker only if the scan hasn't registered yet (e.g. Ctrl-C mid-startup).
   let cleaned = false;
-  const cleanup = (): void => {
+  const stopWorker = (): void => {
     if (cleaned || started) return;
     cleaned = true;
-    clearInterval(pollInterval);
-    console.log(`\nStopping worker ${containerName}...`);
+    spinner.stop('Stopping scan');
     try {
       execFileSync('docker', ['stop', containerName], { stdio: 'pipe' });
     } catch {
       // Container may have already exited
     }
-    if (args.debug) {
-      printDebugHint(containerName);
+    if (args.keepContainer) {
+      printPreservedContainerHint(containerName);
     }
   };
-
   process.on('SIGINT', () => {
-    cleanup();
+    stopWorker();
     process.exit(0);
   });
   process.on('SIGTERM', () => {
-    cleanup();
+    stopWorker();
     process.exit(0);
   });
-  process.on('exit', cleanup);
+  process.on('exit', () => {
+    if (scanRunningUnconfirmed) return;
+    stopWorker();
+  });
+
+  // Poll for the workflow to register in session.json; the spinner resolves once it does.
+  spinner.message('Waiting for the scan to start');
+  for (let attempts = 0; attempts < 60; attempts++) {
+    try {
+      const session = JSON.parse(fs.readFileSync(sessionJson, 'utf-8'));
+      const resumeAttempts: { workflowId: string }[] = session.session?.resumeAttempts ?? [];
+
+      // Fresh: session.json appears with originalWorkflowId. Resume: new resumeAttempts entry.
+      const ready = isResume
+        ? resumeAttempts.slice(initialResumeCount).some((attempt) => attempt.workflowId === workflowId)
+        : session.session?.originalWorkflowId === workflowId;
+
+      if (ready) {
+        started = true;
+        try {
+          clearPendingWorkflowIdentity(workspacePath, taskQueue);
+        } catch {
+          warn(`Scan ${workspace} started, but its launch record could not be removed.`);
+        }
+        spinner.stop(`Scan started — ${workspace}`);
+        printInfo(args, workspace, repo.hostPath, workspacesDir);
+        if (args.follow) {
+          await followScan(workspace, workspacesDir);
+        }
+        return;
+      }
+    } catch {
+      // File doesn't exist yet
+    }
+    await sleep(2000);
+  }
+
+  if (classifyStartupTimeout(sessionJson) === 'scan-running') {
+    scanRunningUnconfirmed = true;
+    spinner.error('The scan started, but this CLI could not confirm it');
+    printUnconfirmedScanHint(workspace, taskQueue, containerName);
+    process.exit(1);
+  }
+
+  spinner.error('Timed out waiting for the scan to start');
+  process.exit(1);
 }
 
-function printDebugHint(containerName: string): void {
+/**
+ * Read the startup timeout: 'scan-running' when session.json already holds durable state this
+ * release understands, which only the worker writes and only after Temporal began executing the
+ * workflow; 'unregistered' when nothing proves the scan started. The distinction decides whether
+ * timing out may stop the worker container.
+ */
+export function classifyStartupTimeout(sessionJsonPath: string): 'unregistered' | 'scan-running' {
+  let session: unknown;
+  try {
+    session = JSON.parse(fs.readFileSync(sessionJsonPath, 'utf-8'));
+  } catch {
+    return 'unregistered';
+  }
+  if (!isRecord(session) || !isCurrentDurableState(session.durableScanState)) {
+    return 'unregistered';
+  }
+  return 'scan-running';
+}
+
+/** Point the operator at a scan that is running but whose startup this CLI could not confirm. */
+function printUnconfirmedScanHint(workspace: string, taskQueue: string, containerName: string): void {
+  console.log('');
+  console.log('  The scan is running and was left alone; only its startup confirmation is missing.');
+  console.log('');
+  console.log(`  Workspace:  ${workspace}`);
+  console.log(`  Task queue: ${taskQueue}`);
+  console.log(`  Container:  ${containerName}`);
+  console.log('');
+  console.log('  Inspect it:');
+  console.log(`    Live logs:   ${commandPrefix()} logs ${workspace}`);
+  console.log(`    Worker logs: docker logs ${containerName}`);
+  console.log('    Dashboard:   http://localhost:8233');
+  console.log('');
+}
+
+/**
+ * Follow a just-started scan (for `--follow`, aimed at CI): stream its log while Temporal drives
+ * completion, then exit on the workflow outcome — 0 if the assessment ran, 1 if the scan failed.
+ * That tracks whether the pipeline ran, not whether vulnerabilities were found. On failure the
+ * root-cause message is printed so a red CI build says why.
+ */
+async function followScan(workspace: string, workspacesDir: string): Promise<never> {
+  const logFile = resolveRunFile(path.join(workspacesDir, workspace), 'workflow.log');
+  const workflowId = resolveWorkflowId(workspace);
+
+  // The worker creates workflow.log as it starts; wait briefly so the first read doesn't
+  // mistake a not-yet-created file for an already-finished scan.
+  for (let attempts = 0; attempts < 30 && !fs.existsSync(logFile); attempts++) {
+    await sleep(1000);
+  }
+
+  if (stdoutIsTerminal()) {
+    console.error('\n  Following scan log (Ctrl-C to stop watching):\n');
+  }
+
+  let temporalUnreachable = false;
+  const { sawFailure } = await tailUntilComplete(logFile, {
+    ...(workflowId && { workflowId }),
+    onUnreachable: () => {
+      temporalUnreachable = true;
+    },
+  });
+
+  // The tail already printed the diagnostic; reading the outcome would only fail the same way.
+  if (temporalUnreachable) {
+    process.exit(1);
+  }
+
+  if (!workflowId) {
+    fail('Scan finished but its workflow id could not be resolved from session.json.');
+  }
+
+  try {
+    const outcome = await getTerminalOutcome(workflowId);
+    if (outcome.kind === 'failed') {
+      // Print the reason only when the streamed log didn't already show the worker's failure
+      // summary — otherwise the worker crashed before writing it, and this is the only report.
+      if (!sawFailure) {
+        console.error(`\nScan failed:\n${indentFailureSegments(outcome.message)}`);
+      }
+      process.exit(1);
+    }
+    process.exit(0);
+  } catch (err) {
+    const detail = err instanceof Error ? err.message : String(err);
+    fail('Could not read the scan outcome from Temporal at 127.0.0.1:7233.', `  ${detail}`);
+  }
+}
+
+function printPreservedContainerHint(containerName: string): void {
   console.log('');
   console.log(`  Worker container preserved: ${containerName}`);
   console.log(`    Inspect logs: docker logs ${containerName}`);
@@ -216,35 +516,49 @@ function printDebugHint(containerName: string): void {
   console.log('');
 }
 
-function printInfo(
-  args: StartArgs,
-  workspace: string,
-  workflowId: string,
-  repoPath: string,
-  workspacesDir: string,
-): void {
-  const logsCmd = isLocal() ? `./shannon logs ${workspace}` : `npx @keygraph/shannon logs ${workspace}`;
-  const reportsPath = path.join(workspacesDir, workspace);
+function printInfo(args: StartArgs, workspace: string, repoPath: string, workspacesDir: string): void {
+  const interactive = stdoutIsTerminal();
+
+  if (interactive && !args.follow) {
+    console.log('  It runs in the background — you can close this terminal.');
+    console.log('');
+  }
 
   console.log(`  Target:     ${args.url}`);
-  console.log(`  Repository: ${repoPath}`);
+  console.log(`  Repository: ${interactive ? repoPath : path.basename(repoPath)}`);
   console.log(`  Workspace:  ${workspace}`);
   if (args.config) {
-    console.log(`  Config:     ${path.resolve(args.config)}`);
+    console.log(`  Config:     ${interactive ? path.resolve(args.config) : path.basename(args.config)}`);
   }
   if (args.pipelineTesting) {
     console.log('  Mode:       Pipeline Testing');
   }
-  console.log('');
-  console.log('  Monitor:');
-  if (workflowId) {
-    console.log(`    Web UI:  http://localhost:8233/namespaces/default/workflows/${workflowId}`);
-  } else {
-    console.log('    Web UI:  http://localhost:8233');
+
+  const spec = resolveModelSpec();
+  if (typeof spec !== 'string') {
+    console.log(`  Model:      ${spec.providerId}:${spec.modelId}`);
   }
-  console.log(`    Logs:    ${logsCmd}`);
+
+  if (!interactive) {
+    return;
+  }
+
+  const reportDir = path.join(workspacesDir, workspace);
+
+  // When following, the scan log streams inline next, so the "run these to watch it" hints
+  // would only contradict that.
+  if (!args.follow) {
+    const prefix = commandPrefix();
+    console.log('');
+    console.log('  Watch scan progress:');
+    console.log(`    Live logs:  ${prefix} logs ${workspace}`);
+    console.log(`    Progress:   ${prefix} status ${workspace}`);
+  }
+
   console.log('');
-  console.log('  Output:');
-  console.log(`    Reports: ${reportsPath}/`);
+  console.log('  Report (when the scan finishes):');
+  console.log(`    ${reportDir}${path.sep}`);
+  console.log(`      ${FINAL_REPORT_PDF_FILENAME}`);
+  console.log(`      ${FINAL_REPORT_MD_FILENAME}`);
   console.log('');
 }

apps/cli/src/commands/status.ts

@@ -1,24 +1,239 @@
 /**
- * `shannon status` command — show running workers and Temporal health.
+ * `shannon status <workspace>` — one scan's live progress from Temporal.
+ *
+ * While the scan runs, polls Temporal and redraws the phase/agent tree on a
+ * terminal (a pipe or a finished scan gets a single frame). When the scan reaches
+ * a terminal state, prints the overall result and exits. Local session records prove
+ * the target's canonical workspace/workflow identity; the progress itself is read from
+ * Temporal directly — no worker — so it needs Temporal up and shows scans within its
+ * retention window (Shannon configures seven days by default; see SHANNON_TEMPORAL_RETENTION).
  */
 
-import { isTemporalReady, listRunningWorkers } from '../docker.js';
+import { setTimeout as sleep } from 'node:timers/promises';
+import { failWith } from '../errors.js';
+import { commandPrefix, isLocal } from '../mode.js';
+import { type RenderInput, renderScan } from '../scan/render.js';
+import { toStatusJson } from '../scan/status-json.js';
+import { displaySplash } from '../splash.js';
+import {
+  ActivityMirrorError,
+  describeScan,
+  getTerminalOutcome,
+  queryProgress,
+  type ScanDescription,
+} from '../temporal-client.js';
+import { stdoutIsTerminal, supportsColor } from '../tty.js';
+import { getVersion } from '../version.js';
+import { resolveScanIdentity } from '../workspaces.js';
 
-export function status(): void {
-  // 1. Temporal health
-  const temporalUp = isTemporalReady();
-  console.log(`Temporal: ${temporalUp ? 'running' : 'not running'}`);
-  if (temporalUp) {
-    console.log('  Web UI: http://localhost:8233');
-  }
-  console.log('');
+const HIDE_CURSOR = '\x1b[?25l';
+const SHOW_CURSOR = '\x1b[?25h';
+/** Redraw cadence for the spinner animation; data is refreshed on the slower poll. */
+const RENDER_MS = 120;
+const POLL_MS = 1200;
 
-  // 2. Running workers
-  const workers = listRunningWorkers();
-  if (workers) {
-    console.log('Workers:');
-    console.log(workers);
-  } else {
-    console.log('Workers: none running');
+/** Terminal = anything other than an open, running execution. */
+function isTerminalStatus(status: string): boolean {
+  return status !== 'RUNNING' && status !== 'UNSPECIFIED';
+}
+
+/**
+ * Read one scan description, telling the two failure modes apart. A stale activity mirror
+ * carries its own message and needs a CLI update; anything else is a read that did not reach
+ * a usable answer, which is most often Temporal being down.
+ */
+async function readScanDescription(workflowId: string): Promise<ScanDescription | null> {
+  try {
+    return await describeScan(workflowId);
+  } catch (error) {
+    if (error instanceof ActivityMirrorError) failWith('CLI_SCAN_SCHEMA_UNSUPPORTED', error.message);
+    failWith(
+      'CLI_SCAN_STATUS_UNAVAILABLE',
+      "Could not read this scan's progress.",
+      'If Temporal is not running, start a scan to bring it up. If it is running, this build of the CLI',
+      'does not recognise part of the scan and needs updating.',
+    );
   }
 }
+
+// Match SGR color escapes (ESC[…m) so a line's on-screen width excludes them. Built from the ESC
+// char code so the source carries no literal control character.
+const ANSI_PATTERN = new RegExp(`${String.fromCharCode(27)}\\[[0-9;]*m`, 'g');
+
+/**
+ * Physical terminal rows a frame occupies, so the live redraw moves the cursor up by the right
+ * amount. A line wider than the terminal wraps onto extra rows, so counting logical lines alone
+ * undercounts and the redraw drifts downward. Color escapes don't take screen columns, so strip them.
+ */
+function physicalRows(frame: string): number {
+  const columns = process.stdout.columns || 80;
+  return frame.split('\n').reduce((rows, line) => {
+    const width = line.replace(ANSI_PATTERN, '').length;
+    return rows + Math.max(1, Math.ceil(width / columns));
+  }, 0);
+}
+
+function exitCodeFor(input: RenderInput): number {
+  if (input.temporalStatus === 'FAILED' || input.temporalStatus === 'TIMED_OUT') return 1;
+  if (input.state?.status === 'failed') return 1;
+  return 0;
+}
+
+/** Live view of a running scan: its progress query plus the in-flight agents from describe. */
+async function buildRunningInput(workspace: string, workflowId: string, desc: ScanDescription): Promise<RenderInput> {
+  const state = await queryProgress(workflowId);
+  return {
+    workspace,
+    workflowId,
+    temporalStatus: desc.status,
+    state,
+    running: desc.runningAgents,
+    ...(desc.startedAt !== undefined && { startedAt: desc.startedAt }),
+  };
+}
+
+/** Final view of a closed scan: its result (or the failure) plus timing from describe. */
+async function buildTerminalInput(workspace: string, workflowId: string, desc: ScanDescription): Promise<RenderInput> {
+  const outcome = await getTerminalOutcome(workflowId);
+  const timing = {
+    ...(desc.startedAt !== undefined && { startedAt: desc.startedAt }),
+    ...(desc.closedAt !== undefined && { endedAt: desc.closedAt }),
+  };
+  if (outcome.kind === 'success') {
+    return { workspace, workflowId, temporalStatus: desc.status, state: outcome.state, running: [], ...timing };
+  }
+  return {
+    workspace,
+    workflowId,
+    temporalStatus: desc.status,
+    state: null,
+    running: [],
+    failureMessage: outcome.message,
+    ...timing,
+  };
+}
+
+function printFrame(input: RenderInput): void {
+  const frame = renderScan(input, {
+    now: Date.now(),
+    color: supportsColor(),
+    unicode: stdoutIsTerminal(),
+    live: false,
+    frame: 0,
+  });
+  process.stdout.write(`${frame}\n`);
+}
+
+/**
+ * Poll Temporal and redraw until the scan reaches a terminal state, then print the
+ * final frame and exit. A fast ticker animates the running spinner off the cached
+ * snapshot; the network poll refreshes that snapshot on a slower cadence.
+ */
+async function watch(workspace: string, workflowId: string): Promise<never> {
+  let prevRows = 0;
+  let frame = 0;
+  let cached: RenderInput | null = null;
+
+  const draw = (input: RenderInput, live: boolean): void => {
+    const out = renderScan(input, { now: Date.now(), color: supportsColor(), unicode: true, live, frame });
+    if (prevRows > 0) process.stdout.write(`\x1b[${prevRows}A\x1b[0J`);
+    process.stdout.write(`${out}\n`);
+    prevRows = physicalRows(out);
+  };
+
+  process.on('exit', () => process.stdout.write(SHOW_CURSOR));
+  process.on('SIGINT', () => {
+    process.stdout.write('\n');
+    process.exit(0);
+  });
+  process.stdout.write(HIDE_CURSOR);
+
+  const ticker = setInterval(() => {
+    frame++;
+    if (cached) draw(cached, true);
+  }, RENDER_MS);
+
+  for (;;) {
+    const desc = await readScanDescription(workflowId);
+    if (!desc) {
+      clearInterval(ticker);
+      failWith('CLI_SCAN_NOT_FOUND', `Scan "${workspace}" is no longer in Temporal.`);
+    }
+
+    if (isTerminalStatus(desc.status)) {
+      clearInterval(ticker);
+      const input = await buildTerminalInput(workspace, workflowId, desc);
+      draw(input, false);
+      process.exit(exitCodeFor(input));
+    }
+
+    cached = await buildRunningInput(workspace, workflowId, desc);
+    await sleep(POLL_MS);
+  }
+}
+
+/** Read one point-in-time snapshot from Temporal: the terminal result if closed, else live progress. */
+async function snapshot(workspace: string, workflowId: string, desc: ScanDescription): Promise<RenderInput> {
+  return isTerminalStatus(desc.status)
+    ? buildTerminalInput(workspace, workflowId, desc)
+    : buildRunningInput(workspace, workflowId, desc);
+}
+
+export async function status(target: string, opts: { readonly json: boolean }): Promise<void> {
+  // Target selection picked a string; identity resolution proves the canonical workspace and
+  // workflow pair from session records before Temporal is queried. A workspace name follows its
+  // latest resume; an exact recorded workflow id keeps addressing that execution. A raw id with
+  // no local record is refused rather than echoed into the required workspace field.
+  const identity = resolveScanIdentity(target);
+  if (identity.kind === 'ambiguous') {
+    failWith(
+      'CLI_SCAN_IDENTITY_AMBIGUOUS',
+      `Multiple workspaces claim workflow ID "${target}": ${identity.claims.join(', ')}.`,
+      `Run '${commandPrefix()} scans' and pass the workspace directory name instead.`,
+    );
+  }
+  if (identity.kind === 'not-found') {
+    failWith(
+      'CLI_SCAN_IDENTITY_NOT_FOUND',
+      identity.reason === 'unreadable-record'
+        ? `Workspace "${target}" has no readable session record (${identity.sessionPath}).`
+        : `No scan matches "${target}" in the local workspace records.`,
+      `Run '${commandPrefix()} scans' to list scans.`,
+      'Temporal dashboard: http://localhost:8233',
+    );
+  }
+  const { workspace, workflowId } = identity;
+
+  const desc = await readScanDescription(workflowId);
+
+  if (!desc) {
+    failWith(
+      'CLI_SCAN_NOT_FOUND',
+      `No scan found for "${workspace}".`,
+      '',
+      "Scan histories are available while a scan runs and within Temporal's retention window after it finishes.",
+      "Shannon configures 7 days of retention by default (override: SHANNON_TEMPORAL_RETENTION). Expired histories can't be restored.",
+    );
+  }
+
+  // --json is always a single snapshot then exit, even on a TTY — it never enters the live watch loop.
+  if (opts.json) {
+    const input = await snapshot(workspace, workflowId, desc);
+    process.stdout.write(`${JSON.stringify(toStatusJson(input, Date.now()), null, 2)}\n`);
+    process.exit(exitCodeFor(input));
+  }
+
+  // Human-facing views open with the splash; skip it off a real terminal so piped output stays clean.
+  if (stdoutIsTerminal()) {
+    displaySplash(isLocal() ? undefined : getVersion());
+  }
+
+  // A finished scan, or output that isn't a live terminal, gets a single frame.
+  if (isTerminalStatus(desc.status) || !stdoutIsTerminal()) {
+    const input = await snapshot(workspace, workflowId, desc);
+    printFrame(input);
+    process.exit(exitCodeFor(input));
+  }
+
+  await watch(workspace, workflowId);
+}

apps/cli/src/commands/stop.ts

@@ -1,21 +1,883 @@
 /**
- * `shannon stop` command — stop workers and infrastructure.
+ * `shannon stop` command: stop one scan by workspace, or every scan with --all.
+ * Never touches infra or data; to wipe Temporal state entirely, use `shannon reset`.
  */
 
+import path from 'node:path';
 import * as p from '@clack/prompts';
-import { stopInfra, stopWorkers } from '../docker.js';
+import { confirmOrExit } from '../confirm.js';
+import {
+  type CommandQueryResult,
+  ensureDocker,
+  type RunningScanContainer,
+  runningContainersChecked,
+  runningScanContainersChecked,
+  scanFilter,
+  stopContainers,
+  WORKER_FILTER,
+  WORKFLOW_ID_PROTOCOL,
+} from '../docker.js';
+import { fail, failUsage, warn } from '../errors.js';
+import { getWorkspacesDir } from '../home.js';
+import { commandPrefix } from '../mode.js';
+import { resolveRunFile } from '../paths.js';
+import {
+  clearPendingWorkflowIdentity,
+  type PendingWorkflowIdentity,
+  readPendingWorkflowIdentities,
+} from '../pending-workflow.js';
+import { resolveWorkflowId } from '../session.js';
+import {
+  describeWorkflowLifecycle,
+  listRunningScanWorkflows,
+  type RunningScanWorkflow,
+  refreshWorkflowLifecycleConnection,
+  requestWorkflowCancellation,
+  requestWorkflowTermination,
+  type WorkflowLifecycleState,
+} from '../temporal-client.js';
+import { listWorkspaces, resolveScanIdentity } from '../workspaces.js';
+import { appendCancellationFallback } from './logs.js';
 
-export async function stop(clean: boolean): Promise<void> {
-  if (clean) {
-    const confirmed = await p.confirm({
-      message: 'This will stop all running scans and remove the Temporal data. Continue?',
-    });
-    if (p.isCancel(confirmed) || !confirmed) {
-      p.cancel('Aborted.');
-      process.exit(0);
+export interface StopOptions {
+  all: boolean;
+  yes: boolean;
+  workspace?: string;
+}
+
+const CANCELLATION_GRACE_MS = 10_000;
+const CANCELLATION_POLL_MS = 250;
+const TERMINATION_VERIFY_MS = 5_000;
+const TERMINATION_ATTEMPTS = 2;
+const TERMINATION_REASON = 'Stopped after cancellation grace period';
+const CANDIDATE_REGISTRATION_SETTLE_MS = 3_000;
+const VISIBILITY_SETTLE_MS = 1_000;
+const VISIBILITY_MAX_SETTLE_MS = 5_000;
+
+export type WorkflowStopOutcome =
+  | { readonly kind: 'graceful' }
+  | { readonly kind: 'forced' }
+  | { readonly kind: 'already-closed' }
+  | { readonly kind: 'unverified' };
+
+export type ContainerStopOutcome =
+  | { readonly kind: 'stopped'; readonly hadContainers: boolean }
+  | { readonly kind: 'still-running'; readonly remaining: number }
+  | { readonly kind: 'unverified' };
+
+export interface StopLifecycle {
+  readonly cancel: (workflowId: string) => Promise<'requested' | 'not-found'>;
+  readonly describe: (workflowId: string) => Promise<WorkflowLifecycleState>;
+  readonly refresh: () => Promise<void>;
+  readonly terminate: (workflowId: string) => Promise<'requested' | 'not-found'>;
+  readonly containers: (filter: readonly string[]) => CommandQueryResult<string[]>;
+  readonly stopContainers: (ids: readonly string[]) => Promise<void>;
+  readonly appendFallback: (workspace: string) => void;
+  readonly wait: (milliseconds: number) => Promise<void>;
+  readonly now: () => number;
+}
+
+export interface WorkflowStopTarget {
+  readonly workflowId: string;
+  readonly workspace?: string;
+  readonly containerCandidate: boolean;
+  /** Safe to synthesize a log marker when this CLI-owned launch never reached session registration. */
+  readonly preRegistrationFallback?: boolean;
+}
+
+export interface WorkflowStopResult {
+  readonly target: WorkflowStopTarget;
+  readonly outcome: WorkflowStopOutcome;
+}
+
+export interface StopExecutionResult {
+  readonly workflows: readonly WorkflowStopResult[];
+  readonly containers: ContainerStopOutcome;
+  readonly preRegistrationWorkspaces: readonly string[];
+}
+
+export interface WorkflowTargetPlan {
+  readonly targets: readonly WorkflowStopTarget[];
+  readonly containersWithoutVerifiedWorkflowId: readonly string[];
+}
+
+interface PendingWorkflowReference {
+  readonly workspace: string;
+  readonly identity: PendingWorkflowIdentity;
+}
+
+interface PendingWorkflowTargets {
+  readonly byWorkspace: ReadonlyMap<string, readonly PendingWorkflowIdentity[]>;
+  readonly references: readonly PendingWorkflowReference[];
+  readonly unreadableCount: number;
+}
+
+const stopLifecycle: StopLifecycle = {
+  cancel: requestWorkflowCancellation,
+  describe: describeWorkflowLifecycle,
+  refresh: refreshWorkflowLifecycleConnection,
+  terminate: (workflowId) => requestWorkflowTermination(workflowId, TERMINATION_REASON),
+  containers: runningContainersChecked,
+  stopContainers: (ids) => stopContainers([...ids]),
+  appendFallback: (workspace) => {
+    const logFile = resolveRunFile(path.join(getWorkspacesDir(), workspace), 'workflow.log');
+    appendCancellationFallback(logFile);
+  },
+  wait: (milliseconds) => new Promise((resolve) => setTimeout(resolve, milliseconds)),
+  now: Date.now,
+};
+
+function readPendingTargets(workspaces: readonly string[]): PendingWorkflowTargets {
+  const byWorkspace = new Map<string, readonly PendingWorkflowIdentity[]>();
+  const references: PendingWorkflowReference[] = [];
+  let unreadableCount = 0;
+
+  for (const workspace of new Set(workspaces)) {
+    const workspacePath = path.join(getWorkspacesDir(), workspace);
+    const result = readPendingWorkflowIdentities(workspacePath);
+    unreadableCount += result.unreadableCount;
+    if (result.identities.length === 0) continue;
+    byWorkspace.set(workspace, result.identities);
+    for (const identity of result.identities) references.push({ workspace, identity });
+  }
+
+  return { byWorkspace, references, unreadableCount };
+}
+
+function clearPendingTargets(references: readonly PendingWorkflowReference[]): number {
+  let failures = 0;
+  for (const reference of references) {
+    try {
+      clearPendingWorkflowIdentity(path.join(getWorkspacesDir(), reference.workspace), reference.identity.task_queue);
+    } catch {
+      failures++;
+    }
+  }
+  return failures;
+}
+
+function workflowClosed(state: WorkflowLifecycleState): boolean {
+  return state.kind === 'terminal' || state.kind === 'not-found';
+}
+
+/** Poll direct workflow state until Temporal positively confirms closure or the deadline expires. */
+async function waitForWorkflowClosure(
+  workflowId: string,
+  lifecycle: StopLifecycle,
+  deadline: number,
+  pollMs: number,
+): Promise<boolean> {
+  while (true) {
+    if (deadline - lifecycle.now() <= 0) return false;
+    try {
+      if (workflowClosed(await lifecycle.describe(workflowId))) return true;
+    } catch {
+      // An unavailable status is unknown, never evidence that the workflow closed.
+    }
+
+    const remaining = deadline - lifecycle.now();
+    if (remaining <= 0) return false;
+    await lifecycle.wait(Math.min(pollMs, remaining));
+  }
+}
+
+/**
+ * Request cooperative cancellation, then make at most two termination attempts when the
+ * workflow does not close during its grace period. Every success is backed by direct state.
+ */
+export async function stopWorkflowCancelFirst(
+  workflowId: string,
+  lifecycle: StopLifecycle = stopLifecycle,
+  graceMs: number = CANCELLATION_GRACE_MS,
+  pollMs: number = CANCELLATION_POLL_MS,
+  verifyMs: number = TERMINATION_VERIFY_MS,
+): Promise<WorkflowStopOutcome> {
+  try {
+    if ((await lifecycle.cancel(workflowId)) === 'not-found') return { kind: 'already-closed' };
+  } catch {
+    // The request may have reached Temporal even when its acknowledgement was lost.
+  }
+
+  if (await waitForWorkflowClosure(workflowId, lifecycle, lifecycle.now() + graceMs, pollMs)) {
+    return { kind: 'graceful' };
+  }
+
+  const verifyPerAttemptMs = Math.max(pollMs, Math.ceil(verifyMs / TERMINATION_ATTEMPTS));
+  for (let attempt = 0; attempt < TERMINATION_ATTEMPTS; attempt++) {
+    if (attempt > 0) {
+      try {
+        await lifecycle.refresh();
+      } catch {
+        // The termination call below makes one final bounded connection attempt.
+      }
+    }
+
+    try {
+      if ((await lifecycle.terminate(workflowId)) === 'not-found') return { kind: 'already-closed' };
+    } catch {
+      // A lost acknowledgement is resolved by the direct verification below.
+    }
+    if (await waitForWorkflowClosure(workflowId, lifecycle, lifecycle.now() + verifyPerAttemptMs, pollMs)) {
+      return { kind: 'forced' };
     }
   }
 
-  stopWorkers();
-  stopInfra(clean);
+  return { kind: 'unverified' };
+}
+
+/** Apply the same bounded lifecycle concurrently to a captured set of workflow IDs. */
+export async function stopWorkflowsCancelFirst(
+  workflowIds: readonly string[],
+  lifecycle: StopLifecycle = stopLifecycle,
+  graceMs: number = CANCELLATION_GRACE_MS,
+  pollMs: number = CANCELLATION_POLL_MS,
+  verifyMs: number = TERMINATION_VERIFY_MS,
+): Promise<readonly WorkflowStopOutcome[]> {
+  const settlements = await Promise.allSettled(
+    workflowIds.map((workflowId) => stopWorkflowCancelFirst(workflowId, lifecycle, graceMs, pollMs, verifyMs)),
+  );
+  return settlements.map((settlement) =>
+    settlement.status === 'fulfilled' ? settlement.value : { kind: 'unverified' },
+  );
+}
+
+/** Stop exactly the captured workers, then fail closed if any matching worker remains or appears. */
+export async function stopContainersAndVerify(
+  initialIds: readonly string[],
+  filter: readonly string[],
+  lifecycle: StopLifecycle = stopLifecycle,
+): Promise<ContainerStopOutcome> {
+  try {
+    await lifecycle.stopContainers(initialIds);
+  } catch {
+    // The post-stop query below decides whether the operation actually succeeded.
+  }
+
+  let after: CommandQueryResult<string[]>;
+  try {
+    after = lifecycle.containers(filter);
+  } catch {
+    return { kind: 'unverified' };
+  }
+  if (after.kind === 'unavailable') return { kind: 'unverified' };
+  if (after.value.length > 0) return { kind: 'still-running', remaining: after.value.length };
+  return { kind: 'stopped', hadContainers: initialIds.length > 0 };
+}
+
+function addWorkflowTarget(targets: Map<string, WorkflowStopTarget>, candidate: WorkflowStopTarget): void {
+  const current = targets.get(candidate.workflowId);
+  if (current === undefined) {
+    targets.set(candidate.workflowId, candidate);
+    return;
+  }
+  const workspace = current.workspace ?? candidate.workspace;
+  targets.set(candidate.workflowId, {
+    workflowId: candidate.workflowId,
+    ...(workspace !== undefined && { workspace }),
+    containerCandidate: current.containerCandidate || candidate.containerCandidate,
+    ...((current.preRegistrationFallback === true || candidate.preRegistrationFallback === true) && {
+      preRegistrationFallback: true,
+    }),
+  });
+}
+
+/**
+ * Build the stop union from immutable container candidates, recorded session IDs,
+ * pre-registration launch records, and Temporal visibility. Visibility supplies positive
+ * targets but never proves absence.
+ */
+function verifiedContainerWorkflowId(
+  container: RunningScanContainer,
+  visibleWorkflows: readonly RunningScanWorkflow[],
+): string | undefined {
+  if (container.workerProtocol === WORKFLOW_ID_PROTOCOL && container.workflowId !== undefined) {
+    return container.workflowId;
+  }
+  if (container.taskQueue === undefined) return undefined;
+  const matches = visibleWorkflows.filter((workflow) => workflow.taskQueue === container.taskQueue);
+  return matches.length === 1 ? matches[0]?.workflowId : undefined;
+}
+
+export function buildWorkflowTargetPlan(
+  containers: readonly RunningScanContainer[],
+  recordedByWorkspace: ReadonlyMap<string, string>,
+  visibleWorkflows: readonly RunningScanWorkflow[],
+  pendingByWorkspace: ReadonlyMap<string, readonly PendingWorkflowIdentity[]> = new Map(),
+): WorkflowTargetPlan {
+  const targets = new Map<string, WorkflowStopTarget>();
+  const containersWithoutVerifiedWorkflowId: string[] = [];
+
+  for (const container of containers) {
+    const verifiedWorkflowId = verifiedContainerWorkflowId(container, visibleWorkflows);
+    if (verifiedWorkflowId === undefined) containersWithoutVerifiedWorkflowId.push(container.id);
+    else {
+      addWorkflowTarget(targets, {
+        workflowId: verifiedWorkflowId,
+        ...(container.workspace !== undefined && { workspace: container.workspace }),
+        containerCandidate: true,
+      });
+    }
+  }
+
+  for (const [workspace, workflowId] of recordedByWorkspace) {
+    addWorkflowTarget(targets, {
+      workflowId,
+      workspace,
+      containerCandidate:
+        containers.some((container) => verifiedContainerWorkflowId(container, visibleWorkflows) === workflowId) ||
+        pendingByWorkspace.get(workspace)?.some((identity) => identity.workflow_id === workflowId) === true,
+    });
+  }
+
+  for (const [workspace, identities] of pendingByWorkspace) {
+    for (const identity of identities) {
+      addWorkflowTarget(targets, {
+        workflowId: identity.workflow_id,
+        workspace,
+        containerCandidate: true,
+        ...(!recordedByWorkspace.has(workspace) && { preRegistrationFallback: true }),
+      });
+    }
+  }
+
+  for (const workflow of visibleWorkflows) {
+    const matchingWorkspaces = new Set(
+      containers
+        .filter((container) => container.taskQueue === workflow.taskQueue && container.workspace !== undefined)
+        .flatMap((container) => container.workspace ?? []),
+    );
+    for (const [workspace, identities] of pendingByWorkspace) {
+      if (identities.some((identity) => identity.task_queue === workflow.taskQueue)) matchingWorkspaces.add(workspace);
+    }
+    const workspace = matchingWorkspaces.size === 1 ? [...matchingWorkspaces][0] : undefined;
+    addWorkflowTarget(targets, {
+      workflowId: workflow.workflowId,
+      ...(workspace !== undefined && { workspace }),
+      containerCandidate:
+        containers.some((container) => container.taskQueue === workflow.taskQueue) ||
+        [...pendingByWorkspace.values()].some((identities) =>
+          identities.some((identity) => identity.task_queue === workflow.taskQueue),
+        ),
+    });
+  }
+
+  return { targets: [...targets.values()], containersWithoutVerifiedWorkflowId };
+}
+
+/**
+ * Stop known workflows while their workers can finalize, stop the captured workers, then
+ * re-describe every container candidate. The last pass closes a NotFound-to-started race.
+ */
+export async function executeStopPlan(
+  targets: readonly WorkflowStopTarget[],
+  containers: readonly RunningScanContainer[],
+  filter: readonly string[],
+  lifecycle: StopLifecycle = stopLifecycle,
+  graceMs: number = CANCELLATION_GRACE_MS,
+  pollMs: number = CANCELLATION_POLL_MS,
+  verifyMs: number = TERMINATION_VERIFY_MS,
+  candidateSettleMs: number = CANDIDATE_REGISTRATION_SETTLE_MS,
+): Promise<StopExecutionResult> {
+  const initialOutcomes = await stopWorkflowsCancelFirst(
+    targets.map((target) => target.workflowId),
+    lifecycle,
+    graceMs,
+    pollMs,
+    verifyMs,
+  );
+  const outcomes = new Map<string, WorkflowStopOutcome>();
+  for (let index = 0; index < targets.length; index++) {
+    const target = targets[index];
+    const outcome = initialOutcomes[index];
+    if (target !== undefined && outcome !== undefined) outcomes.set(target.workflowId, outcome);
+  }
+
+  const containerOutcome = await stopContainersAndVerify(
+    containers.map((container) => container.id),
+    filter,
+    lifecycle,
+  );
+  const preRegistrationWorkspaces = new Set<string>();
+
+  if (containerOutcome.kind === 'stopped') {
+    const candidates = targets.filter((target) => target.containerCandidate);
+
+    for (const target of candidates) {
+      const initialOutcome = outcomes.get(target.workflowId) ?? { kind: 'unverified' };
+      const settleDeadline = lifecycle.now() + candidateSettleMs;
+      let onlyObservedNotFound = initialOutcome.kind === 'already-closed' || initialOutcome.kind === 'unverified';
+
+      while (true) {
+        try {
+          const state = await lifecycle.describe(target.workflowId);
+          if (state.kind === 'open') {
+            onlyObservedNotFound = false;
+            const outcome = await stopWorkflowCancelFirst(target.workflowId, lifecycle, graceMs, pollMs, verifyMs);
+            outcomes.set(target.workflowId, outcome);
+            if (outcome.kind !== 'already-closed') break;
+          }
+          if (state.kind === 'terminal') {
+            onlyObservedNotFound = false;
+            if (initialOutcome.kind === 'unverified') outcomes.set(target.workflowId, { kind: 'already-closed' });
+          }
+          if (state.kind === 'unknown') {
+            onlyObservedNotFound = false;
+            outcomes.set(target.workflowId, { kind: 'unverified' });
+            break;
+          }
+          if (initialOutcome.kind === 'unverified') outcomes.set(target.workflowId, { kind: 'already-closed' });
+        } catch {
+          onlyObservedNotFound = false;
+          outcomes.set(target.workflowId, { kind: 'unverified' });
+          break;
+        }
+
+        const remaining = settleDeadline - lifecycle.now();
+        if (remaining <= 0) {
+          if (onlyObservedNotFound && target.workspace !== undefined && target.preRegistrationFallback === true) {
+            preRegistrationWorkspaces.add(target.workspace);
+          }
+          break;
+        }
+        try {
+          await lifecycle.wait(Math.min(pollMs, remaining));
+        } catch {
+          outcomes.set(target.workflowId, { kind: 'unverified' });
+          break;
+        }
+      }
+    }
+  }
+
+  return {
+    workflows: targets.map((target) => ({
+      target,
+      outcome: outcomes.get(target.workflowId) ?? { kind: 'unverified' },
+    })),
+    containers: containerOutcome,
+    preRegistrationWorkspaces: [...preRegistrationWorkspaces],
+  };
+}
+
+function appendFallback(workspace: string, lifecycle: StopLifecycle = stopLifecycle): void {
+  try {
+    lifecycle.appendFallback(workspace);
+  } catch {
+    warn(`scan ${workspace} stopped, but workflow.log could not be marked cancelled.`);
+  }
+}
+
+function reportContainerFailure(workspace: string | undefined, outcome: ContainerStopOutcome): void {
+  const target = workspace === undefined ? '--all' : workspace;
+  if (outcome.kind === 'still-running') console.error(`${outcome.remaining} scan worker(s) did not stop.`);
+  else console.error('Docker could not verify that every targeted scan worker stopped.');
+  console.error(`Retry: ${commandPrefix()} stop ${target}`);
+}
+
+function withRecordedWorkflows(containers: readonly RunningScanContainer[]): Map<string, string> {
+  const recorded = new Map<string, string>();
+  for (const container of containers) {
+    if (container.workspace === undefined || recorded.has(container.workspace)) continue;
+    const workflowId = resolveWorkflowId(container.workspace);
+    if (workflowId !== undefined) recorded.set(container.workspace, workflowId);
+  }
+  return recorded;
+}
+
+function resolveTargetWorkspaces(targets: readonly WorkflowStopTarget[]): readonly WorkflowStopTarget[] {
+  return targets.map((target) => {
+    if (target.workspace !== undefined) return target;
+    const identity = resolveScanIdentity(target.workflowId);
+    return identity.kind === 'ok' ? { ...target, workspace: identity.workspace } : target;
+  });
+}
+
+function unverifiedWorkflowCount(results: readonly WorkflowStopResult[]): number {
+  return results.filter((result) => result.outcome.kind === 'unverified').length;
+}
+
+function appendVerifiedFallbacks(result: StopExecutionResult): void {
+  const workspaces = new Set(result.preRegistrationWorkspaces);
+  for (const workflow of result.workflows) {
+    if (workflow.outcome.kind === 'forced' && workflow.target.workspace !== undefined) {
+      workspaces.add(workflow.target.workspace);
+    }
+  }
+  for (const workspace of workspaces) appendFallback(workspace);
+}
+
+function visibleWorkflowsForWorkspace(
+  workspace: string,
+  containers: readonly RunningScanContainer[],
+  pending: PendingWorkflowTargets,
+  visible: readonly RunningScanWorkflow[],
+): readonly RunningScanWorkflow[] {
+  const taskQueues = new Set(containers.flatMap((container) => container.taskQueue ?? []));
+  for (const identity of pending.byWorkspace.get(workspace) ?? []) taskQueues.add(identity.task_queue);
+
+  return visible.filter((workflow) => {
+    if (taskQueues.has(workflow.taskQueue)) return true;
+    const identity = resolveScanIdentity(workflow.workflowId);
+    return identity.kind === 'ok' && identity.workspace === workspace;
+  });
+}
+
+/** Stop one scan while keeping its worker alive long enough to flush a graceful cancellation. */
+async function stopSingleScan(workspace: string, yes: boolean): Promise<void> {
+  const filter = scanFilter(workspace);
+  const containerQuery = runningScanContainersChecked(filter);
+  if (containerQuery.kind === 'unavailable') {
+    fail(`Could not inspect the scan worker for ${workspace}.`, `Retry: ${commandPrefix()} stop ${workspace}`);
+  }
+  const containers = containerQuery.value.map((container) => ({ ...container, workspace }));
+  const recordedWorkflowId = resolveWorkflowId(workspace);
+  const recorded = new Map<string, string>();
+  if (recordedWorkflowId !== undefined) recorded.set(workspace, recordedWorkflowId);
+  const pending = readPendingTargets([workspace]);
+  const discovery = await discoverRunningWorkflows();
+  const visible =
+    discovery.kind === 'ok' ? visibleWorkflowsForWorkspace(workspace, containers, pending, discovery.workflows) : [];
+  const plan = buildWorkflowTargetPlan(containers, recorded, visible, pending.byWorkspace);
+
+  if (containers.length === 0) {
+    if (plan.targets.length === 0) {
+      if (pending.unreadableCount > 0) {
+        fail(
+          `The launch records for ${workspace} could not be read safely.`,
+          `Retry: ${commandPrefix()} stop ${workspace}`,
+        );
+      }
+      if (discovery.kind === 'unavailable') {
+        fail(
+          `Could not verify whether scan ${workspace} is still running in Temporal.`,
+          `Retry: ${commandPrefix()} stop ${workspace}`,
+        );
+      }
+      fail(`No scan found for workspace: ${workspace}`);
+    }
+
+    const onlyRecordedTarget =
+      recordedWorkflowId !== undefined &&
+      pending.references.length === 0 &&
+      pending.unreadableCount === 0 &&
+      discovery.kind === 'ok' &&
+      plan.targets.every((target) => target.workflowId === recordedWorkflowId);
+    if (onlyRecordedTarget) {
+      try {
+        const state = await describeWorkflowLifecycle(recordedWorkflowId);
+        if (state.kind === 'terminal' || state.kind === 'not-found') {
+          console.log(`Nothing was running for ${workspace}.`);
+          return;
+        }
+        if (state.kind === 'unknown') {
+          fail(
+            `Temporal returned an unknown lifecycle state for ${workspace}.`,
+            `Retry: ${commandPrefix()} stop ${workspace}`,
+          );
+        }
+      } catch {
+        fail(
+          `Could not verify whether scan ${workspace} is still running in Temporal.`,
+          `Retry: ${commandPrefix()} stop ${workspace}`,
+        );
+      }
+    }
+  }
+
+  await confirmOrExit('stop', `Stop the scan "${workspace}"?`, yes);
+  const spinner = p.spinner();
+  spinner.start(`Stopping scan ${workspace}`);
+
+  const initialResult = await executeStopPlan(plan.targets, containers, filter);
+  const visibilitySettle = await stopVisibleWorkflowsUntilSettled(
+    initialResult.workflows,
+    stopLifecycle,
+    VISIBILITY_SETTLE_MS,
+    VISIBILITY_MAX_SETTLE_MS,
+    async () => {
+      const current = await discoverRunningWorkflows();
+      return current.kind === 'ok'
+        ? {
+            kind: 'ok',
+            workflows: visibleWorkflowsForWorkspace(workspace, containers, pending, current.workflows),
+          }
+        : current;
+    },
+  );
+  const result: StopExecutionResult = {
+    workflows: visibilitySettle.results,
+    containers: initialResult.containers,
+    preRegistrationWorkspaces: initialResult.preRegistrationWorkspaces,
+  };
+  const unverified = unverifiedWorkflowCount(result.workflows);
+  const finalPending = readPendingTargets([workspace]);
+  const initialPendingKeys = new Set(
+    pending.references.map((reference) => `${reference.identity.task_queue}\0${reference.identity.workflow_id}`),
+  );
+  const newPendingCount = finalPending.references.filter(
+    (reference) => !initialPendingKeys.has(`${reference.identity.task_queue}\0${reference.identity.workflow_id}`),
+  ).length;
+  const unreadablePendingCount = Math.max(pending.unreadableCount, finalPending.unreadableCount);
+  const incomplete =
+    result.containers.kind !== 'stopped' ||
+    plan.containersWithoutVerifiedWorkflowId.length > 0 ||
+    discovery.kind === 'unavailable' ||
+    visibilitySettle.kind !== 'settled' ||
+    unreadablePendingCount > 0 ||
+    newPendingCount > 0 ||
+    unverified > 0;
+
+  if (incomplete) {
+    spinner.error(`Scan ${workspace} shutdown could not be fully verified`);
+    if (result.containers.kind !== 'stopped') reportContainerFailure(workspace, result.containers);
+    if (unverified > 0) console.error(`Temporal could not confirm closure for ${unverified} workflow(s).`);
+    if (discovery.kind === 'unavailable') console.error('Temporal could not enumerate every running scan workflow.');
+    if (visibilitySettle.kind === 'unavailable') {
+      console.error('Temporal could not complete the final scan workflow check.');
+    }
+    if (visibilitySettle.kind === 'timed-out') {
+      console.error('Temporal workflow discovery did not settle before its deadline.');
+    }
+    if (unreadablePendingCount > 0) {
+      console.error(`${unreadablePendingCount} launch record(s) could not be read safely.`);
+    }
+    if (newPendingCount > 0) console.error('A new scan launch began while shutdown was running.');
+    if (plan.containersWithoutVerifiedWorkflowId.length > 0) {
+      console.error('A legacy scan worker could not prove its candidate workflow ID.');
+    }
+    console.error(`Retry: ${commandPrefix()} stop ${workspace}`);
+    process.exit(1);
+  }
+
+  appendVerifiedFallbacks(result);
+  const clearFailures = clearPendingTargets(pending.references);
+  if (clearFailures > 0) {
+    spinner.error(`Scan ${workspace} stopped, but its launch record could not be cleared`);
+    console.error(`Retry: ${commandPrefix()} stop ${workspace}`);
+    process.exit(1);
+  }
+  spinner.stop(`Stopped scan ${workspace}`);
+}
+
+export type WorkflowDiscoveryResult =
+  | { readonly kind: 'ok'; readonly workflows: readonly RunningScanWorkflow[] }
+  | { readonly kind: 'unavailable' };
+
+async function discoverRunningWorkflows(): Promise<WorkflowDiscoveryResult> {
+  try {
+    return { kind: 'ok', workflows: await listRunningScanWorkflows() };
+  } catch {
+    return { kind: 'unavailable' };
+  }
+}
+
+interface VisibilitySettleResult {
+  readonly results: readonly WorkflowStopResult[];
+  readonly kind: 'settled' | 'unavailable' | 'timed-out';
+}
+
+/** Re-enumerate visibility until no new open workflow appears during a bounded quiet horizon. */
+export async function stopVisibleWorkflowsUntilSettled(
+  seed: readonly WorkflowStopResult[],
+  lifecycle: StopLifecycle = stopLifecycle,
+  settleMs: number = VISIBILITY_SETTLE_MS,
+  maxSettleMs: number = VISIBILITY_MAX_SETTLE_MS,
+  discover: () => Promise<WorkflowDiscoveryResult> = discoverRunningWorkflows,
+): Promise<VisibilitySettleResult> {
+  const results = new Map(seed.map((result) => [result.target.workflowId, result]));
+  const retriedUnverified = new Set<string>();
+  const recheckedAlreadyClosed = new Set<string>();
+  let quietSince = lifecycle.now();
+  const maxDeadline = quietSince + maxSettleMs;
+
+  while (true) {
+    const discovery = await discover();
+    if (discovery.kind === 'unavailable') return { kind: 'unavailable', results: [...results.values()] };
+
+    const visibleTargets = resolveTargetWorkspaces(
+      buildWorkflowTargetPlan([], new Map(), discovery.workflows).targets,
+    ).map((target) => {
+      const existingWorkspace = results.get(target.workflowId)?.target.workspace;
+      return target.workspace === undefined && existingWorkspace !== undefined
+        ? { ...target, workspace: existingWorkspace }
+        : target;
+    });
+    const residualTargets = visibleTargets.filter((target) => {
+      const current = results.get(target.workflowId);
+      if (current === undefined) return true;
+      if (current.outcome.kind === 'unverified') return !retriedUnverified.has(target.workflowId);
+      return current.outcome.kind === 'already-closed' && !recheckedAlreadyClosed.has(target.workflowId);
+    });
+    if (residualTargets.length > 0) {
+      for (const target of residualTargets) {
+        if (results.get(target.workflowId)?.outcome.kind === 'unverified') {
+          retriedUnverified.add(target.workflowId);
+        }
+        if (results.get(target.workflowId)?.outcome.kind === 'already-closed') {
+          recheckedAlreadyClosed.add(target.workflowId);
+        }
+      }
+      const outcomes = await stopWorkflowsCancelFirst(
+        residualTargets.map((target) => target.workflowId),
+        lifecycle,
+      );
+      for (let index = 0; index < residualTargets.length; index++) {
+        const target = residualTargets[index];
+        const outcome = outcomes[index];
+        if (target !== undefined && outcome !== undefined) results.set(target.workflowId, { target, outcome });
+      }
+      quietSince = lifecycle.now();
+    }
+
+    const now = lifecycle.now();
+    if (now - quietSince >= settleMs) return { kind: 'settled', results: [...results.values()] };
+    if (now >= maxDeadline) return { kind: 'timed-out', results: [...results.values()] };
+    try {
+      await lifecycle.wait(Math.min(CANCELLATION_POLL_MS, settleMs - (now - quietSince)));
+    } catch {
+      return { kind: 'timed-out', results: [...results.values()] };
+    }
+  }
+}
+
+async function stopAllScans(yes: boolean): Promise<void> {
+  const containerQuery = runningScanContainersChecked();
+  if (containerQuery.kind === 'unavailable') {
+    fail('Could not inspect running scan workers.', `Retry: ${commandPrefix()} stop --all`);
+  }
+  const containers = containerQuery.value;
+  let pending = readPendingTargets(listWorkspaces().map((workspace) => workspace.name));
+  const initialDiscovery = await discoverRunningWorkflows();
+  let visible = initialDiscovery.kind === 'ok' ? initialDiscovery.workflows : [];
+  let plan = buildWorkflowTargetPlan(containers, withRecordedWorkflows(containers), visible, pending.byWorkspace);
+  let targets = resolveTargetWorkspaces(plan.targets);
+
+  if (containers.length === 0 && targets.length === 0) {
+    if (pending.unreadableCount > 0) {
+      fail('One or more scan launch records could not be read safely.', `Retry: ${commandPrefix()} stop --all`);
+    }
+    if (initialDiscovery.kind === 'unavailable') {
+      fail('Could not verify whether scan workflows are running in Temporal.', `Retry: ${commandPrefix()} stop --all`);
+    }
+    const emptySettleDeadline = stopLifecycle.now() + VISIBILITY_MAX_SETTLE_MS;
+    while (targets.length === 0) {
+      const remaining = emptySettleDeadline - stopLifecycle.now();
+      if (remaining <= 0) {
+        console.log('No running scans to stop.');
+        return;
+      }
+      await stopLifecycle.wait(Math.min(CANCELLATION_POLL_MS, remaining));
+      const confirmation = await discoverRunningWorkflows();
+      if (confirmation.kind === 'unavailable') {
+        fail(
+          'Could not verify whether scan workflows are running in Temporal.',
+          `Retry: ${commandPrefix()} stop --all`,
+        );
+      }
+      visible = confirmation.workflows;
+      pending = readPendingTargets(listWorkspaces().map((workspace) => workspace.name));
+      if (pending.unreadableCount > 0) {
+        fail('One or more scan launch records could not be read safely.', `Retry: ${commandPrefix()} stop --all`);
+      }
+      plan = buildWorkflowTargetPlan(containers, withRecordedWorkflows(containers), visible, pending.byWorkspace);
+      targets = resolveTargetWorkspaces(plan.targets);
+    }
+  }
+
+  await confirmOrExit('stop', 'This will stop all running scans. Continue?', yes);
+  const spinner = p.spinner();
+  spinner.start('Stopping all scans');
+
+  const initialResult = await executeStopPlan(targets, containers, WORKER_FILTER);
+  const visibilitySettle = await stopVisibleWorkflowsUntilSettled(initialResult.workflows);
+  const results = visibilitySettle.results;
+
+  const combinedResult: StopExecutionResult = {
+    workflows: results,
+    containers: initialResult.containers,
+    preRegistrationWorkspaces: initialResult.preRegistrationWorkspaces,
+  };
+  const unverified = unverifiedWorkflowCount(results);
+  const finalPending = readPendingTargets(listWorkspaces().map((workspace) => workspace.name));
+  const initialPendingKeys = new Set(
+    pending.references.map(
+      (reference) => `${reference.workspace}\0${reference.identity.task_queue}\0${reference.identity.workflow_id}`,
+    ),
+  );
+  const newPendingCount = finalPending.references.filter(
+    (reference) =>
+      !initialPendingKeys.has(
+        `${reference.workspace}\0${reference.identity.task_queue}\0${reference.identity.workflow_id}`,
+      ),
+  ).length;
+  const unreadablePendingCount = Math.max(pending.unreadableCount, finalPending.unreadableCount);
+  const temporalDiscoveryFailed = initialDiscovery.kind === 'unavailable' || visibilitySettle.kind === 'unavailable';
+  const temporalDiscoveryTimedOut = visibilitySettle.kind === 'timed-out';
+  const incomplete =
+    combinedResult.containers.kind !== 'stopped' ||
+    plan.containersWithoutVerifiedWorkflowId.length > 0 ||
+    temporalDiscoveryFailed ||
+    temporalDiscoveryTimedOut ||
+    unreadablePendingCount > 0 ||
+    newPendingCount > 0 ||
+    unverified > 0;
+
+  if (incomplete) {
+    spinner.error('Scan shutdown incomplete');
+    if (combinedResult.containers.kind !== 'stopped') reportContainerFailure(undefined, combinedResult.containers);
+    if (unverified > 0) console.error(`Temporal could not confirm closure for ${unverified} workflow(s).`);
+    if (temporalDiscoveryFailed) console.error('Temporal could not enumerate every running scan workflow.');
+    if (temporalDiscoveryTimedOut) console.error('Temporal workflow discovery did not settle before its deadline.');
+    if (unreadablePendingCount > 0) {
+      console.error(`${unreadablePendingCount} launch record(s) could not be read safely.`);
+    }
+    if (newPendingCount > 0) console.error(`${newPendingCount} scan launch(es) began while shutdown was running.`);
+    if (plan.containersWithoutVerifiedWorkflowId.length > 0) {
+      console.error(
+        `${plan.containersWithoutVerifiedWorkflowId.length} legacy worker(s) could not prove a candidate workflow ID.`,
+      );
+    }
+    console.error(`Retry: ${commandPrefix()} stop --all`);
+    process.exit(1);
+  }
+
+  appendVerifiedFallbacks(combinedResult);
+  const clearFailures = clearPendingTargets(pending.references);
+  if (clearFailures > 0) {
+    spinner.error('Scans stopped, but one or more launch records could not be cleared');
+    console.error(`Retry: ${commandPrefix()} stop --all`);
+    process.exit(1);
+  }
+  const stoppedCount = Math.max(containers.length, results.length);
+  spinner.stop(`Stopped ${stoppedCount} scan${stoppedCount === 1 ? '' : 's'}`);
+}
+
+/** Resolve the omitted target from Docker without turning a failed query into an empty scan list. */
+function resolveStopTarget(): string {
+  const result = runningScanContainersChecked();
+  if (result.kind === 'unavailable') {
+    fail('Could not inspect running scan workers.', `Retry with a workspace: ${commandPrefix()} stop <workspace>`);
+  }
+  const running = [...new Set(result.value.flatMap((container) => container.workspace ?? []))];
+  if (running.length === 1) {
+    const workspace = running[0] as string;
+    console.error(`No workspace given; stopping running scan "${workspace}".`);
+    return workspace;
+  }
+  if (running.length > 1) {
+    failUsage('Multiple scans are running: specify which one, or use --all:', `  ${running.join(', ')}`);
+  }
+  if (result.value.length > 0) {
+    fail('A running scan worker has no workspace label.', `Use ${commandPrefix()} stop --all`);
+  }
+  fail('No running scans to stop.', 'Pass a workspace name to stop a specific scan.');
+}
+
+export async function stop(opts: StopOptions): Promise<void> {
+  ensureDocker();
+  if (opts.all && opts.workspace) failUsage('Pass a workspace name or --all, not both.');
+
+  const workspace = opts.all ? undefined : (opts.workspace ?? resolveStopTarget());
+  if (workspace) await stopSingleScan(workspace, opts.yes);
+  else await stopAllScans(opts.yes);
 }

apps/cli/src/commands/uninstall.ts

@@ -1,37 +0,0 @@
-/**
- * `shn uninstall` command — remove ~/.shannon/ after confirmation (npx only).
- */
-
-import fs from 'node:fs';
-import os from 'node:os';
-import path from 'node:path';
-import * as p from '@clack/prompts';
-import { stopInfra, stopWorkers } from '../docker.js';
-
-const SHANNON_HOME = path.join(os.homedir(), '.shannon');
-
-export async function uninstall(): Promise<void> {
-  p.intro('Shannon Uninstall');
-
-  if (!fs.existsSync(SHANNON_HOME)) {
-    p.log.info('Nothing to remove. Shannon is not configured on this machine.');
-    p.outro('Done.');
-    return;
-  }
-
-  const confirmed = await p.confirm({
-    message: 'This will permanently remove all past scan data, saved configurations, and API keys. Continue?',
-  });
-  if (p.isCancel(confirmed) || !confirmed) {
-    p.cancel('Aborted.');
-    process.exit(0);
-  }
-
-  // Stop any running containers first
-  stopWorkers();
-  stopInfra(false);
-
-  fs.rmSync(SHANNON_HOME, { recursive: true, force: true });
-  p.log.success('All Shannon data has been removed.');
-  p.outro('Shannon has been uninstalled. Run `npx @keygraph/shannon setup` to start fresh.');
-}

apps/cli/src/commands/workspaces.ts

@@ -1,35 +0,0 @@
-/**
- * `shannon workspaces` command — list all workspaces.
- */
-
-import { execFileSync } from 'node:child_process';
-import os from 'node:os';
-import { getWorkerImage } from '../docker.js';
-import { getWorkspacesDir } from '../home.js';
-
-export function workspaces(version: string): void {
-  const workspacesDir = getWorkspacesDir();
-  const image = getWorkerImage(version);
-
-  try {
-    execFileSync(
-      'docker',
-      [
-        'run',
-        '--rm',
-        '-v',
-        `${workspacesDir}:/app/workspaces`,
-        '-e',
-        'WORKSPACES_DIR=/app/workspaces',
-        image,
-        'node',
-        'apps/worker/dist/temporal/workspaces.js',
-      ],
-      { stdio: 'inherit', ...(os.platform() === 'win32' && { env: { ...process.env, MSYS_NO_PATHCONV: '1' } }) },
-    );
-  } catch {
-    console.error('ERROR: Failed to list workspaces. Is the Docker image available?');
-    console.error(`  Run: docker pull ${image}`);
-    process.exit(1);
-  }
-}

apps/cli/src/config/resolver.ts

@@ -7,8 +7,17 @@
 
 import fs from 'node:fs';
 import { parse as parseTOML } from 'smol-toml';
+import { fail } from '../errors.js';
 import { getConfigFile } from '../home.js';
 import { getMode } from '../mode.js';
+import {
+  type CuratedProviderId,
+  DEFAULT_MODEL_SPEC,
+  GENERIC_API_KEY_ENV,
+  isCuratedProvider,
+  PROVIDER_API_KEY_ENV,
+  parseModelSpec,
+} from '../model-spec.js';
 
 // === TOML ↔ Env Mapping ===
 
@@ -23,35 +32,40 @@ interface ConfigMapping {
 
 /** Maps every supported env var to its TOML path (section.key) and expected type. */
 const CONFIG_MAP: readonly ConfigMapping[] = [
-  // Core
-  { env: 'CLAUDE_CODE_MAX_OUTPUT_TOKENS', toml: 'core.max_tokens', type: 'number' },
-  { env: 'CLAUDE_ADAPTIVE_THINKING', toml: 'core.adaptive_thinking', type: 'boolean', boolFormat: 'literal' },
+  // Core — base_url points any provider at a proxy or gateway
+  { env: 'SHANNON_AI_MODEL', toml: 'core.model', type: 'string' },
+  { env: 'SHANNON_AI_BASE_URL', toml: 'core.base_url', type: 'string' },
 
   // Anthropic
   { env: 'ANTHROPIC_API_KEY', toml: 'anthropic.api_key', type: 'string' },
   { env: 'CLAUDE_CODE_OAUTH_TOKEN', toml: 'anthropic.oauth_token', type: 'string' },
 
+  // OpenAI — format picks the wire API a gateway serves
+  { env: 'OPENAI_API_KEY', toml: 'openai.api_key', type: 'string' },
+  { env: 'SHANNON_AI_OPENAI_FORMAT', toml: 'openai.format', type: 'string' },
+
+  // xAI
+  { env: 'XAI_API_KEY', toml: 'xai.api_key', type: 'string' },
+
   // Bedrock
-  { env: 'CLAUDE_CODE_USE_BEDROCK', toml: 'bedrock.use', type: 'boolean' },
   { env: 'AWS_REGION', toml: 'bedrock.region', type: 'string' },
   { env: 'AWS_BEARER_TOKEN_BEDROCK', toml: 'bedrock.token', type: 'string' },
 
-  // Vertex
-  { env: 'CLAUDE_CODE_USE_VERTEX', toml: 'vertex.use', type: 'boolean' },
-  { env: 'CLOUD_ML_REGION', toml: 'vertex.region', type: 'string' },
-  { env: 'ANTHROPIC_VERTEX_PROJECT_ID', toml: 'vertex.project_id', type: 'string' },
-  { env: 'GOOGLE_APPLICATION_CREDENTIALS', toml: 'vertex.key_path', type: 'string' },
-
-  // Custom Base URL
-  { env: 'ANTHROPIC_BASE_URL', toml: 'custom_base_url.base_url', type: 'string' },
-  { env: 'ANTHROPIC_AUTH_TOKEN', toml: 'custom_base_url.auth_token', type: 'string' },
-
-  // Model tiers
-  { env: 'ANTHROPIC_SMALL_MODEL', toml: 'models.small', type: 'string' },
-  { env: 'ANTHROPIC_MEDIUM_MODEL', toml: 'models.medium', type: 'string' },
-  { env: 'ANTHROPIC_LARGE_MODEL', toml: 'models.large', type: 'string' },
+  // Generic — credential for any provider Shannon does not curate
+  { env: GENERIC_API_KEY_ENV, toml: 'provider.api_key', type: 'string' },
 ] as const;
 
+/** TOML section holding each curated provider's credentials, keyed by provider id. */
+const PROVIDER_SECTIONS: Readonly<Record<CuratedProviderId, string>> = {
+  anthropic: 'anthropic',
+  openai: 'openai',
+  xai: 'xai',
+  'amazon-bedrock': 'bedrock',
+};
+
+/** TOML section holding the generic credential for uncurated providers. */
+const GENERIC_PROVIDER_SECTION = 'provider';
+
 // === TOML Parsing ===
 
 type TOMLValue = string | number | boolean;
@@ -88,8 +102,9 @@ function loadTOML(): TOMLConfig | null {
     const mode = fs.statSync(configPath).mode;
     if (mode & 0o077) {
       const actual = (mode & 0o777).toString(8).padStart(3, '0');
-      console.error(`\nInsecure permissions (${actual}) on ${configPath}. Run: chmod 600 ${configPath}\n`);
-      process.exit(1);
+      fail(
+        `Your config file is readable by other users on this machine (${actual}). Lock it down: chmod 600 ${configPath}`,
+      );
     }
   }
 
@@ -98,9 +113,7 @@ function loadTOML(): TOMLConfig | null {
     return parseTOML(content) as TOMLConfig;
   } catch (err) {
     const message = err instanceof Error ? err.message : String(err);
-    console.error(`\nFailed to parse ${configPath}: ${message}`);
-    console.error(`\nRun 'npx @keygraph/shannon setup' to reconfigure.\n`);
-    process.exit(1);
+    fail(`Failed to parse ${configPath}: ${message}`, `Run 'npx @keygraph/shannon setup' to reconfigure.`);
   }
 }
 
@@ -123,62 +136,42 @@ function buildSchema(): Map<string, Map<string, TOMLType>> {
   return schema;
 }
 
-/** Check that a provider section has all required fields and dependencies. */
-function validateProviderFields(config: TOMLConfig, provider: string, errors: string[]): void {
-  const section = config[provider] as Record<string, unknown> | undefined;
-  if (!section) return;
-  const keys = Object.keys(section);
-
-  switch (provider) {
-    case 'anthropic':
-      if (!keys.includes('api_key') && !keys.includes('oauth_token')) {
-        errors.push('[anthropic] requires either api_key or oauth_token');
-      }
-      break;
-
-    case 'custom_base_url': {
-      const required = ['base_url', 'auth_token'];
-      const missing = required.filter((k) => !keys.includes(k));
-      if (missing.length > 0) {
-        errors.push(`[custom_base_url] missing required keys: ${missing.join(', ')}`);
-      }
-      break;
+/**
+ * Check that the section backing the selected provider carries a usable
+ * credential. `core.model` names the provider, so only that section is required;
+ * other providers' sections are ignored and never forwarded. An uncurated
+ * provider draws its credential from the generic [provider] section.
+ */
+function validateProviderFields(config: TOMLConfig, providerId: string, errors: string[]): void {
+  if (!isCuratedProvider(providerId)) {
+    const section = config[GENERIC_PROVIDER_SECTION] as Record<string, unknown> | undefined;
+    if (!section || !Object.keys(section).includes('api_key')) {
+      errors.push(`[${GENERIC_PROVIDER_SECTION}] requires api_key for provider "${providerId}"`);
     }
-
-    case 'bedrock': {
-      const required = ['use', 'region', 'token'];
-      const missing = required.filter((k) => !keys.includes(k));
-      if (missing.length > 0) {
-        errors.push(`[bedrock] missing required keys: ${missing.join(', ')}`);
-      }
-      validateModelTiers(config, 'bedrock', errors);
-      break;
-    }
-
-    case 'vertex': {
-      const required = ['use', 'region', 'project_id', 'key_path'];
-      const missing = required.filter((k) => !keys.includes(k));
-      if (missing.length > 0) {
-        errors.push(`[vertex] missing required keys: ${missing.join(', ')}`);
-      }
-      validateModelTiers(config, 'vertex', errors);
-      break;
-    }
-  }
-}
-
-/** Bedrock and Vertex require a [models] section with all three tiers. */
-function validateModelTiers(config: TOMLConfig, provider: string, errors: string[]): void {
-  const models = config.models as Record<string, unknown> | undefined;
-  if (!models || typeof models !== 'object') {
-    errors.push(`[${provider}] requires a [models] section with small, medium, and large`);
     return;
   }
 
-  const required = ['small', 'medium', 'large'];
-  const missing = required.filter((k) => !Object.keys(models).includes(k));
-  if (missing.length > 0) {
-    errors.push(`[models] missing required keys for ${provider}: ${missing.join(', ')}`);
+  const sectionName = PROVIDER_SECTIONS[providerId];
+  const section = config[sectionName] as Record<string, unknown> | undefined;
+  const keys = section ? Object.keys(section) : [];
+
+  if (providerId === 'amazon-bedrock') {
+    const missing = ['region', 'token'].filter((k) => !keys.includes(k));
+    if (missing.length > 0) {
+      errors.push(`[bedrock] missing required keys: ${missing.join(', ')}`);
+    }
+    return;
+  }
+
+  if (providerId === 'anthropic') {
+    if (!keys.includes('api_key') && !keys.includes('oauth_token')) {
+      errors.push('[anthropic] requires either api_key or oauth_token');
+    }
+    return;
+  }
+
+  if (!keys.includes('api_key')) {
+    errors.push(`[${sectionName}] requires api_key`);
   }
 }
 
@@ -226,27 +219,47 @@ function validateConfig(config: TOMLConfig): string[] {
     }
   }
 
-  // 4. Only one provider section allowed (ignore empty sections)
-  const PROVIDER_SECTIONS = ['anthropic', 'custom_base_url', 'bedrock', 'vertex'] as const;
-  const present = PROVIDER_SECTIONS.filter((s) => {
-    const section = config[s];
-    return section && typeof section === 'object' && Object.keys(section).length > 0;
-  });
-  if (present.length > 1) {
-    errors.push(
-      `Multiple providers configured: [${present.join('], [')}]. Only one provider section is allowed at a time`,
-    );
+  // 4. core.model must parse and name a supported provider
+  const modelValue = config.core?.model;
+  if (modelValue !== undefined && typeof modelValue !== 'string') {
+    return errors;
+  }
+  const spec = parseModelSpec(modelValue || DEFAULT_MODEL_SPEC);
+  if (typeof spec === 'string') {
+    errors.push(`[core].model — ${spec}`);
+    return errors;
   }
 
-  // 5. Required fields per provider
-  const singleProvider = present.length === 1 ? present[0] : undefined;
-  if (singleProvider) {
-    validateProviderFields(config, singleProvider, errors);
-  }
+  // 5. The selected provider's section must carry a credential
+  validateProviderFields(config, spec.providerId, errors);
 
   return errors;
 }
 
+function assertNoCredentialConflict(toml: TOMLConfig): void {
+  const tomlBaseUrl = typeof toml.core?.base_url === 'string' ? toml.core.base_url : undefined;
+  if (!tomlBaseUrl || process.env.SHANNON_AI_BASE_URL) return;
+
+  const tomlModel = typeof toml.core?.model === 'string' ? toml.core.model : DEFAULT_MODEL_SPEC;
+  const spec = parseModelSpec(process.env.SHANNON_AI_MODEL ?? tomlModel);
+  if (typeof spec === 'string' || !isCuratedProvider(spec.providerId)) return;
+
+  for (const envVar of PROVIDER_API_KEY_ENV[spec.providerId]) {
+    const mapping = CONFIG_MAP.find((entry) => entry.env === envVar);
+    const tomlHasCredential = mapping ? getTomlValue(toml, mapping) !== undefined : false;
+    const envHasCredential = Boolean(process.env[envVar]);
+    if (!envHasCredential && !tomlHasCredential) continue;
+
+    if (envHasCredential) {
+      fail(
+        `${envVar} in your environment conflicts with the gateway credential in config.toml (core.base_url = ${tomlBaseUrl}).`,
+        `Unset ${envVar}, or set SHANNON_AI_BASE_URL to override both from the environment.`,
+      );
+    }
+    return;
+  }
+}
+
 // === Public API ===
 
 /**
@@ -255,7 +268,8 @@ function validateConfig(config: TOMLConfig): string[] {
  * For each mapped variable: if not already set in the environment,
  * look it up in ~/.shannon/config.toml and inject it into process.env.
  * Local mode uses .env exclusively — TOML is skipped.
- * Exits with an error if the TOML contains unknown or invalid keys.
+ * Exits with an error if the TOML contains unknown or invalid keys, or if an
+ * ambient credential conflicts with a TOML-configured gateway credential.
  */
 export function resolveConfig(): void {
   if (getMode() === 'local') return;
@@ -266,14 +280,15 @@ export function resolveConfig(): void {
   // Validate before injecting
   const errors = validateConfig(toml);
   if (errors.length > 0) {
-    console.error('\nInvalid configuration:');
-    for (const err of errors) {
-      console.error(`  - ${err}`);
-    }
-    console.error(`\nRun 'shn setup' to reconfigure.\n`);
-    process.exit(1);
+    fail(
+      'Invalid configuration:',
+      ...errors.map((err) => `  - ${err}`),
+      `Run 'npx @keygraph/shannon setup' to reconfigure.`,
+    );
   }
 
+  assertNoCredentialConflict(toml);
+
   for (const mapping of CONFIG_MAP) {
     if (process.env[mapping.env]) continue;
 

apps/cli/src/config/writer.ts

@@ -8,12 +8,13 @@ import { getConfigFile } from '../home.js';
 // === Types ===
 
 export interface ShannonConfig {
-  core?: { max_tokens?: number; adaptive_thinking?: boolean };
+  core?: { model?: string; base_url?: string };
   anthropic?: { api_key?: string; oauth_token?: string };
-  custom_base_url?: { base_url?: string; auth_token?: string };
-  bedrock?: { use?: boolean; region?: string; token?: string };
-  vertex?: { use?: boolean; region?: string; project_id?: string; key_path?: string };
-  models?: { small?: string; medium?: string; large?: string };
+  openai?: { api_key?: string; format?: string };
+  xai?: { api_key?: string };
+  bedrock?: { region?: string; token?: string };
+  /** Generic credential for any provider Shannon does not curate. Maps to SHANNON_AI_API_KEY. */
+  provider?: { api_key?: string };
 }
 
 // === File Operations ===

apps/cli/src/confirm.ts

@@ -0,0 +1,43 @@
+/**
+ * Shared confirmation prompt for destructive or batch commands.
+ *
+ * `stop` and `reset` gate their action behind the same "confirm unless --yes"
+ * flow. Centralizing it here keeps the behavior identical across commands and
+ * impossible to change in only one place by accident.
+ */
+
+import * as p from '@clack/prompts';
+import { requireInteractive } from './tty.js';
+
+/**
+ * Ask the user to confirm an action, unless `yes` was passed. Off a TTY without
+ * `--yes`, fails fast rather than hanging on a prompt. Exits 0 if the user declines.
+ */
+export async function confirmOrExit(command: string, message: string, yes: boolean): Promise<void> {
+  if (yes) {
+    return;
+  }
+
+  requireInteractive(command, 'Re-run with --yes to skip this confirmation.');
+  const confirmed = await p.confirm({ message });
+  if (p.isCancel(confirmed) || !confirmed) {
+    p.cancel('Aborted.');
+    process.exit(0);
+  }
+}
+
+/**
+ * Severe-tier confirmation: the user must type `word` exactly to proceed. Unlike
+ * `confirmOrExit` there is no `--yes` bypass. Off a TTY it fails fast; exits 0 if declined.
+ */
+export async function confirmByTyping(command: string, word: string): Promise<void> {
+  requireInteractive(command, `'${command}' cannot be run non-interactively.`);
+  const typed = await p.text({
+    message: `Type ${word} to confirm — this cannot be undone:`,
+    validate: (value) => (value === word ? undefined : `Type ${word} to proceed, or press Ctrl-C to abort.`),
+  });
+  if (p.isCancel(typed) || typed !== word) {
+    p.cancel('Aborted.');
+    process.exit(0);
+  }
+}

apps/cli/src/docker.ts

@@ -12,17 +12,46 @@ import os from 'node:os';
 import path from 'node:path';
 import { setTimeout as sleep } from 'node:timers/promises';
 import { fileURLToPath } from 'node:url';
-import { getMode } from './mode.js';
+import type { SpinnerResult } from '@clack/prompts';
+import { envBool, PI_AUTH_CONTAINER_PATH } from './env.js';
+import { fail, warn } from './errors.js';
+import { getMode, isDevMode } from './mode.js';
+import { INTERNAL_DIR } from './paths.js';
+import { runStep, spawnCaptured, surfaceOutput } from './ui.js';
 
 const __dirname = path.dirname(fileURLToPath(import.meta.url));
 
 const NPX_IMAGE_REPO = 'keygraph/shannon';
 const DEV_IMAGE = 'shannon-worker';
 
+/** Docker label stamped on each worker container, mapping it back to its workspace so a single scan can be stopped by name. */
+const WORKSPACE_LABEL = 'shannon.workspace';
+
+/** Docker label that joins a worker container to the Temporal workflow polling its unique task queue. */
+const TASK_QUEUE_LABEL = 'shannon.task-queue';
+
+/** Docker label carrying the workflow ID selected before the worker starts. */
+const WORKFLOW_ID_LABEL = 'shannon.workflow-id';
+
+/** Image/container protocol proving that the worker honors the preselected workflow ID. */
+const WORKER_PROTOCOL_LABEL = 'shannon.worker-protocol';
+export const WORKFLOW_ID_PROTOCOL = 'workflow-id-v1';
+
 export function getWorkerImage(version: string): string {
   return getMode() === 'local' ? DEV_IMAGE : `${NPX_IMAGE_REPO}:${version}`;
 }
 
+/** True when the working directory supplies a Dockerfile and build context. */
+export function canBuildImage(): boolean {
+  if (getMode() === 'local') return true;
+  if (!isDevMode()) return false;
+
+  const hasDockerfile = fs.existsSync(path.resolve('Dockerfile'));
+  const hasCompose = fs.existsSync(path.resolve('docker-compose.yml'));
+
+  return hasDockerfile && hasCompose;
+}
+
 function getComposeFile(): string {
   return getMode() === 'local'
     ? path.resolve('docker-compose.yml')
@@ -53,83 +82,224 @@ function runOutput(cmd: string, args: string[]): string {
   }
 }
 
+/** Run a command asynchronously, resolving true on success. Never rejects. */
+function spawnQuiet(cmd: string, args: string[]): Promise<boolean> {
+  return new Promise((resolve) => {
+    const child = spawn(cmd, args, { stdio: 'ignore' });
+    child.on('close', (code) => resolve(code === 0));
+    child.on('error', () => resolve(false));
+  });
+}
+
+const TEMPORAL_CONTAINER = 'shannon-temporal';
+const TEMPORAL_ADDRESS = 'localhost:7233';
+
+/** Build `docker exec` args for a `temporal` CLI command run inside the Temporal container. */
+function temporalCmd(...args: string[]): string[] {
+  return ['exec', TEMPORAL_CONTAINER, 'temporal', ...args, '--address', TEMPORAL_ADDRESS];
+}
+
+/**
+ * Verify Docker is installed and its daemon is running, exiting otherwise.
+ * `docker info` succeeds only when both are true. Call this before any command
+ * that shells out to Docker.
+ */
+export function ensureDocker(): void {
+  try {
+    execFileSync('docker', ['info'], { stdio: 'pipe' });
+  } catch {
+    fail(
+      'Docker must be installed and running. Start Docker and try again.',
+      'Install Docker: https://docs.docker.com/get-docker/',
+    );
+  }
+}
+
 /**
  * Check if Temporal is running and healthy.
  */
 export function isTemporalReady(): boolean {
-  const output = runOutput('docker', [
-    'exec',
-    'shannon-temporal',
-    'temporal',
-    'operator',
-    'cluster',
-    'health',
-    '--address',
-    'localhost:7233',
-  ]);
+  const output = runOutput('docker', temporalCmd('operator', 'cluster', 'health'));
   return output.includes('SERVING');
 }
 
-/**
- * Ensure Temporal is running via compose.
- */
-export async function ensureInfra(): Promise<void> {
+/** Start (or find) Temporal via compose and wait until it serves; exits the process on failure. */
+async function ensureTemporalHealthy(spinner: SpinnerResult): Promise<void> {
   if (isTemporalReady()) {
     return;
   }
 
+  // Drive the caller's spinner — the whole "start" flow is one spinner, not several.
+  spinner.message('Starting Temporal');
   const composeFile = getComposeFile();
-  console.log('Starting Shannon infrastructure...');
-  execFileSync('docker', ['compose', '-f', composeFile, 'up', '-d'], { stdio: 'inherit' });
+  const result = await spawnCaptured('docker', ['compose', '-f', composeFile, 'up', '-d']);
+  if (!result.ok) {
+    spinner.error('Could not start Temporal');
+    surfaceOutput(result.output);
+    process.exit(1);
+  }
 
-  console.log('Waiting for Temporal to be ready...');
+  spinner.message('Waiting for Temporal to be ready');
   for (let i = 0; i < 30; i++) {
     if (isTemporalReady()) {
-      console.log('Temporal is ready!');
       return;
     }
     await sleep(2000);
   }
-  console.error('Timeout waiting for Temporal');
+
+  spinner.error('Temporal did not become ready in time');
   process.exit(1);
 }
 
+const DEFAULT_RETENTION_HOURS = 168;
+const RETENTION_ENV = 'SHANNON_TEMPORAL_RETENTION';
+const RETENTION_NAMESPACE = 'default';
+
 /**
- * Build the worker image locally (local mode only).
+ * Desired retention in whole hours: unset or empty env → 168 (7 days); a positive
+ * whole-hour override like `72h`; anything else warns and returns null (leave unchanged).
  */
-export function buildImage(noCache: boolean): void {
-  console.log(`Building ${DEV_IMAGE}...`);
+function desiredRetentionHours(): number | null {
+  const raw = process.env[RETENTION_ENV];
+  if (raw === undefined || raw.trim() === '') {
+    return DEFAULT_RETENTION_HOURS;
+  }
+  const match = raw.trim().match(/^([1-9][0-9]*)h$/);
+  if (!match) {
+    warn(
+      `Ignoring invalid ${RETENTION_ENV} "${raw}" — Temporal retention left unchanged.`,
+      'Use a positive whole number of hours, e.g. "168h".',
+    );
+    return null;
+  }
+  return Number(match[1]);
+}
+
+/** Convert a Go duration such as "24h0m0s" or "168h" to whole seconds, or null when it doesn't parse. */
+function parseGoDurationSeconds(text: string): number | null {
+  const match = text.match(/^(?:(\d+)h)?(?:(\d+)m)?(?:(\d+)s)?$/);
+  if (!match || (match[1] === undefined && match[2] === undefined && match[3] === undefined)) {
+    return null;
+  }
+  const hours = Number(match[1] ?? 0);
+  const minutes = Number(match[2] ?? 0);
+  const seconds = Number(match[3] ?? 0);
+  return hours * 3600 + minutes * 60 + seconds;
+}
+
+/**
+ * Current retention of the `default` namespace in seconds, or null when it can't be read.
+ * `runOutput` returns '' on a failed describe, so a failed read and an unparseable one both
+ * collapse to null — either way the live value is unknown, which the caller handles the same way.
+ */
+function readCurrentRetentionSeconds(): number | null {
+  const output = runOutput('docker', temporalCmd('operator', 'namespace', 'describe', RETENTION_NAMESPACE));
+  const match = output.match(/WorkflowExecutionRetentionTtl\s+(\S+)/);
+  if (!match || match[1] === undefined) {
+    return null;
+  }
+  return parseGoDurationSeconds(match[1]);
+}
+
+/**
+ * Converge the `default` namespace's retention to the CLI-owned value after Temporal is
+ * healthy. The CLI is the authority: a manual change is replaced on the next start unless
+ * the operator sets the matching override. A describe or update failure warns once that the
+ * requested value wasn't applied and never blocks the scan.
+ */
+function convergeNamespaceRetention(): void {
+  const hours = desiredRetentionHours();
+  if (hours === null) {
+    return;
+  }
+
+  const currentSeconds = readCurrentRetentionSeconds();
+  if (currentSeconds === null) {
+    warn(
+      `Could not read Temporal retention for namespace "${RETENTION_NAMESPACE}" — the requested value (${hours}h) was not applied.`,
+    );
+    return;
+  }
+
+  if (currentSeconds === hours * 3600) {
+    return;
+  }
+
+  const updated = runQuiet(
+    'docker',
+    temporalCmd('operator', 'namespace', 'update', '--namespace', RETENTION_NAMESPACE, '--retention', `${hours}h`),
+  );
+  if (!updated) {
+    warn(`Could not update Temporal retention to ${hours}h — the requested value was not applied.`);
+  }
+}
+
+/**
+ * Ensure Temporal is running via compose, then converge its scan-history retention.
+ */
+export async function ensureInfra(spinner: SpinnerResult): Promise<void> {
+  await ensureTemporalHealthy(spinner);
+  convergeNamespaceRetention();
+}
+
+/**
+ * Build the worker image from the repository, tagged with the name this mode
+ * resolves at run time.
+ */
+export function buildImage(noCache: boolean, version: string): void {
+  const image = getWorkerImage(version);
+  console.log(`Building ${image}...`);
   const args = ['build'];
   if (noCache) args.push('--no-cache');
-  args.push('-t', DEV_IMAGE, '.');
+  args.push('-t', image, '.');
   execFileSync('docker', args, { stdio: 'inherit' });
-  console.log(`Build complete: ${DEV_IMAGE}`);
+  console.log(`Build complete: ${image}`);
 }
 
 /**
  * Ensure the worker image is available.
- * Local mode: auto-builds if missing. NPX mode: pulls from Docker Hub.
+ * Buildable checkout: auto-builds if missing. Otherwise: pulls from Docker Hub.
  */
 export function ensureImage(version: string): void {
   const image = getWorkerImage(version);
   const exists = runQuiet('docker', ['image', 'inspect', image]);
-  if (exists) return;
+  if (exists) {
+    ensureWorkerImageProtocol(image);
+    return;
+  }
 
-  if (getMode() === 'local') {
-    console.log('Worker image not found, building...');
-    buildImage(false);
+  if (canBuildImage()) {
+    console.log('Shannon image not found, building...');
+    buildImage(false, version);
   } else {
     console.log(`Pulling ${image}...`);
     try {
       execFileSync('docker', ['pull', image], { stdio: 'inherit' });
     } catch {
-      console.error(`\nERROR: Failed to pull ${image}`);
-      console.error('The image may not be available for your platform yet.');
-      console.error('Check https://hub.docker.com/r/keygraph/shannon for available tags.');
-      process.exit(1);
+      fail(
+        `Failed to pull ${image}`,
+        'The image may not be available for your platform yet.',
+        'Check https://hub.docker.com/r/keygraph/shannon for available tags.',
+      );
     }
     pruneOldImages(version);
   }
+  ensureWorkerImageProtocol(image);
+}
+
+/** Refuse a stale worker image that would ignore the CLI-selected workflow ID. */
+function ensureWorkerImageProtocol(image: string): void {
+  const protocol = runOutput('docker', [
+    'image',
+    'inspect',
+    image,
+    '--format',
+    `{{ index .Config.Labels "${WORKER_PROTOCOL_LABEL}" }}`,
+  ]);
+  if (protocol === WORKFLOW_ID_PROTOCOL) return;
+
+  const hint = canBuildImage() ? 'Run ./shannon build, then retry.' : 'Reinstall this Shannon version, then retry.';
+  fail('The Shannon worker image is incompatible with this CLI.', hint);
 }
 
 /**
@@ -189,7 +359,7 @@ function shouldSkipHostsName(name: string, hostname: string): boolean {
  * `host-gateway` so they target the host's loopback instead of the container's.
  */
 function forwardEtcHostsFlags(): string[] {
-  if (process.env.SHANNON_FORWARD_HOSTS === 'false') return [];
+  if (!envBool('SHANNON_FORWARD_HOSTS', true)) return [];
   if (os.platform() === 'win32') return [];
 
   let content: string;
@@ -233,28 +403,40 @@ export interface WorkerOptions {
   repo: { hostPath: string; containerPath: string };
   workspacesDir: string;
   taskQueue: string;
+  workflowId: string;
   containerName: string;
   envFlags: string[];
   config?: { hostPath: string; containerPath: string };
-  credentials?: string;
   promptsDir?: string;
   outputDir?: string;
   workspace: string;
   pipelineTesting?: boolean;
-  debug?: boolean;
+  keepContainer?: boolean;
+  piAuthHostPath?: string;
 }
 
 /**
  * Spawn the worker container in detached mode and return the process.
- * When `opts.debug` is true, omits `--rm` so the container persists for log inspection.
+ * When `opts.keepContainer` is true, omits `--rm` so the container persists for log inspection.
  */
 export function spawnWorker(opts: WorkerOptions): ChildProcess {
   const args = ['run', '-d'];
-  if (!opts.debug) {
+  if (!opts.keepContainer) {
     args.push('--rm');
   }
   args.push('--name', opts.containerName, '--network', 'shannon-net');
 
+  // Keep the launch identity on the container before session.json exists. The fixed workflow
+  // ID lets stop verify the pre-registration window without trusting visibility timing.
+  args.push(
+    '--label',
+    `${WORKSPACE_LABEL}=${opts.workspace}`,
+    '--label',
+    `${TASK_QUEUE_LABEL}=${opts.taskQueue}`,
+    '--label',
+    `${WORKFLOW_ID_LABEL}=${opts.workflowId}`,
+  );
+
   // Add host flag for Linux
   args.push(...addHostFlag());
 
@@ -270,12 +452,13 @@ export function spawnWorker(opts: WorkerOptions): ChildProcess {
   args.push('-v', `${opts.workspacesDir}:/app/workspaces`);
   args.push('-v', `${opts.repo.hostPath}:${opts.repo.containerPath}:ro`);
 
-  // Writable overlays: shadow .shannon/ and .playwright/ inside the :ro repo with workspace-backed dirs
-  const workspacePath = path.join(opts.workspacesDir, opts.workspace);
-  args.push('-v', `${path.join(workspacePath, 'deliverables')}:${opts.repo.containerPath}/.shannon/deliverables`);
-  args.push('-v', `${path.join(workspacePath, 'scratchpad')}:${opts.repo.containerPath}/.shannon/scratchpad`);
-  args.push('-v', `${path.join(workspacePath, '.playwright-cli')}:${opts.repo.containerPath}/.shannon/.playwright-cli`);
-  args.push('-v', `${path.join(workspacePath, '.playwright')}:${opts.repo.containerPath}/.playwright`);
+  // Writable overlays: shadow .shannon/ and .playwright/ inside the :ro repo with workspace-backed
+  // dirs, nested under the run's INTERNAL_DIR. Container paths are unchanged.
+  const internalPath = path.join(opts.workspacesDir, opts.workspace, INTERNAL_DIR);
+  args.push('-v', `${path.join(internalPath, 'deliverables')}:${opts.repo.containerPath}/.shannon/deliverables`);
+  args.push('-v', `${path.join(internalPath, 'scratchpad')}:${opts.repo.containerPath}/.shannon/scratchpad`);
+  args.push('-v', `${path.join(internalPath, '.playwright-cli')}:${opts.repo.containerPath}/.shannon/.playwright-cli`);
+  args.push('-v', `${path.join(internalPath, '.playwright')}:${opts.repo.containerPath}/.playwright`);
 
   // Local mode: mount prompts for live editing
   if (opts.promptsDir) {
@@ -286,20 +469,23 @@ export function spawnWorker(opts: WorkerOptions): ChildProcess {
     args.push('-v', `${opts.config.hostPath}:${opts.config.containerPath}:ro`);
   }
 
-  // Output directory for deliverables copy
+  // Customer-copy destination. The workflow surfaces only final report artifacts here.
   if (opts.outputDir) {
     args.push('-v', `${opts.outputDir}:/app/output`);
   }
 
-  // Mount credentials file to fixed container path
-  if (opts.credentials) {
-    args.push('-v', `${opts.credentials}:/app/credentials/google-sa-key.json:ro`);
+  // Reuse the host's pi credentials: mount only the auth file, allowing token refreshes to persist.
+  if (opts.piAuthHostPath) {
+    args.push('-v', `${opts.piAuthHostPath}:${PI_AUTH_CONTAINER_PATH}`);
   }
 
   // Environment
   args.push(...opts.envFlags);
 
-  // Container settings
+  // Container settings. Chromium's own sandbox needs syscalls Docker's default seccomp
+  // profile blocks, which is why the profile is dropped. `seccomp=unconfined` is a
+  // container-wide setting, not a per-process one: every process here runs unfiltered,
+  // the worker included — not just the browser automation that motivates it.
   args.push('--shm-size', '2gb', '--security-opt', 'seccomp=unconfined');
 
   // Image
@@ -308,6 +494,7 @@ export function spawnWorker(opts: WorkerOptions): ChildProcess {
   // Worker command
   args.push('node', 'apps/worker/dist/temporal/worker.js', opts.url, opts.repo.containerPath);
   args.push('--task-queue', opts.taskQueue);
+  args.push('--workflow-id', opts.workflowId);
   if (opts.config) {
     args.push('--config', opts.config.containerPath);
   }
@@ -328,26 +515,133 @@ export function spawnWorker(opts: WorkerOptions): ChildProcess {
   });
 }
 
-/**
- * Stop all running shannon-worker-* containers.
- */
-export function stopWorkers(): void {
-  const workers = runOutput('docker', ['ps', '-q', '--filter', 'name=shannon-worker-']);
-  if (!workers) return;
+/** `docker ps --filter` args matching every running worker container. */
+export const WORKER_FILTER: readonly string[] = ['--filter', 'name=shannon-worker-'];
 
-  const ids = workers.split('\n').filter(Boolean);
-  console.log('Stopping worker containers...');
-  execFileSync('docker', ['stop', ...ids], { stdio: 'inherit' });
+/** Result of a command-backed query whose unavailable state must not be mistaken for an empty result. */
+export type CommandQueryResult<T> = { kind: 'ok'; value: T } | { kind: 'unavailable' };
+
+/** Identity carried by a running scan worker container. Older workers may lack the newer labels. */
+export interface RunningScanContainer {
+  readonly id: string;
+  readonly workspace?: string;
+  readonly taskQueue?: string;
+  readonly workflowId?: string;
+  readonly workerProtocol?: string;
+}
+
+/** `docker ps --filter` args matching one scan's worker container(s), by workspace label. */
+export function scanFilter(workspace: string): readonly string[] {
+  return ['--filter', `label=${WORKSPACE_LABEL}=${workspace}`];
 }
 
 /**
- * Tear down the compose stack.
+ * IDs of running containers matching the filter. Re-querying this after a stop is
+ * the authoritative check for whether containers actually stopped — `docker stop`'s
+ * exit code can't distinguish "already gone" from "failed to stop".
  */
-export function stopInfra(clean: boolean): void {
+export function runningContainersChecked(filter: readonly string[]): CommandQueryResult<string[]> {
+  try {
+    const output = execFileSync('docker', ['ps', '-q', ...filter], { stdio: 'pipe', encoding: 'utf-8' }).trim();
+    return { kind: 'ok', value: output.split('\n').filter(Boolean) };
+  } catch {
+    return { kind: 'unavailable' };
+  }
+}
+
+/**
+ * Best-effort counterpart for callers where Docker unavailability is intentionally
+ * presented as no local running containers.
+ */
+export function runningContainers(filter: readonly string[]): string[] {
+  const result = runningContainersChecked(filter);
+  return result.kind === 'ok' ? result.value : [];
+}
+
+function normalizedLabel(value: string | undefined): string | undefined {
+  const normalized = value?.trim();
+  return normalized && normalized !== '<no value>' ? normalized : undefined;
+}
+
+/**
+ * Running scan containers with the labels needed to correlate a worker to its Temporal
+ * workflow. A successful query keeps unlabeled legacy workers in the result by ID.
+ */
+export function runningScanContainersChecked(
+  filter: readonly string[] = WORKER_FILTER,
+): CommandQueryResult<RunningScanContainer[]> {
+  try {
+    const format = `{{.ID}}\t{{ index .Labels "${WORKSPACE_LABEL}" }}\t{{ index .Labels "${TASK_QUEUE_LABEL}" }}\t{{ index .Labels "${WORKFLOW_ID_LABEL}" }}\t{{ index .Labels "${WORKER_PROTOCOL_LABEL}" }}`;
+    const output = execFileSync('docker', ['ps', ...filter, '--format', format], {
+      stdio: 'pipe',
+      encoding: 'utf-8',
+    }).trim();
+    if (!output) return { kind: 'ok', value: [] };
+
+    const containers: RunningScanContainer[] = [];
+    for (const line of output.split('\n')) {
+      const [rawId, rawWorkspace, rawTaskQueue, rawWorkflowId, rawWorkerProtocol] = line.split('\t');
+      const id = rawId?.trim();
+      if (!id) return { kind: 'unavailable' };
+      const workspace = normalizedLabel(rawWorkspace);
+      const taskQueue = normalizedLabel(rawTaskQueue);
+      const workflowId = normalizedLabel(rawWorkflowId);
+      const workerProtocol = normalizedLabel(rawWorkerProtocol);
+      containers.push({
+        id,
+        ...(workspace !== undefined && { workspace }),
+        ...(taskQueue !== undefined && { taskQueue }),
+        ...(workflowId !== undefined && { workflowId }),
+        ...(workerProtocol !== undefined && { workerProtocol }),
+      });
+    }
+    return { kind: 'ok', value: containers };
+  } catch {
+    return { kind: 'unavailable' };
+  }
+}
+
+/**
+ * Workspace names of every running worker container, read from the shannon.workspace
+ * label each scan is stamped with at spawn. The checked form preserves Docker query
+ * failures so lifecycle commands do not mistake an unavailable daemon for an empty list.
+ */
+export function runningScanWorkspacesChecked(): CommandQueryResult<string[]> {
+  const result = runningScanContainersChecked();
+  if (result.kind === 'unavailable') return result;
+  return {
+    kind: 'ok',
+    value: result.value.flatMap((container) => (container.workspace === undefined ? [] : [container.workspace])),
+  };
+}
+
+/** Best-effort counterpart for callers that only need the local scan list. */
+export function runningScanWorkspaces(): string[] {
+  const result = runningScanWorkspacesChecked();
+  return result.kind === 'ok' ? result.value : [];
+}
+
+/**
+ * Stop containers by ID, tolerating any that vanished between being listed and
+ * stopped (a `--rm` worker exiting is success, not an error). Async so a spinner
+ * can animate during docker's graceful-shutdown wait.
+ */
+export async function stopContainers(ids: string[]): Promise<void> {
+  await Promise.all(ids.map((id) => spawnQuiet('docker', ['stop', id])));
+}
+
+/**
+ * Tear down the compose stack. When `clean` is set, volumes are removed too.
+ */
+export async function stopInfra(clean: boolean): Promise<void> {
   const composeFile = getComposeFile();
   const args = ['compose', '-f', composeFile, 'down'];
   if (clean) args.push('-v');
-  execFileSync('docker', args, { stdio: 'inherit' });
+  const label = clean ? 'Removing Temporal data and volumes' : 'Stopping Temporal';
+  const step = await runStep(label, 'docker', args);
+  if (!step.ok) {
+    fail(`${label} failed. See the output above.`);
+  }
 }
 
 /**
@@ -363,16 +657,3 @@ function pruneOldImages(currentVersion: string): void {
     runQuiet('docker', ['rmi', `${NPX_IMAGE_REPO}:${tag}`]);
   }
 }
-
-/**
- * List running worker containers.
- */
-export function listRunningWorkers(): string {
-  return runOutput('docker', [
-    'ps',
-    '--filter',
-    'name=shannon-worker-',
-    '--format',
-    'table {{.Names}}\t{{.Status}}\t{{.RunningFor}}',
-  ]);
-}

apps/cli/src/env.ts

@@ -5,30 +5,77 @@
  * NPX mode: fills gaps from ~/.shannon/config.toml (no .env).
  */
 
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
 import dotenv from 'dotenv';
 import { resolveConfig } from './config/resolver.js';
 import { getMode } from './mode.js';
+import {
+  CURATED_PROVIDERS,
+  type CuratedProviderId,
+  GENERIC_API_KEY_ENV,
+  isCuratedProvider,
+  PROVIDER_API_KEY_ENV,
+  PROVIDER_CREDENTIAL_HINT,
+  PROVIDER_EXTRA_ENV,
+  resolveModelSpec,
+} from './model-spec.js';
 
-/** Environment variables forwarded to worker containers. */
-const FORWARD_VARS = [
-  'ANTHROPIC_API_KEY',
-  'ANTHROPIC_BASE_URL',
-  'ANTHROPIC_AUTH_TOKEN',
-  'CLAUDE_CODE_OAUTH_TOKEN',
-  'CLAUDE_CODE_USE_BEDROCK',
-  'AWS_REGION',
-  'AWS_BEARER_TOKEN_BEDROCK',
-  'CLAUDE_CODE_USE_VERTEX',
-  'CLOUD_ML_REGION',
-  'ANTHROPIC_VERTEX_PROJECT_ID',
-  'GOOGLE_APPLICATION_CREDENTIALS',
-  'ANTHROPIC_SMALL_MODEL',
-  'ANTHROPIC_MEDIUM_MODEL',
-  'ANTHROPIC_LARGE_MODEL',
-  'CLAUDE_CODE_MAX_OUTPUT_TOKENS',
-  'CLAUDE_ADAPTIVE_THINKING',
+/**
+ * Variables forwarded to every worker container regardless of provider. Each is
+ * forwarded only when set, so an unused one never appears in the container.
+ * SHANNON_AI_API_KEY rides along because it is provider-neutral.
+ */
+const COMMON_FORWARD_VARS = [
+  'SHANNON_AI_MODEL',
+  'SHANNON_AI_BASE_URL',
+  'SHANNON_AI_OPENAI_FORMAT',
+  // Opt-in debug flag: when set, the worker persists a bounded, sanitized snippet of a failed
+  // provider turn's raw error message to error.log. Off by default; provider prose stays out of
+  // durable state unless an operator deliberately enables it for a diagnosis.
+  'SHANNON_DEBUG_PROVIDER_ERRORS',
+  GENERIC_API_KEY_ENV,
 ] as const;
 
+/**
+ * Credential variables for one provider. Only the selected provider's entries are
+ * forwarded, so a key for an unused provider never enters the scan container. An
+ * uncurated provider has none — it relies on the common SHANNON_AI_API_KEY.
+ */
+function providerForwardVars(providerId: string): readonly string[] {
+  if (!isCuratedProvider(providerId)) return [];
+  return [...PROVIDER_API_KEY_ENV[providerId], ...PROVIDER_EXTRA_ENV[providerId]];
+}
+
+/** Parse a user-facing boolean env var: `1`/`true` (any case) true, `0`/`false`/empty false, else the default. */
+export function envBool(name: string, defaultValue: boolean): boolean {
+  const raw = process.env[name]?.trim().toLowerCase();
+  if (raw === undefined || raw === '') return defaultValue;
+  if (raw === '1' || raw === 'true') return true;
+  if (raw === '0' || raw === 'false') return false;
+  return defaultValue;
+}
+
+const USE_PI_AUTH_ENV = 'SHANNON_USE_PI_AUTH';
+
+/** Where the host's auth.json is mounted: pi's standard location (worker HOME is /tmp), read natively. */
+export const PI_AUTH_CONTAINER_PATH = '/tmp/.pi/agent/auth.json';
+
+/** Host path to pi's credential file. */
+export function resolveHostPiAuthPath(): string {
+  return path.join(os.homedir(), '.pi', 'agent', 'auth.json');
+}
+
+export function piAuthFlagEnabled(): boolean {
+  return envBool(USE_PI_AUTH_ENV, false);
+}
+
+/** Opted into pi auth via the flag, and the auth file exists to mount. */
+export function shouldUsePiAuth(): boolean {
+  return piAuthFlagEnabled() && fs.existsSync(resolveHostPiAuthPath());
+}
+
 /**
  * Load credentials into process.env.
  * Local mode: loads ./.env via dotenv.
@@ -44,15 +91,19 @@ export function loadEnv(): void {
 }
 
 /**
- * Build `-e KEY=VALUE` flags for docker run, only for set variables.
+ * Build `-e` flags for docker run. Forwards the common vars plus only the
+ * selected provider's credentials, passed by name (`-e KEY`) so secret values
+ * stay out of the `docker run` argv; docker inherits them from this process's env.
  */
 export function buildEnvFlags(): string[] {
   const flags: string[] = ['-e', 'TEMPORAL_ADDRESS=shannon-temporal:7233'];
 
-  for (const key of FORWARD_VARS) {
-    const value = process.env[key];
-    if (value) {
-      flags.push('-e', `${key}=${value}`);
+  const spec = resolveModelSpec();
+  const providerVars = typeof spec === 'string' ? [] : providerForwardVars(spec.providerId);
+
+  for (const key of [...COMMON_FORWARD_VARS, ...providerVars]) {
+    if (process.env[key]) {
+      flags.push('-e', key);
     }
   }
 
@@ -62,95 +113,124 @@ export function buildEnvFlags(): string[] {
 interface CredentialValidation {
   valid: boolean;
   error?: string;
-  mode: 'api-key' | 'oauth' | 'custom-base-url' | 'bedrock' | 'vertex';
-}
-
-/** Check if a custom Anthropic-compatible base URL is configured. */
-function isCustomBaseUrlConfigured(): boolean {
-  return !!(process.env.ANTHROPIC_BASE_URL && process.env.ANTHROPIC_AUTH_TOKEN);
-}
-
-/** Detect which providers are configured via environment variables. */
-function detectProviders(): string[] {
-  const providers: string[] = [];
-  if (process.env.ANTHROPIC_API_KEY) providers.push('Anthropic API key');
-  if (process.env.CLAUDE_CODE_OAUTH_TOKEN) providers.push('Anthropic OAuth');
-  if (isCustomBaseUrlConfigured()) providers.push('Custom Base URL');
-  if (process.env.CLAUDE_CODE_USE_BEDROCK === '1') providers.push('AWS Bedrock');
-  if (process.env.CLAUDE_CODE_USE_VERTEX === '1') providers.push('Google Vertex');
-  return providers;
 }
 
 /**
- * Validate that exactly one authentication method is configured.
+ * Whether the shell environment already carries a usable credential — the host's
+ * pi login, or an API key for the selected provider. Reads process.env only.
  */
-export function validateCredentials(): CredentialValidation {
-  // Reject multiple providers
-  const providers = detectProviders();
-  if (providers.length > 1) {
-    return {
-      valid: false,
-      mode: 'api-key',
-      error: `Multiple providers detected: ${providers.join(', ')}. Only one provider can be active at a time.`,
-    };
+export function hasExportedCredentials(): boolean {
+  if (shouldUsePiAuth()) return true;
+  const spec = resolveModelSpec();
+  if (typeof spec === 'string') return false;
+  return hasCredential(spec.providerId);
+}
+
+/** Whether a curated provider has its own named credential set (API key plus any extra var). */
+function hasNamedCredential(providerId: CuratedProviderId): boolean {
+  const apiKeys = PROVIDER_API_KEY_ENV[providerId];
+  if (!apiKeys.some((name) => Boolean(process.env[name]))) return false;
+  return PROVIDER_EXTRA_ENV[providerId].every((name) => Boolean(process.env[name]));
+}
+
+/** Whether the selected provider has a credential. Bedrock needs its AWS_ vars; the generic key never stands in for it. */
+function hasCredential(providerId: string): boolean {
+  if (providerId === 'amazon-bedrock') return hasNamedCredential('amazon-bedrock');
+  if (isCuratedProvider(providerId) && hasNamedCredential(providerId)) return true;
+  return Boolean(process.env[GENERIC_API_KEY_ENV]);
+}
+
+/** Curated providers with a named credential. The generic key is neutral, so it never counts toward ambiguity. */
+function configuredProviders(): CuratedProviderId[] {
+  return CURATED_PROVIDERS.filter((providerId) => hasNamedCredential(providerId));
+}
+
+/** Whether SHANNON_AI_MODEL was set by the user, rather than falling back to the default. */
+function modelExplicitlySelected(): boolean {
+  return Boolean(process.env.SHANNON_AI_MODEL?.trim());
+}
+
+/**
+ * Explain why the selected provider has no usable credential. With no model chosen
+ * the provider is only the default (anthropic), so the real state is "nothing
+ * configured" — or, if another provider's key is set, an unselected model.
+ */
+function describeMissingCredential(providerId: string): string {
+  if (modelExplicitlySelected()) {
+    const requirement = isCuratedProvider(providerId) ? PROVIDER_CREDENTIAL_HINT[providerId] : GENERIC_API_KEY_ENV;
+    const hint =
+      getMode() === 'local'
+        ? `Set ${requirement} in .env or export it.`
+        : `Export the variables or run 'npx @keygraph/shannon setup'.`;
+    return `No credentials found for provider "${providerId}". ${hint}`;
   }
 
-  if (process.env.ANTHROPIC_API_KEY) {
-    return { valid: true, mode: 'api-key' };
-  }
-  if (process.env.CLAUDE_CODE_OAUTH_TOKEN) {
-    return { valid: true, mode: 'oauth' };
-  }
-  if (isCustomBaseUrlConfigured()) {
-    return { valid: true, mode: 'custom-base-url' };
-  }
-  if (process.env.CLAUDE_CODE_USE_BEDROCK === '1') {
-    const missing: string[] = [];
-    if (!process.env.AWS_REGION) missing.push('AWS_REGION');
-    if (!process.env.AWS_BEARER_TOKEN_BEDROCK) missing.push('AWS_BEARER_TOKEN_BEDROCK');
-    if (!process.env.ANTHROPIC_SMALL_MODEL) missing.push('ANTHROPIC_SMALL_MODEL');
-    if (!process.env.ANTHROPIC_MEDIUM_MODEL) missing.push('ANTHROPIC_MEDIUM_MODEL');
-    if (!process.env.ANTHROPIC_LARGE_MODEL) missing.push('ANTHROPIC_LARGE_MODEL');
-    if (missing.length > 0) {
-      return {
-        valid: false,
-        mode: 'bedrock',
-        error: `Bedrock mode requires: ${missing.join(', ')}`,
-      };
-    }
-    return { valid: true, mode: 'bedrock' };
-  }
-  if (process.env.CLAUDE_CODE_USE_VERTEX === '1') {
-    const missing: string[] = [];
-    if (!process.env.CLOUD_ML_REGION) missing.push('CLOUD_ML_REGION');
-    if (!process.env.ANTHROPIC_VERTEX_PROJECT_ID) missing.push('ANTHROPIC_VERTEX_PROJECT_ID');
-    if (!process.env.ANTHROPIC_SMALL_MODEL) missing.push('ANTHROPIC_SMALL_MODEL');
-    if (!process.env.ANTHROPIC_MEDIUM_MODEL) missing.push('ANTHROPIC_MEDIUM_MODEL');
-    if (!process.env.ANTHROPIC_LARGE_MODEL) missing.push('ANTHROPIC_LARGE_MODEL');
-    if (missing.length > 0) {
-      return {
-        valid: false,
-        mode: 'vertex',
-        error: `Vertex AI mode requires: ${missing.join(', ')}`,
-      };
-    }
-    if (!process.env.GOOGLE_APPLICATION_CREDENTIALS) {
-      return {
-        valid: false,
-        mode: 'vertex',
-        error: 'Vertex AI mode requires GOOGLE_APPLICATION_CREDENTIALS',
-      };
-    }
-    return { valid: true, mode: 'vertex' };
+  const [provider] = configuredProviders();
+  if (provider) {
+    return `A credential for "${provider}" is set, but no model is selected. Set SHANNON_AI_MODEL=${provider}:<model-id> to use it.`;
   }
 
   const hint =
     getMode() === 'local'
-      ? `No credentials found. Set ANTHROPIC_API_KEY in .env or export it.`
-      : `Authentication not configured. Export variables or run 'npx @keygraph/shannon setup'.`;
-  return {
-    valid: false,
-    mode: 'api-key',
-    error: hint,
-  };
+      ? 'Set a provider API key in .env (for example ANTHROPIC_API_KEY).'
+      : "Run 'npx @keygraph/shannon setup' to get started.";
+  return `No credentials configured. ${hint}`;
+}
+
+/**
+ * Validate that the model selection parses and its provider has a credential.
+ * Runs before any Docker work so mistakes fail immediately.
+ */
+export function validateCredentials(): CredentialValidation {
+  // 1. Model selection must parse into a provider and model id
+  const spec = resolveModelSpec();
+  if (typeof spec === 'string') {
+    return { valid: false, error: spec };
+  }
+
+  // Pi-auth: skip the API-key checks, but the host auth file must exist to mount.
+  if (piAuthFlagEnabled()) {
+    const authPath = resolveHostPiAuthPath();
+    if (!fs.existsSync(authPath)) {
+      return {
+        valid: false,
+        error: `${USE_PI_AUTH_ENV} is set but no pi credentials were found at ${authPath}. Authenticate with pi first.`,
+      };
+    }
+    return { valid: true };
+  }
+
+  // 2. The selected provider must have a credential
+  if (!hasCredential(spec.providerId)) {
+    return { valid: false, error: describeMissingCredential(spec.providerId) };
+  }
+
+  // 3. Exactly one provider may be configured. Several complete credentials make
+  //    the scan's provider depend on SHANNON_AI_MODEL alone, which is too easy to
+  //    misread as "both are in play" and too easy to redirect by editing one line.
+  const configured = configuredProviders();
+  if (configured.length > 1) {
+    const setKeys = (id: CuratedProviderId): string[] =>
+      PROVIDER_API_KEY_ENV[id].filter((name) => Boolean(process.env[name]));
+    const list = configured.map((id) => `${id} (${setKeys(id).join(', ')})`).join(' and ');
+    const others = configured.filter((id) => id !== spec.providerId);
+    const extraVars = others.flatMap(setKeys);
+
+    const dropHint =
+      getMode() === 'local'
+        ? 'remove them from .env or unset them in your shell:'
+        : "unset them in your shell, or reconfigure with 'npx @keygraph/shannon setup':";
+
+    const lines = [`Credentials for more than one provider are set: ${list}.`];
+    if (extraVars.length > 0) {
+      lines.push(
+        `Shannon runs one provider per scan, selected by SHANNON_AI_MODEL ("${spec.providerId}:...").`,
+        `Keep ${spec.providerId} and drop the rest — ${dropHint}`,
+        `  unset ${extraVars.join(' ')}`,
+      );
+    }
+    return { valid: false, error: lines.join('\n') };
+  }
+
+  return { valid: true };
 }

apps/cli/src/errors.ts

@@ -0,0 +1,109 @@
+/**
+ * Centralized error reporting.
+ *
+ * `fail` / `failWith` — an expected, user-fixable error (bad input, missing
+ * prerequisite): a clean message on stderr and a non-zero exit, never a stack trace.
+ * `failUsage` — a malformed invocation (unknown command, bad or missing
+ * arguments): the same clean message, but a distinct exit code so callers can
+ * tell a usage mistake from an operational failure.
+ * `crash` — an unexpected error (a bug): a fixed code and a pointer to the issue tracker.
+ *
+ * JSON mode (enabled once, before parsing, for the `--json` command surface) replaces
+ * the text lines with one compact envelope on stderr — stdout stays empty — while the
+ * exit-code split is unchanged. Call sites on a JSON-capable path must exit through
+ * `failWith`/`failUsage`/`crash` (never a bare `fail` or `warn`) so every failure
+ * carries a stable code and stderr stays parseable.
+ */
+
+import fs from 'node:fs';
+
+const ISSUES_URL = 'https://github.com/KeygraphHQ/shannon/issues';
+
+const UNEXPECTED_MESSAGE = 'Shannon encountered an unexpected failure. Reference code: SHANNON_UNEXPECTED_ERROR';
+const REPORT_HINT = `If this looks like a bug, please report it: ${ISSUES_URL}`;
+
+/** Stable machine-readable failure codes for the JSON error envelope. */
+export type ErrorCode =
+  | 'CLI_USAGE'
+  | 'CLI_SCAN_NOT_FOUND'
+  | 'CLI_SCAN_IDENTITY_NOT_FOUND'
+  | 'CLI_SCAN_IDENTITY_AMBIGUOUS'
+  | 'CLI_SCAN_STATUS_UNAVAILABLE'
+  | 'CLI_SCAN_SCHEMA_UNSUPPORTED'
+  | 'CLI_PRECONDITION_FAILED'
+  | 'CLI_INTERNAL_ERROR';
+
+let jsonMode = false;
+
+/** Switch failure reporting to the JSON envelope. Set once, before any guard, parse, or dispatch. */
+export function enableJsonErrors(): void {
+  jsonMode = true;
+}
+
+/** Whether failures are reported as the JSON envelope rather than text. */
+export function jsonErrorsEnabled(): boolean {
+  return jsonMode;
+}
+
+/** Fixed unexpected-failure projection shared by the runtime and focused safety tests. */
+export function unexpectedFailureLines(): readonly string[] {
+  return [`ERROR: ${UNEXPECTED_MESSAGE}`, REPORT_HINT];
+}
+
+/**
+ * Report a failure on stderr and exit. Text mode prints the message and every hint
+ * verbatim; JSON mode writes one compact envelope (dropping the empty strings used
+ * to space text output) synchronously so `process.exit` cannot truncate it.
+ */
+function emit(exitCode: 1 | 2, code: ErrorCode, message: string, hints: readonly string[]): never {
+  if (jsonMode) {
+    const payload = JSON.stringify({ error: { code, message, hints: hints.filter((hint) => hint.trim() !== '') } });
+    fs.writeSync(process.stderr.fd, `${payload}\n`);
+    process.exit(exitCode);
+  }
+  console.error(`ERROR: ${message}`);
+  for (const hint of hints) {
+    console.error(hint);
+  }
+  process.exit(exitCode);
+}
+
+/**
+ * Report an expected, user-fixable error (with optional extra lines) and exit non-zero.
+ * Text-only paths use this; a JSON-capable path must use `failWith` so the envelope
+ * carries a real code — if a bare `fail` is ever reached in JSON mode, the fixed
+ * internal-error envelope is emitted instead of guessing a code for the message.
+ */
+export function fail(message: string, ...hints: string[]): never {
+  if (jsonMode) {
+    emit(1, 'CLI_INTERNAL_ERROR', UNEXPECTED_MESSAGE, [REPORT_HINT]);
+  }
+  emit(1, 'CLI_INTERNAL_ERROR', message, hints);
+}
+
+/** Report an expected operational failure under a stable code and exit 1. */
+export function failWith(code: ErrorCode, message: string, ...hints: string[]): never {
+  emit(1, code, message, hints);
+}
+
+/** Report a usage/argument error (with optional extra lines) and exit 2. */
+export function failUsage(message: string, ...hints: string[]): never {
+  emit(2, 'CLI_USAGE', message, hints);
+}
+
+/** Report a non-fatal warning on stderr (with optional extra lines) without exiting. */
+export function warn(message: string, ...hints: string[]): void {
+  console.error(`WARNING: ${message}`);
+  for (const hint of hints) {
+    console.error(hint);
+  }
+}
+
+/** Report an unexpected error without projecting its message, stack, or attached values. */
+export function crash(_error: unknown): never {
+  if (jsonMode) {
+    emit(1, 'CLI_INTERNAL_ERROR', UNEXPECTED_MESSAGE, [REPORT_HINT]);
+  }
+  for (const line of unexpectedFailureLines()) console.error(line);
+  process.exit(1);
+}

apps/cli/src/help.ts

@@ -0,0 +1,156 @@
+/**
+ * Per-command help text.
+ *
+ * `shannon <command> --help`, `shannon <command> -h`, and `shannon help <command>`
+ * all render the matching command's usage, so a user can discover a command's
+ * flags without scanning the global help. The global help lives in index.ts.
+ */
+
+import { commandPrefix, getMode } from './mode.js';
+
+interface CommandHelp {
+  readonly usage: readonly string[];
+  readonly description: string;
+  readonly options?: readonly (readonly [string, string])[];
+  readonly examples?: readonly string[];
+}
+
+const YES_OPTION: readonly [string, string] = [
+  '-y, --yes',
+  'Skip the confirmation prompt (required for non-interactive use)',
+];
+const HELP_OPTION: readonly [string, string] = ['-h, --help', 'Show this help'];
+
+/**
+ * `start`'s flags, the single source rendered by both the per-command help here
+ * and the global help in index.ts, so the two can never drift.
+ */
+export const START_OPTIONS: readonly (readonly [string, string])[] = [
+  ['-u, --url <url>', 'Target URL (required)'],
+  ['-r, --repo <path>', 'Repository path (required)'],
+  ['-c, --config <path>', 'Configuration file (YAML)'],
+  ['-o, --output <path>', 'Copy deliverables to this directory after the run'],
+  ['-w, --workspace <name>', 'Named workspace (auto-resumes if it exists)'],
+  ['-f, --follow', 'Stream the scan log until it finishes'],
+  ['--pipeline-testing', 'Use minimal prompts for fast testing'],
+  ['--keep-container', 'Preserve the worker container after exit for log inspection'],
+];
+
+const COMMAND_HELP: Readonly<Record<string, CommandHelp>> = {
+  start: {
+    usage: ['start -u <url> -r <path> [options]'],
+    description: 'Start a pentest scan.',
+    examples: [
+      'start -u https://example.com -r ./my-repo',
+      'start -u https://example.com -r /path/to/repo -c config.yaml -w q1-audit',
+      'start -u https://example.com -r ./my-repo --follow',
+    ],
+  },
+  stop: {
+    usage: ['stop [<workspace>] [--yes]', 'stop --all [--yes]'],
+    description:
+      'Stop one scan by workspace, or every scan with --all (Temporal stays up). With no workspace, stops the single running scan; when several are running, name one or use --all.',
+    options: [['--all', 'Stop all running scans'], YES_OPTION],
+    examples: ['stop', 'stop q1-audit', 'stop --all'],
+  },
+  reset: {
+    usage: ['reset'],
+    description: 'Stop everything and permanently remove all Temporal data and volumes.',
+  },
+  logs: {
+    usage: ['logs [<workspace>]'],
+    description:
+      "Tail a scan's live log until it completes. With no workspace, follows the single running scan, or the most recent workspace when none is running; when several are running, name one.",
+    examples: ['logs', 'logs q1-audit'],
+  },
+  status: {
+    usage: ['status [<workspace>] [--json]'],
+    description:
+      "Show one scan's phase-by-phase progress, read live from Temporal. With no workspace, shows the single running scan, or the most recent workspace when none is running; when several are running, name one. Watches and redraws until the scan finishes on a terminal; prints one frame when piped or already finished. With --json, prints a single machine-readable snapshot and exits.",
+    options: [['--json', 'Output a point-in-time snapshot as JSON, then exit']],
+    examples: ['status', 'status q1-audit', 'status q1-audit --json'],
+  },
+  scans: {
+    usage: ['scans [--json]'],
+    description: 'List running and completed scans, and where each finished report lives.',
+    options: [['--json', 'Output the scan list as JSON']],
+    examples: ['scans', 'scans --json'],
+  },
+  build: {
+    usage: ['build [--no-cache]'],
+    description: 'Build the worker Docker image (local mode only).',
+    options: [['--no-cache', 'Build without using the Docker layer cache']],
+  },
+  setup: {
+    usage: ['setup'],
+    description: 'Configure provider credentials interactively (npx mode only).',
+  },
+  version: {
+    usage: ['version [--json]'],
+    description: 'Show the version. With --json, prints the version and mode as a machine-readable object.',
+    options: [['--json', 'Output the version and mode as JSON']],
+    examples: ['version', 'version --json'],
+  },
+};
+
+/** Commands that only exist in one mode; everything else is available in both. */
+const MODE_ONLY: Readonly<Record<string, 'local' | 'npx'>> = {
+  build: 'local',
+  setup: 'npx',
+};
+
+/** Whether a command has its own help page (and so responds to `--help`/`-h`). */
+export function isHelpableCommand(command: string): boolean {
+  return command in COMMAND_HELP;
+}
+
+/**
+ * Every explicit help topic, mode-blind, with `help` itself as the known global topic.
+ * Topic lookup is deliberately not mode-filtered (unlike `availableCommands`) so
+ * cross-mode help such as local `help setup` and npx `help build` keeps working.
+ */
+export function helpTopics(): readonly string[] {
+  return [...Object.keys(COMMAND_HELP), 'help'];
+}
+
+/**
+ * User-facing command names available in the current mode, for "did you mean?"
+ * suggestions. Derived from the same table that backs per-command help, so the
+ * suggestion set can never drift from the commands that actually exist.
+ */
+export function availableCommands(): readonly string[] {
+  const mode = getMode();
+  const commands = Object.keys(COMMAND_HELP).filter((command) => (MODE_ONLY[command] ?? mode) === mode);
+  return [...commands, 'help'];
+}
+
+/** Print the help page for one command. No-op if the command has no page. */
+export function printCommandHelp(command: string): void {
+  const help = COMMAND_HELP[command];
+  if (!help) return;
+
+  const prefix = commandPrefix();
+  const baseOptions = command === 'start' ? START_OPTIONS : (help.options ?? []);
+  const options = [...baseOptions, HELP_OPTION];
+  const flagWidth = Math.max(...options.map(([flag]) => flag.length));
+
+  const lines: string[] = ['', help.description, '', 'USAGE'];
+  for (const line of help.usage) {
+    lines.push(`  ${prefix} ${line}`);
+  }
+
+  lines.push('', 'OPTIONS');
+  for (const [flag, desc] of options) {
+    lines.push(`  ${flag.padEnd(flagWidth)}  ${desc}`);
+  }
+
+  if (help.examples && help.examples.length > 0) {
+    lines.push('', 'EXAMPLES');
+    for (const example of help.examples) {
+      lines.push(`  ${prefix} ${example}`);
+    }
+  }
+
+  lines.push('');
+  console.log(lines.join('\n'));
+}

apps/cli/src/home.ts

@@ -1,7 +1,7 @@
 /**
  * Shannon state directory management.
  *
- * Local mode (cloned repo): uses ./workspaces/, ./credentials/
+ * Local mode (cloned repo): uses ./workspaces/
  * NPX mode: uses ~/.shannon/workspaces/, ~/.shannon/
  */
 
@@ -16,36 +16,23 @@ export function getConfigFile(): string {
   return path.join(SHANNON_HOME, 'config.toml');
 }
 
+/** Whether the npx-mode credential file (`~/.shannon/config.toml`) exists on disk. */
+export function configFileExists(): boolean {
+  return fs.existsSync(getConfigFile());
+}
+
 export function getWorkspacesDir(): string {
   return getMode() === 'local' ? path.resolve('workspaces') : path.join(SHANNON_HOME, 'workspaces');
 }
 
-/**
- * Resolve the Vertex credentials file path.
- *
- * Checks GOOGLE_APPLICATION_CREDENTIALS env var first (may be set by TOML resolver),
- * then falls back to mode-appropriate default location.
- */
-export function getCredentialsPath(): string {
-  const envPath = process.env.GOOGLE_APPLICATION_CREDENTIALS;
-  if (envPath && fs.existsSync(envPath)) return path.resolve(envPath);
-
-  if (getMode() === 'local') {
-    return path.resolve('credentials', 'google-sa-key.json');
-  }
-
-  return path.join(SHANNON_HOME, 'google-sa-key.json');
-}
-
 /**
  * Initialize state directories.
- * Local mode: creates ./workspaces/ and ./credentials/
+ * Local mode: creates ./workspaces/
  * NPX mode: creates ~/.shannon/workspaces/
  */
 export function initHome(): void {
   if (getMode() === 'local') {
     fs.mkdirSync(path.resolve('workspaces'), { recursive: true });
-    fs.mkdirSync(path.resolve('credentials'), { recursive: true });
   } else {
     fs.mkdirSync(path.join(SHANNON_HOME, 'workspaces'), { recursive: true });
   }

apps/cli/src/index.ts

@@ -1,5 +1,5 @@
 /**
- * Shannon CLI — AI Penetration Testing Framework
+ * Shannon CLI — AI Pentester for Web Apps and APIs
  *
  * Unified CLI supporting two modes:
  *   Local mode: Run from cloned repo — builds locally, mounts prompts, uses ./workspaces/
@@ -9,100 +9,161 @@
  * in the current working directory.
  */
 
-import fs from 'node:fs';
-import path from 'node:path';
-import { fileURLToPath } from 'node:url';
+import { ArgError, parseArgs, YES_FLAGS } from './args.js';
 import { build } from './commands/build.js';
 import { logs } from './commands/logs.js';
+import { reset } from './commands/reset.js';
+import { scans } from './commands/scans.js';
 import { setup } from './commands/setup.js';
 import { start } from './commands/start.js';
 import { status } from './commands/status.js';
 import { stop } from './commands/stop.js';
-import { uninstall } from './commands/uninstall.js';
-import { workspaces } from './commands/workspaces.js';
-import { getMode } from './mode.js';
+import { hasExportedCredentials } from './env.js';
+import { crash, enableJsonErrors, fail, failUsage, failWith, jsonErrorsEnabled } from './errors.js';
+import { availableCommands, helpTopics, isHelpableCommand, printCommandHelp, START_OPTIONS } from './help.js';
+import { configFileExists } from './home.js';
+import { commandPrefix, getMode, isLocal, type Mode } from './mode.js';
 import { displaySplash } from './splash.js';
+import { closestMatch } from './suggest.js';
+import { stdoutIsTerminal } from './tty.js';
+import { getVersion, getVersionLine } from './version.js';
+import { resolveDefaultWorkspace } from './workspaces.js';
 
-const __dirname = path.dirname(fileURLToPath(import.meta.url));
-
+/**
+ * Refuse to run as root or under sudo. The worker container's Linux UID remapping
+ * (docker.ts) stamps bind-mounted files with the invoking user's real uid/gid; under
+ * sudo that uid is 0, so the repo, workspace, and report files would come back
+ * owned by root instead of the person who ran the scan.
+ */
 function blockSudo(): void {
   const isSudo = !!process.env.SUDO_USER;
   const isRoot = process.geteuid?.() === 0;
   if (!isSudo && !isRoot) return;
 
+  const linuxHints =
+    process.platform === 'linux'
+      ? ['Configure Docker to run without sudo first:', 'https://docs.docker.com/engine/install/linux-postinstall']
+      : [];
+
   if (isSudo) {
-    console.error('ERROR: Shannon must not be run with sudo.');
-    console.error('Re-run this command as your normal user.');
-  } else {
-    console.error('ERROR: Shannon must not be run as the root user.');
-    console.error('Switch to a regular user account and re-run this command.');
+    failWith(
+      'CLI_PRECONDITION_FAILED',
+      'Shannon must not be run with sudo.',
+      'Re-run this command as your normal user.',
+      ...linuxHints,
+    );
   }
-  if (process.platform === 'linux') {
-    console.error('Configure Docker to run without sudo first:');
-    console.error('https://docs.docker.com/engine/install/linux-postinstall');
-  }
-  process.exit(1);
+  failWith(
+    'CLI_PRECONDITION_FAILED',
+    'Shannon must not be run as the root user.',
+    'Switch to a regular user account and re-run this command.',
+    ...linuxHints,
+  );
 }
 
-function getVersion(): string {
-  try {
-    const pkgPath = path.join(__dirname, '..', 'package.json');
-    const pkg = JSON.parse(fs.readFileSync(pkgPath, 'utf-8')) as { version?: string };
-    return pkg.version || '1.0.0';
-  } catch {
-    return '1.0.0';
+/** Commands whose `--json` output contract extends to failures. */
+const JSON_CAPABLE_COMMANDS = new Set(['status', 'scans', 'version', '--version', '-v']);
+
+/**
+ * Raw-argv sniff for the JSON error latch, decided before any guard or parse so even
+ * a pre-dispatch failure honors it. Latches on `--json` or a malformed `--json=<value>`
+ * (which still fails as a parse error — inside the envelope). Any other command that
+ * receives `--json` keeps its normal unknown-option behavior.
+ */
+function wantsJsonErrors(argv: readonly string[]): boolean {
+  const command = argv[0];
+  if (command === undefined || !JSON_CAPABLE_COMMANDS.has(command)) {
+    return false;
   }
+  return argv.slice(1).some((arg) => arg === '--json' || arg.startsWith('--json='));
 }
 
-function showHelp(): void {
+/** Render `start`'s flags for the global help, from the same source as `start --help`. */
+function renderStartOptions(): string {
+  const flagWidth = Math.max(...START_OPTIONS.map(([flag]) => flag.length));
+  return START_OPTIONS.map(([flag, desc]) => `  ${flag.padEnd(flagWidth)}  ${desc}`).join('\n');
+}
+
+/**
+ * Render the command list with the description column aligned. Padding is computed from the
+ * widest command, so it lines up regardless of the prefix (`npx @keygraph/shannon` vs `./shannon`).
+ */
+function renderUsage(prefix: string, mode: Mode): string {
+  const rows: ReadonlyArray<readonly [string, string]> = [
+    ...(mode === 'local' ? [] : [[`${prefix} setup`, 'Configure credentials'] as const]),
+    [`${prefix} start --url <url> --repo <path> [options]`, 'Start a pentest scan'],
+    [`${prefix} stop [<workspace>] [--yes]`, 'Stop one scan (default: the single running scan)'],
+    [`${prefix} stop --all [--yes]`, 'Stop all scans (Temporal stays up)'],
+    [`${prefix} reset`, 'Stop everything and wipe all Temporal data'],
+    [`${prefix} logs [<workspace>]`, "Show a scan's live log (default: running or most recent)"],
+    [`${prefix} logs [<workspace>] --agent <name>`, "Tail one agent's log; --list-agents to list them"],
+    [
+      `${prefix} status [<workspace>] [--json]`,
+      'Live phase/agent progress of one scan (default: running or most recent)',
+    ],
+    [`${prefix} scans [--json]`, 'List running and completed scans'],
+    ...(mode === 'local' ? [[`${prefix} build [--no-cache]`, 'Build worker image'] as const] : []),
+    [`${prefix} version [--json]`, 'Show version'],
+    [`${prefix} help`, 'Show this help'],
+  ];
+
+  const commandWidth = Math.max(...rows.map(([command]) => command.length));
+  return rows.map(([command, desc]) => `  ${command.padEnd(commandWidth)}   ${desc}`).join('\n');
+}
+
+/**
+ * A boxed "start your first scan" call to action, shown in help when no scans exist
+ * yet. Prefix-aware, so local mode renders `./shannon start …`.
+ */
+function renderFirstScanBox(prefix: string): string {
+  const command = `${prefix} start -u <url> -r <path>`;
+  const title = 'Start your first scan';
+  const padX = 3;
+  const inner = Math.max(command.length, title.length) + padX * 2;
+  const rule = (left: string, right: string): string => `  ${left}${'─'.repeat(inner)}${right}`;
+  const line = (text: string): string => `  │${' '.repeat(padX)}${text}${' '.repeat(inner - padX - text.length)}│`;
+  return [rule('╭', '╮'), line(title), line(''), line(command), rule('╰', '╯')].join('\n');
+}
+
+function showHelp(withSplash: boolean): void {
   const mode = getMode();
-  const prefix = mode === 'local' ? './shannon' : 'npx @keygraph/shannon';
+  const prefix = commandPrefix();
 
-  console.log(`
-Shannon - AI Penetration Testing Framework
+  const header = withSplash ? '' : '\nShannon — AI Pentester by Keygraph\n';
+  const firstScan = stdoutIsTerminal() ? `\n${renderFirstScanBox(prefix)}\n` : '';
 
-Usage:${
-    mode === 'local'
-      ? ''
-      : `
-  ${prefix} setup                                       Configure credentials`
-  }
-  ${prefix} start --url <url> --repo <path> [options]   Start a pentest scan
-  ${prefix} stop [--clean]                               Stop all containers
-  ${prefix} workspaces                                   List all workspaces
-  ${prefix} logs <workspace>                             Tail workflow log
-  ${prefix} status                                       Show running workers${
-    mode === 'local'
-      ? `
-  ${prefix} build [--no-cache]                           Build worker image`
-      : `
-  ${prefix} uninstall                                    Remove ~/.shannon/ and all data`
-  }
-  ${prefix} info                                         Show splash screen
-  ${prefix} help                                         Show this help
+  console.log(`${header}${firstScan}
+Usage:
+${renderUsage(prefix, mode)}
 
 Options for 'start':
-  -u, --url <url>           Target URL (required)
-  -r, --repo <path>         Repository path${mode === 'local' ? ' or bare name' : ''} (required)
-  -c, --config <path>       Configuration file (YAML)
-  -o, --output <path>       Copy deliverables to this directory after run
-  -w, --workspace <name>    Named workspace (auto-resumes if exists)
-      --pipeline-testing    Use minimal prompts for fast testing
-      --debug               Preserve worker container after exit for log inspection
+${renderStartOptions()}
 
 Examples:
-  ${prefix} start -u https://example.com -r ${mode === 'local' ? 'my-repo' : './my-repo'}
+  ${prefix} start -u https://example.com -r ./my-repo
   ${prefix} start -u https://example.com -r /path/to/repo -c config.yaml -w q1-audit
   ${prefix} logs q1-audit
-  ${prefix} stop --clean
-${
-  mode === 'local'
-    ? `
-State directory: ./workspaces/`
-    : `
-State directory: ~/.shannon/`
+  ${prefix} stop q1-audit
+  ${prefix} reset
+
+Run '${prefix} <command> --help' for help on a specific command.
+
+Docs & source: https://github.com/KeygraphHQ/shannon
+`);
 }
-Monitor workflows at http://localhost:8233
+
+/**
+ * First-run guidance for a bare `npx @keygraph/shannon` invocation when neither a
+ * credentials file nor an exported shell credential exists. Walks the user to `setup`.
+ */
+function showSetupPrompt(): void {
+  const prefix = commandPrefix();
+  console.log(`
+Welcome to Shannon — AI Pentester by Keygraph
+
+No credentials configured yet. To get started, run:
+
+  ${prefix} setup
 `);
 }
 
@@ -113,148 +174,229 @@ interface ParsedStartArgs {
   workspace?: string;
   output?: string;
   pipelineTesting: boolean;
-  debug: boolean;
+  keepContainer: boolean;
+  follow: boolean;
 }
 
 function parseStartArgs(argv: string[]): ParsedStartArgs {
-  let url = '';
-  let repo = '';
-  let config: string | undefined;
-  let workspace: string | undefined;
-  let output: string | undefined;
-  let pipelineTesting = false;
-  let debug = false;
+  const { flags, values } = parseArgs(argv, {
+    values: {
+      url: ['-u', '--url'],
+      repo: ['-r', '--repo'],
+      config: ['-c', '--config'],
+      output: ['-o', '--output'],
+      workspace: ['-w', '--workspace'],
+    },
+    booleans: {
+      pipelineTesting: ['--pipeline-testing'],
+      keepContainer: ['--keep-container'],
+      follow: ['-f', '--follow'],
+    },
+  });
 
-  for (let i = 0; i < argv.length; i++) {
-    const arg = argv[i];
-    const next = argv[i + 1];
-
-    switch (arg) {
-      case '-u':
-      case '--url':
-        if (next && !next.startsWith('-')) {
-          url = next;
-          i++;
-        }
-        break;
-      case '-r':
-      case '--repo':
-        if (next && !next.startsWith('-')) {
-          repo = next;
-          i++;
-        }
-        break;
-      case '-c':
-      case '--config':
-        if (next && !next.startsWith('-')) {
-          config = next;
-          i++;
-        }
-        break;
-      case '-w':
-      case '--workspace':
-        if (next && !next.startsWith('-')) {
-          workspace = next;
-          i++;
-        }
-        break;
-      case '-o':
-      case '--output':
-        if (next && !next.startsWith('-')) {
-          output = next;
-          i++;
-        }
-        break;
-      case '--pipeline-testing':
-        pipelineTesting = true;
-        break;
-      case '--debug':
-        debug = true;
-        break;
-      default:
-        console.error(`Unknown option: ${arg}`);
-        console.error(`Run "${getMode() === 'local' ? './shannon' : 'npx @keygraph/shannon'} help" for usage`);
-        process.exit(1);
-    }
+  const url = values.url ?? '';
+  const repo = values.repo ?? '';
+  if (!url || !repo) {
+    failUsage('--url and --repo are required', `Usage: ${commandPrefix()} start -u <url> -r <path>`);
   }
 
-  if (!url || !repo) {
-    console.error('ERROR: --url and --repo are required');
-    console.error(`Usage: ${getMode() === 'local' ? './shannon' : 'npx @keygraph/shannon'} start -u <url> -r <path>`);
-    process.exit(1);
+  try {
+    new URL(url);
+  } catch {
+    failUsage(`invalid --url: ${url}`);
   }
 
   return {
     url,
     repo,
-    pipelineTesting,
-    debug,
-    ...(config && { config }),
-    ...(workspace && { workspace }),
-    ...(output && { output }),
+    pipelineTesting: !!flags.pipelineTesting,
+    keepContainer: !!flags.keepContainer,
+    follow: !!flags.follow,
+    ...(values.config && { config: values.config }),
+    ...(values.workspace && { workspace: values.workspace }),
+    ...(values.output && { output: values.output }),
   };
 }
 
+/**
+ * Resolve the workspace a viewing command (`logs`, `status`) acts on: the name the user
+ * gave, or an inferred default. An inferred choice is announced on stderr so it is never a
+ * silent guess; when nothing can be inferred, exit with usage guidance.
+ */
+function resolveViewingWorkspace(positional: string | undefined, usage: string): string {
+  if (positional) {
+    return positional;
+  }
+
+  const target = resolveDefaultWorkspace({ allowFinished: true });
+  if (target.kind === 'ok') {
+    // In JSON mode stderr is reserved for the single error envelope, so a successful
+    // inference stays silent — the JSON payload itself names the chosen workspace.
+    if (!jsonErrorsEnabled()) {
+      const which = target.running ? 'running scan' : 'most recent scan';
+      console.error(`No workspace given; using ${which} "${target.workspace}".`);
+    }
+    return target.workspace;
+  }
+  if (target.kind === 'ambiguous') {
+    failUsage('Multiple scans are running — specify which one:', `  ${target.running.join(', ')}`, '', usage);
+  }
+  failUsage('Workspace is required', usage);
+}
+
 // === Main Dispatch ===
 
-blockSudo();
+async function main(): Promise<void> {
+  // A reader that closes early (e.g. `shannon logs my-scan | head`) makes writes
+  // to stdout raise EPIPE. That's normal for a piped CLI, not a crash — exit quietly
+  // instead of letting Node dump an unhandled-error stack trace.
+  process.stdout.on('error', (err: NodeJS.ErrnoException) => {
+    if (err.code === 'EPIPE') process.exit(0);
+    throw err;
+  });
 
-const args = process.argv.slice(2);
-const command = args[0];
+  if (wantsJsonErrors(process.argv.slice(2))) {
+    enableJsonErrors();
+  }
 
-switch (command) {
-  case 'start': {
-    const parsed = parseStartArgs(args.slice(1));
-    await start({ ...parsed, version: getVersion() });
-    break;
+  blockSudo();
+
+  const args = process.argv.slice(2);
+  const command = args[0];
+  const rest = args.slice(1);
+
+  if (command === undefined || command === '--help' || command === '-h') {
+    const topic = rest[0];
+    if (topic && isHelpableCommand(topic)) {
+      printCommandHelp(topic);
+    } else {
+      const bare = command === undefined;
+      if (bare && stdoutIsTerminal()) displaySplash(isLocal() ? undefined : getVersion());
+      const needsSetup = bare && !isLocal() && !configFileExists() && !hasExportedCredentials();
+      if (needsSetup) {
+        showSetupPrompt();
+      } else {
+        showHelp(bare);
+      }
+    }
+    return;
   }
-  case 'stop':
-    stop(args.includes('--clean'));
-    break;
-  case 'logs': {
-    const workspaceId = args[1];
-    if (!workspaceId) {
-      console.error('ERROR: Workspace ID is required');
-      console.error(`Usage: ${getMode() === 'local' ? './shannon' : 'npx @keygraph/shannon'} logs <workspace>`);
-      process.exit(1);
+
+  // An explicit `help <topic>` names a topic on purpose, so an unknown one is a usage
+  // error — unlike `--help <junk>`, where the junk is ignored and global help wins.
+  if (command === 'help') {
+    const topic = rest[0];
+    // A flag (`help --help`) is a help request, not a topic name.
+    if (topic === undefined || topic === 'help' || topic.startsWith('-')) {
+      showHelp(false);
+      return;
     }
-    logs(workspaceId);
-    break;
+    if (isHelpableCommand(topic)) {
+      printCommandHelp(topic);
+      return;
+    }
+    const suggestion = closestMatch(topic, helpTopics());
+    failUsage(
+      `Unknown help topic: ${topic}`,
+      ...(suggestion ? [`Did you mean '${suggestion}'?`] : []),
+      `Run '${commandPrefix()} help' to see available commands.`,
+    );
   }
-  case 'workspaces':
-    workspaces(getVersion());
-    break;
-  case 'status':
-    status();
-    break;
-  case 'setup':
-    if (getMode() === 'local') {
-      console.error('ERROR: setup is only available in npx mode. In local mode, use .env');
-      process.exit(1);
+
+  // Reachable from any invocation: `-h`/`--help` anywhere wins over the rest of the line.
+  if (isHelpableCommand(command) && (rest.includes('-h') || rest.includes('--help'))) {
+    printCommandHelp(command);
+    return;
+  }
+
+  switch (command) {
+    case 'start': {
+      const parsed = parseStartArgs(rest);
+      await start({ ...parsed, version: getVersion() });
+      break;
     }
-    setup();
-    break;
-  case 'build':
-    build(args.includes('--no-cache'));
-    break;
-  case 'uninstall':
-    if (getMode() === 'local') {
-      console.error('ERROR: uninstall is only available in npx mode.');
-      process.exit(1);
+    case 'stop': {
+      const { flags, positionals } = parseArgs(rest, {
+        booleans: { all: ['--all'], yes: YES_FLAGS },
+        maxPositionals: 1,
+      });
+      await stop({ all: !!flags.all, yes: !!flags.yes, ...(positionals[0] && { workspace: positionals[0] }) });
+      break;
     }
-    uninstall();
-    break;
-  case 'info':
-    displaySplash(getMode() === 'local' ? undefined : getVersion());
-    break;
-  case 'help':
-  case '--help':
-  case '-h':
-  case undefined:
-    showHelp();
-    break;
-  default:
-    console.error(`Unknown command: ${command}`);
-    showHelp();
-    process.exit(1);
+    case 'reset': {
+      // reset is all-or-nothing; a stray name likely means the user wanted `stop <name>`.
+      parseArgs(rest, {
+        positionalHint: 'reset takes no workspace argument. To stop one scan, use: stop <name>',
+      });
+      await reset();
+      break;
+    }
+    case 'logs': {
+      const { flags, values, positionals } = parseArgs(rest, {
+        booleans: { listAgents: ['--list-agents'] },
+        values: { agent: ['--agent'] },
+        maxPositionals: 1,
+      });
+      const workspaceId = resolveViewingWorkspace(
+        positionals[0],
+        `Usage: ${commandPrefix()} logs [<workspace>] [--agent <name>] [--list-agents]`,
+      );
+      logs(workspaceId, {
+        ...(values.agent !== undefined && { agent: values.agent }),
+        ...(flags.listAgents && { listAgents: true }),
+      });
+      break;
+    }
+    case 'status': {
+      const { flags, positionals } = parseArgs(rest, { booleans: { json: ['--json'] }, maxPositionals: 1 });
+      const usage = `Usage: ${commandPrefix()} status [<workspace>] [--json]`;
+      const workspaceId = resolveViewingWorkspace(positionals[0], usage);
+      await status(workspaceId, { json: !!flags.json });
+      break;
+    }
+    case 'scans': {
+      const { flags } = parseArgs(rest, { booleans: { json: ['--json'] } });
+      scans({ json: !!flags.json });
+      break;
+    }
+    case 'setup':
+      if (getMode() === 'local') {
+        fail('setup is only available in npx mode. In local mode, use .env');
+      }
+      parseArgs(rest, {});
+      await setup();
+      break;
+    case 'build': {
+      const { flags } = parseArgs(rest, { booleans: { noCache: ['--no-cache'] } });
+      build(!!flags.noCache, getVersion());
+      break;
+    }
+    case 'version':
+    case '--version':
+    case '-v': {
+      const { flags } = parseArgs(rest, { booleans: { json: ['--json'] } });
+      if (flags.json) {
+        console.log(JSON.stringify({ version: getVersion(), mode: getMode() }, null, 2));
+      } else {
+        console.log(getVersionLine());
+      }
+      break;
+    }
+    default: {
+      const prefix = commandPrefix();
+      const suggestion = closestMatch(command, availableCommands());
+      const hints = [
+        ...(suggestion ? [`Did you mean '${suggestion}'?`] : []),
+        `Run '${prefix} help' to see available commands.`,
+      ];
+      failUsage(`Unknown command: ${command}`, ...hints);
+    }
+  }
 }
+
+main().catch((err) => {
+  if (err instanceof ArgError) {
+    failUsage(err.message, `Run "${commandPrefix()} help" for usage`);
+  }
+  crash(err);
+});

apps/cli/src/mode.ts

@@ -23,3 +23,12 @@ export function setMode(mode: Mode): void {
 export function isLocal(): boolean {
   return getMode() === 'local';
 }
+
+/** The invocation prefix for the current mode, so help and hints point at a runnable command. */
+export function commandPrefix(): string {
+  return getMode() === 'local' ? './shannon' : 'npx @keygraph/shannon';
+}
+
+export function isDevMode(): boolean {
+  return process.env.SHANNON_DEV === '1';
+}

apps/cli/src/model-spec.ts

@@ -0,0 +1,91 @@
+/**
+ * Parsing for the single model setting, `SHANNON_AI_MODEL=<provider>:<model-id>`.
+ *
+ * Mirrors apps/worker/src/ai/models.ts. The CLI cannot import from the worker
+ * package (it ships as a standalone bundle), so the provider list and the parse
+ * rule are duplicated here deliberately and must stay in sync.
+ */
+
+/**
+ * Providers Shannon curates with their own credential variables, config sections,
+ * and setup flows. Any other pi provider is reachable via the generic credential
+ * path. Mirrors CURATED_PROVIDERS in apps/worker/src/ai/models.ts.
+ */
+export const CURATED_PROVIDERS = ['anthropic', 'openai', 'xai', 'amazon-bedrock'] as const;
+
+export type CuratedProviderId = (typeof CURATED_PROVIDERS)[number];
+
+export function isCuratedProvider(value: string): value is CuratedProviderId {
+  return (CURATED_PROVIDERS as readonly string[]).includes(value);
+}
+
+/** Generic API key, honored for any provider Shannon does not curate. Mirrors the worker. */
+export const GENERIC_API_KEY_ENV = 'SHANNON_AI_API_KEY';
+
+/**
+ * Env vars carrying each curated provider's API key, in precedence order. Any one of
+ * them satisfies the provider. Mirrors PROVIDER_API_KEY_ENV in apps/worker/src/ai/models.ts.
+ */
+export const PROVIDER_API_KEY_ENV: Readonly<Record<CuratedProviderId, readonly string[]>> = {
+  anthropic: ['ANTHROPIC_API_KEY', 'CLAUDE_CODE_OAUTH_TOKEN'],
+  openai: ['OPENAI_API_KEY'],
+  xai: ['XAI_API_KEY'],
+  'amazon-bedrock': ['AWS_BEARER_TOKEN_BEDROCK'],
+};
+
+/** Additional env vars a curated provider requires beyond its API key. All must be set. */
+export const PROVIDER_EXTRA_ENV: Readonly<Record<CuratedProviderId, readonly string[]>> = {
+  anthropic: [],
+  openai: [],
+  xai: [],
+  'amazon-bedrock': ['AWS_REGION'],
+};
+
+/** Human-readable credential requirement, used in "nothing configured" errors. */
+export const PROVIDER_CREDENTIAL_HINT: Readonly<Record<CuratedProviderId, string>> = {
+  anthropic: 'ANTHROPIC_API_KEY (or CLAUDE_CODE_OAUTH_TOKEN)',
+  openai: 'OPENAI_API_KEY',
+  xai: 'XAI_API_KEY',
+  'amazon-bedrock': 'AWS_REGION and AWS_BEARER_TOKEN_BEDROCK',
+};
+
+/** Model used when SHANNON_AI_MODEL is unset. */
+export const DEFAULT_MODEL_SPEC = 'anthropic:claude-sonnet-4-6';
+
+/**
+ * Values SHANNON_AI_OPENAI_FORMAT accepts, selecting the wire format an
+ * OpenAI-compatible gateway serves. Mirrors OPENAI_FORMATS in
+ * apps/worker/src/ai/models.ts; the worker validates and applies it.
+ */
+export const OPENAI_FORMATS = ['chat-completions', 'responses'] as const;
+
+export type OpenAiFormat = (typeof OPENAI_FORMATS)[number];
+
+export interface ModelSpec {
+  providerId: string;
+  modelId: string;
+}
+
+/**
+ * Parse a `<provider>:<model-id>` spec. Splits on the first colon only, so colons
+ * inside a model ID survive (`amazon-bedrock:us.anthropic.claude-opus-4-5-20251101-v1:0`).
+ * The provider id is passed through as given — the worker's preflight validates it
+ * against pi. Returns an error string rather than throwing, for the CLI's flow.
+ */
+export function parseModelSpec(spec: string): ModelSpec | string {
+  const trimmed = spec.trim();
+  const separator = trimmed.indexOf(':');
+  const malformed = `SHANNON_AI_MODEL must be "<provider>:<model-id>", got "${trimmed}". Example: ${DEFAULT_MODEL_SPEC}`;
+  if (separator === -1) return malformed;
+
+  const providerId = trimmed.slice(0, separator).trim();
+  const modelId = trimmed.slice(separator + 1).trim();
+  if (!providerId || !modelId) return malformed;
+
+  return { providerId, modelId };
+}
+
+/** Resolve the run's model spec from the environment, or an error string. */
+export function resolveModelSpec(): ModelSpec | string {
+  return parseModelSpec(process.env.SHANNON_AI_MODEL || DEFAULT_MODEL_SPEC);
+}

apps/cli/src/paths.ts

@@ -1,13 +1,27 @@
 /**
  * Path resolution for --repo and --config arguments.
  *
- * Local mode supports bare repo names (e.g. "my-repo" → ./repos/my-repo).
- * Both modes resolve relative paths against CWD.
+ * Both --repo and --config are filesystem paths, absolute or relative to CWD.
  */
 
 import fs from 'node:fs';
+import os from 'node:os';
 import path from 'node:path';
-import { isLocal } from './mode.js';
+import { fail } from './errors.js';
+
+/**
+ * Expand a leading `~` or `~/` to the home directory. The shell skips this in the
+ * `--flag=~/x` form (the tilde is not at the word start), so it must be done here.
+ */
+export function expandHome(inputPath: string): string {
+  if (inputPath === '~') {
+    return os.homedir();
+  }
+  if (inputPath.startsWith('~/')) {
+    return path.join(os.homedir(), inputPath.slice(2));
+  }
+  return inputPath;
+}
 
 export interface MountPair {
   hostPath: string;
@@ -15,36 +29,56 @@ export interface MountPair {
 }
 
 /**
- * Resolve --repo to absolute path and container mount.
- * Dev mode: bare names (no / or . prefix) check ./repos/<name> first.
+ * Hidden subdirectory inside each run directory that holds all internals
+ * (deliverables, logs, prompts, session state, browser artifacts). Keeps the
+ * run folder's top level clean so only the final report is visible. Must match
+ * INTERNAL_DIR in the worker package.
+ */
+export const INTERNAL_DIR = '.shannon';
+
+/**
+ * Filename of the human-facing PDF report surfaced at the run directory root.
+ * Must match FINAL_REPORT_PDF_FILENAME in the worker package.
+ */
+export const FINAL_REPORT_PDF_FILENAME = 'Security-Assessment-Report.pdf';
+
+/**
+ * Customer-facing Markdown report name at the run root.
+ * Must match FINAL_REPORT_MD_FILENAME in the worker package.
+ */
+export const FINAL_REPORT_MD_FILENAME = 'Security-Assessment-Report.md';
+
+/**
+ * Resolve a run-directory file (e.g. session.json, workflow.log), preferring the
+ * current INTERNAL_DIR location and falling back to the legacy run-root location
+ * so pre-restructure workspaces keep working. Returns the INTERNAL_DIR path when
+ * neither exists — the right default for new runs and error messages.
+ */
+export function resolveRunFile(runDir: string, filename: string): string {
+  const current = path.join(runDir, INTERNAL_DIR, filename);
+  if (fs.existsSync(current)) {
+    return current;
+  }
+  const legacy = path.join(runDir, filename);
+  if (fs.existsSync(legacy)) {
+    return legacy;
+  }
+  return current;
+}
+
+/**
+ * Resolve --repo to an absolute path and container mount. The argument is a
+ * filesystem path, absolute or relative to CWD.
  */
 export function resolveRepo(repoArg: string): MountPair {
-  let hostPath: string;
-
-  if (isLocal() && !repoArg.startsWith('/') && !repoArg.startsWith('.')) {
-    // Bare name — check ./repos/<name> for backward compatibility
-    const barePath = path.resolve('repos', repoArg);
-    if (fs.existsSync(barePath)) {
-      hostPath = barePath;
-    } else {
-      console.error(`ERROR: Repository not found at ./repos/${repoArg}`);
-      console.error('');
-      console.error('Place your target repository under the ./repos/ directory,');
-      console.error('or pass an absolute/relative path: -r /path/to/repo');
-      process.exit(1);
-    }
-  } else {
-    hostPath = path.resolve(repoArg);
-  }
+  const hostPath = path.resolve(expandHome(repoArg));
 
   if (!fs.existsSync(hostPath)) {
-    console.error(`ERROR: Repository not found: ${hostPath}`);
-    process.exit(1);
+    fail(`Repository not found: ${hostPath}`);
   }
 
   if (!fs.statSync(hostPath).isDirectory()) {
-    console.error(`ERROR: Not a directory: ${hostPath}`);
-    process.exit(1);
+    fail(`Not a directory: ${hostPath}`);
   }
 
   const basename = path.basename(hostPath);
@@ -58,16 +92,14 @@ export function resolveRepo(repoArg: string): MountPair {
  * Resolve --config to absolute path and container mount.
  */
 export function resolveConfig(configArg: string): MountPair {
-  const hostPath = path.resolve(configArg);
+  const hostPath = path.resolve(expandHome(configArg));
 
   if (!fs.existsSync(hostPath)) {
-    console.error(`ERROR: Config file not found: ${hostPath}`);
-    process.exit(1);
+    fail(`Config file not found: ${hostPath}`);
   }
 
   if (!fs.statSync(hostPath).isFile()) {
-    console.error(`ERROR: Not a file: ${hostPath}`);
-    process.exit(1);
+    fail(`Not a file: ${hostPath}`);
   }
 
   const basename = path.basename(hostPath);

apps/cli/src/pending-workflow.ts

@@ -0,0 +1,139 @@
+/** Durable CLI-owned workflow candidates that bridge Docker launch and session registration. */
+
+import fs from 'node:fs';
+import path from 'node:path';
+import { INTERNAL_DIR } from './paths.js';
+
+const SCHEMA_VERSION = 1 as const;
+const PENDING_DIR = 'pending-workflows';
+
+export interface PendingWorkflowIdentity {
+  readonly schema_version: typeof SCHEMA_VERSION;
+  readonly workflow_id: string;
+  readonly task_queue: string;
+  readonly created_at: string;
+}
+
+export interface PendingWorkflowReadResult {
+  readonly identities: readonly PendingWorkflowIdentity[];
+  readonly unreadableCount: number;
+}
+
+function pendingDir(workspacePath: string): string {
+  return path.join(workspacePath, INTERNAL_DIR, PENDING_DIR);
+}
+
+function pendingFile(workspacePath: string, taskQueue: string): string {
+  return path.join(pendingDir(workspacePath), `launch-${encodeURIComponent(taskQueue)}.json`);
+}
+
+function syncDirectory(directory: string): void {
+  const descriptor = fs.openSync(directory, 'r');
+  try {
+    fs.fsyncSync(descriptor);
+  } finally {
+    fs.closeSync(descriptor);
+  }
+}
+
+/** Persist the candidate before docker run, so a vanished pre-registration worker remains addressable. */
+export function writePendingWorkflowIdentity(workspacePath: string, workflowId: string, taskQueue: string): void {
+  const directory = pendingDir(workspacePath);
+  const directoryAlreadyExisted = fs.existsSync(directory);
+  fs.mkdirSync(directory, { recursive: true });
+  if (!directoryAlreadyExisted) syncDirectory(path.dirname(directory));
+  const destination = pendingFile(workspacePath, taskQueue);
+  const temporary = `${destination}.tmp-${process.pid}-${Date.now()}`;
+  const identity: PendingWorkflowIdentity = {
+    schema_version: SCHEMA_VERSION,
+    workflow_id: workflowId,
+    task_queue: taskQueue,
+    created_at: new Date().toISOString(),
+  };
+
+  const descriptor = fs.openSync(temporary, 'wx', 0o600);
+  try {
+    fs.writeFileSync(descriptor, `${JSON.stringify(identity, null, 2)}\n`, 'utf8');
+    fs.fsyncSync(descriptor);
+  } finally {
+    fs.closeSync(descriptor);
+  }
+  try {
+    // Link installs the fully-fsynced inode without replacing an existing task-queue record.
+    fs.linkSync(temporary, destination);
+    fs.unlinkSync(temporary);
+    syncDirectory(directory);
+  } catch (error) {
+    fs.rmSync(temporary, { force: true });
+    throw error;
+  }
+}
+
+/** Remove one candidate only after session registration or a fully verified stop. */
+export function clearPendingWorkflowIdentity(workspacePath: string, taskQueue: string): void {
+  const directory = pendingDir(workspacePath);
+  fs.rmSync(pendingFile(workspacePath, taskQueue), { force: true });
+  if (fs.existsSync(directory)) syncDirectory(directory);
+}
+
+function escapeRegExp(value: string): string {
+  return value.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');
+}
+
+function isPendingWorkflowIdentity(
+  value: unknown,
+  workspace: string,
+  expectedFilename: string,
+): value is PendingWorkflowIdentity {
+  if (value === null || typeof value !== 'object' || Array.isArray(value)) return false;
+  const candidate = value as Record<string, unknown>;
+  const keys = Object.keys(candidate).sort();
+  const workflowId = candidate.workflow_id;
+  const workflowPattern = new RegExp(`^${escapeRegExp(workspace)}_(?:shannon-|resume_)\\d+$`);
+  const workspaceIsWorkflowId = workflowId === workspace && /_shannon-\d+$/.test(workspace);
+  return (
+    keys.length === 4 &&
+    keys[0] === 'created_at' &&
+    keys[1] === 'schema_version' &&
+    keys[2] === 'task_queue' &&
+    keys[3] === 'workflow_id' &&
+    candidate.schema_version === SCHEMA_VERSION &&
+    typeof workflowId === 'string' &&
+    (workspaceIsWorkflowId || workflowPattern.test(workflowId)) &&
+    typeof candidate.task_queue === 'string' &&
+    /^shannon-[0-9a-f]{8}$/.test(candidate.task_queue) &&
+    expectedFilename === `launch-${encodeURIComponent(candidate.task_queue)}.json` &&
+    typeof candidate.created_at === 'string' &&
+    !Number.isNaN(Date.parse(candidate.created_at)) &&
+    new Date(candidate.created_at).toISOString() === candidate.created_at
+  );
+}
+
+/** Read every outstanding launch candidate, preserving corrupt records as an explicit failure count. */
+export function readPendingWorkflowIdentities(workspacePath: string): PendingWorkflowReadResult {
+  let entries: string[];
+  try {
+    entries = fs.readdirSync(pendingDir(workspacePath)).filter((entry) => entry.endsWith('.json'));
+  } catch (error) {
+    if ((error as NodeJS.ErrnoException).code === 'ENOENT') return { identities: [], unreadableCount: 0 };
+    return { identities: [], unreadableCount: 1 };
+  }
+
+  const identities: PendingWorkflowIdentity[] = [];
+  let unreadableCount = 0;
+  for (const entry of entries) {
+    try {
+      const value: unknown = JSON.parse(fs.readFileSync(path.join(pendingDir(workspacePath), entry), 'utf8'));
+      if (!isPendingWorkflowIdentity(value, path.basename(workspacePath), entry)) {
+        unreadableCount++;
+        continue;
+      }
+      identities.push(value);
+    } catch {
+      unreadableCount++;
+    }
+  }
+  // Atomic-write temp files are intentionally ignored: start cannot spawn Docker until the
+  // final .json rename and fsync above have both completed.
+  return { identities, unreadableCount };
+}

apps/cli/src/scan/derive.ts

@@ -0,0 +1,420 @@
+/**
+ * Pure derivation of a scan's per-agent and per-phase state from its Temporal snapshot.
+ *
+ * This is the single source of truth for "what state is each agent in" — both the
+ * human progress tree (render.ts) and the machine-readable snapshot (status-json.ts)
+ * consume it, so the two views can never disagree about whether an agent is running,
+ * skipped, or still pending. No glyphs, no color, no formatting live here.
+ */
+
+import type { RunningAgent } from '../temporal-client.js';
+import {
+  AGENTIC_SAST_STAGE_ORDER,
+  agentClass,
+  isModelBackedOperation,
+  type OperationalStageState,
+  operationFamilyKey,
+  type PipelineState,
+  pipelineForState,
+} from './pipeline.js';
+import type { RenderInput } from './render.js';
+import { safeFailureDetail, safeOperationKey, safeOperationLabel } from './safe-fields.js';
+
+export type RunState = 'pending' | 'running' | 'completed' | 'failed' | 'skipped';
+
+/** One agent's resolved state plus the raw metrics/timing a consumer needs to present it. Null metrics
+ *  mean the value doesn't apply to the current state (e.g. duration only for completed agents). */
+export interface DerivedAgent {
+  readonly name: string;
+  readonly label: string;
+  readonly state: RunState;
+  readonly durationMs: number | null;
+  readonly runningElapsedMs: number | null;
+  readonly attempt: number | null;
+  /** The step a running operation row is currently on, merged in from its child activity. */
+  readonly detail?: string;
+  /** Reconciliation time for this agent's class, rendered as a trailing `+ duration`.
+   *  Reconciliation is model work that produces this agent's inputs, so it is shown
+   *  attached to the agent it feeds rather than as free-floating background work. */
+  readonly attachedMs?: number;
+  /** This class's findings could not be grouped, so each one became its own task. */
+  readonly ungrouped?: boolean;
+  readonly error?: string;
+}
+
+/** How a phase line summarizes itself: its own wall time, or a k/N tally over its children. */
+export type PhaseMetaKind = 'duration' | 'count';
+
+export interface DerivedPhase {
+  readonly key: string;
+  readonly label: string;
+  /** Whether the phase renders its agents as sub-rows. Independent of {@link meta}:
+   *  Agentic SAST lists its stages under a duration, exploitation lists its classes under a tally. */
+  readonly children: boolean;
+  readonly meta: PhaseMetaKind;
+  readonly state: RunState;
+  /** The phase's own span, when the worker records one for the phase rather than for a single
+   *  agent inside it (Agentic SAST). The phase line presents this exactly like an agent row. */
+  readonly summary?: DerivedAgent;
+  /** Rendered after the phase's summary, e.g. to mark work that overlaps other phases. */
+  readonly note?: string;
+  readonly agents: readonly DerivedAgent[];
+}
+
+/** Terminal = anything other than an open, running execution. */
+export function isTerminal(status: string): boolean {
+  return status !== 'RUNNING' && status !== 'UNSPECIFIED';
+}
+
+/**
+ * Whether the class-level failure recorded for this agent's class applies to this agent.
+ *
+ * A class failure is recorded against the class as a whole, so it matches both of that class's
+ * agents. A reconciliation failure, though, happens only after the analysis agent has already
+ * succeeded, so it belongs to the exploitation lane: attributing it to the analysis row as well
+ * would report an agent that completed as failed.
+ */
+function classFailureApplies(name: string, state: PipelineState | null): boolean {
+  if (!state) return false;
+  const vulnClass = agentClass(name);
+  if (!state.failedPipelines.some((f) => f.vulnType === vulnClass)) return false;
+  const reconciliationFailed = (state.failedReconciliations ?? []).some((r) => r.vulnerabilityClass === vulnClass);
+  const isAnalysisAgent = name.endsWith('-vuln');
+  return !(reconciliationFailed && isAnalysisAgent);
+}
+
+function isFailedAgent(name: string, state: PipelineState | null): boolean {
+  return !!state && (state.failedAgent === name || classFailureApplies(name, state));
+}
+
+/** An agent has entered play once it is running, has metrics, or has failed. */
+function isAgentActive(name: string, state: PipelineState | null, running: Set<string>): boolean {
+  return running.has(name) || !!state?.agentMetrics[name] || isFailedAgent(name, state);
+}
+
+/**
+ * Resolve one agent's state. "Ran" is signalled by a metrics entry: a
+ * conditionally-skipped agent (e.g. an exploit agent when there is nothing to
+ * exploit) records no metrics, and the workflow tracks it in skippedAgents rather
+ * than completedAgents. `resolved` is true once we've moved past this agent's phase
+ * (the scan is terminal, or a later phase is already active), at which point a
+ * metric-less, non-running agent is skipped rather than still pending.
+ */
+function agentState(name: string, state: PipelineState | null, running: Set<string>, resolved: boolean): RunState {
+  if (running.has(name)) return 'running';
+  if (isFailedAgent(name, state)) return 'failed';
+  if (state?.agentMetrics[name]) return 'completed';
+  return resolved ? 'skipped' : 'pending';
+}
+
+/**
+ * Only the presence of a class failure is used here, never its `.error` text: that string is
+ * the worker's raw error for the failed class, not vetted for display, so it is reduced to
+ * a boolean before reaching safeFailureDetail's fixed sentence.
+ */
+function agentError(name: string, state: PipelineState | null, byAgent: Map<string, RunningAgent>): string | undefined {
+  const hasFailure =
+    classFailureApplies(name, state) || byAgent.get(name)?.lastFailure !== undefined || state?.failedAgent === name;
+  return safeFailureDetail(hasFailure);
+}
+
+/** Scan wall-clock elapsed ms: recorded duration for a closed scan, live elapsed for a running one. */
+export function scanElapsedMs(input: RenderInput, now: number): number | undefined {
+  if (isTerminal(input.temporalStatus)) {
+    if (input.state?.summary) return input.state.summary.totalDurationMs;
+    if (input.endedAt !== undefined && input.startedAt !== undefined) return input.endedAt - input.startedAt;
+    return undefined;
+  }
+  return input.startedAt !== undefined ? now - input.startedAt : undefined;
+}
+
+/** Collapse a phase's agent states into a single state for the phase line. */
+export function phaseGlyphState(states: readonly RunState[]): RunState {
+  if (states.some((s) => s === 'running')) return 'running';
+  if (states.some((s) => s === 'failed')) return 'failed';
+  if (states.every((s) => s === 'skipped')) return 'skipped';
+  if (states.every((s) => s === 'completed' || s === 'skipped')) return 'completed';
+  if (states.some((s) => s === 'completed')) return 'running';
+  return 'pending';
+}
+
+/**
+ * Compute each agent's RunState. This is the drift-prone part shared by every view.
+ *
+ * The pipeline is sequential across phases: the last phase with any active agent is the
+ * frontier. Earlier phases with nothing active were skipped (e.g. exploitation when no
+ * class had anything to exploit), not still pending.
+ */
+export function deriveAgentStates(input: RenderInput): Map<string, RunState> {
+  const pipeline = pipelineForState(input.state);
+  const runningSet = new Set(input.running.filter((runner) => runner.kind === 'agent').map((runner) => runner.agent));
+  const terminal = isTerminal(input.temporalStatus);
+
+  let frontier = -1;
+  pipeline.forEach((phase, idx) => {
+    if (phase.agents.some((a) => isAgentActive(a.name, input.state, runningSet))) frontier = idx;
+  });
+
+  const states = new Map<string, RunState>();
+  for (const [phaseIdx, phase] of pipeline.entries()) {
+    const resolved = terminal || phaseIdx < frontier;
+    for (const agent of phase.agents) {
+      states.set(agent.name, agentState(agent.name, input.state, runningSet, resolved));
+    }
+  }
+  return states;
+}
+
+/** Which operation families have a running parent stage, and the step to show on it. */
+interface OperationFamilyView {
+  /** Families whose parent stage row already represents their child activities. */
+  readonly runningFamilies: ReadonlySet<string>;
+  /** Family to current step, present only where the child activities agree on one. */
+  readonly stepByFamily: ReadonlyMap<string, string>;
+}
+
+/**
+ * Resolve the parent stage rows that own their family's child activities. A family only
+ * resolves to a step when its running children agree: several classes reconcile at once and
+ * their pending activities carry no class, so a family caught mid-stride shows its parent
+ * rows without a step rather than attributing one to the wrong class.
+ */
+function operationFamilyView(
+  running: readonly RunningAgent[],
+  persistedOperations: readonly OperationalStageState[],
+): OperationFamilyView {
+  const runningFamilies = new Set(
+    persistedOperations
+      .filter((operation) => operation.status === 'running')
+      .map((operation) => operationFamilyKey(operation.key)),
+  );
+
+  const labelsByFamily = new Map<string, Set<string>>();
+  for (const runner of running) {
+    if (runner.kind !== 'operation' || runner.parentKey === undefined) continue;
+    if (!runningFamilies.has(runner.parentKey)) continue;
+    const labels = labelsByFamily.get(runner.parentKey) ?? new Set<string>();
+    labels.add(runner.label);
+    labelsByFamily.set(runner.parentKey, labels);
+  }
+
+  const stepByFamily = new Map<string, string>();
+  for (const [family, labels] of labelsByFamily) {
+    const [onlyLabel] = labels;
+    if (labels.size === 1 && onlyLabel !== undefined) stepByFamily.set(family, lowercaseFirst(onlyLabel));
+  }
+  return { runningFamilies, stepByFamily };
+}
+
+/** Progress labels are written to start a row; as a detail they continue a sentence. */
+function lowercaseFirst(label: string): string {
+  return label.charAt(0).toLowerCase() + label.slice(1);
+}
+
+/**
+ * Full structured view of the pipeline: every agent's state plus the raw
+ * metrics/timing needed to present it, and each phase's collapsed state.
+ */
+export function derivePipeline(input: RenderInput, now: number): DerivedPhase[] {
+  const states = deriveAgentStates(input);
+  const byAgent = new Map(input.running.map((r) => [r.agent, r]));
+  const pipeline = pipelineForState(input.state);
+
+  const agentPhases = pipeline.map((phase) => {
+    const agents = phase.agents.map((a): DerivedAgent => {
+      const state = states.get(a.name) ?? 'pending';
+      const metrics = input.state?.agentMetrics[a.name];
+      const runner = byAgent.get(a.name);
+      const error = agentError(a.name, input.state, byAgent);
+      return {
+        name: a.name,
+        label: a.label,
+        state,
+        durationMs: state === 'completed' && metrics ? metrics.durationMs : null,
+        runningElapsedMs: state === 'running' && runner?.startedAt !== undefined ? now - runner.startedAt : null,
+        attempt: state === 'running' && runner ? runner.attempt : null,
+        ...(error !== undefined && { error }),
+      };
+    });
+
+    return {
+      key: phase.key,
+      label: phase.label,
+      children: phase.parallel,
+      meta: phase.parallel ? ('count' as const) : ('duration' as const),
+      state: phaseGlyphState(agents.map((ag) => ag.state)),
+      agents,
+    };
+  });
+
+  // Operational rows merge two sources: stages the worker has persisted (durable truth,
+  // including terminal outcomes) and pending activities whose stage record has not landed
+  // yet. Persisted keys win, so a stage is never listed twice while the two views overlap.
+  const persistedOperations = Object.values(input.state?.operationalStages ?? {});
+  const persistedKeys = new Set(persistedOperations.map((operation) => operation.key));
+  const { runningFamilies, stepByFamily } = operationFamilyView(input.running, persistedOperations);
+  const unpersistedRunning = input.running
+    .filter((runner) => runner.kind === 'operation' && !persistedKeys.has(runner.agent))
+    // A child activity whose family already has a running parent stage is that stage's current
+    // step, not separate work: the parent row below represents it, with the step as its detail
+    // where the family's children agree on one. Without such a parent it keeps its own row.
+    .filter((runner) => runner.parentKey === undefined || !runningFamilies.has(runner.parentKey))
+    .map((runner) => ({
+      key: runner.agent,
+      label: runner.label,
+      status: 'running' as const,
+      ...(runner.startedAt !== undefined && { startedAt: runner.startedAt }),
+      ...(runner.lastFailure !== undefined && { error: safeFailureDetail(true) }),
+    }));
+  const operationalAgents: DerivedAgent[] = [...persistedOperations, ...unpersistedRunning].map((operation) => {
+    const runner = byAgent.get(operation.key);
+    const operationState = operation.status as RunState;
+    const persistedDurationMs = 'durationMs' in operation ? (operation.durationMs ?? null) : null;
+    const detail = operationState === 'running' ? stepByFamily.get(operationFamilyKey(operation.key)) : undefined;
+    return {
+      name: safeOperationKey(operation.key),
+      label: safeOperationLabel(operation.label),
+      state: operationState,
+      durationMs: operationState === 'completed' ? persistedDurationMs : null,
+      runningElapsedMs:
+        operationState === 'running' && operation.startedAt !== undefined ? now - operation.startedAt : null,
+      attempt: operationState === 'running' ? (runner?.attempt ?? null) : null,
+      ...(detail !== undefined && { detail }),
+      ...(operation.error !== undefined && { error: safeFailureDetail(true) }),
+    };
+  });
+
+  // Operational rows are not peers of the agents. Each one is either model work that
+  // belongs to an agent (reconciliation), model work that belongs to the SAST engine
+  // (its stages), or bookkeeping that only earns a row when it is stuck or broken.
+  return assemblePhases(agentPhases, operationalAgents);
+}
+
+/** Reconciliation wall time per vulnerability class, plus the classes whose grouping degraded. */
+interface ReconciliationView {
+  readonly durationByClass: ReadonlyMap<string, number>;
+  readonly ungroupedClasses: ReadonlySet<string>;
+}
+
+function reconciliationView(operations: readonly DerivedAgent[]): ReconciliationView {
+  const durationByClass = new Map<string, number>();
+  const ungroupedClasses = new Set<string>();
+  for (const operation of operations) {
+    if (operationFamilyKey(operation.name) !== 'reconciliation') continue;
+    const [, vulnerabilityClass] = operation.name.split(':');
+    if (vulnerabilityClass === undefined) continue;
+    if (operation.name.endsWith(':fallback')) {
+      ungroupedClasses.add(vulnerabilityClass);
+      continue;
+    }
+    if (operation.durationMs !== null) durationByClass.set(vulnerabilityClass, operation.durationMs);
+  }
+  return { durationByClass, ungroupedClasses };
+}
+
+/** Attach each class's reconciliation time to the agent row it feeds. */
+function withReconciliation(phase: DerivedPhase, view: ReconciliationView): DerivedPhase {
+  const agents = phase.agents.map((agent): DerivedAgent => {
+    const vulnerabilityClass = agentClass(agent.name);
+    const attachedMs = view.durationByClass.get(vulnerabilityClass);
+    const ungrouped = view.ungroupedClasses.has(vulnerabilityClass);
+    return {
+      ...agent,
+      ...(attachedMs !== undefined && { attachedMs }),
+      ...(ungrouped && { ungrouped }),
+    };
+  });
+  return { ...phase, agents };
+}
+
+/**
+ * Build the Agentic SAST phase from the aggregate span the parent workflow records and the
+ * per-stage rows the SAST child signals up. Scans that predate stage signalling have the
+ * aggregate but no stages, and render as a bare phase line rather than an error.
+ */
+function agenticSastPhase(operations: readonly DerivedAgent[]): DerivedPhase | undefined {
+  const aggregate = operations.find((operation) => operation.name === 'agentic-sast');
+  if (aggregate === undefined) return undefined;
+
+  const byStage = new Map<string, DerivedAgent>();
+  for (const operation of operations) {
+    const [family, stage] = operation.name.split(':');
+    if (family !== 'agentic-sast' || stage === undefined) continue;
+    // The worker's label is the scan log's Title Case form. These rows sit beside the
+    // lowercase class rows below them, so they read in the same register here.
+    byStage.set(stage, { ...operation, label: lowercaseFirst(operation.label) });
+  }
+  // Run order, not insertion order: a resumed or replayed run can persist stages out of order.
+  const stages = AGENTIC_SAST_STAGE_ORDER.map((stage) => byStage.get(stage)).filter(
+    (stage): stage is DerivedAgent => stage !== undefined,
+  );
+
+  return {
+    key: 'agentic-sast',
+    label: 'Agentic SAST',
+    children: stages.length > 0,
+    meta: 'duration',
+    state: aggregate.state,
+    summary: aggregate,
+    // It shares wall time with the pentest phases below it, so the times do not add up
+    // in sequence. Saying so is cheaper than a layout that pretends to be two columns.
+    note: 'concurrent',
+    agents: stages,
+  };
+}
+
+/**
+ * Bookkeeping rows worth showing. A deterministic stage that has completed says nothing —
+ * it can only ever read 0s — but one that is still running, or that failed, is exactly what
+ * an operator needs to see, so those keep a row under the phase they belong to.
+ */
+function troubledReportSteps(operations: readonly DerivedAgent[]): readonly DerivedAgent[] {
+  return operations.filter((operation) => {
+    if (isModelBackedOperation(operation.name)) return false;
+    if (operationFamilyKey(operation.name) !== 'report') return false;
+    return operation.state === 'running' || operation.state === 'failed';
+  });
+}
+
+/**
+ * Fold operational rows into the agent phases. Nothing here becomes a bucket of its own:
+ * every surviving row is either a SAST stage, time attached to an agent, or a report step
+ * that is currently in trouble.
+ */
+function assemblePhases(agentPhases: readonly DerivedPhase[], operations: readonly DerivedAgent[]): DerivedPhase[] {
+  const view = reconciliationView(operations);
+  // Reconciliation produces the exploitation queue, so its time belongs on the exploitation
+  // row it feeds. With exploitation off there is no such row, and it falls back to the
+  // analysis row for the same class so the time is never silently dropped.
+  const attachTo = agentPhases.some((phase) => phase.key === 'exploitation')
+    ? 'exploitation'
+    : 'vulnerability-analysis';
+  const reportSteps = troubledReportSteps(operations);
+
+  const phases = agentPhases.map((phase) => {
+    if (phase.key === attachTo) return withReconciliation(phase, view);
+    if (phase.key === 'reporting' && reportSteps.length > 0) {
+      // The report agent stays on the phase line it already titles; the steps in trouble
+      // become its children, so nothing is listed twice.
+      const summary = phase.agents[0];
+      return {
+        ...phase,
+        children: true,
+        ...(summary !== undefined && { summary }),
+        state: phaseGlyphState([...phase.agents, ...reportSteps].map((row) => row.state)),
+        agents: reportSteps,
+      };
+    }
+    return phase;
+  });
+
+  const sast = agenticSastPhase(operations);
+  if (sast === undefined) return phases;
+
+  // Agentic SAST starts with the scan and runs alongside the pentest, so it reads after
+  // the login check rather than appended past Reporting where it never ran.
+  const afterAuth = phases.findIndex((phase) => phase.key === 'auth-validation') + 1;
+  return [...phases.slice(0, afterAuth), sast, ...phases.slice(afterAuth)];
+}
+
+export { agentError };

apps/cli/src/scan/failure.ts

@@ -0,0 +1,31 @@
+/**
+ * Rendering for the worker's '|'-delimited failure string.
+ *
+ * `formatWorkflowError` in the worker joins error segments — phase context, error type,
+ * message, and remediation hint — with '|' as a delimiter. These helpers turn that raw
+ * string into readable output for the CLI's own surfaces.
+ */
+
+/**
+ * Split the failure string into trimmed, non-empty lines. Segments are delimited by '|', and a
+ * segment's own embedded newlines (e.g. a multi-line validation message) become their own lines so
+ * each aligns with the rest of the block.
+ */
+export function parseFailureSegments(message: string): string[] {
+  return message
+    .split(/[|\n]/)
+    .map((segment) => segment.trim())
+    .filter((segment) => segment.length > 0);
+}
+
+/** Multi-line block: one segment per indented line (the caller prints the header). */
+export function indentFailureSegments(message: string, indent = '  '): string {
+  return parseFailureSegments(message)
+    .map((segment) => `${indent}${segment}`)
+    .join('\n');
+}
+
+/** Single-line summary for compact contexts like the status footer. */
+export function inlineFailureReason(message: string): string {
+  return parseFailureSegments(message).join(' — ');
+}

apps/cli/src/scan/pipeline.ts

@@ -0,0 +1,383 @@
+/**
+ * Static description of the Shannon scan pipeline, plus the worker types the CLI
+ * reads back from Temporal.
+ *
+ * The CLI cannot import from the worker package, so this mirrors it. Keep in sync with:
+ *   - apps/worker/src/types/agents.ts        (agent names / ordering)
+ *   - apps/worker/src/session-manager.ts     (phase membership)
+ *   - apps/worker/src/temporal/activities.ts (the run*Agent activity names → `activityType`)
+ *   - apps/worker/src/temporal/shared.ts     (PipelineState / PipelineSummary)
+ *   - apps/worker/src/types/metrics.ts       (AgentMetrics)
+ *   - apps/worker/src/types/run-state.ts     (PartialReasonView)
+ */
+
+export interface AgentSpec {
+  /** Canonical agent name as it appears in PipelineState.completedAgents / agentMetrics. */
+  readonly name: string;
+  /** Short label for the progress tree. */
+  readonly label: string;
+  /** Temporal activity type name — how a running agent shows up in pendingActivities. */
+  readonly activityType: string;
+}
+
+export interface PhaseSpec {
+  readonly key: string;
+  readonly label: string;
+  readonly parallel: boolean;
+  readonly agents: readonly AgentSpec[];
+}
+
+export interface ActivityProgressSpec {
+  readonly key: string;
+  readonly label: string;
+  readonly kind: 'agent' | 'operation';
+  /**
+   * Operation rows whose work is already represented by a persisted parent stage. The parent
+   * owns the row; this activity supplies the step shown as its detail. Parent stage keys are
+   * the family key itself or the family key followed by ':' and a class or stage suffix.
+   */
+  readonly parentKey?: string;
+}
+
+/** The pipeline phases in execution order, each with its agents. */
+export const PIPELINE: readonly PhaseSpec[] = [
+  {
+    // Preflight login check. Only authenticated scans record metrics here; a non-auth scan
+    // records none, so it renders as skipped — like Exploitation when nothing is exploitable.
+    key: 'auth-validation',
+    label: 'Authentication',
+    parallel: false,
+    agents: [{ name: 'validate-authentication', label: 'auth', activityType: 'runAuthenticationValidation' }],
+  },
+  {
+    key: 'pre-recon',
+    label: 'Pre-Recon',
+    parallel: false,
+    agents: [{ name: 'pre-recon', label: 'pre-recon', activityType: 'runPreReconAgent' }],
+  },
+  {
+    key: 'recon',
+    label: 'Recon',
+    parallel: false,
+    agents: [{ name: 'recon', label: 'recon', activityType: 'runReconAgent' }],
+  },
+  {
+    key: 'vulnerability-analysis',
+    label: 'Vulnerability Analysis',
+    parallel: true,
+    agents: [
+      { name: 'injection-vuln', label: 'injection', activityType: 'runInjectionVulnAgent' },
+      { name: 'xss-vuln', label: 'xss', activityType: 'runXssVulnAgent' },
+      { name: 'auth-vuln', label: 'auth', activityType: 'runAuthVulnAgent' },
+      { name: 'ssrf-vuln', label: 'ssrf', activityType: 'runSsrfVulnAgent' },
+      { name: 'authz-vuln', label: 'authz', activityType: 'runAuthzVulnAgent' },
+    ],
+  },
+  {
+    key: 'exploitation',
+    label: 'Exploitation',
+    parallel: true,
+    agents: [
+      { name: 'injection-exploit', label: 'injection', activityType: 'runInjectionExploitAgent' },
+      { name: 'xss-exploit', label: 'xss', activityType: 'runXssExploitAgent' },
+      { name: 'auth-exploit', label: 'auth', activityType: 'runAuthExploitAgent' },
+      { name: 'ssrf-exploit', label: 'ssrf', activityType: 'runSsrfExploitAgent' },
+      { name: 'authz-exploit', label: 'authz', activityType: 'runAuthzExploitAgent' },
+    ],
+  },
+  {
+    key: 'reporting',
+    label: 'Reporting',
+    parallel: false,
+    agents: [{ name: 'report', label: 'report', activityType: 'runReportAgent' }],
+  },
+];
+
+const MISCELLANEOUS_EXPLOIT_AGENT: AgentSpec = {
+  name: 'miscellaneous-exploit',
+  label: 'miscellaneous',
+  activityType: 'runMiscellaneousExploitAgent',
+};
+
+/**
+ * Shape the static PIPELINE to one scan's durable truth. expectedAgents, persisted by the
+ * worker at scan start, names every exploit agent the scan can ever run: exploit rows it
+ * excludes are dropped, 'miscellaneous-exploit' is appended only once the miscellaneous pipeline has
+ * admitted findings, and a phase left with no agents disappears entirely. Without state
+ * (the scan has not initialized durable state yet) the full static pipeline is the best
+ * available guess.
+ */
+export function pipelineForState(state: PipelineState | null): readonly PhaseSpec[] {
+  if (state?.expectedAgents === undefined) return PIPELINE;
+  const expected = new Set(state.expectedAgents);
+  return PIPELINE.map((phase) => {
+    if (phase.key !== 'exploitation') return phase;
+    const agents = phase.agents.filter((agent) => expected.has(agent.name));
+    if (expected.has(MISCELLANEOUS_EXPLOIT_AGENT.name)) agents.push(MISCELLANEOUS_EXPLOIT_AGENT);
+    return { ...phase, agents };
+  }).filter((phase) => phase.agents.length > 0);
+}
+
+const AGENT_ACTIVITY_PROGRESS: Readonly<Record<string, ActivityProgressSpec>> = Object.fromEntries(
+  [...PIPELINE.flatMap((phase) => phase.agents), MISCELLANEOUS_EXPLOIT_AGENT].map((agent) => [
+    agent.activityType,
+    { key: agent.name, label: agent.label, kind: 'agent' },
+  ]),
+);
+
+/** Families whose per-class or per-stage work is already carried by one persisted stage row. */
+const RECONCILIATION_PARENT_KEY = 'reconciliation';
+const AGENTIC_SAST_PARENT_KEY = 'agentic-sast';
+
+// Every production activity that is not an agent run must have a row here. describeScan
+// throws on an unmapped activity type, so adding a worker activity without updating this
+// table breaks `shannon status` loudly instead of hiding the new work. The authoritative
+// name lists live in apps/worker/src/temporal/worker.ts,
+// apps/worker/src/temporal/reconcile-activity-types.ts, and
+// apps/worker/src/ai/sast/capella/temporal/activity-types.ts.
+const OPERATION_ACTIVITY_PROGRESS: Readonly<Record<string, ActivityProgressSpec>> = {
+  runPreflightValidation: { key: 'preflight', label: 'Preflight validation', kind: 'operation' },
+  syncPlaywrightStealthConfig: { key: 'preflight', label: 'Browser setup', kind: 'operation' },
+  initDeliverableGit: { key: 'scan-initialization', label: 'Initialize deliverables', kind: 'operation' },
+  syncCodePathDenyRules: { key: 'scan-initialization', label: 'Apply source rules', kind: 'operation' },
+  initializeDurableScanState: { key: 'durable-state', label: 'Saving scan state', kind: 'operation' },
+  persistMiscellaneousOutcome: {
+    key: 'miscellaneous-pipeline',
+    label: 'Including miscellaneous findings',
+    kind: 'operation',
+  },
+  initializeReportProgress: { key: 'report:initialize', label: 'Initialize report state', kind: 'operation' },
+  renumberClassFindings: { key: 'report:renumber', label: 'Renumber findings', kind: 'operation' },
+  assembleReportActivity: { key: 'report:assemble', label: 'Assemble report inputs', kind: 'operation' },
+  compactReportFindings: { key: 'report:compact', label: 'Compact report findings', kind: 'operation' },
+  persistCanonicalReportProgress: { key: 'report:checkpoint', label: 'Saving report progress', kind: 'operation' },
+  finalizeReportOutputs: { key: 'report:finalize', label: 'Finalize report outputs', kind: 'operation' },
+  persistFinalizedReportProgress: { key: 'report:terminal', label: 'Saving final report state', kind: 'operation' },
+  surfaceReportOutputs: { key: 'report:surface', label: 'Surface customer report', kind: 'operation' },
+  checkExploitationQueue: { key: 'queue-check', label: 'Check exploitation queue', kind: 'operation' },
+  loadResumeState: { key: 'resume-validation', label: 'Validate resume state', kind: 'operation' },
+  restoreGitCheckpoint: { key: 'resume-restore', label: 'Restore checkpoint', kind: 'operation' },
+  registerResumeAttempt: { key: 'resume-registration', label: 'Register resume', kind: 'operation' },
+  recordResumeAttempt: { key: 'resume-registration', label: 'Record resume', kind: 'operation' },
+  logPhaseTransition: { key: 'audit-log', label: 'Update audit log', kind: 'operation' },
+  logWorkflowComplete: { key: 'audit-log', label: 'Finalize audit log', kind: 'operation' },
+  saveCheckpoint: { key: 'checkpoint', label: 'Save checkpoint', kind: 'operation' },
+  seedEmptyProducerQueue: {
+    key: 'miscellaneous-pipeline',
+    label: 'Preparing miscellaneous findings',
+    kind: 'operation',
+  },
+  prepareClassReconciliation: {
+    key: 'reconciliation',
+    label: 'Preparing findings',
+    kind: 'operation',
+    parentKey: RECONCILIATION_PARENT_KEY,
+  },
+  enrichClassSastObservations: {
+    key: 'reconciliation',
+    label: 'Adding code context',
+    kind: 'operation',
+    parentKey: RECONCILIATION_PARENT_KEY,
+  },
+  formClassExploitTasks: {
+    key: 'reconciliation',
+    label: 'Grouping into test cases',
+    kind: 'operation',
+    parentKey: RECONCILIATION_PARENT_KEY,
+  },
+  materializeClassExploitTasks: {
+    key: 'reconciliation',
+    label: 'Writing test cases',
+    kind: 'operation',
+    parentKey: RECONCILIATION_PARENT_KEY,
+  },
+  publishClassReconciliationOss: {
+    key: 'reconciliation',
+    label: 'Saving results',
+    kind: 'operation',
+    parentKey: RECONCILIATION_PARENT_KEY,
+  },
+  capellaArchitecture: {
+    key: 'agentic-sast:architecture',
+    label: 'Mapping architecture',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+  capellaThreatModel: {
+    key: 'agentic-sast:threat-model',
+    label: 'Modelling threats',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+  capellaPlan: {
+    key: 'agentic-sast:plan',
+    label: 'Planning the review',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+  capellaResearch: {
+    key: 'agentic-sast:research',
+    label: 'Researching code',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+  capellaDedupe: {
+    key: 'agentic-sast:dedupe',
+    label: 'Merging duplicates',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+  capellaReview: {
+    key: 'agentic-sast:review',
+    label: 'Reviewing findings',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+  capellaCritic: {
+    key: 'agentic-sast:critic',
+    label: 'Critiquing findings',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+  capellaConfirm: {
+    key: 'agentic-sast:confirm',
+    label: 'Confirming findings',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+  capellaCalibrate: {
+    key: 'agentic-sast:calibrate',
+    label: 'Calibrating risk',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+  capellaExport: {
+    key: 'agentic-sast:export',
+    label: 'Exporting findings',
+    kind: 'operation',
+    parentKey: AGENTIC_SAST_PARENT_KEY,
+  },
+};
+
+/** Complete production activity mirror. Unknown names are errors, never hidden progress. */
+export const ACTIVITY_TO_PROGRESS: Readonly<Record<string, ActivityProgressSpec>> = Object.freeze({
+  ...AGENT_ACTIVITY_PROGRESS,
+  ...OPERATION_ACTIVITY_PROGRESS,
+});
+
+/** Agent-only projection of ACTIVITY_TO_PROGRESS: activity type name to canonical agent name. */
+export const ACTIVITY_TO_AGENT: Readonly<Record<string, string>> = Object.fromEntries(
+  Object.entries(ACTIVITY_TO_PROGRESS)
+    .filter(([, progress]) => progress.kind === 'agent')
+    .map(([activityType, progress]) => [activityType, progress.key]),
+);
+
+/** The vuln/exploit class of an agent (e.g. "authz-vuln" → "authz"), for failedPipelines matching. */
+export function agentClass(name: string): string {
+  return name.replace(/-(vuln|exploit)$/, '');
+}
+
+// === Worker types read back from Temporal (mirror of shared.ts / metrics.ts) ===
+
+export interface AgentMetrics {
+  readonly durationMs: number;
+  readonly costUsd: number | null;
+  readonly numTurns: number | null;
+  readonly model?: string;
+  readonly skipped?: boolean;
+}
+
+export interface OperationalStageState {
+  readonly key: string;
+  readonly label: string;
+  readonly status: 'pending' | 'running' | 'completed' | 'failed' | 'skipped';
+  readonly startedAt?: number;
+  readonly durationMs?: number;
+  readonly error?: string;
+}
+
+/** Family key a persisted operational stage belongs to, e.g. `reconciliation:xss` to `reconciliation`. */
+export function operationFamilyKey(stageKey: string): string {
+  const separator = stageKey.indexOf(':');
+  return separator === -1 ? stageKey : stageKey.slice(0, separator);
+}
+
+/** The Capella stages that get a progress row, in run order. Mirrors CAPELLA_PROGRESS_STAGES
+ *  in apps/worker/src/ai/sast/types.ts — the deterministic `export` stage is not among them. */
+export const AGENTIC_SAST_STAGE_ORDER: readonly string[] = [
+  'architecture',
+  'threat-model',
+  'plan',
+  'research',
+  'dedupe',
+  'review',
+  'critic',
+  'confirm',
+  'calibrate',
+];
+
+/**
+ * Whether an operational stage represents model work rather than bookkeeping.
+ *
+ * Only the agentic-SAST stages and per-class reconciliation run a model; every other
+ * operational stage is a git commit or a durable-state write that can only ever record
+ * sub-second wall time. The progress tree shows model work, so this is what decides
+ * whether a stage is worth a row at all.
+ */
+export function isModelBackedOperation(stageKey: string): boolean {
+  const family = operationFamilyKey(stageKey);
+  if (family === 'agentic-sast') return true;
+  // A `reconciliation:<class>:fallback` marker records a degradation, not a model span.
+  return family === 'reconciliation' && !stageKey.endsWith(':fallback');
+}
+
+export interface PipelineSummary {
+  readonly totalCostUsd: number;
+  readonly totalDurationMs: number; // Wall-clock (end - start)
+  readonly totalTurns: number;
+  readonly agentCount: number;
+  /** False when operational (Capella/reconciliation) spend is known to be incomplete. */
+  readonly usageAccountingComplete?: boolean;
+}
+
+/** One durable degradation reason with its derived safe message (mirror of PartialReasonView). */
+export interface PartialReasonView {
+  readonly code: string;
+  readonly vulnerabilityClass?: string;
+  readonly stage?: string;
+  readonly message: string;
+}
+
+export type PipelineStatus = 'running' | 'completed' | 'failed' | 'cancelled' | 'partial';
+
+export interface PipelineState {
+  readonly status: PipelineStatus;
+  readonly currentPhase: string | null;
+  readonly currentAgent: string | null;
+  readonly completedAgents: string[];
+  readonly expectedAgents?: string[];
+  readonly participatingClasses?: string[];
+  readonly failedPipelines: { vulnType: string; error: string }[];
+  readonly failedReconciliations?: { vulnerabilityClass: string; error: string }[];
+  readonly failedAgent: string | null;
+  readonly error: string | null;
+  readonly startTime: number;
+  readonly agentMetrics: Record<string, AgentMetrics>;
+  readonly operationalMetrics?: Record<string, AgentMetrics>;
+  readonly operationalStages?: Record<string, OperationalStageState>;
+  /** `error` is the worker's sanitized failure sentence, safe to print verbatim. */
+  readonly agenticSast?: {
+    readonly status: string;
+    readonly durationMs?: number;
+    /** Reader-facing name of the failed stage, already projected by the worker. */
+    readonly failedStageLabel?: string;
+    readonly error?: string;
+    readonly errorCode?: string;
+    /** Usage-accounting warnings projected by the worker; empty when the ledger reconciled. */
+    readonly warnings?: readonly string[];
+  };
+  readonly nonFatalFailures?: { readonly phase: string; readonly error: string }[];
+  /** Ordered durable degradation reasons with safe messages; empty or absent for full success. */
+  readonly partialReasons?: readonly PartialReasonView[];
+  readonly summary: PipelineSummary | null;
+}

apps/cli/src/scan/render.ts

@@ -0,0 +1,321 @@
+/**
+ * Renders a scan's Temporal state into the terminal progress tree.
+ *
+ * The same PipelineState drives both the live view (from the getProgress query) and
+ * the final view (from the workflow result); the running-agents overlay (from
+ * pendingActivities) supplies the in-flight set and retry counts the state lacks.
+ * Colors and Unicode glyphs are gated by the caller so the frame degrades off a TTY.
+ */
+
+import { BOLD, DIM, GOLD, paint, RED, YELLOW } from '../colors.js';
+import { commandPrefix } from '../mode.js';
+import type { RunningAgent } from '../temporal-client.js';
+import { derivePipeline, isTerminal, type RunState, scanElapsedMs } from './derive.js';
+import type { PipelineState } from './pipeline.js';
+import { safeAgenticSast, safeCliIdentifier, safePartialReasons, safeTerminalFailure } from './safe-fields.js';
+
+export interface RenderInput {
+  readonly workspace: string;
+  /** Temporal workflow id backing this scan (differs from workspace on a resume); used for the dashboard link. */
+  readonly workflowId?: string;
+  /** Temporal WorkflowExecutionStatusName: RUNNING | COMPLETED | FAILED | CANCELLED | TERMINATED | … */
+  readonly temporalStatus: string;
+  /** Progress (live) or result (terminal). Null when unavailable, e.g. a hard failure with no result. */
+  readonly state: PipelineState | null;
+  readonly running: readonly RunningAgent[];
+  readonly startedAt?: number;
+  readonly endedAt?: number;
+  /** Failure text when a failed scan has no readable state. */
+  readonly failureMessage?: string;
+}
+
+export interface RenderOptions {
+  readonly now: number;
+  readonly color: boolean;
+  readonly unicode: boolean;
+  /** True for the live view (adds a watch footer); false for the final/one-shot frame. */
+  readonly live: boolean;
+  /** Animation tick — advances the running-agent spinner. Ignored for static frames. */
+  readonly frame: number;
+}
+
+const COLORS = {
+  red: RED,
+  gold: GOLD,
+  yellow: YELLOW,
+  dim: DIM,
+  bold: BOLD,
+} as const;
+
+// === Formatting ===
+
+function formatDuration(ms: number): string {
+  const seconds = Math.max(0, Math.floor(ms / 1000));
+  const hours = Math.floor(seconds / 3600);
+  const minutes = Math.floor((seconds % 3600) / 60);
+  const secs = seconds % 60;
+
+  if (hours > 0) return `${hours}h ${minutes}m`;
+  if (minutes > 0) return `${minutes}m ${secs}s`;
+  return `${secs}s`;
+}
+
+function truncate(text: string, max: number): string {
+  const flat = text.replace(/\s+/g, ' ').trim();
+  return flat.length <= max ? flat : `${flat.slice(0, max - 1)}…`;
+}
+
+/** Temporal Web UI, published by compose on 8233; deep-links to the workflow when its id is known. */
+function temporalDashboardUrl(workflowId: string | undefined): string {
+  const base = 'http://localhost:8233';
+  return workflowId ? `${base}/namespaces/default/workflows/${safeCliIdentifier(workflowId)}` : base;
+}
+
+// === Glyphs & status ===
+
+const GLYPH_UNICODE: Record<RunState, string> = {
+  pending: '○',
+  running: '⟳',
+  completed: '●',
+  failed: '✗',
+  skipped: '·',
+};
+const GLYPH_ASCII: Record<RunState, string> = {
+  pending: '.',
+  running: '>',
+  completed: '+',
+  failed: 'x',
+  skipped: '-',
+};
+const STATE_COLOR: Record<RunState, string> = {
+  pending: COLORS.dim,
+  running: COLORS.gold,
+  completed: COLORS.gold,
+  failed: COLORS.red,
+  skipped: COLORS.dim,
+};
+
+/** Column width for an agent or background-work label inside a phase. */
+const AGENT_LABEL_WIDTH = 18;
+
+/** Inline budget for a failure sentence, wide enough to carry a whole first sentence. */
+const FAILURE_DETAIL_WIDTH = 120;
+
+/** Braille spinner frames for running agents — the clack loader style. */
+const SPINNER_FRAMES = ['⠋', '⠙', '⠹', '⠸', '⠼', '⠴', '⠦', '⠧', '⠇', '⠏'] as const;
+
+function glyph(state: RunState, opts: RenderOptions): string {
+  if (state === 'running' && opts.unicode) {
+    const spin = SPINNER_FRAMES[opts.frame % SPINNER_FRAMES.length] ?? SPINNER_FRAMES[0];
+    return paint(spin, STATE_COLOR.running, opts.color);
+  }
+  const symbol = opts.unicode ? GLYPH_UNICODE[state] : GLYPH_ASCII[state];
+  return paint(symbol, STATE_COLOR[state], opts.color);
+}
+
+/** Badge text + color for the scan as a whole, preferring the workflow's own status when known. */
+function statusBadge(input: RenderInput, opts: RenderOptions): string {
+  const workflowStatus = input.state?.status;
+  if (!isTerminal(input.temporalStatus)) return paint('running', COLORS.gold, opts.color);
+  if (workflowStatus === 'partial') return paint('partial', COLORS.yellow, opts.color);
+  if (workflowStatus === 'cancelled') return paint('cancelled', COLORS.yellow, opts.color);
+  if (input.temporalStatus === 'COMPLETED') return paint('completed', COLORS.gold, opts.color);
+  if (input.temporalStatus === 'TERMINATED') return paint('stopped', COLORS.yellow, opts.color);
+  if (input.temporalStatus === 'CANCELLED' || input.temporalStatus === 'CANCELED') {
+    return paint('cancelled', COLORS.yellow, opts.color);
+  }
+  if (input.temporalStatus === 'TIMED_OUT') return paint('timed out', COLORS.red, opts.color);
+  return paint('failed', COLORS.red, opts.color);
+}
+
+// === Line builders ===
+
+/** The parts of a derived row agentMeta reads beyond its state and metrics. */
+interface RowExtras {
+  readonly runningElapsedMs?: number | null;
+  readonly attachedMs?: number;
+  readonly ungrouped?: boolean;
+}
+
+function agentMeta(
+  state: RunState,
+  metrics: { durationMs: number } | undefined,
+  runner: RunningAgent | undefined,
+  error: string | undefined,
+  opts: RenderOptions,
+  step?: string,
+  extras?: RowExtras,
+): string {
+  if (state === 'completed') {
+    const duration = metrics?.durationMs != null ? formatDuration(metrics.durationMs) : 'done';
+    return paint(`${duration}${attachedSuffix(extras)}`, COLORS.dim, opts.color);
+  }
+  if (state === 'running') {
+    const parts = ['running'];
+    if (step !== undefined) parts.push(step);
+    // An operational row carries its own elapsed time: it is derived from the persisted stage
+    // span, and has no pending activity on the parent workflow to read a start time from.
+    const elapsedMs =
+      runner?.startedAt !== undefined ? opts.now - runner.startedAt : (extras?.runningElapsedMs ?? null);
+    if (elapsedMs !== null) parts.push(formatDuration(elapsedMs));
+    if (runner && runner.attempt > 1) parts.push(`retry ${runner.attempt}`);
+    return paint(parts.join(' · '), COLORS.gold, opts.color);
+  }
+  if (state === 'failed') {
+    const detail = error ? ` · ${truncate(error, FAILURE_DETAIL_WIDTH)}` : '';
+    return paint(`failed${detail}`, COLORS.red, opts.color);
+  }
+  if (state === 'skipped') return paint('skipped', COLORS.dim, opts.color);
+  return paint('queued', COLORS.dim, opts.color);
+}
+
+/**
+ * Time a reconciliation lane contributed to this agent's class, shown as `+ duration` on the
+ * row it feeds. `ungrouped` marks a class whose findings could not be grouped, so each one
+ * was tested separately and duplicates are expected.
+ */
+function attachedSuffix(extras: RowExtras | undefined): string {
+  if (extras === undefined) return '';
+  const time = extras.attachedMs === undefined ? '' : ` + ${formatDuration(extras.attachedMs)}`;
+  return extras.ungrouped ? `${time} · ungrouped` : time;
+}
+
+function phaseMeta(states: readonly RunState[], inPlay: number, parallel: boolean, opts: RenderOptions): string {
+  if (states.every((s) => s === 'pending')) return paint('pending', COLORS.dim, opts.color);
+  if (states.every((s) => s === 'skipped')) return paint('skipped', COLORS.dim, opts.color);
+  if (states.some((s) => s === 'failed') && !states.some((s) => s === 'running')) {
+    return paint('failed', COLORS.red, opts.color);
+  }
+  if (!parallel) return '';
+  const done = states.filter((s) => s === 'completed').length;
+  const allDone = states.every((s) => s === 'completed' || s === 'skipped');
+  return paint(`${done}/${inPlay} done`, allDone ? COLORS.gold : COLORS.dim, opts.color);
+}
+
+/** Render the full progress frame as one string (no trailing newline). */
+export function renderScan(input: RenderInput, opts: RenderOptions): string {
+  const byAgent = new Map(input.running.map((r) => [r.agent, r]));
+  const phases = derivePipeline(input, opts.now);
+  const lines: string[] = ['', ...headerLines(input, opts), ''];
+
+  // Only agents that have actually entered play are shown; pending/skipped ones stay hidden.
+  const inPlay = (s: RunState): boolean => s === 'running' || s === 'completed' || s === 'failed';
+
+  for (const phase of phases) {
+    const states = phase.agents.map((agent) => agent.state);
+    const playing = states.filter(inPlay).length;
+    const phaseRunState = phase.state;
+    const metaFor = (agent: (typeof phase.agents)[number]): string => {
+      const metrics = agent.durationMs === null ? undefined : { durationMs: agent.durationMs };
+      return agentMeta(agent.state, metrics, byAgent.get(agent.name), agent.error, opts, agent.detail, agent);
+    };
+
+    // A phase summarizes itself by wall time or by a "k/N done" tally. A phase with its own
+    // recorded span (Agentic SAST) presents it like any agent row; otherwise a single-agent
+    // phase borrows its one agent's duration once that agent starts.
+    const first = phase.agents[0];
+    const firstState = states[0];
+    const borrowed = first && firstState && inPlay(firstState) ? metaFor(first) : undefined;
+    const durationMeta = phase.summary === undefined ? borrowed : metaFor(phase.summary);
+    const summaryMeta =
+      phase.meta === 'duration' && durationMeta !== undefined
+        ? durationMeta
+        : phaseMeta(states, playing, phase.meta === 'count', opts);
+    const note = phase.note === undefined ? '' : paint(` · ${phase.note}`, COLORS.dim, opts.color);
+    lines.push(`  ${glyph(phaseRunState, opts)}  ${phase.label.padEnd(26)}${summaryMeta}${note}`);
+
+    if (!phase.children) continue;
+    for (let i = 0; i < phase.agents.length; i++) {
+      const agent = phase.agents[i];
+      const state = states[i];
+      if (!agent || !state || !inPlay(state)) continue;
+      // Two trailing spaces before padding, so a label wider than the column still separates
+      // from its meta text; a label inside the column pads to the same width as before.
+      lines.push(`       ${glyph(state, opts)} ${`${agent.label}  `.padEnd(AGENT_LABEL_WIDTH)}${metaFor(agent)}`);
+    }
+  }
+
+  lines.push(...footerLines(input, opts));
+  return lines.join('\n');
+}
+
+function headerLines(input: RenderInput, opts: RenderOptions): string[] {
+  const elapsedMs = scanElapsedMs(input, opts.now);
+  const meta = [statusBadge(input, opts), elapsedMs !== undefined ? formatDuration(elapsedMs) : '—'].join(' · ');
+  return [`  ${paint('Scan:', COLORS.bold, opts.color)} ${safeCliIdentifier(input.workspace).padEnd(22)} ${meta}`];
+}
+
+/** Aligned label column for the footer's Logs / Temporal rows. */
+const FOOTER_LABEL_WIDTH = 12;
+
+/** A thin rule that sets the footer apart from the phase list above it. */
+function footerDivider(opts: RenderOptions): string {
+  return paint(`  ${(opts.unicode ? '─' : '-').repeat(60)}`, COLORS.dim, opts.color);
+}
+
+/** One footer row: an accent-colored label in a fixed column, then its value in the default color. */
+function footerRow(label: string, value: string, opts: RenderOptions): string {
+  return `  ${paint(label.padEnd(FOOTER_LABEL_WIDTH), COLORS.gold, opts.color)}${value}`;
+}
+
+function footerLines(input: RenderInput, opts: RenderOptions): string[] {
+  const prefix = commandPrefix();
+
+  if (isTerminal(input.temporalStatus) && input.state?.summary) {
+    const wall = formatDuration(input.state.summary.totalDurationMs);
+    const lines = ['', `  Time Taken   ${wall}`];
+
+    // A partial scan names each durable degradation reason through its safe message,
+    // so the operator never has to guess why the badge is not "completed".
+    const reasons = safePartialReasons(input.state.partialReasons ?? []);
+    if (reasons.length > 0) {
+      lines.push('', `  ${paint('Why this scan is partial:', COLORS.yellow, opts.color)}`);
+      for (const reason of reasons) {
+        lines.push(paint(`    - ${reason.message}`, COLORS.dim, opts.color));
+      }
+      // The safe message names what degraded; these three name the agentic-SAST failure
+      // behind it, under the same labels the scan log and worker output use.
+      const agenticSast = safeAgenticSast(input.state.agenticSast);
+      if (agenticSast?.status === 'failed') {
+        if (agenticSast.failedStageLabel !== undefined) {
+          lines.push(paint(`    Agentic SAST stopped at: ${agenticSast.failedStageLabel}`, COLORS.dim, opts.color));
+        }
+        if (agenticSast.error !== undefined) {
+          lines.push(paint(`    What happened: ${agenticSast.error}`, COLORS.dim, opts.color));
+        }
+        if (agenticSast.errorCode !== undefined) {
+          lines.push(paint(`    Reference code (for a bug report): ${agenticSast.errorCode}`, COLORS.dim, opts.color));
+        }
+      }
+    }
+    if (input.state.summary.usageAccountingComplete === false) {
+      lines.push(
+        paint('  Cost is incomplete — some background work is not included in this total.', COLORS.dim, opts.color),
+      );
+    }
+    return lines;
+  }
+
+  const logsValue = `${prefix} logs ${safeCliIdentifier(input.workspace)}`;
+  const temporalValue = temporalDashboardUrl(input.workflowId);
+
+  if (isTerminal(input.temporalStatus)) {
+    const hasRecordedFailure =
+      input.failureMessage !== undefined || (input.state !== null && input.state.error !== null);
+    const reason = safeTerminalFailure(hasRecordedFailure) ?? 'no result recorded';
+    return [
+      footerDivider(opts),
+      paint(
+        `  ${input.temporalStatus === 'TERMINATED' ? 'Stopped' : 'Ended'} — ${truncate(reason, 240)}`,
+        COLORS.dim,
+        opts.color,
+      ),
+      footerRow('Logs', logsValue, opts),
+      footerRow('Temporal', temporalValue, opts),
+    ];
+  }
+
+  const lines = [footerDivider(opts), footerRow('Logs', logsValue, opts), footerRow('Temporal', temporalValue, opts)];
+  if (opts.live) lines.push('', paint('  Ctrl-C stops watching — the scan keeps running.', COLORS.dim, opts.color));
+  return lines;
+}

apps/cli/src/scan/safe-fields.ts

@@ -0,0 +1,278 @@
+/**
+ * Closed-field projection for Temporal values displayed by the CLI.
+ *
+ * PipelineState travels through Temporal from a worker container this process does not
+ * control, so free-text fields are treated as unvetted: this module either matches a
+ * value against a known closed set (safe to print as-is) or collapses it to a fixed,
+ * bounded message. A value with no case here should fail closed to something generic,
+ * never pass through untouched.
+ */
+
+import type { PartialReasonView, PipelineState } from './pipeline.js';
+
+const CLASS_NAMES: Readonly<Record<string, string>> = Object.freeze({
+  injection: 'Injection',
+  xss: 'Cross-Site Scripting',
+  auth: 'Authentication',
+  authz: 'Authorization',
+  ssrf: 'Server-Side Request Forgery',
+  miscellaneous: 'Miscellaneous',
+});
+
+const STAGE_NAMES: Readonly<Record<string, string>> = Object.freeze({
+  architecture: 'architecture mapping',
+  'threat-model': 'threat modelling',
+  plan: 'review planning',
+  research: 'deep code research',
+  dedupe: 'duplicate merging',
+  review: 'independent review',
+  critic: 'viability critique',
+  confirm: 'static confirmation',
+  calibrate: 'risk calibration',
+  export: 'findings export',
+  workflow: 'orchestration',
+});
+
+const TERMINAL_STAGE_NAMES = new Set([
+  'architecture',
+  'threat model',
+  'planning',
+  'audit wave',
+  'deduplication',
+  'review',
+  'critic',
+  'confirmation',
+  'calibration',
+  'export',
+  'orchestration',
+]);
+
+const CAPELLA_FAILURE_MESSAGES = new Set([
+  'Provider authentication failed. Verify the configured credential.',
+  'Agentic SAST configuration is invalid.',
+  'Agentic SAST received invalid input.',
+  'An agentic SAST step returned an unusable result.',
+  'An agentic SAST step failed.',
+  'Agentic SAST infrastructure failed before producing a usable result.',
+  'Agentic SAST had not finished when the scan stopped.',
+]);
+
+// Mirrors apps/worker/src/types/errors.ts. The CLI cannot import from the worker package,
+// so keep this exact closed set in sync with ProviderFailureCategory.
+const PROVIDER_FAILURE_CATEGORIES = new Set([
+  'rate_limit',
+  'overloaded',
+  'transport',
+  'context_limit',
+  'quota',
+  'authentication',
+  'configuration',
+  'unknown',
+]);
+
+function isProviderFailureCategory(value: unknown): value is string {
+  return typeof value === 'string' && PROVIDER_FAILURE_CATEGORIES.has(value);
+}
+
+const OPERATION_LABELS = new Set([
+  'Agentic SAST',
+  // Capella stage rows, signalled up from the SAST child workflow. Mirrors
+  // CAPELLA_STAGE_LABELS in apps/worker/src/ai/sast/types.ts, minus the deterministic
+  // export stage, which never becomes a row.
+  'Architecture',
+  'Threat model',
+  'Plan',
+  'Research',
+  'Dedupe',
+  'Review',
+  'Critique',
+  'Confirm',
+  'Calibrate',
+  'Reconcile injection',
+  'Reconcile xss',
+  'Reconcile auth',
+  'Reconcile authz',
+  'Reconcile ssrf',
+  'Reconcile miscellaneous',
+  'Prepare reconciliation',
+  'Enrich observations',
+  'Form exploit tasks',
+  'Materialize exploit tasks',
+  'Publish reconciliation',
+  'Renumber injection',
+  'Renumber xss',
+  'Renumber auth',
+  'Renumber authz',
+  'Renumber ssrf',
+  'Renumber miscellaneous',
+  'Initialize report state',
+  'Assemble report inputs',
+  'Compact report findings',
+  'Saving report progress',
+  'Finalize report outputs',
+  'Finalize report without SARIF',
+  'Saving final report state',
+  'Surface customer report',
+]);
+
+function safeClassName(value: string | undefined): string | undefined {
+  return value === undefined ? undefined : CLASS_NAMES[value];
+}
+
+function safeStageName(value: string | undefined): string | undefined {
+  return value === undefined ? undefined : STAGE_NAMES[value];
+}
+
+function reasonMessage(reason: PartialReasonView): string | undefined {
+  const className = safeClassName(reason.vulnerabilityClass);
+  switch (reason.code) {
+    case 'agentic_sast_failed': {
+      const stageName = safeStageName(reason.stage);
+      return stageName === undefined
+        ? 'Agentic SAST failed, so the pentest continued without its findings.'
+        : `Agentic SAST failed during ${stageName}, so the pentest continued without its findings.`;
+    }
+    case 'agentic_sast_reduced':
+      return 'Agentic SAST completed with reduced coverage.';
+    case 'class_pipeline_failed':
+      return className === undefined
+        ? undefined
+        : `${className} could not be fully assessed. The other classes completed. Re-running this workspace retries only the part that failed.`;
+    case 'class_reconciliation_failed':
+      return className === undefined
+        ? undefined
+        : `${className} findings could not be grouped into test cases, so that class was not exploited and its findings are not in the report.`;
+    case 'report_renumber_failed':
+      return className === undefined
+        ? undefined
+        : `${className} findings kept their working reference numbers, so numbering in the report may have gaps. The findings themselves are complete.`;
+    case 'report_compaction_failed':
+      return 'Finding reference numbers in the report may have gaps. Every finding is present; only the numbering is affected.';
+    case 'report_class_omitted':
+      return className === undefined
+        ? undefined
+        : `${className} was assessed but could not be included in the final report.`;
+    case 'report_sarif_failed':
+      return 'Report SARIF could not be generated. JSON and Markdown remain available.';
+    default:
+      return undefined;
+  }
+}
+
+export function safePartialReasons(reasons: readonly PartialReasonView[]): readonly PartialReasonView[] {
+  return reasons.flatMap((reason) => {
+    const message = reasonMessage(reason);
+    if (message === undefined) return [];
+    const vulnerabilityClass =
+      safeClassName(reason.vulnerabilityClass) === undefined ? undefined : reason.vulnerabilityClass;
+    const stage = safeStageName(reason.stage) === undefined ? undefined : reason.stage;
+    return [
+      {
+        code: reason.code,
+        message,
+        ...(vulnerabilityClass !== undefined && { vulnerabilityClass }),
+        ...(stage !== undefined && { stage }),
+      },
+    ];
+  });
+}
+
+/** Upper bounds on the warning array crossing into cli.status.json, so a malformed state cannot bloat it. */
+const MAX_AGENTIC_SAST_WARNINGS = 20;
+const MAX_AGENTIC_SAST_WARNING_LENGTH = 2_000;
+
+/** Sanitize the worker's usage-accounting warnings: strings only, bounded count and length. */
+function safeAgenticSastWarnings(value: PipelineState['agenticSast']): readonly string[] {
+  const warnings = value?.warnings;
+  if (!Array.isArray(warnings)) return [];
+  return warnings
+    .filter((warning): warning is string => typeof warning === 'string')
+    .slice(0, MAX_AGENTIC_SAST_WARNINGS)
+    .map((warning) => warning.slice(0, MAX_AGENTIC_SAST_WARNING_LENGTH));
+}
+
+export function safeAgenticSast(value: PipelineState['agenticSast']):
+  | {
+      readonly status: string;
+      readonly failedStageLabel?: string;
+      readonly error?: string;
+      readonly errorCode?: string;
+      readonly warnings: readonly string[];
+    }
+  | undefined {
+  if (value === undefined || !['disabled', 'running', 'succeeded', 'failed'].includes(value.status)) return undefined;
+  const failedStageLabel = TERMINAL_STAGE_NAMES.has(value.failedStageLabel ?? '') ? value.failedStageLabel : undefined;
+  let error: string | undefined;
+  if (value.error !== undefined && CAPELLA_FAILURE_MESSAGES.has(value.error)) {
+    error = value.error;
+  } else if (value.status === 'failed') {
+    error = 'An agentic SAST step failed.';
+  }
+  const errorCode =
+    value.errorCode !== undefined &&
+    (/^[A-Z][A-Z0-9_]{0,63}$/u.test(value.errorCode) || isProviderFailureCategory(value.errorCode))
+      ? value.errorCode
+      : undefined;
+  return {
+    status: value.status,
+    ...(failedStageLabel !== undefined && { failedStageLabel }),
+    ...(error !== undefined && { error }),
+    ...(errorCode !== undefined && { errorCode }),
+    warnings: safeAgenticSastWarnings(value),
+  };
+}
+
+export function safeOperationLabel(value: string): string {
+  return OPERATION_LABELS.has(value) ? value : 'Background task';
+}
+
+export function safeOperationKey(value: string): string {
+  if (
+    /^(?:agentic-sast|miscellaneous-pipeline|report:(?:initialize|assemble|compact|checkpoint|finalize|finalize-degraded|terminal|surface))$/u.test(
+      value,
+    ) ||
+    /^agentic-sast:(?:architecture|threat-model|plan|research|dedupe|review|critic|confirm|calibrate)$/u.test(value) ||
+    /^(?:reconciliation|report:renumber):(?:injection|xss|auth|authz|ssrf|miscellaneous)$/u.test(value) ||
+    /^reconciliation:(?:injection|xss|auth|authz|ssrf|miscellaneous):fallback$/u.test(value)
+  ) {
+    return value;
+  }
+  return 'background-task';
+}
+
+/**
+ * A workspace or workflow id is printed straight into the progress display, so this
+ * confines it to a plain identifier charset before that happens: no control or escape
+ * characters survive to reach the terminal.
+ */
+export function safeCliIdentifier(value: string): string {
+  return /^[A-Za-z0-9][A-Za-z0-9._-]{0,127}$/u.test(value) ? value : 'unknown';
+}
+
+export function safeTemporalStatus(value: string): string {
+  return [
+    'RUNNING',
+    'UNSPECIFIED',
+    'COMPLETED',
+    'FAILED',
+    'CANCELLED',
+    'CANCELED',
+    'TERMINATED',
+    'TIMED_OUT',
+    'CONTINUED_AS_NEW',
+  ].includes(value)
+    ? value
+    : 'UNKNOWN';
+}
+
+export function safeFailureDetail(hasFailure: true): string;
+export function safeFailureDetail(hasFailure: false): undefined;
+export function safeFailureDetail(hasFailure: boolean): string | undefined;
+export function safeFailureDetail(hasFailure: boolean): string | undefined {
+  return hasFailure ? 'This scan step could not be completed.' : undefined;
+}
+
+/** Same closed-set trade-off as safeFailureDetail, for the scan-level (not per-agent) failure. */
+export function safeTerminalFailure(hasFailure: boolean): string | undefined {
+  return hasFailure ? 'The scan could not be completed.' : undefined;
+}

apps/cli/src/scan/status-json.ts

@@ -0,0 +1,105 @@
+/**
+ * Machine-readable snapshot of one scan, for `shannon status --json`.
+ *
+ * A point-in-time view built from the same derivation the human progress tree uses
+ * (derive.ts), so the JSON and the rendered tree can never disagree about an agent's
+ * state. One invocation is one snapshot — callers that want to track progress poll it.
+ */
+
+import type { DerivedPhase } from './derive.js';
+import { derivePipeline, isTerminal, scanElapsedMs } from './derive.js';
+import type { PartialReasonView } from './pipeline.js';
+import type { RenderInput } from './render.js';
+import {
+  safeAgenticSast,
+  safeCliIdentifier,
+  safePartialReasons,
+  safeTemporalStatus,
+  safeTerminalFailure,
+} from './safe-fields.js';
+
+/** Coarse scan status token, mirroring the human status badge in machine-friendly form. */
+export type ScanStatus = 'running' | 'completed' | 'partial' | 'failed' | 'stopped' | 'cancelled' | 'timed_out';
+
+export interface StatusJson {
+  readonly workspace: string;
+  /** Temporal workflow id backing this scan (differs from workspace on a resume). */
+  readonly workflowId?: string;
+  /** Coarse outcome: `running` until the scan closes, then its terminal status. */
+  readonly status: ScanStatus;
+  /** Raw Temporal WorkflowExecutionStatusName, for callers that need the source status. */
+  readonly temporalStatus: string;
+  /** Wall-clock elapsed ms (live for a running scan, final for a closed one), or null when unknown. */
+  readonly elapsedMs: number | null;
+  readonly startedAt?: string;
+  readonly endedAt?: string;
+  /** Failure text when a failed scan left no readable state. */
+  readonly failureMessage?: string;
+  /** Ordered durable degradation reasons with safe messages; present only when non-empty. */
+  readonly partialReasons?: readonly PartialReasonView[];
+  /** Agentic SAST outcome, with the worker's sanitized failure sentence and bounded code. */
+  readonly agenticSast?: {
+    readonly status: string;
+    readonly error?: string;
+    readonly errorCode?: string;
+    /** Usage-accounting warnings; always present (empty when the ledger reconciled) so it is never null. */
+    readonly warnings: readonly string[];
+  };
+  /** False when operational (Capella/reconciliation) spend is known to be incomplete. */
+  readonly usageAccountingComplete?: boolean;
+  readonly phases: readonly DerivedPhase[];
+}
+
+/** Map the raw Temporal status (and workflow status) onto the coarse machine token. */
+function deriveStatus(input: RenderInput): ScanStatus {
+  if (!isTerminal(input.temporalStatus)) return 'running';
+  if (input.state?.status === 'partial') return 'partial';
+  if (input.state?.status === 'cancelled') return 'cancelled';
+
+  switch (input.temporalStatus) {
+    case 'COMPLETED':
+      return 'completed';
+    case 'TERMINATED':
+      return 'stopped';
+    case 'CANCELLED':
+    case 'CANCELED':
+      return 'cancelled';
+    case 'TIMED_OUT':
+      return 'timed_out';
+    default:
+      return 'failed';
+  }
+}
+
+/** Build the JSON snapshot for a scan at instant `now`. */
+export function toStatusJson(input: RenderInput, now: number): StatusJson {
+  const elapsedMs = scanElapsedMs(input, now);
+  const partialReasons = safePartialReasons(input.state?.partialReasons ?? []);
+  const agenticSast = safeAgenticSast(input.state?.agenticSast);
+  const usageAccountingComplete = input.state?.summary?.usageAccountingComplete;
+  const failureMessage = safeTerminalFailure(input.failureMessage !== undefined);
+
+  return {
+    workspace: safeCliIdentifier(input.workspace),
+    ...(input.workflowId !== undefined && { workflowId: safeCliIdentifier(input.workflowId) }),
+    status: deriveStatus(input),
+    temporalStatus: safeTemporalStatus(input.temporalStatus),
+    elapsedMs: elapsedMs ?? null,
+    ...(input.startedAt !== undefined && { startedAt: new Date(input.startedAt).toISOString() }),
+    ...(input.endedAt !== undefined && { endedAt: new Date(input.endedAt).toISOString() }),
+    ...(failureMessage !== undefined && { failureMessage }),
+    ...(partialReasons.length > 0 && { partialReasons }),
+    // Present only when agentic SAST actually ran; a disabled scan omits the key entirely.
+    ...(agenticSast !== undefined &&
+      agenticSast.status !== 'disabled' && {
+        agenticSast: {
+          status: agenticSast.status,
+          ...(agenticSast.error !== undefined && { error: agenticSast.error }),
+          ...(agenticSast.errorCode !== undefined && { errorCode: agenticSast.errorCode }),
+          warnings: [...agenticSast.warnings],
+        },
+      }),
+    ...(usageAccountingComplete !== undefined && { usageAccountingComplete }),
+    phases: derivePipeline(input, now),
+  };
+}

apps/cli/src/session.ts

@@ -0,0 +1,27 @@
+/**
+ * Workspace → Temporal workflow-id resolution.
+ *
+ * A workspace name is not always its workflow id: a fresh named workspace gets
+ * `<workspace>_shannon-<timestamp>` as its workflow id (only an auto-named workspace's
+ * directory name equals its original id), and each resume spawns a new workflow
+ * (`<workspace>_resume_<ts>`). The workspace's session.json records the authoritative
+ * id — the latest resume attempt, or the original — so commands that query Temporal
+ * (status, stop) resolve through here instead of assuming the name is the id.
+ */
+
+import fs from 'node:fs';
+import path from 'node:path';
+import { getWorkspacesDir } from './home.js';
+import { resolveRunFile } from './paths.js';
+
+/** Latest workflow id recorded for a workspace: last resume attempt, else the original. */
+export function resolveWorkflowId(workspace: string): string | undefined {
+  const sessionPath = resolveRunFile(path.join(getWorkspacesDir(), workspace), 'session.json');
+  try {
+    const session = JSON.parse(fs.readFileSync(sessionPath, 'utf-8'));
+    const resumeAttempts: { workflowId?: string }[] = session.session?.resumeAttempts ?? [];
+    return resumeAttempts.at(-1)?.workflowId ?? session.session?.originalWorkflowId ?? undefined;
+  } catch {
+    return undefined;
+  }
+}

apps/cli/src/splash.ts

@@ -1,50 +1,94 @@
 /**
  * Splash screen display — pure terminal output, no npm dependencies.
+ * Color escapes are gated on terminal support; the Unicode art is always kept.
  */
 
-export function displaySplash(version?: string): void {
-  const GOLD = '\x1b[38;2;244;197;66m';
-  const CYAN = '\x1b[36;1m';
-  const WHITE = '\x1b[1;37m';
-  const GRAY = '\x1b[0;37m';
-  const YELLOW = '\x1b[1;33m';
-  const RESET = '\x1b[0m';
+import { supportsColor } from './tty.js';
 
-  const B = `${CYAN}\u2551${RESET}`;
-  const S67 = ' '.repeat(67);
-  const HR = '\u2550'.repeat(67);
+/** SHANNON wordmark. Block glyphs take the row fill; box-drawing strokes take the deeper edge shade. */
+const SHANNON = [
+  '███████╗██╗  ██╗ █████╗ ███╗   ██╗███╗   ██╗ ██████╗ ███╗   ██╗',
+  '██╔════╝██║  ██║██╔══██╗████╗  ██║████╗  ██║██╔═══██╗████╗  ██║',
+  '███████╗███████║███████║██╔██╗ ██║██╔██╗ ██║██║   ██║██╔██╗ ██║',
+  '╚════██║██╔══██║██╔══██║██║╚██╗██║██║╚██╗██║██║   ██║██║╚██╗██║',
+  '███████║██║  ██║██║  ██║██║ ╚████║██║ ╚████║╚██████╔╝██║ ╚████║',
+  '╚══════╝╚═╝  ╚═╝╚═╝  ╚═╝╚═╝  ╚═══╝╚═╝  ╚═══╝ ╚═════╝ ╚═╝  ╚═══╝',
+];
+
+/**
+ * Sunset ramp, yellow at the top row down to burnt orange at the base.
+ * Wordmark row i is filled with stop i and edged with stop i + 1, so the
+ * box-drawing strokes read as a shadow one shade deeper than their row.
+ * `xterm` is the 256-color approximation for terminals without 24-bit color.
+ */
+const SUNSET: ReadonlyArray<{ rgb: readonly [number, number, number]; xterm: number }> = [
+  { rgb: [247, 203, 45], xterm: 220 },
+  { rgb: [246, 182, 38], xterm: 220 },
+  { rgb: [245, 160, 32], xterm: 214 },
+  { rgb: [242, 141, 28], xterm: 214 },
+  { rgb: [238, 121, 24], xterm: 208 },
+  { rgb: [231, 100, 21], xterm: 208 },
+  { rgb: [222, 82, 19], xterm: 202 },
+];
+
+export function displaySplash(version?: string): void {
+  const color = supportsColor();
+  const truecolor = color && /truecolor|24bit/i.test(process.env.COLORTERM ?? '');
+  const RESET = color ? '\x1b[0m' : '';
+  const WHITE = color ? '\x1b[1;97m' : '';
+  const GRAY = color ? '\x1b[0;37m' : '';
+  const DIM = color ? '\x1b[90m' : '';
+
+  const ramp = SUNSET.map(({ rgb: [r, g, b], xterm }) => {
+    if (!color) return '';
+    return truecolor ? `\x1b[38;2;${r};${g};${b}m` : `\x1b[38;5;${xterm}m`;
+  });
+
+  /** Color one wordmark row, emitting an escape only where the run changes. Spaces stay unpainted. */
+  const paint = (row: string, fill: string, edge: string): string => {
+    if (!color) return row;
+    let out = '';
+    let open = '';
+    for (const ch of row) {
+      const want = ch === ' ' ? '' : ch === '█' ? fill : edge;
+      if (want !== open) {
+        if (open) out += RESET;
+        out += want;
+        open = want;
+      }
+      out += ch;
+    }
+    return open ? out + RESET : out;
+  };
 
   const lines = [
     '',
-    `  ${CYAN}\u2554${HR}\u2557${RESET}`,
-    `  ${B}${S67}${B}`,
-    `  ${B}  ${GOLD}\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2557  \u2588\u2588\u2557 \u2588\u2588\u2588\u2588\u2588\u2557 \u2588\u2588\u2588\u2557   \u2588\u2588\u2557\u2588\u2588\u2588\u2557   \u2588\u2588\u2557 \u2588\u2588\u2588\u2588\u2588\u2588\u2557 \u2588\u2588\u2588\u2557   \u2588\u2588\u2557${RESET}  ${B}`,
-    `  ${B}  ${GOLD}\u2588\u2588\u2554\u2550\u2550\u2550\u2550\u255D\u2588\u2588\u2551  \u2588\u2588\u2551\u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2557\u2588\u2588\u2588\u2588\u2557  \u2588\u2588\u2551\u2588\u2588\u2588\u2588\u2557  \u2588\u2588\u2551\u2588\u2588\u2554\u2550\u2550\u2550\u2588\u2588\u2557\u2588\u2588\u2588\u2588\u2557  \u2588\u2588\u2551${RESET}  ${B}`,
-    `  ${B}  ${GOLD}\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2557\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2551\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2551\u2588\u2588\u2554\u2588\u2588\u2557 \u2588\u2588\u2551\u2588\u2588\u2554\u2588\u2588\u2557 \u2588\u2588\u2551\u2588\u2588\u2551   \u2588\u2588\u2551\u2588\u2588\u2554\u2588\u2588\u2557 \u2588\u2588\u2551${RESET}  ${B}`,
-    `  ${B}  ${GOLD}\u255A\u2550\u2550\u2550\u2550\u2588\u2588\u2551\u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2551\u2588\u2588\u2554\u2550\u2550\u2588\u2588\u2551\u2588\u2588\u2551\u255A\u2588\u2588\u2557\u2588\u2588\u2551\u2588\u2588\u2551\u255A\u2588\u2588\u2557\u2588\u2588\u2551\u2588\u2588\u2551   \u2588\u2588\u2551\u2588\u2588\u2551\u255A\u2588\u2588\u2557\u2588\u2588\u2551${RESET}  ${B}`,
-    `  ${B}  ${GOLD}\u2588\u2588\u2588\u2588\u2588\u2588\u2588\u2551\u2588\u2588\u2551  \u2588\u2588\u2551\u2588\u2588\u2551  \u2588\u2588\u2551\u2588\u2588\u2551 \u255A\u2588\u2588\u2588\u2588\u2551\u2588\u2588\u2551 \u255A\u2588\u2588\u2588\u2588\u2551\u255A\u2588\u2588\u2588\u2588\u2588\u2588\u2554\u255D\u2588\u2588\u2551 \u255A\u2588\u2588\u2588\u2588\u2551${RESET}  ${B}`,
-    `  ${B}  ${GOLD}\u255A\u2550\u2550\u2550\u2550\u2550\u2550\u255D\u255A\u2550\u255D  \u255A\u2550\u255D\u255A\u2550\u255D  \u255A\u2550\u255D\u255A\u2550\u255D  \u255A\u2550\u2550\u2550\u255D\u255A\u2550\u255D  \u255A\u2550\u2550\u2550\u255D \u255A\u2550\u2550\u2550\u2550\u2550\u255D \u255A\u2550\u255D  \u255A\u2550\u2550\u2550\u255D${RESET}  ${B}`,
-    `  ${B}${S67}${B}`,
-    `  ${B}              ${CYAN}\u2554\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2557${RESET}               ${B}`,
-    `  ${B}              ${CYAN}\u2551${RESET}  ${WHITE}AI Penetration Testing Framework${RESET}  ${CYAN}\u2551${RESET}               ${B}`,
-    `  ${B}              ${CYAN}\u255A\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u2550\u255D${RESET}               ${B}`,
-    `  ${B}${S67}${B}`,
-  ];
-
-  if (version) {
-    const verStr = `v${version}`;
-    const verPadLeft = Math.floor((67 - verStr.length) / 2);
-    const verPadRight = 67 - verStr.length - verPadLeft;
-    lines.push(`  ${B}${' '.repeat(verPadLeft)}${GRAY}${verStr}${RESET}${' '.repeat(verPadRight)}${B}`);
-  }
-
-  lines.push(
-    `  ${B}${S67}${B}`,
-    `  ${B}                    ${YELLOW}\uD83D\uDD10 DEFENSIVE SECURITY ONLY \uD83D\uDD10${RESET}                  ${B}`,
-    `  ${B}${S67}${B}`,
-    `  ${CYAN}\u255A${HR}\u255D${RESET}`,
+    `  ${WHITE}Keygraph${RESET}${version ? `  ${DIM}v${version}${RESET}` : ''}`,
     '',
-  );
+    ...SHANNON.map((row, i) => `  ${paint(row, ramp[i] ?? '', ramp[i + 1] ?? '')}`),
+    '',
+    `  ${WHITE}AI Pentester for Web Apps and APIs${RESET}`,
+    '',
+    `  ${GRAY}-Authorized Security Testing Only-${RESET}`,
+    '',
+  ];
 
   console.log(lines.join('\n'));
 }
+
+/** Matches the divider width the CI wrappers and the scan renderer already use. */
+const RULE_WIDTH = 60;
+
+/**
+ * Plain-text banner for non-terminal output (CI logs, pipes, redirects).
+ * Drops the wordmark but keeps the authorized-use notice, which a reader of
+ * someone else's pipeline log still needs to see.
+ */
+export function displayPlainBanner(version?: string): void {
+  const rule = '─'.repeat(RULE_WIDTH);
+  console.log(rule);
+  console.log(version ? ` Shannon v${version}` : ' Shannon');
+  console.log(' AI Pentester for Web Apps and APIs, by Keygraph');
+  console.log(' Authorized security testing only.');
+  console.log(rule);
+}

apps/cli/src/suggest.ts

@@ -0,0 +1,58 @@
+/**
+ * "Did you mean?" suggestions for mistyped commands and flags.
+ *
+ * A single Levenshtein-based matcher powers both the unknown-command path in the
+ * dispatcher and the unknown-option path in `parseArgs`, so a typo like `statsu`
+ * or `--workspce` points the user at the closest real name instead of just failing.
+ */
+
+/** Levenshtein edit distance between two strings (insertions, deletions, substitutions). */
+export function editDistance(a: string, b: string): number {
+  if (a.length === 0) return b.length;
+  if (b.length === 0) return a.length;
+
+  // Rolling single row; `diagonal` and `above` carry the two neighbours a full grid would.
+  const row = Array.from({ length: b.length + 1 }, (_, j) => j);
+
+  for (let i = 1; i <= a.length; i++) {
+    let diagonal = row[0] as number;
+    row[0] = i;
+    for (let j = 1; j <= b.length; j++) {
+      const above = row[j] as number;
+      const cost = a[i - 1] === b[j - 1] ? 0 : 1;
+      row[j] = Math.min(above + 1, (row[j - 1] as number) + 1, diagonal + cost);
+      diagonal = above;
+    }
+  }
+  return row[b.length] as number;
+}
+
+/**
+ * The candidate closest to `input`, or undefined if none is near enough.
+ *
+ * A prefix match ("stat" -> "status") wins first; otherwise the lowest edit
+ * distance within a length-scaled threshold, so unrelated words don't match.
+ */
+export function closestMatch(input: string, candidates: readonly string[]): string | undefined {
+  if (input.length >= 2) {
+    const prefix = candidates.find((candidate) => candidate.startsWith(input));
+    if (prefix) return prefix;
+  }
+
+  let best: string | undefined;
+  let bestDistance = Number.POSITIVE_INFINITY;
+  for (const candidate of candidates) {
+    if (candidate.length <= 3) continue;
+
+    const distance = editDistance(input, candidate);
+    if (distance < bestDistance) {
+      bestDistance = distance;
+      best = candidate;
+    }
+  }
+
+  if (best === undefined) return undefined;
+
+  const threshold = Math.max(2, Math.floor(best.length / 3));
+  return bestDistance <= threshold ? best : undefined;
+}

apps/cli/src/temporal-client.ts

@@ -0,0 +1,374 @@
+/**
+ * Thin Temporal client for reading scan state and controlling scan workflow lifecycle.
+ *
+ * A running scan is queried live (getProgress) and read via pendingActivities for
+ * the in-flight agents; a closed scan is read once from its result. Everything goes
+ * straight to the frontend on 127.0.0.1:7233 — the gRPC port the compose file
+ * publishes — so this needs Temporal up, but no worker of its own.
+ */
+
+import { setTimeout as sleep } from 'node:timers/promises';
+import { Client, Connection, WorkflowFailedError, WorkflowNotFoundError } from '@temporalio/client';
+import { ACTIVITY_TO_PROGRESS, type PipelineState } from './scan/pipeline.js';
+
+const ADDRESS = '127.0.0.1:7233';
+const NAMESPACE = 'default';
+const LIFECYCLE_RPC_DEADLINE_MS = 3_000;
+const OPEN_SCAN_WORKFLOW_QUERY =
+  "WorkflowType = 'pentestPipelineWorkflow' AND (ExecutionStatus = 'Running' OR ExecutionStatus = 'Paused')";
+
+// WorkflowExecutionStatusName values that positively prove this execution has closed.
+// PAUSED is open; UNSPECIFIED and UNKNOWN are not safe closure evidence.
+const TERMINAL_STATUSES: ReadonlySet<string> = new Set([
+  'COMPLETED',
+  'FAILED',
+  'CANCELLED',
+  'TERMINATED',
+  'CONTINUED_AS_NEW',
+  'TIMED_OUT',
+]);
+
+export interface RunningAgent {
+  readonly agent: string;
+  readonly label: string;
+  /** 'agent' rows join the static pipeline tree; 'operation' rows feed the background-work phase. */
+  readonly kind: 'agent' | 'operation';
+  /** Set when a persisted parent stage owns this row; the label then reads as that stage's step. */
+  readonly parentKey?: string;
+  readonly attempt: number;
+  readonly startedAt?: number;
+  readonly lastFailure?: string;
+}
+
+/**
+ * The CLI's activity mirror does not know an activity type the running scan is using, so the
+ * progress tree cannot be rendered completely. Distinct from a Temporal connection failure.
+ */
+export class ActivityMirrorError extends Error {
+  override name = 'ActivityMirrorError' as const;
+
+  constructor(activityType: string) {
+    super(
+      `This version of the Shannon command line does not recognise part of the running scan\n(${activityType}). Update Shannon, or watch the scan with: shannon logs <workspace>`,
+    );
+  }
+}
+
+/** Convert a proto ITimestamp (seconds is a Long) to epoch millis. */
+function timestampMs(
+  ts: { seconds?: { toString(): string } | number | null; nanos?: number | null } | null,
+): number | undefined {
+  const seconds = ts?.seconds;
+  if (seconds == null) return undefined;
+  const secNum = typeof seconds === 'number' ? seconds : Number(seconds.toString());
+  return secNum * 1000 + (ts?.nanos ?? 0) / 1e6;
+}
+
+export interface ScanDescription {
+  /** WorkflowExecutionStatusName: RUNNING | COMPLETED | FAILED | CANCELLED | TERMINATED | TIMED_OUT | … */
+  readonly status: string;
+  readonly startedAt?: number;
+  readonly closedAt?: number;
+  readonly runningAgents: readonly RunningAgent[];
+}
+
+export type TerminalOutcome =
+  | { readonly kind: 'success'; readonly state: PipelineState }
+  | { readonly kind: 'failed'; readonly message: string };
+
+/**
+ * The authoritative Temporal state used by lifecycle commands. Transport failures deliberately
+ * remain errors instead of being represented as a closed workflow: callers must not report a
+ * scan stopped unless Temporal has positively confirmed it.
+ */
+export type WorkflowLifecycleState =
+  | { readonly kind: 'open'; readonly status: 'RUNNING' | 'PAUSED' }
+  | { readonly kind: 'terminal'; readonly status: string }
+  | { readonly kind: 'unknown'; readonly status: string }
+  | { readonly kind: 'not-found' };
+
+/** A scan workflow returned by Temporal's eventually consistent open-workflow visibility query. */
+export interface RunningScanWorkflow {
+  readonly workflowId: string;
+  readonly taskQueue: string;
+}
+
+let clientPromise: Promise<Client> | null = null;
+
+function getClient(): Promise<Client> {
+  if (!clientPromise) {
+    const pending = Connection.connect({ address: ADDRESS, connectTimeout: LIFECYCLE_RPC_DEADLINE_MS }).then(
+      (connection) => new Client({ connection, namespace: NAMESPACE }),
+    );
+    // A rejected connect must not be cached forever: clear the memo so the next call rebuilds
+    // instead of replaying the same failure. Scoped to `pending` so a later successful reconnect
+    // that replaced the memo is left untouched.
+    pending.catch(() => resetClient(pending));
+    clientPromise = pending;
+  }
+  return clientPromise;
+}
+
+/**
+ * Drop the memoized client so the next {@link getClient} builds a fresh Connection. The underlying
+ * gRPC channel can wedge such that every reused call fails identically ("Unexpected error while
+ * making gRPC request"), and only a new Connection recovers. Best-effort closes the old channel.
+ * When `only` is given, the memo is cleared only if it still holds that exact promise.
+ */
+function resetClient(only?: Promise<Client>): void {
+  if (only !== undefined && clientPromise !== only) return;
+  const previous = clientPromise;
+  clientPromise = null;
+  previous?.then((client) => client.connection.close()).catch(() => {});
+}
+
+/** Close the current channel and establish another before a termination retry. */
+export async function refreshWorkflowLifecycleConnection(): Promise<void> {
+  const previous = clientPromise;
+  if (previous !== null) {
+    if (clientPromise === previous) clientPromise = null;
+    try {
+      const client = await previous;
+      await client.connection.close();
+    } catch {
+      // A failed prior connection is already detached. The new connection below is authoritative.
+    }
+  }
+  await getClient();
+}
+
+/**
+ * Run a bounded lifecycle RPC and discard the connection when Temporal did not positively say
+ * that the workflow is absent. A fresh connection is important after a gRPC timeout or transport
+ * failure: reusing a wedged channel can turn a recoverable stop into an indefinitely ambiguous one.
+ */
+async function runLifecycleRpc<T>(operation: (client: Client) => Promise<T>): Promise<T> {
+  const pending = getClient();
+  try {
+    const client = await pending;
+    return await client.withDeadline(Date.now() + LIFECYCLE_RPC_DEADLINE_MS, () => operation(client));
+  } catch (err) {
+    if (!(err instanceof WorkflowNotFoundError)) resetClient(pending);
+    throw err;
+  }
+}
+
+/** Describe a workflow for lifecycle control without reading its progress or pending activities. */
+export async function describeWorkflowLifecycle(workflowId: string): Promise<WorkflowLifecycleState> {
+  try {
+    const desc = await runLifecycleRpc((client) => client.workflow.getHandle(workflowId).describe());
+    if (desc.status.name === 'RUNNING' || desc.status.name === 'PAUSED') {
+      return { kind: 'open', status: desc.status.name };
+    }
+    if (TERMINAL_STATUSES.has(desc.status.name)) return { kind: 'terminal', status: desc.status.name };
+    return { kind: 'unknown', status: desc.status.name };
+  } catch (err) {
+    if (err instanceof WorkflowNotFoundError) return { kind: 'not-found' };
+    throw err;
+  }
+}
+
+/** Request cooperative cancellation. This confirms request acceptance, not workflow closure. */
+export async function requestWorkflowCancellation(workflowId: string): Promise<'requested' | 'not-found'> {
+  try {
+    await runLifecycleRpc((client) => client.workflow.getHandle(workflowId).cancel());
+    return 'requested';
+  } catch (err) {
+    if (err instanceof WorkflowNotFoundError) return 'not-found';
+    throw err;
+  }
+}
+
+/** Request forced termination. This confirms request acceptance, not workflow closure. */
+export async function requestWorkflowTermination(
+  workflowId: string,
+  reason: string,
+): Promise<'requested' | 'not-found'> {
+  try {
+    await runLifecycleRpc((client) => client.workflow.getHandle(workflowId).terminate(reason));
+    return 'requested';
+  } catch (err) {
+    if (err instanceof WorkflowNotFoundError) return 'not-found';
+    throw err;
+  }
+}
+
+/** List currently open Shannon scan workflows through Temporal visibility. */
+export async function listRunningScanWorkflows(): Promise<readonly RunningScanWorkflow[]> {
+  return runLifecycleRpc(async (client) => {
+    const workflows: RunningScanWorkflow[] = [];
+    for await (const execution of client.workflow.list({ query: OPEN_SCAN_WORKFLOW_QUERY })) {
+      // Visibility is eventually consistent. Keep only the open scan rows returned by this page;
+      // each discovered workflow is described directly before `stop` accepts its closure.
+      if (
+        (execution.status.name === 'RUNNING' || execution.status.name === 'PAUSED') &&
+        execution.type === 'pentestPipelineWorkflow'
+      ) {
+        workflows.push({ workflowId: execution.workflowId, taskQueue: execution.taskQueue });
+      }
+    }
+    return workflows;
+  });
+}
+
+/** Describe a scan: status, timing, and the agents currently running (from pendingActivities). Null if not found. */
+export async function describeScan(workflowId: string): Promise<ScanDescription | null> {
+  const client = await getClient();
+  try {
+    const desc = await client.workflow.getHandle(workflowId).describe();
+
+    const runningAgents: RunningAgent[] = [];
+    for (const pending of desc.raw.pendingActivities ?? []) {
+      const activityType = pending.activityType?.name ?? '';
+      const progress = ACTIVITY_TO_PROGRESS[activityType];
+      // Fail closed: skipping an unknown activity would render a quietly incomplete tree.
+      if (!progress) {
+        throw new ActivityMirrorError(activityType || 'unknown activity');
+      }
+      // Temporal's own failure message is never forwarded verbatim: it can carry raw
+      // exception text from inside the activity, which this client has no way to vet
+      // before painting it into a terminal. Only its presence is kept; the boolean feeds
+      // a fixed sentence downstream (see safeFailureDetail), and the real detail stays
+      // one `shannon logs` away.
+      // NOTE: the proto decoder writes an absent lastFailure as null, not undefined, so a
+      // loose check is what distinguishes a healthy attempt from a failed one.
+      const lastFailure = pending.lastFailure == null ? undefined : 'This activity attempt failed.';
+      const startedAt = timestampMs(pending.scheduledTime ?? pending.lastStartedTime ?? null);
+      runningAgents.push({
+        agent: progress.key,
+        label: progress.label,
+        kind: progress.kind,
+        ...(progress.parentKey !== undefined ? { parentKey: progress.parentKey } : {}),
+        attempt: pending.attempt ?? 1,
+        ...(startedAt !== undefined ? { startedAt } : {}),
+        ...(lastFailure ? { lastFailure } : {}),
+      });
+    }
+
+    return {
+      status: desc.status.name,
+      runningAgents,
+      ...(desc.startTime ? { startedAt: desc.startTime.getTime() } : {}),
+      ...(desc.closeTime ? { closedAt: desc.closeTime.getTime() } : {}),
+    };
+  } catch (err) {
+    if (err instanceof WorkflowNotFoundError) return null;
+    throw err;
+  }
+}
+
+/** Live progress of a running scan via the getProgress query. Null if the query can't be served (no worker). */
+export async function queryProgress(workflowId: string): Promise<PipelineState | null> {
+  const client = await getClient();
+  try {
+    return await client.workflow.getHandle(workflowId).query<PipelineState>('getProgress');
+  } catch {
+    // The query needs a live worker; a just-closed scan may have none. Caller falls back to the result.
+    return null;
+  }
+}
+
+/**
+ * Deepest message in a Temporal failure's cause chain — the real reason nested under generic
+ * wrappers (WorkflowFailedError → ActivityFailure → ApplicationFailure). Covers failed, cancelled,
+ * and terminated alike. Mirrors the SDK's `rootCause` (only exported from @temporalio/common).
+ */
+function rootFailureMessage(err: WorkflowFailedError): string {
+  let message = err.message;
+  let cause: unknown = err.cause;
+  while (cause instanceof Error && cause.message) {
+    message = cause.message;
+    cause = cause.cause;
+  }
+  return message;
+}
+
+/** How a {@link waitForWorkflowClose} watch ended. */
+export type WatchEnd = { readonly reason: 'closed' } | { readonly reason: 'unreachable'; readonly lastError: string };
+
+export interface WatchOptions {
+  /** Poll interval in ms (default 3000). */
+  readonly pollMs?: number;
+  /** Consecutive connection failures before giving up (default 10 → ~30s at the default interval). */
+  readonly maxConnectFailures?: number;
+  /** Consecutive connection failures before {@link onConnectionTrouble} fires once (default 3). */
+  readonly warnAfterFailures?: number;
+  /** Abort the watch (the caller stopped for another reason, e.g. Ctrl-C). */
+  readonly signal?: AbortSignal;
+  /** Called once when contact is first lost, so a live follower's log isn't silent during the outage. */
+  readonly onConnectionTrouble?: (lastError: string) => void;
+  /** Called once when contact is regained after {@link onConnectionTrouble} fired. */
+  readonly onReconnected?: () => void;
+}
+
+/**
+ * Resolve once the scan is no longer running, using the workflow's Temporal status as the
+ * completion signal. Ends on a terminal status, a not-found workflow (closed past retention), or
+ * maxConnectFailures consecutive unreachable polls (a scan can't progress while its Temporal is
+ * down, so sustained no-contact is a safe stop). Never rejects; connection errors surface via the
+ * callbacks and the returned {@link WatchEnd}.
+ */
+export async function waitForWorkflowClose(workflowId: string, opts: WatchOptions = {}): Promise<WatchEnd> {
+  const pollMs = opts.pollMs ?? 3000;
+  const maxConnectFailures = opts.maxConnectFailures ?? 10;
+  const warnAfterFailures = opts.warnAfterFailures ?? 3;
+  const signal = opts.signal;
+
+  let connectFailures = 0;
+  let lastError = '';
+  let warned = false;
+
+  while (!signal?.aborted) {
+    try {
+      const client = await getClient();
+      const desc = await client.workflow.getHandle(workflowId).describe();
+      if (TERMINAL_STATUSES.has(desc.status.name)) {
+        return { reason: 'closed' };
+      }
+      // Reachable and still RUNNING — reset the failure streak and note any recovery.
+      if (warned) {
+        warned = false;
+        opts.onReconnected?.();
+      }
+      connectFailures = 0;
+    } catch (err) {
+      if (err instanceof WorkflowNotFoundError) {
+        return { reason: 'closed' };
+      }
+      // Drop the wedged channel so the next poll dials a fresh one; a cached dead channel would
+      // otherwise fail every retry identically and never recover.
+      resetClient();
+      connectFailures++;
+      lastError = err instanceof Error ? err.message : String(err);
+      if (!warned && connectFailures >= warnAfterFailures) {
+        warned = true;
+        opts.onConnectionTrouble?.(lastError);
+      }
+      if (connectFailures >= maxConnectFailures) {
+        return { reason: 'unreachable', lastError };
+      }
+    }
+
+    try {
+      await sleep(pollMs, undefined, { signal });
+    } catch {
+      break; // Aborted mid-wait by the caller.
+    }
+  }
+
+  return { reason: 'closed' };
+}
+
+/** Final state of a closed scan: success carries the full PipelineState, failure carries the message. */
+export async function getTerminalOutcome(workflowId: string): Promise<TerminalOutcome> {
+  const client = await getClient();
+  try {
+    const state = (await client.workflow.getHandle(workflowId).result()) as PipelineState;
+    return { kind: 'success', state };
+  } catch (err) {
+    if (err instanceof WorkflowFailedError) {
+      return { kind: 'failed', message: rootFailureMessage(err) };
+    }
+    throw err;
+  }
+}

apps/cli/src/tty.ts

@@ -0,0 +1,34 @@
+/**
+ * Terminal capability detection — output coloring, cursor animation, and
+ * whether the user can be prompted interactively.
+ */
+
+import { fail } from './errors.js';
+
+/** True when stdout is a real terminal — safe for color, cursor moves, and spinners. */
+export function stdoutIsTerminal(): boolean {
+  return !!process.stdout.isTTY;
+}
+
+/** True when both stdin and stdout are terminals, so interactive prompts can run. */
+function isInteractive(): boolean {
+  return !!process.stdin.isTTY && !!process.stdout.isTTY;
+}
+
+/** True when color escapes should be emitted. NO_COLOR disables; FORCE_COLOR overrides (0/false/empty = off). */
+export function supportsColor(): boolean {
+  if (process.env.NO_COLOR !== undefined) return false;
+
+  const force = process.env.FORCE_COLOR;
+  if (force !== undefined) {
+    return force !== '0' && force !== 'false' && force !== '';
+  }
+
+  return stdoutIsTerminal();
+}
+
+/** Exit with a clear error when an interactive-only command has no terminal, instead of hanging on a prompt. */
+export function requireInteractive(command: string, alternative: string): void {
+  if (isInteractive()) return;
+  fail(`'${command}' needs an interactive terminal.`, alternative);
+}

apps/cli/src/ui.ts

@@ -0,0 +1,60 @@
+/**
+ * Terminal status output for long-running steps.
+ *
+ * Commands are run with their output captured rather than inherited, so raw docker
+ * plumbing never floods the terminal. Progress is shown with a `@clack/prompts`
+ * spinner. On failure the captured output is printed so the error stays visible
+ * instead of being swallowed.
+ */
+
+import { spawn } from 'node:child_process';
+import * as p from '@clack/prompts';
+
+export interface StepResult {
+  ok: boolean;
+  output: string;
+}
+
+/**
+ * Run a command capturing stdout and stderr. Resolves the exit result and combined
+ * output; never rejects. Callers that want a spinner wrap this in one themselves.
+ */
+export function spawnCaptured(cmd: string, args: string[]): Promise<StepResult> {
+  return new Promise((resolve) => {
+    let output = '';
+    const child = spawn(cmd, args, { stdio: ['ignore', 'pipe', 'pipe'] });
+    child.stdout?.on('data', (chunk) => {
+      output += chunk.toString();
+    });
+    child.stderr?.on('data', (chunk) => {
+      output += chunk.toString();
+    });
+    child.on('close', (code) => resolve({ ok: code === 0, output }));
+    child.on('error', () => resolve({ ok: false, output }));
+  });
+}
+
+/** Print captured command output to stderr, so a failure is never swallowed. */
+export function surfaceOutput(output: string): void {
+  const trimmed = output.trim();
+  if (trimmed) process.stderr.write(`${trimmed}\n`);
+}
+
+/**
+ * Run a command as a labeled step, with a spinner over it. On failure the captured
+ * output is surfaced. Returns the exit result and captured output.
+ */
+export async function runStep(label: string, cmd: string, args: string[]): Promise<StepResult> {
+  const spinner = p.spinner();
+  spinner.start(label);
+
+  const result = await spawnCaptured(cmd, args);
+  if (result.ok) {
+    spinner.stop(label);
+  } else {
+    spinner.error(label);
+    surfaceOutput(result.output);
+  }
+
+  return result;
+}

apps/cli/src/version.ts

@@ -0,0 +1,59 @@
+/**
+ * Version reporting — mode-aware.
+ *
+ * NPX mode:   the published package.json version (stamped by CI at release).
+ * Local mode: the git commit SHA of the checked-out clone (`git-<full-sha>`).
+ *             A clone has no meaningful semver, so the commit is the honest identifier.
+ */
+
+import { execFileSync } from 'node:child_process';
+import fs from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { getMode } from './mode.js';
+
+const __dirname = path.dirname(fileURLToPath(import.meta.url));
+
+function readPackageVersion(): string {
+  try {
+    const pkgPath = path.join(__dirname, '..', 'package.json');
+    const pkg = JSON.parse(fs.readFileSync(pkgPath, 'utf-8')) as { version?: string };
+    return pkg.version || '1.0.0';
+  } catch {
+    return '1.0.0';
+  }
+}
+
+/** Run a git command in the CLI's own repo; returns trimmed stdout or null on any failure. */
+function git(...args: string[]): string | null {
+  try {
+    return execFileSync('git', args, { cwd: __dirname, encoding: 'utf-8', stdio: ['ignore', 'pipe', 'ignore'] }).trim();
+  } catch {
+    return null;
+  }
+}
+
+function readGitSha(): string | null {
+  return git('rev-parse', 'HEAD');
+}
+
+/**
+ * Version identifier. NPX: package.json version. Local: `git-<full-sha>`,
+ * falling back to the package version if git is unavailable.
+ */
+export function getVersion(): string {
+  if (getMode() !== 'local') return readPackageVersion();
+
+  const sha = readGitSha();
+  if (!sha) return readPackageVersion();
+
+  return `git-${sha}`;
+}
+
+/**
+ * Human-facing version line printed by `--version`.
+ * NPX: `shannon <version>`. Local: `shannon git-<full-sha>`.
+ */
+export function getVersionLine(): string {
+  return `shannon ${getVersion()}`;
+}

apps/cli/src/workspaces.ts

@@ -0,0 +1,172 @@
+/**
+ * Workspace enumeration, default-target resolution, and scan identity proof.
+ *
+ * The action commands (`logs`, `status`, `stop`) each take a workspace name. When one
+ * is omitted, `resolveDefaultWorkspace` picks the obvious candidate — the single running
+ * scan, or the most recent workspace — so the common "I just started one scan, show me
+ * its logs" path doesn't require retyping an auto-generated name. Target selection and
+ * identity proof are separate steps: `resolveScanIdentity` turns a selected or explicit
+ * string into the one canonical (workspace, workflowId) pair the session records prove.
+ *
+ * Running workers are identified by Docker workspace label for default-target selection.
+ * `stop` supplements that local discovery with Temporal lifecycle state. Recency for
+ * finished scans comes from each run's session.json createdAt,
+ * with the workspace directory mtime as the fallback for runs that predate it.
+ */
+
+import fs from 'node:fs';
+import path from 'node:path';
+import { runningScanWorkspaces } from './docker.js';
+import { getWorkspacesDir } from './home.js';
+import { resolveRunFile } from './paths.js';
+import { resolveWorkflowId } from './session.js';
+
+export interface WorkspaceInfo {
+  readonly name: string;
+  /** Creation time in ms — the recency sort key. Null when neither session.json nor stat is readable. */
+  readonly createdMs: number | null;
+}
+
+/** Creation time of a workspace: session.json createdAt, else directory mtime, else null. */
+function readCreatedMs(runDir: string): number | null {
+  try {
+    const parsed = JSON.parse(fs.readFileSync(resolveRunFile(runDir, 'session.json'), 'utf-8'));
+    const createdMs = Date.parse(parsed?.session?.createdAt ?? '');
+    if (!Number.isNaN(createdMs)) {
+      return createdMs;
+    }
+  } catch {
+    // Fall through to the directory mtime.
+  }
+
+  try {
+    return fs.statSync(runDir).mtimeMs;
+  } catch {
+    return null;
+  }
+}
+
+/** Every workspace directory, newest-first by createdAt (directory mtime fallback). */
+export function listWorkspaces(): WorkspaceInfo[] {
+  const workspacesDir = getWorkspacesDir();
+
+  let entries: fs.Dirent[];
+  try {
+    entries = fs.readdirSync(workspacesDir, { withFileTypes: true });
+  } catch {
+    // Workspaces directory does not exist yet — no scans have ever run.
+    return [];
+  }
+
+  const workspaces: WorkspaceInfo[] = [];
+  for (const entry of entries) {
+    if (!entry.isDirectory()) {
+      continue;
+    }
+    workspaces.push({ name: entry.name, createdMs: readCreatedMs(path.join(workspacesDir, entry.name)) });
+  }
+
+  // Newest first; workspaces with no known time sort last.
+  workspaces.sort((a, b) => (b.createdMs ?? 0) - (a.createdMs ?? 0));
+  return workspaces;
+}
+
+export type ScanIdentity =
+  | { readonly kind: 'ok'; readonly workspace: string; readonly workflowId: string }
+  | {
+      readonly kind: 'not-found';
+      readonly reason: 'no-match' | 'unreadable-record';
+      /** For 'unreadable-record': the session.json path that could not prove the identity. */
+      readonly sessionPath?: string;
+    }
+  | { readonly kind: 'ambiguous'; readonly claims: readonly string[] };
+
+/** Every workflow id a run's session record has ever claimed: the original plus each resume. */
+function readRecordedWorkflowIds(runDir: string): readonly string[] {
+  try {
+    const session = JSON.parse(fs.readFileSync(resolveRunFile(runDir, 'session.json'), 'utf-8'));
+    const resumeAttempts: { workflowId?: string }[] = session.session?.resumeAttempts ?? [];
+    const ids = [session.session?.originalWorkflowId, ...resumeAttempts.map((attempt) => attempt.workflowId)];
+    return ids.filter((id): id is string => typeof id === 'string' && id.length > 0);
+  } catch {
+    return [];
+  }
+}
+
+/**
+ * Prove the canonical (workspace, workflowId) pair for a status target.
+ *
+ * A directory with a readable session record takes precedence and follows its latest
+ * resume (an auto-named directory whose name equals its original workflow id resolves
+ * here). Otherwise the input is matched exactly against every workflow id the session
+ * records have claimed — session.json is the only trustworthy reverse mapping, and a
+ * valid workspace name may itself end in `_shannon-<digits>`, so the id naming
+ * convention is never used to guess.
+ */
+export function resolveScanIdentity(input: string): ScanIdentity {
+  const runDir = path.join(getWorkspacesDir(), input);
+  let isDirectory = false;
+  try {
+    isDirectory = fs.statSync(runDir).isDirectory();
+  } catch {
+    // Not a workspace directory — fall through to the exact workflow-id match.
+  }
+
+  if (isDirectory) {
+    const workflowId = resolveWorkflowId(input);
+    if (workflowId !== undefined) {
+      return { kind: 'ok', workspace: input, workflowId };
+    }
+    return { kind: 'not-found', reason: 'unreadable-record', sessionPath: resolveRunFile(runDir, 'session.json') };
+  }
+
+  const claims: string[] = [];
+  for (const workspace of listWorkspaces()) {
+    const recorded = readRecordedWorkflowIds(path.join(getWorkspacesDir(), workspace.name));
+    if (recorded.includes(input)) {
+      claims.push(workspace.name);
+    }
+  }
+
+  if (claims.length === 1) {
+    // The exact requested id is kept, so an older workflow id keeps addressing that older execution.
+    return { kind: 'ok', workspace: claims[0] as string, workflowId: input };
+  }
+  if (claims.length > 1) {
+    return { kind: 'ambiguous', claims: [...claims].sort() };
+  }
+  return { kind: 'not-found', reason: 'no-match' };
+}
+
+export type DefaultTarget =
+  | { readonly kind: 'ok'; readonly workspace: string; readonly running: boolean }
+  | { readonly kind: 'none' }
+  | { readonly kind: 'ambiguous'; readonly running: readonly string[] };
+
+/**
+ * Pick the default workspace when the user gave none.
+ *
+ * Exactly one scan running → that scan. Multiple running → ambiguous, so the caller can
+ * list them and ask for an explicit name. None running → the most recent workspace when
+ * `allowFinished` (viewing commands), otherwise none (stopping a finished scan is a no-op).
+ */
+export function resolveDefaultWorkspace(opts: { readonly allowFinished: boolean }): DefaultTarget {
+  const running = runningScanWorkspaces();
+  if (running.length === 1) {
+    return { kind: 'ok', workspace: running[0] as string, running: true };
+  }
+  if (running.length > 1) {
+    return { kind: 'ambiguous', running };
+  }
+
+  if (!opts.allowFinished) {
+    return { kind: 'none' };
+  }
+
+  const workspaces = listWorkspaces();
+  const mostRecent = workspaces[0];
+  if (!mostRecent) {
+    return { kind: 'none' };
+  }
+  return { kind: 'ok', workspace: mostRecent.name, running: false };
+}

apps/worker/configs/config-schema.json

@@ -102,23 +102,6 @@
       "required": ["login_type", "login_url", "credentials", "success_condition"],
       "additionalProperties": false
     },
-    "pipeline": {
-      "type": "object",
-      "description": "Pipeline execution settings for retry behavior and concurrency",
-      "properties": {
-        "retry_preset": {
-          "type": "string",
-          "enum": ["default", "subscription"],
-          "description": "Retry preset. 'subscription' extends timeouts for Anthropic subscription rate limit windows (5h+)."
-        },
-        "max_concurrent_pipelines": {
-          "type": "string",
-          "pattern": "^[1-5]$",
-          "description": "Max concurrent vulnerability pipelines (1-5, default: 5)"
-        }
-      },
-      "additionalProperties": false
-    },
     "rules": {
       "type": "object",
       "description": "Testing rules that define what to focus on or avoid during penetration testing",
@@ -142,16 +125,18 @@
       },
       "additionalProperties": false
     },
-    "vuln_classes": {
-      "type": "array",
-      "description": "Vulnerability classes to test. When omitted, all five classes run. When set, only listed classes run; their vuln+exploit agents and report sections are included.",
-      "items": {
-        "type": "string",
-        "enum": ["injection", "xss", "auth", "authz", "ssrf"]
+    "agentic_sast": {
+      "type": "object",
+      "description": "Opt in to agentic static analysis, which reads the repository for vulnerabilities before the pentest and feeds what it finds into the exploitation phase. Off by default. It does not change which vulnerability classes run. If agentic static analysis fails, the pentest continues without its findings and the scan finishes as \"partial\".",
+      "properties": {
+        "enabled": {
+          "type": "string",
+          "enum": ["true", "false"],
+          "description": "Set to \"true\" to run agentic static analysis. Defaults to \"false\"."
+        }
       },
-      "minItems": 1,
-      "maxItems": 5,
-      "uniqueItems": true
+      "required": ["enabled"],
+      "additionalProperties": false
     },
     "exploit": {
       "type": "string",
@@ -177,6 +162,11 @@
           "minLength": 1,
           "maxLength": 500,
           "description": "Free-text guidance to the report agent (e.g., 'Drop findings about missing security headers')."
+        },
+        "sarif": {
+          "type": "string",
+          "enum": ["true", "false"],
+          "description": "Emit a SARIF 2.1.0 log (report.sarif) beside the report. On by default for exploit runs; set \"false\" to opt out. Ignored when exploit=false."
         }
       },
       "additionalProperties": false
@@ -205,7 +195,7 @@
     { "required": ["rules"] },
     { "required": ["authentication", "rules"] },
     { "required": ["description"] },
-    { "required": ["vuln_classes"] },
+    { "required": ["agentic_sast"] },
     { "required": ["exploit"] },
     { "required": ["report"] },
     { "required": ["rules_of_engagement"] }
@@ -218,7 +208,6 @@
       "properties": {
         "description": {
           "type": "string",
-          "minLength": 1,
           "maxLength": 200,
           "description": "Human-readable description of the rule"
         },
@@ -234,7 +223,7 @@
           "description": "Value to match"
         }
       },
-      "required": ["description", "type", "value"],
+      "required": ["type", "value"],
       "additionalProperties": false
     }
   }

apps/worker/configs/example-config.yaml

@@ -4,8 +4,14 @@
 # Description of the target environment (optional, max 500 chars)
 description: "Next.js e-commerce app on PostgreSQL. Local dev environment — .env files contain local-only credentials, not deployed to production."
 
-# Limit which vulnerability classes run end-to-end (optional, default: all five)
-# vuln_classes: [injection, xss, auth, authz, ssrf]
+# Every scan runs all five vulnerability classes: injection, xss, auth, authz, and ssrf.
+# There is no setting to narrow that.
+
+# Agentic static analysis (optional, default: "false").
+# Reads the repository for vulnerabilities before the pentest and feeds them into exploitation.
+# It costs extra model time, and if it fails the scan finishes as "partial" without its findings.
+# agentic_sast:
+#   enabled: "true"
 
 # Skip the exploitation phase (optional, default: "true")
 # exploit: "false"
@@ -96,13 +102,11 @@ rules:
 # Report filters applied by the report agent when assembling the final report (optional).
 # Example below is illustrative; edit, remove, or add sections as needed.
 # report:
+#   # SARIF 2.1.0 log (report.sarif) beside the report. On by default for exploit runs;
+#   # set "false" to opt out. Ignored when exploit is "false".
+#   sarif: "false"
 #   min_severity: low
 #   min_confidence: low
 #   guidance: |
 #     Drop findings about missing security headers and rate-limit gaps.
 #     ...
-
-# Pipeline execution settings (optional)
-# pipeline:
-#   retry_preset: subscription          # 'default' or 'subscription' (6h max retry for rate limit recovery)
-#   max_concurrent_pipelines: 2         # 1-5, default: 5 (reduce to lower API usage spikes)

apps/worker/package.json

@@ -10,7 +10,27 @@
     "./types/agents": "./dist/types/agents.js",
     "./pipeline": "./dist/temporal/pipeline.js",
     "./activities": "./dist/temporal/activities.js",
+    "./temporal/reconcile-activity-types": "./dist/temporal/reconcile-activity-types.js",
     "./services": "./dist/services/index.js",
+    "./services/queue-validation": "./dist/services/queue-validation.js",
+    "./services/renumber-core": "./dist/services/renumber-core.js",
+    "./services/compaction-core": "./dist/services/compaction-core.js",
+    "./services/finding-order": "./dist/services/finding-order.js",
+    "./ai/structured-generation": "./dist/ai/structured-generation.js",
+    "./ai/pi/source-jail": "./dist/ai/pi/source-jail.js",
+    "./ai/reconciliation/contracts": "./dist/ai/reconciliation/contracts.js",
+    "./ai/reconciliation/stage-contracts": "./dist/ai/reconciliation/stage-contracts.js",
+    "./ai/reconciliation/artifact-store": "./dist/ai/reconciliation/artifact-store.js",
+    "./ai/reconciliation/schema-version": "./dist/ai/reconciliation/schema-version.js",
+    "./ai/reconciliation/manifest": "./dist/ai/reconciliation/manifest.js",
+    "./ai/reconciliation/prepare": "./dist/ai/reconciliation/prepare.js",
+    "./ai/reconciliation/enrich": "./dist/ai/reconciliation/enrich.js",
+    "./ai/reconciliation/form": "./dist/ai/reconciliation/form.js",
+    "./ai/reconciliation/materialize": "./dist/ai/reconciliation/materialize.js",
+    "./ai/reconciliation/observation-view": "./dist/ai/reconciliation/observation-view.js",
+    "./ai/reconciliation/labels": "./dist/ai/reconciliation/labels.js",
+    "./ai/reconciliation/submit-validation": "./dist/ai/reconciliation/submit-validation.js",
+    "./ai/reconciliation/refs": "./dist/ai/reconciliation/refs.js",
     "./config": "./dist/config-parser.js"
   },
   "scripts": {
@@ -19,16 +39,20 @@
     "clean": "rm -rf dist"
   },
   "dependencies": {
-    "@anthropic-ai/claude-agent-sdk": "catalog:",
-    "@temporalio/activity": "^1.11.0",
-    "@temporalio/client": "^1.11.0",
-    "@temporalio/worker": "^1.11.0",
-    "@temporalio/workflow": "^1.11.0",
+    "@earendil-works/pi-agent-core": "^0.84.2",
+    "@earendil-works/pi-ai": "^0.84.2",
+    "@earendil-works/pi-coding-agent": "^0.84.2",
+    "@gotgenes/pi-permission-system": "^10.9.0",
+    "@temporalio/activity": "1.15.0",
+    "@temporalio/client": "1.15.0",
+    "@temporalio/worker": "1.15.0",
+    "@temporalio/workflow": "1.15.0",
     "ajv": "^8.12.0",
     "ajv-formats": "^2.1.1",
     "dotenv": "^16.4.5",
+    "handlebars": "^4.7.9",
     "js-yaml": "^4.1.0",
-    "zod": "^4.3.6",
+    "typebox": "1.1.38",
     "zx": "^8.0.0"
   },
   "devDependencies": {

apps/worker/prompts/exploit-auth.txt

@@ -8,7 +8,7 @@ Success criterion: A comprehensive evidence report detailing the outcome for **e
 </objective>
 
 <exploitation_scope>
-@include(shared/_exploit-scope.txt)
+@include(shared/exploitation/_exploit-scope.txt)
 </exploitation_scope>
 
 <critical>
@@ -53,21 +53,23 @@ Remember: An unproven vulnerability is worse than no finding at all - it wastes
 
 **Level 1 - Authentication Weakness Identified**
 - Evidence: Theoretical bypass found but not yet exploited
-- Classification: POTENTIAL (Low Confidence)
+- Classification: `status: "blocked"`, `confidence: "low"`
 
 **Level 2 - Partial Bypass Achieved** 
 - Evidence: Access gained to non-critical protected resources or UI elements
-- Classification: POTENTIAL (Medium Confidence)
+- Classification: `status: "blocked"`, `confidence: "medium"`
 
 **Level 3 - Authentication Bypass Confirmed**
 - Evidence: Successfully accessed protected functionality as unauthorized user
-- Classification: EXPLOITED
+- Classification: `status: "exploited"`
 
 **Level 4 - Account Takeover Demonstrated**
 - Evidence: Complete user impersonation with access to sensitive data/actions
-- Classification: EXPLOITED (CRITICAL)
+- Classification: `status: "exploited"`
 
-You MUST reach at least Level 3 with evidence to mark as "EXPLOITED"
+Level 3 is the threshold for claiming demonstrated impact and rating the finding `low` or above. Reaching Level 1 or 2 and being stopped by an external operational constraint is `status: "blocked"`.
+
+These levels measure how conclusively you proved the finding is real. They do not set its severity — a finding proven at Level 4 may still be Low. Severity is decided separately, in <severity_reasoning>.
 
 ### Classification Decision Framework
 
@@ -116,7 +118,7 @@ Before beginning exploitation, read these strategic intelligence files in order:
 2.  `.shannon/deliverables/recon_deliverable.md` - Complete API inventory, user roles, and data flow maps.
 3.  `.shannon/deliverables/auth_analysis_deliverable.md` - Strategic context from the Auth analysis specialist, including notes on session mechanisms, password policies, and flawed logic paths.
 
-- You will manage your work using the **TodoWrite tool** to track your exploitation tasks and progress. The todo list is your private workbench for organizing and tracking all exploitation attempts.
+- You will manage your work using the **`todo_write` tool** to track your exploitation tasks and progress. The todo list is your private workbench for organizing and tracking all exploitation attempts.
 </starting_context>
 
 <system_architecture>
@@ -145,18 +147,18 @@ You are the **Identity Compromise Specialist** - proving tangible impact of brok
 
 <cli_tools>
 - **Browser Automation (playwright-cli skill):** Essential for interacting with multi-step authentication flows, injecting stolen session cookies, and verifying account takeover in a real browser context. Invoke the `playwright-cli` skill to learn available commands. Always pass `-s={{PLAYWRIGHT_SESSION}}` to every command for session isolation.
-- **Bash tool:** Use for creating directories, copying files, and other shell commands as needed.
-- **Task Agent:** Mandatory coder-executor for brute force batches, credential stuffing, token replay automation, and any scripted workflow.
-- **TodoWrite tool:** Essential for managing your exploitation task list. You MUST use this tool to track all tasks, mark progress, and document your exploitation workflow.
-- **Read tool:** To read false positives from your tracking file at `workspace/auth_false_positives.md`.
+- **`bash` tool:** Use for creating directories, copying files, and other shell commands as needed.
+- **`task` agent:** Mandatory coder-executor for brute force batches, credential stuffing, token replay automation, and any scripted workflow.
+- **`todo_write` tool:** Essential for managing your exploitation task list. You MUST use this tool to track all tasks, mark progress, and document your exploitation workflow.
+- **`read` tool:** To read false positives from your tracking file at `workspace/auth_false_positives.md`.
 
 **CRITICAL TASK AGENT WORKFLOW:**
-- Delegate every custom script or multi-step automation to the Task Agent; do not hand-write scripts in Bash.
-- Keep requests ≤15 lines and use the template below so the Task Agent has clear inputs and success criteria.
+- Delegate every custom script or multi-step automation to the `task` agent; do not hand-write scripts in Bash.
+- Keep requests ≤15 lines and use the template below so the `task` agent has clear inputs and success criteria.
 </cli_tools>
 
-## Task Agent Scripting Requirements
-Use the Task Agent to execute all bulk or repetitive auth attacks. Manual commands are for quick verification only.
+## `task` agent Scripting Requirements
+Use the `task` agent to execute all bulk or repetitive auth attacks. Manual commands are for quick verification only.
 
 **MANDATORY TEMPLATE:**
 "
@@ -172,7 +174,7 @@ Inputs:
 Success Criteria: Print each attempt with status code and key response text so I can evaluate impact.
 "
 
-**DO NOT:** exceed 5–6 inputs per run, ask for advanced parsing, or bypass the Task Agent for scripting.
+**DO NOT:** exceed 5–6 inputs per run, ask for advanced parsing, or bypass the `task` agent for scripting.
 
 <methodology_and_domain_expertise>
 
@@ -183,11 +185,11 @@ Your primary goal is to achieve a definitive, evidence-backed conclusion for **e
 - **Complete the Workflow:** For each vulnerability, you must follow the full Exploitation Workflow from Confirmation to Impact Demonstration.
 
 ## **The Task-Based Attack Plan**
-Your entire exploitation process will be managed through the TodoWrite tool. You MUST adhere to this process.
+Your entire exploitation process will be managed through the `todo_write` tool. You MUST adhere to this process.
 
 **1. Initialization:**
 - Read `.shannon/deliverables/auth_exploitation_queue.json`.
-- Use the TodoWrite tool to create tasks for each vulnerability in the queue.
+- Use the `todo_write` tool to create tasks for each vulnerability in the queue.
 - Each task should include the vulnerability ID and the hypothesis to test.
     Example tasks:
     - "AUTH-VULN-01: Exploit Session Fixation on /login (Hypothesis: Session ID not rotated)"
@@ -195,10 +197,10 @@ Your entire exploitation process will be managed through the TodoWrite tool. You
 
 **2. The Execution Loop:**
 - **A. Get Next Task:** Review your todo list and identify the next pending task to work on.
-- **B. Plan & Decompose:** Mark the task as in_progress using TodoWrite. Read the `suggested_exploit_technique` from the vulnerability details. This is your primary directive. Plan the specific commands and tools needed to execute this attack pattern. If the attack is complex (e.g., a multi-step logic flaw), use TodoWrite to add smaller tasks.
+- **B. Plan & Decompose:** Mark the task as in_progress using `todo_write`. Read the `suggested_exploit_technique` from the vulnerability details. This is your primary directive. Plan the specific commands and tools needed to execute this attack pattern. If the attack is complex (e.g., a multi-step logic flaw), use `todo_write` to add smaller tasks.
 - **C. Execute:** Run the command for the current task.
-- **D. Update & Document:** Based on the result, update the TodoWrite status to completed if successful. Document the outcome in your memory for the final report.
-- **E. Expand the Plan (Branching):** If successful and more steps are needed, use TodoWrite to add the next logical tasks required by the OWASP workflow.
+- **D. Update & Document:** Based on the result, update the `todo_write` status to completed if successful. Document the outcome in your memory for the final report.
+- **E. Expand the Plan (Branching):** If successful and more steps are needed, use `todo_write` to add the next logical tasks required by the OWASP workflow.
 
 ## **OWASP Authentication Exploitation Workflow**
 For each root vulnerability, follow this systematic, two-stage attack flow.
@@ -231,26 +233,26 @@ Only after exhausting bypass attempts can you proceed to classification analysis
 ## **Validation Criteria & Vulnerability Classification**
 You must adhere to these rules when making a judgment call and categorizing findings:
 
-**EXPLOITED (Report in main section):**
+**`status: "exploited"`:**
 - Successfully achieved account takeover or authentication bypass with concrete evidence (impersonated user, accessed protected resources)
 - OR vulnerability confirmed with active exploitation evidence (e.g., "Performed 200 brute force attempts without rate limiting")
 - Definitive proof of access to a protected resource as another authenticated user OR concrete demonstration of security control bypass
 - Must have completed weaponization and demonstrated impact through impersonation OR provided evidence of exploitable vulnerability
 
-POTENTIAL (Report in separate subsection):
-- This classification is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
+**`status: "blocked"`:**
+- This status is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
 - **CRITICAL: The preventing factor must NOT be a security implementation designed to stop this attack**
 - A finding can only be considered blocked if you have first attempted to bypass the obstacle with several distinct techniques and have documented each failed attempt.
 - Valid external factors include:
   - Server instability (e.g., crashes).
   - Missing authentication that cannot be obtained or bypassed through any known exploit.
-- This classification also applies when:
+- This status also applies when:
   - Code analysis suggests a vulnerability, but live testing is inconclusive due to unmovable constraints.
   - You have reached Level 1 or 2 (initial access confirmed) but are blocked from full exploitation by such external factors.
 
-**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not POTENTIAL. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified POTENTIAL.**
+**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not `blocked`. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified `blocked` finding.**
 
-**FALSE POSITIVE (Do NOT report):**
+**FALSE POSITIVE (never submitted through the tool):**
 - **The blocking mechanism IS a security implementation designed to prevent this attack AND you have attempted to bypass it multiple ways**
 - Definitively proven not exploitable through exhaustive testing
 - Documented attack plan showing all weaponization attempts failed
@@ -260,7 +262,7 @@ POTENTIAL (Report in separate subsection):
 
 **Important:** Rate limiting and account lockouts are obstacles to work around, not grounds for immediate false positive classification. Document constraints and attempt alternative approaches before making final determination.
 
-**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop authentication attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = POTENTIAL.
+**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop authentication attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = `blocked`.
 
 ### CRITICAL: REPRODUCIBILITY REQUIREMENTS
 
@@ -289,7 +291,7 @@ Prioritize these based on the intelligence from your queue and vulnerability ana
 **2. Credential-Based Attacks**
 - **Credential Stuffing with cURL:**
   `curl -X POST -H "Content-Type: application/json" -d '{"username":"admin", "password":"CrackedPassword123"}' https://target.com/api/login`
-- **Brute Force (Example with Task Agent):**
+- **Brute Force (Example with `task` agent):**
   "Task: Write a python script to POST to /api/login with username 'admin' and passwords from this list: ['admin', 'password', '123456']. Print the status code for each."
 
 **3. Logic Flaws**
@@ -304,19 +306,25 @@ Prioritize these based on the intelligence from your queue and vulnerability ana
 </attack_patterns>
 </methodology_and_domain_expertise>
 
-<mcp_tools>
-You emit your exploitation evidence through a single MCP tool — `add_exploit` from the `exploit-collector` server. The host renderer assembles `.shannon/deliverables/auth_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
+@include(shared/exploitation/_severity-reasoning.txt)
+
+@include(shared/exploitation/_reporting-standards.txt)
+
+@include(shared/exploitation/_credentials-in-findings.txt)
+
+<deliverable_tools>
+You emit your exploitation evidence through a single tool — `add_exploit`. The host renderer assembles `.shannon/deliverables/auth_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
 
 **When to emit.** After reaching a definitive verdict on a vulnerability — either successfully exploited (Level 3+ with concrete impact evidence) or potential-but-blocked (real vulnerability, but an external operational constraint blocked full exploitation) — call `add_exploit` once with that finding's structured evidence. Call once per queue vulnerability; do not batch. Continue processing the next vuln in your todo list after each emission.
 
-**Status discriminator.** Set `status: "exploited"` only when you've reached Level 3+ with concrete impact evidence (account takeover demonstrated, session hijacked end-to-end, password reset abused, MFA bypassed). Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. See the Classification Decision Framework in this prompt. Do NOT call `add_exploit` for findings classified FALSE POSITIVE; those go in your `workspace/auth_false_positives.md` tracking file, not the deliverable.
+**Status, for this class.** `status: "exploited"` means your own testing settled the question, and it carries a `severity`. Level 3+ with concrete impact evidence (account takeover demonstrated, session hijacked end-to-end, password reset abused, MFA bypassed) is what you need in order to claim demonstrated impact and rate the finding `low` or above. Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. False positives are recorded in `workspace/auth_false_positives.md`, never through this tool.
 
 **ID alignment.** `vulnerability_id` must match an ID from `.shannon/deliverables/auth_exploitation_queue.json` exactly (e.g. `AUTH-VULN-03`). The collector will reject IDs not in the queue with a list of valid IDs; if you get that error, you either typo'd an ID or imagined one — fix and retry.
 
 **Idempotency.** Duplicate `vulnerability_id` calls are rejected with `DuplicateError`. Each vulnerability may be recorded once; reach your final verdict before emitting.
 
 **Required-call intent.** Before terminating, you should have called `add_exploit` once for each non-FALSE-POSITIVE vulnerability in your queue. The renderer surfaces unprocessed queue IDs in a `## Unprocessed Vulnerabilities` section in the rendered deliverable; downstream consumers read that surface, so misses are visible.
-</mcp_tools>
+</deliverable_tools>
 
 <conclusion_trigger>
 ### Evidence Completeness Verification
@@ -328,7 +336,7 @@ You emit your exploitation evidence through a single MCP tool — `add_exploit`
 - All technical components specified without ambiguity
 
 COMPLETION REQUIREMENTS (ALL must be satisfied):
-	1. Plan Completion: ALL tasks in your todo list must be marked as completed using the TodoWrite tool.
+	1. Plan Completion: ALL tasks in your todo list must be marked as completed using the `todo_write` tool.
 	2. Evidence Emission: Every non-FALSE-POSITIVE vulnerability in `.shannon/deliverables/auth_exploitation_queue.json` has been recorded with one `add_exploit` call carrying the appropriate `status` ("exploited" or "blocked") and required fields. FALSE POSITIVE findings are recorded only in `workspace/auth_false_positives.md`, not via `add_exploit`.
 
 CRITICAL WARNING: Announcing completion before every item in .shannon/deliverables/auth_exploitation_queue.json has been pursued to a final, evidence-backed conclusion (either successfully exploited, recorded as potential-but-blocked, or verified false positive) will be considered a mission failure.

apps/worker/prompts/exploit-authz.txt

@@ -8,7 +8,7 @@ Success criterion: A comprehensive evidence report detailing the outcome for **e
 </objective>
 
 <exploitation_scope>
-@include(shared/_exploit-scope.txt)
+@include(shared/exploitation/_exploit-scope.txt)
 </exploitation_scope>
 
 <critical>
@@ -53,21 +53,23 @@ Remember: An unproven vulnerability is worse than no finding at all - it wastes
 
 **Level 1 - Authorization Weakness Identified**
 - Evidence: Theoretical bypass found but not yet exploited
-- Classification: POTENTIAL (Low Confidence)
+- Classification: `status: "blocked"`, `confidence: "low"`
 
 **Level 2 - Partial Access Control Bypass** 
 - Evidence: Access gained to some protected resources or partial privilege escalation
-- Classification: POTENTIAL (Medium Confidence)
+- Classification: `status: "blocked"`, `confidence: "medium"`
 
 **Level 3 - Access Control Bypass Confirmed**
 - Evidence: Successfully accessed protected functionality or data as unauthorized user
-- Classification: EXPLOITED
+- Classification: `status: "exploited"`
 
 **Level 4 - Critical Privilege Escalation Demonstrated**
 - Evidence: Complete privilege escalation with access to administrative functions or highly sensitive data
-- Classification: EXPLOITED (CRITICAL)
+- Classification: `status: "exploited"`
 
-You MUST reach at least Level 3 with evidence to mark as "EXPLOITED"
+Level 3 is the threshold for claiming demonstrated impact and rating the finding `low` or above. Reaching Level 1 or 2 and being stopped by an external operational constraint is `status: "blocked"`.
+
+These levels measure how conclusively you proved the finding is real. They do not set its severity — a finding proven at Level 4 may still be Low. Severity is decided separately, in <severity_reasoning>.
 
 ### Classification Decision Framework
 
@@ -105,7 +107,7 @@ Before beginning exploitation, read these strategic intelligence files in order:
 2.  `.shannon/deliverables/recon_deliverable.md` - Complete API inventory, user roles, and permission models.
 3.  `.shannon/deliverables/authz_analysis_deliverable.md` - Strategic context from the Authz analysis specialist, including notes on access control patterns, role hierarchies, and flawed logic paths.
 
-- You will manage your work using the **TodoWrite tool** to track your exploitation tasks and progress. The todo list is your private workbench for organizing and tracking all exploitation attempts.
+- You will manage your work using the **`todo_write` tool** to track your exploitation tasks and progress. The todo list is your private workbench for organizing and tracking all exploitation attempts.
 </starting_context>
 
 <system_architecture>
@@ -134,18 +136,18 @@ You are the **Privilege Escalation Specialist** - proving tangible impact of bro
 
 <cli_tools>
 - **Browser Automation (playwright-cli skill):** Essential for interacting with complex authorization flows, testing role-based access controls in browser contexts, and verifying privilege escalation through UI elements. Invoke the `playwright-cli` skill to learn available commands. Always pass `-s={{PLAYWRIGHT_SESSION}}` to every command for session isolation.
-- **Bash tool:** Use for creating directories, copying files, and other shell commands as needed.
-- **Task Agent:** Mandatory coder-executor for IDOR sweeps, role escalation loops, and workflow bypass automation.
-- **TodoWrite tool:** Essential for managing your exploitation task list. You MUST use this tool to track all tasks, mark progress, and document your exploitation workflow.
-- **Read tool:** To read false positives from your tracking file at `workspace/authz_false_positives.md`.
+- **`bash` tool:** Use for creating directories, copying files, and other shell commands as needed.
+- **`task` agent:** Mandatory coder-executor for IDOR sweeps, role escalation loops, and workflow bypass automation.
+- **`todo_write` tool:** Essential for managing your exploitation task list. You MUST use this tool to track all tasks, mark progress, and document your exploitation workflow.
+- **`read` tool:** To read false positives from your tracking file at `workspace/authz_false_positives.md`.
 
 **CRITICAL TASK AGENT WORKFLOW:**
-- Delegate every multi-user iteration, role toggle test, or workflow automation script to the Task Agent—never handcraft these scripts yourself.
-- Keep requests ≤15 lines and adhere to the template below so the Task Agent can act deterministically.
+- Delegate every multi-user iteration, role toggle test, or workflow automation script to the `task` agent—never handcraft these scripts yourself.
+- Keep requests ≤15 lines and adhere to the template below so the `task` agent can act deterministically.
 </cli_tools>
 
-## Task Agent Scripting Requirements
-All repeated authorization tests must run through the Task Agent.
+## `task` agent Scripting Requirements
+All repeated authorization tests must run through the `task` agent.
 
 **MANDATORY TEMPLATE:**
 "
@@ -161,7 +163,7 @@ Inputs:
 Success Criteria: Execute one request per identity, logging status code and key response text so I can confirm access levels.
 "
 
-**DO NOT:** exceed 5 identities per run, ask for complex diffing, or bypass the Task Agent for scripting.
+**DO NOT:** exceed 5 identities per run, ask for complex diffing, or bypass the `task` agent for scripting.
 
 <methodology_and_domain_expertise>
 
@@ -172,11 +174,11 @@ Your primary goal is to achieve a definitive, evidence-backed conclusion for **e
 - **Complete the Workflow:** For each vulnerability, you must follow the full Exploitation Workflow from Confirmation to Impact Demonstration.
 
 ## **The Task-Based Attack Plan**
-Your entire exploitation process will be managed through the TodoWrite tool. You MUST adhere to this process.
+Your entire exploitation process will be managed through the `todo_write` tool. You MUST adhere to this process.
 
 **1. Initialization:**
 - Read `.shannon/deliverables/authz_exploitation_queue.json`.
-- Use the TodoWrite tool to create tasks for each vulnerability in the queue.
+- Use the `todo_write` tool to create tasks for each vulnerability in the queue.
 - Each task should include the vulnerability ID, type, and the hypothesis to test.
     Example tasks:
     - "AUTHZ-VULN-01 (Horizontal): Exploit ownership bypass on /api/user/{id} (Hypothesis: Access to other users' data)"
@@ -185,10 +187,10 @@ Your entire exploitation process will be managed through the TodoWrite tool. You
 
 **2. The Execution Loop:**
 - **A. Get Next Task:** Review your todo list and identify the next pending task to work on.
-- **B. Plan & Decompose:** Mark the task as in_progress using TodoWrite. Read the vulnerability type (`Horizontal`, `Vertical`, or `Context_Workflow`) and the `minimal_witness` from the vulnerability details. This is your primary directive. Plan the specific commands and tools needed to execute this attack pattern. If the attack is complex (e.g., a multi-step privilege escalation), use TodoWrite to add smaller tasks.
+- **B. Plan & Decompose:** Mark the task as in_progress using `todo_write`. Read the vulnerability type (`Horizontal`, `Vertical`, or `Context_Workflow`) and the `minimal_witness` from the vulnerability details. This is your primary directive. Plan the specific commands and tools needed to execute this attack pattern. If the attack is complex (e.g., a multi-step privilege escalation), use `todo_write` to add smaller tasks.
 - **C. Execute:** Run the command for the current task.
-- **D. Update & Document:** Based on the result, update the TodoWrite status to completed if successful. Document the outcome in your memory for the final report.
-- **E. Expand the Plan (Branching):** If successful and more steps are needed, use TodoWrite to add the next logical tasks required by the OWASP workflow.
+- **D. Update & Document:** Based on the result, update the `todo_write` status to completed if successful. Document the outcome in your memory for the final report.
+- **E. Expand the Plan (Branching):** If successful and more steps are needed, use `todo_write` to add the next logical tasks required by the OWASP workflow.
 
 ## **OWASP Authorization Exploitation Workflow**
 For each root vulnerability, follow this systematic, two-stage attack flow.
@@ -230,26 +232,26 @@ Only after exhausting bypass attempts can you proceed to classification analysis
 ## **Validation Criteria & Vulnerability Classification**
 You must adhere to these rules when making a judgment call and categorizing findings:
 
-**EXPLOITED (Report in main section):**
+**`status: "exploited"`:**
 - Successfully achieved privilege escalation or unauthorized access with concrete evidence (accessed protected data, modified unauthorized resources)
 - Definitive proof of access to resources that should be restricted based on role or ownership
 - Must have completed weaponization and demonstrated impact through unauthorized access
 - MUST have reached Level 3 or 4 with evidence
 
-POTENTIAL (Report in separate subsection):
-- This classification is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
+**`status: "blocked"`:**
+- This status is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
 - **CRITICAL: The preventing factor must NOT be a security implementation designed to stop this attack**
 - A finding can only be considered blocked if you have first attempted to bypass the obstacle with several distinct techniques and have documented each failed attempt.
 - Valid external factors include:
   - Server instability (e.g., crashes).
   - Missing authentication that cannot be obtained or bypassed through any known exploit.
-- This classification also applies when:
+- This status also applies when:
   - Code analysis suggests a vulnerability, but live testing is inconclusive due to unmovable constraints.
   - You have reached Level 1 or 2 (initial access confirmed) but are blocked from full exploitation by such external factors.
 
-**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not POTENTIAL. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified POTENTIAL.**
+**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not `blocked`. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified `blocked` finding.**
 
-**FALSE POSITIVE (Do NOT report):**
+**FALSE POSITIVE (never submitted through the tool):**
 - **The blocking mechanism IS a security implementation designed to prevent this attack AND you have attempted to bypass it multiple ways**
 - Definitively proven not exploitable through exhaustive testing
 - Documented attack plan showing all weaponization attempts failed
@@ -259,7 +261,7 @@ POTENTIAL (Report in separate subsection):
 
 **Important:** Authentication requirements and access restrictions are obstacles to work around, not grounds for immediate false positive classification. Document constraints and attempt alternative approaches before making final determination.
 
-**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop authorization attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = POTENTIAL.
+**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop authorization attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = `blocked`.
 
 ### CRITICAL: REPRODUCIBILITY REQUIREMENTS
 
@@ -312,19 +314,25 @@ Remember: The most effective attacks often come from understanding the specific
 </attack_patterns>
 </methodology_and_domain_expertise>
 
-<mcp_tools>
-You emit your exploitation evidence through a single MCP tool — `add_exploit` from the `exploit-collector` server. The host renderer assembles `.shannon/deliverables/authz_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
+@include(shared/exploitation/_severity-reasoning.txt)
+
+@include(shared/exploitation/_reporting-standards.txt)
+
+@include(shared/exploitation/_credentials-in-findings.txt)
+
+<deliverable_tools>
+You emit your exploitation evidence through a single tool — `add_exploit`. The host renderer assembles `.shannon/deliverables/authz_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
 
 **When to emit.** After reaching a definitive verdict on a vulnerability — either successfully exploited (Level 3+ with concrete impact evidence) or potential-but-blocked (real vulnerability, but an external operational constraint blocked full exploitation) — call `add_exploit` once with that finding's structured evidence. Call once per queue vulnerability; do not batch. Continue processing the next vuln in your todo list after each emission.
 
-**Status discriminator.** Set `status: "exploited"` only when you've reached Level 3+ with concrete impact evidence (cross-tenant data access demonstrated, privilege escalation proven, unauthorized workflow transition executed). Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. See the Classification Decision Framework in this prompt. Do NOT call `add_exploit` for findings classified FALSE POSITIVE; those go in your `workspace/authz_false_positives.md` tracking file, not the deliverable.
+**Status, for this class.** `status: "exploited"` means your own testing settled the question, and it carries a `severity`. Level 3+ with concrete impact evidence (cross-tenant data access demonstrated, privilege escalation proven, unauthorized workflow transition executed) is what you need in order to claim demonstrated impact and rate the finding `low` or above. Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. False positives are recorded in `workspace/authz_false_positives.md`, never through this tool.
 
 **ID alignment.** `vulnerability_id` must match an ID from `.shannon/deliverables/authz_exploitation_queue.json` exactly (e.g. `AUTHZ-VULN-03`). The collector will reject IDs not in the queue with a list of valid IDs; if you get that error, you either typo'd an ID or imagined one — fix and retry.
 
 **Idempotency.** Duplicate `vulnerability_id` calls are rejected with `DuplicateError`. Each vulnerability may be recorded once; reach your final verdict before emitting.
 
 **Required-call intent.** Before terminating, you should have called `add_exploit` once for each non-FALSE-POSITIVE vulnerability in your queue. The renderer surfaces unprocessed queue IDs in a `## Unprocessed Vulnerabilities` section in the rendered deliverable; downstream consumers read that surface, so misses are visible.
-</mcp_tools>
+</deliverable_tools>
 
 <conclusion_trigger>
 ### Evidence Completeness Verification
@@ -336,7 +344,7 @@ You emit your exploitation evidence through a single MCP tool — `add_exploit`
 - All technical components specified without ambiguity
 
 COMPLETION REQUIREMENTS (ALL must be satisfied):
-	1. Plan Completion: ALL tasks in your todo list must be marked as completed using the TodoWrite tool.
+	1. Plan Completion: ALL tasks in your todo list must be marked as completed using the `todo_write` tool.
 	2. Evidence Emission: Every non-FALSE-POSITIVE vulnerability in `.shannon/deliverables/authz_exploitation_queue.json` has been recorded with one `add_exploit` call carrying the appropriate `status` ("exploited" or "blocked") and required fields. FALSE POSITIVE findings are recorded only in `workspace/authz_false_positives.md`, not via `add_exploit`.
 
 CRITICAL WARNING: Announcing completion before every item in .shannon/deliverables/authz_exploitation_queue.json has been pursued to a final, evidence-backed conclusion (either successfully exploited, recorded as potential-but-blocked, or verified false positive) will be considered a mission failure.

apps/worker/prompts/exploit-injection.txt

@@ -8,7 +8,7 @@ Success criterion: A comprehensive evidence report detailing the outcome for **e
 </objective>
 
 <exploitation_scope>
-@include(shared/_exploit-scope.txt)
+@include(shared/exploitation/_exploit-scope.txt)
 </exploitation_scope>
 
 <critical>
@@ -87,7 +87,7 @@ Before beginning exploitation, read these strategic intelligence files in order:
 - Leverage vulnerability details and context from pre-analysis rather than blind testing.
 - Reference provided vulnerable code snippets to avoid blind guessing of query column counts, command syntax, or database/OS types.
 
-- You will manage your work using the **TodoWrite tool** to track your exploitation tasks and progress. The todo list is your private workbench for organizing and tracking all exploitation attempts.
+- You will manage your work using the **`todo_write` tool** to track your exploitation tasks and progress. The todo list is your private workbench for organizing and tracking all exploitation attempts.
 </starting_context>
 
 <system_architecture>
@@ -116,13 +116,13 @@ You are the **Proof-of-Impact Generator** - converting theoretical injection fla
 
 <cli_tools>
 - **Browser Automation (playwright-cli skill):** For testing injection vulnerabilities through browser interactions when needed. Invoke the `playwright-cli` skill to learn available commands. Always pass `-s={{PLAYWRIGHT_SESSION}}` to every command for session isolation.
-- **Bash tool:** Use for creating directories, copying files, and other shell commands as needed.
-- **Task Agent:** Mandatory coder-executor for any custom scripting beyond single ad-hoc commands.
-- **TodoWrite tool:** Essential for managing your exploitation task list. You MUST use this tool to track all tasks, mark progress, and document your exploitation workflow.
-- **Read tool:** To read false positives from your tracking file at `workspace/injection_false_positives.md`.
+- **`bash` tool:** Use for creating directories, copying files, and other shell commands as needed.
+- **`task` agent:** Mandatory coder-executor for any custom scripting beyond single ad-hoc commands.
+- **`todo_write` tool:** Essential for managing your exploitation task list. You MUST use this tool to track all tasks, mark progress, and document your exploitation workflow.
+- **`read` tool:** To read false positives from your tracking file at `workspace/injection_false_positives.md`.
 
 **CRITICAL TASK AGENT WORKFLOW:**
-- Task Agent must author and run every custom script, payload loop, or enumeration workflow. Do not craft standalone scripts in Bash or other tools.
+- `task` agent must author and run every custom script, payload loop, or enumeration workflow. Do not craft standalone scripts in Bash or other tools.
 - Keep requests ≤15 lines and follow the template below; specify targets, payloads, and success criteria.
 </cli_tools>
 
@@ -135,11 +135,11 @@ Your primary goal is to achieve a definitive, evidence-backed conclusion for **e
 - **Complete the Workflow:** For each vulnerability, you must follow the full OWASP Exploitation Workflow from Confirmation to either Exfiltration or a documented conclusion of non-exploitability.
 
 ## **The Task-Based Attack Plan**
-Your entire exploitation process will be managed through the TodoWrite tool. You MUST adhere to this process.
+Your entire exploitation process will be managed through the `todo_write` tool. You MUST adhere to this process.
 
 **1. Initialization:**
 - Read the `.shannon/deliverables/injection_exploitation_queue.json` file.
-- Use the TodoWrite tool to create tasks for each vulnerability in the queue.
+- Use the `todo_write` tool to create tasks for each vulnerability in the queue.
 - Each task should include the vulnerability ID and the hypothesis to test.
     Example tasks:
     - "SQLI-VULN-01: Exploit endpoint /api/search?q= (Hypothesis: Basic UNION injection)"
@@ -150,16 +150,16 @@ You will repeatedly perform the following loop until all tasks are completed:
 
 - **A. Get Next Task:** Review your todo list and identify the next pending task to work on.
 
-- **B. Plan & Decompose:** Mark the task as in_progress using TodoWrite. Decide on the concrete command or action. If the task is complex (e.g., "Enumerate tables"), use TodoWrite to add smaller, actionable tasks.
+- **B. Plan & Decompose:** Mark the task as in_progress using `todo_write`. Decide on the concrete command or action. If the task is complex (e.g., "Enumerate tables"), use `todo_write` to add smaller, actionable tasks.
 
 - **C. Execute:** Run the command for the current task (e.g., run `curl` with an `ORDER BY` payload).
 
-- **D. Update & Document:** Based on the result, update the TodoWrite status:
+- **D. Update & Document:** Based on the result, update the `todo_write` status:
     - Mark the task as completed if successful.
     - Document the outcome in your memory, including the exact command and result for the final report.
     - Example outcome to remember: "Step 1.1: Determined column count is 4 using ORDER BY - Command: curl 'https://target.com/api/search?q=test' ORDER BY 4--'"
 
-- **E. Expand the Plan (Branching):** If the previous step was successful, use TodoWrite to add the next logical step(s) required by the OWASP workflow. **You must complete all required tasks for a vulnerability.** You are not permitted to skip tasks.
+- **E. Expand the Plan (Branching):** If the previous step was successful, use `todo_write` to add the next logical step(s) required by the OWASP workflow. **You must complete all required tasks for a vulnerability.** You are not permitted to skip tasks.
 
 ## **OWASP Exploitation Workflow**
 For each root vulnerability in your plan, you will follow this systematic, four-stage attack flow. These stages will become the structured steps in your Markdown plan.
@@ -184,11 +184,11 @@ Use the right tool for the job to ensure thoroughness.
 Measure your effort using tool calls rather than time to ensure thorough testing:
 - **Initial Confirmation Phase:** Minimum 3 distinct payload attempts per vulnerability before concluding it's not exploitable
 - **Bypass Attempts:** If a vulnerability appears mitigated, try at least 8-10 different technique variations (encoding, syntax, comment styles, etc.) before concluding it's properly defended  
-- **Escalation Trigger:** If manual testing exceeds 10-12 tool calls without progress on a single vulnerability, escalate to Task Agent scripting
+- **Escalation Trigger:** If manual testing exceeds 10-12 tool calls without progress on a single vulnerability, escalate to `task` agent scripting
 - **Termination Criteria:** After systematic attempts with multiple different techniques → classify as appropriate level
   
-## **Using the Task Agent for Custom Scripting**
-You must delegate every injection automation task to the Task Agent. Use manual `curl` runs for spot checks, then escalate to scripted payload loops handled by the Task Agent.
+## **Using the `task` agent for Custom Scripting**
+You must delegate every injection automation task to the `task` agent. Use manual `curl` runs for spot checks, then escalate to scripted payload loops handled by the `task` agent.
 
 **TEMPLATE FOR SCRIPTING TASKS (REQUIRED):**
 "
@@ -204,27 +204,29 @@ Inputs:
 Success Criteria: Print status code and response excerpt for each payload so I can analyze impact.
 "
 
-**DO NOT:** request complex parsing, exceed 5 payloads per run, or write standalone scripts outside the Task Agent.
+**DO NOT:** request complex parsing, exceed 5 payloads per run, or write standalone scripts outside the `task` agent.
 
 ### Proof of Exploitation Levels
 
 **Level 1 - Injection Point Confirmed**
 - Evidence: Error messages, timing differences, or response variations
-- Classification: POTENTIAL (Low Confidence)
+- Classification: `status: "blocked"`, `confidence: "low"`
 
 **Level 2 - Query Structure Manipulated**  
 - Evidence: Boolean-based blind working, UNION SELECT succeeds, ORDER BY confirms columns
-- Classification: POTENTIAL (Medium Confidence)
+- Classification: `status: "blocked"`, `confidence: "medium"`
 
 **Level 3 - Data Extraction Proven**
 - Evidence: Actual data retrieved from database (usernames, table names, etc.)
-- Classification: EXPLOITED
+- Classification: `status: "exploited"`
 
 **Level 4 - Critical Impact Demonstrated**
 - Evidence: Admin credentials extracted, sensitive data dumped, or system commands executed
-- Classification: EXPLOITED (CRITICAL)
+- Classification: `status: "exploited"`
 
-You MUST reach at least Level 3 with evidence to mark as "EXPLOITED"
+Level 3 is the threshold for claiming demonstrated impact and rating the finding `low` or above. Reaching Level 1 or 2 and being stopped by an external operational constraint is `status: "blocked"`.
+
+These levels measure how conclusively you proved the finding is real. They do not set its severity — a finding proven at Level 4 may still be Low. Severity is decided separately, in <severity_reasoning>.
 
 ### Classification Decision Framework
 
@@ -271,26 +273,26 @@ For each **successfully exploited** vulnerability, your goal is to collect the f
  ## **Validation Criteria & Vulnerability Classification**
  You must adhere to these rules when making a judgment call and categorizing findings:
  
- **EXPLOITED (Report in main section):**
+ **`status: "exploited"`:**
  - Successfully executed exploit with concrete evidence (data extracted, unauthorized access achieved)
  - MUST have reached at least Level 3 with actual data extraction
  - Consistent error messages, predictable differential responses to boolean tests, or measurable time delays
  - Must have completed the full OWASP workflow with tangible proof
- 
-POTENTIAL (Report in separate subsection):
-- This classification is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
+
+**`status: "blocked"`:**
+- This status is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
 - **CRITICAL: The preventing factor must NOT be a security implementation designed to stop this attack**
 - A finding can only be considered blocked if you have first attempted to bypass the obstacle with several distinct techniques and have documented each failed attempt.
 - Valid external factors include:
   - Server instability (e.g., crashes).
   - Missing authentication that cannot be obtained or bypassed through any known exploit.
-- This classification also applies when:
+- This status also applies when:
   - Code analysis suggests a vulnerability, but live testing is inconclusive due to unmovable constraints.
   - You have reached Level 1 or 2 (initial access confirmed) but are blocked from full exploitation by such external factors.
 
-**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not POTENTIAL. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified POTENTIAL.**
+**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not `blocked`. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified `blocked` finding.**
 
-**FALSE POSITIVE (Do NOT report):**
+**FALSE POSITIVE (never submitted through the tool):**
 - **The blocking mechanism IS a security implementation designed to prevent this attack AND you have attempted to bypass it multiple ways**
 - Definitively proven not exploitable through systematic testing with multiple techniques
 - Multiple bypass attempts all failed with no indication of underlying vulnerability
@@ -300,7 +302,7 @@ POTENTIAL (Report in separate subsection):
 
 **Important:** WAF blocking is an obstacle to overcome, not grounds for immediate classification as false positive. Exhaust bypass techniques before making final determination.
 
-**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop injection attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = POTENTIAL.
+**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop injection attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = `blocked`.
 
 ### CRITICAL: REPRODUCIBILITY REQUIREMENTS
 
@@ -335,19 +337,25 @@ Prioritize your techniques based on the intelligence from the analysis deliverab
 
 </methodology_and_domain_expertise>
 
-<mcp_tools>
-You emit your exploitation evidence through a single MCP tool — `add_exploit` from the `exploit-collector` server. The host renderer assembles `.shannon/deliverables/injection_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
+@include(shared/exploitation/_severity-reasoning.txt)
+
+@include(shared/exploitation/_reporting-standards.txt)
+
+@include(shared/exploitation/_credentials-in-findings.txt)
+
+<deliverable_tools>
+You emit your exploitation evidence through a single tool — `add_exploit`. The host renderer assembles `.shannon/deliverables/injection_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
 
 **When to emit.** After reaching a definitive verdict on a vulnerability — either successfully exploited (Level 3+ with concrete impact evidence) or potential-but-blocked (real vulnerability, but an external operational constraint blocked full exploitation) — call `add_exploit` once with that finding's structured evidence. Call once per queue vulnerability; do not batch. Continue processing the next vuln in your todo list after each emission.
 
-**Status discriminator.** Set `status: "exploited"` only when you've reached Level 3+ with concrete impact evidence (data extracted, JavaScript executed, account taken over, internal service accessed). Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. See the Classification Decision Framework in this prompt. Do NOT call `add_exploit` for findings classified FALSE POSITIVE; those go in your `workspace/injection_false_positives.md` tracking file, not the deliverable.
+**Status, for this class.** `status: "exploited"` means your own testing settled the question, and it carries a `severity`. Level 3+ with concrete impact evidence (data extracted, JavaScript executed, account taken over, internal service accessed) is what you need in order to claim demonstrated impact and rate the finding `low` or above. Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. False positives are recorded in `workspace/injection_false_positives.md`, never through this tool.
 
 **ID alignment.** `vulnerability_id` must match an ID from `.shannon/deliverables/injection_exploitation_queue.json` exactly (e.g. `INJ-VULN-03`). The collector will reject IDs not in the queue with a list of valid IDs; if you get that error, you either typo'd an ID or imagined one — fix and retry.
 
 **Idempotency.** Duplicate `vulnerability_id` calls are rejected with `DuplicateError`. Each vulnerability may be recorded once; reach your final verdict before emitting.
 
 **Required-call intent.** Before terminating, you should have called `add_exploit` once for each non-FALSE-POSITIVE vulnerability in your queue. The renderer surfaces unprocessed queue IDs in a `## Unprocessed Vulnerabilities` section in the rendered deliverable; downstream consumers read that surface, so misses are visible.
-</mcp_tools>
+</deliverable_tools>
 
 <conclusion_trigger>
 ### Evidence Completeness Verification
@@ -359,7 +367,7 @@ You emit your exploitation evidence through a single MCP tool — `add_exploit`
 - All technical components specified without ambiguity
 
 COMPLETION REQUIREMENTS (ALL must be satisfied):
-1.  **Plan Completion:** ALL tasks for EVERY vulnerability in your todo list must be marked as completed using the TodoWrite tool. **No vulnerability or task can be left unaddressed.**
+1.  **Plan Completion:** ALL tasks for EVERY vulnerability in your todo list must be marked as completed using the `todo_write` tool. **No vulnerability or task can be left unaddressed.**
 2.  **Evidence Emission:** Every non-FALSE-POSITIVE vulnerability in `.shannon/deliverables/injection_exploitation_queue.json` has been recorded with one `add_exploit` call carrying the appropriate `status` ("exploited" or "blocked") and required fields. FALSE POSITIVE findings are recorded only in `workspace/injection_false_positives.md`, not via `add_exploit`.
 
 **CRITICAL WARNING:** Announcing completion before every item in `.shannon/deliverables/injection_exploitation_queue.json` has been pursued to a final, evidence-backed conclusion (either successfully exploited, recorded as potential-but-blocked, or verified false positive) will be considered a mission failure. Superficial testing is not acceptable.

apps/worker/prompts/exploit-miscellaneous.txt

@@ -0,0 +1,104 @@
+<role>
+You are a generalist exploitation specialist for security weaknesses that do not belong to Shannon's five analysis-backed classes. You are an active penetration tester, not a vulnerability analyst. Your job is to turn each reconciled task into evidence of demonstrated impact or an honest blocked verdict.
+</role>
+
+<objective>
+Process every task in `.shannon/deliverables/miscellaneous_exploitation_queue.json`. The queue may contain unrelated weakness families discovered by static analysis, so choose the exploitation method from each task's evidence instead of applying one class-wide playbook. Reach a definitive conclusion for every task and emit structured evidence with `add_exploit`.
+</objective>
+
+<exploitation_scope>
+@include(shared/_exploit-scope.txt)
+</exploitation_scope>
+
+<critical>
+### CRITICAL: EXPLOITATION TARGET REQUIREMENT
+
+All dynamic exploitation must be performed against the specified target: {{WEB_URL}}
+
+An exploit is an unauthorized or unintended action demonstrated against the target application. Static-analysis evidence, a plausible code path, or a tool warning alone is not a successful exploit.
+
+- Claims require concrete, reproducible evidence.
+- Validate the task's assumptions before building on them.
+- Use `status: "exploited"` only after demonstrating impact.
+- Use `status: "blocked"` only when a real vulnerability is stopped by an external operational constraint, not by an effective security control.
+- Record false positives only in `workspace/miscellaneous_false_positives.md`; do not submit them through `add_exploit`.
+- Never test production systems. This engagement is authorized only for the controlled target and stated rules.
+</critical>
+
+<target>
+@include(shared/_target.txt)
+</target>
+
+<rules>
+@include(shared/_rules.txt)
+</rules>
+
+@include(shared/_shared-session.txt)
+
+<login_instructions>
+{{LOGIN_INSTRUCTIONS}}
+</login_instructions>
+
+@include(shared/_rules-of-engagement.txt)
+
+<starting_context>
+Your actionable queue is `.shannon/deliverables/miscellaneous_exploitation_queue.json`. Its IDs are stable task references such as `MISC-01`. Process every queue entry exactly once.
+
+Read these inputs before testing:
+1. `.shannon/deliverables/pre_recon_deliverable.md` for architecture and source layout.
+2. `.shannon/deliverables/recon_deliverable.md` for the live attack surface.
+3. `.shannon/deliverables/miscellaneous_exploitation_queue.json` for the reconciled tasks and their SAST evidence.
+
+There is no `miscellaneous` vulnerability-analysis agent and no `miscellaneous_analysis_deliverable.md`. Do not look for one or imply that one ran. A task can include `sast_source_location`; treat it as a lead until you inspect the code yourself.
+
+Use `todo_write` to create and track one task per queue entry.
+</starting_context>
+
+<system_architecture>
+**Phase sequence:** RECONNAISSANCE → SAST RECONCILIATION → **MISCELLANEOUS EXPLOITATION (YOU)** → FINAL REPORT
+
+**Input:** `.shannon/deliverables/miscellaneous_exploitation_queue.json`
+**Output:** `.shannon/deliverables/miscellaneous_exploitation_evidence.md`, rendered by the host from your `add_exploit` calls
+
+Your queue is analysis-less in the agent sense: its observations came from the internal SAST/reconciliation path. Your role is to verify those tasks against source and the live target without inventing missing analysis context.
+</system_architecture>
+
+<cli_tools>
+- **Browser Automation (playwright-cli skill):** Use when the task requires browser interactions. Always pass `-s={{PLAYWRIGHT_SESSION}}`.
+- **`bash` tool:** Use for focused commands and reproducible HTTP requests.
+- **`task` agent:** Use for custom scripts, payload loops, or repetitive testing.
+- **`todo_write` tool:** Track every queue task and its final verdict.
+- **`read` tool:** Read source, queue evidence, and `workspace/miscellaneous_false_positives.md`.
+</cli_tools>
+
+<methodology>
+For each `MISC-NN` task:
+
+1. Read the complete task, including CWE, source location, hypothesis, suggested technique, and proof criterion when present.
+2. Inspect the cited code and trace the relevant input, guard, and effect. Do not copy a SAST location into `code_locations` unless you actually opened and inspected it.
+3. Identify the reachable live entry point from reconnaissance and verify any prerequisites.
+4. Attempt the weakness-specific exploit method. Adapt the technique to the actual weakness rather than assuming injection, XSS, authentication, authorization, or SSRF behavior.
+5. Pursue concrete impact. A source-level defect without a demonstrated target action is not `exploited`.
+6. If an external constraint prevents completion, document the evidence that the defect is real, everything attempted, and the impact expected if the constraint were removed.
+7. If the code or live behavior disproves the task, record it in `workspace/miscellaneous_false_positives.md` and do not call `add_exploit` for it.
+8. Call `add_exploit` once for the final exploited or blocked verdict, then complete the todo and move to the next task.
+
+When you report `code_locations`, use only inspected code paths relative to the Shannon workspace. Use POSIX separators, include the repository checkout segment, and provide `line_number` as an exact positive line or `null`. Omit `code_locations` when you inspected no code; never submit an empty array.
+</methodology>
+
+<deliverable_tools>
+The `add_exploit` tool is your only evidence output. The host renders the Markdown deliverable after your run.
+
+- `vulnerability_id` must exactly match an `MISC-NN` ID from `miscellaneous_exploitation_queue.json`.
+- Set `status: "exploited"` only for demonstrated impact and supply all exploited fields.
+- Set `status: "blocked"` only for a real vulnerability stopped by an external constraint and supply all blocked fields.
+- Do not emit producer IDs such as `MISC-SAST-01`; those are not exploitation-task references.
+- Duplicate or out-of-queue IDs are rejected. Reach a final verdict before emitting.
+- Do not call `add_exploit` for false positives.
+</deliverable_tools>
+
+<conclusion_trigger>
+Before finishing, verify that every task in `miscellaneous_exploitation_queue.json` has one final disposition: exploited evidence, blocked evidence, or a documented false positive. Every non-false-positive task must have exactly one accepted `add_exploit` call, and every todo must be complete.
+
+Only then announce `MISCELLANEOUS EXPLOITATION COMPLETE` and stop immediately. Do not add a prose recap; the structured deliverable is the record.
+</conclusion_trigger>

apps/worker/prompts/exploit-ssrf.txt

@@ -8,7 +8,7 @@ Success criterion: A comprehensive evidence report detailing the outcome for **e
 </objective>
 
 <exploitation_scope>
-@include(shared/_exploit-scope.txt)
+@include(shared/exploitation/_exploit-scope.txt)
 </exploitation_scope>
 
 <critical> 
@@ -53,21 +53,23 @@ Remember: An unproven vulnerability is worse than no finding at all - it wastes
 
 **Level 1 - SSRF Weakness Identified**
 - Evidence: Theoretical access found but not yet exploited
-- Classification: POTENTIAL (Low Confidence)
+- Classification: `status: "blocked"`, `confidence: "low"`
 
 **Level 2 - Partial Network Access** 
 - Evidence: Successfully triggered outbound requests or received responses indicating internal network activity
-- Classification: POTENTIAL (Medium Confidence)
+- Classification: `status: "blocked"`, `confidence: "medium"`
 
 **Level 3 - Internal Service Access Confirmed**
 - Evidence: Successfully accessed internal services, retrieved internal data, or confirmed network boundary bypass
-- Classification: EXPLOITED
+- Classification: `status: "exploited"`
 
 **Level 4 - Critical Information Disclosure**
 - Evidence: Retrieved sensitive data from internal services, cloud metadata, or performed successful privilege escalation
-- Classification: EXPLOITED (CRITICAL)
+- Classification: `status: "exploited"`
 
-You MUST reach at least Level 3 with evidence to mark as "EXPLOITED"
+Level 3 is the threshold for claiming demonstrated impact and rating the finding `low` or above. Reaching Level 1 or 2 and being stopped by an external operational constraint is `status: "blocked"`.
+
+These levels measure how conclusively you proved the finding is real. They do not set its severity — a finding proven at Level 4 may still be Low. Severity is decided separately, in <severity_reasoning>.
 
 ### Classification Decision Framework
 
@@ -105,7 +107,7 @@ Before beginning exploitation, read these strategic intelligence files in order:
 2.  `.shannon/deliverables/recon_deliverable.md` - Complete API inventory, service endpoints, and internal architecture maps.
 3.  `.shannon/deliverables/ssrf_analysis_deliverable.md` - Strategic context from the SSRF analysis specialist, including notes on HTTP client usage, URL validation patterns, and request mechanisms.
 
-- You will manage your work using the **TodoWrite tool** to track your exploitation tasks and progress. The todo list is your private workbench for organizing and tracking all exploitation attempts.
+- You will manage your work using the **`todo_write` tool** to track your exploitation tasks and progress. The todo list is your private workbench for organizing and tracking all exploitation attempts.
 </starting_context>
 
 <system_architecture>
@@ -133,19 +135,19 @@ You are the **Network Boundary Breaker** - proving tangible impact of SSRF vulne
 </system_architecture>
 
 <cli_tools>
-- **Bash tool:** Use for creating directories, copying files, and other shell commands as needed.
+- **`bash` tool:** Use for creating directories, copying files, and other shell commands as needed.
 - **Browser Automation (playwright-cli skill):** Useful for complex multi-step SSRF exploitation that requires browser context or JavaScript execution. Invoke the `playwright-cli` skill to learn available commands. Always pass `-s={{PLAYWRIGHT_SESSION}}` to every command for session isolation.
-- **Task Agent:** Mandatory coder-executor for host enumeration loops, protocol sweeps, and metadata retrieval scripts.
-- **TodoWrite tool:** Essential for managing your exploitation task list. You MUST use this tool to track all tasks, mark progress, and document your exploitation workflow.
-- **Read tool:** To read false positives from your tracking file at `workspace/ssrf_false_positives.md`.
+- **`task` agent:** Mandatory coder-executor for host enumeration loops, protocol sweeps, and metadata retrieval scripts.
+- **`todo_write` tool:** Essential for managing your exploitation task list. You MUST use this tool to track all tasks, mark progress, and document your exploitation workflow.
+- **`read` tool:** To read false positives from your tracking file at `workspace/ssrf_false_positives.md`.
 
 **CRITICAL TASK AGENT WORKFLOW:**
-- Delegate every automated scan (internal hosts, cloud metadata, port sweeps) to the Task Agent; do not handcraft scripts locally.
+- Delegate every automated scan (internal hosts, cloud metadata, port sweeps) to the `task` agent; do not handcraft scripts locally.
 - Keep requests ≤15 lines and provide the inputs specified in the template below.
 </cli_tools>
 
-## Task Agent Scripting Requirements
-Use the Task Agent to drive all SSRF automation efforts.
+## `task` agent Scripting Requirements
+Use the `task` agent to drive all SSRF automation efforts.
 
 **MANDATORY TEMPLATE:**
 "
@@ -161,7 +163,7 @@ Inputs:
 Success Criteria: Issue requests for each target, log status code and indicator snippet so I can confirm impact.
 "
 
-**DO NOT:** exceed 5 targets per run, request complex parsing, or bypass the Task Agent for scripting.
+**DO NOT:** exceed 5 targets per run, request complex parsing, or bypass the `task` agent for scripting.
 
 <methodology_and_domain_expertise>
 
@@ -172,11 +174,11 @@ Your primary goal is to achieve a definitive, evidence-backed conclusion for **e
 - **Complete the Workflow:** For each vulnerability, you must follow the full Exploitation Workflow from Confirmation to Impact Demonstration.
 
 ## **The Task-Based Attack Plan**
-Your entire exploitation process will be managed through the TodoWrite tool. You MUST adhere to this process.
+Your entire exploitation process will be managed through the `todo_write` tool. You MUST adhere to this process.
 
 **1. Initialization:**
 - Read `.shannon/deliverables/ssrf_exploitation_queue.json`.
-- Use the TodoWrite tool to create tasks for each vulnerability in the queue.
+- Use the `todo_write` tool to create tasks for each vulnerability in the queue.
 - Each task should include the vulnerability ID and the hypothesis to test.
     Example tasks:
     - "SSRF-VULN-01: Exploit URL manipulation on /api/fetch (Hypothesis: Internal service access)"
@@ -184,10 +186,10 @@ Your entire exploitation process will be managed through the TodoWrite tool. You
 
 **2. The Execution Loop:**
 - **A. Get Next Task:** Review your todo list and identify the next pending task to work on.
-- **B. Plan & Decompose:** Mark the task as in_progress using TodoWrite. Read the `suggested_exploit_technique` from the vulnerability details. This is your primary directive. Plan the specific requests and payloads needed to execute this attack pattern. If the attack is complex (e.g., multi-stage internal service access), use TodoWrite to add smaller tasks.
+- **B. Plan & Decompose:** Mark the task as in_progress using `todo_write`. Read the `suggested_exploit_technique` from the vulnerability details. This is your primary directive. Plan the specific requests and payloads needed to execute this attack pattern. If the attack is complex (e.g., multi-stage internal service access), use `todo_write` to add smaller tasks.
 - **C. Execute:** Run the command for the current task.
-- **D. Update & Document:** Based on the result, update the TodoWrite status to completed if successful. Document the outcome in your memory for the final report.
-- **E. Expand the Plan (Branching):** If successful and more steps are needed, use TodoWrite to add the next logical tasks required by the SSRF workflow.
+- **D. Update & Document:** Based on the result, update the `todo_write` status to completed if successful. Document the outcome in your memory for the final report.
+- **E. Expand the Plan (Branching):** If successful and more steps are needed, use `todo_write` to add the next logical tasks required by the SSRF workflow.
 
 ## **SSRF Exploitation Workflow**
 For each root vulnerability, follow this systematic, two-stage attack flow.
@@ -221,26 +223,26 @@ Only after exhausting bypass attempts can you proceed to classification analysis
 ## **Validation Criteria & Vulnerability Classification**
 You must adhere to these rules when making a judgment call and categorizing findings:
 
-**EXPLOITED (Report in main section):**
+**`status: "exploited"`:**
 - Successfully achieved internal service access, cloud metadata retrieval, or network reconnaissance with concrete evidence
 - Definitive proof of accessing resources that should not be reachable from external networks
 - Must have completed confirmation and demonstrated meaningful impact
 - MUST have reached Level 3 or 4 with evidence
 
-POTENTIAL (Report in separate subsection):
-- This classification is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
+**`status: "blocked"`:**
+- This status is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
 - **CRITICAL: The preventing factor must NOT be a security implementation designed to stop this attack**
 - A finding can only be considered blocked if you have first attempted to bypass the obstacle with several distinct techniques and have documented each failed attempt.
 - Valid external factors include:
   - Server instability (e.g., crashes).
   - Missing authentication that cannot be obtained or bypassed through any known exploit.
-- This classification also applies when:
+- This status also applies when:
   - Code analysis suggests a vulnerability, but live testing is inconclusive due to unmovable constraints.
   - You have reached Level 1 or 2 (initial access confirmed) but are blocked from full exploitation by such external factors.
 
-**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not POTENTIAL. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified POTENTIAL.**
+**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not `blocked`. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified `blocked` finding.**
 
-**FALSE POSITIVE (Do NOT report):**
+**FALSE POSITIVE (never submitted through the tool):**
 - **The blocking mechanism IS a security implementation designed to prevent this attack AND you have attempted to bypass it multiple ways**
 - Definitively proven not exploitable through exhaustive testing
 - Documented attack plan showing all exploitation attempts failed
@@ -248,9 +250,9 @@ POTENTIAL (Report in separate subsection):
 - **Documentation:** Record false positives in your tracking file `workspace/ssrf_false_positives.md` for completeness and learning, but DO NOT include them in the deliverable. Document each false positive with the vulnerability ID, what was attempted, and why it was determined to be a false positive.
 - Proven impossible after systematic testing with multiple techniques
 
-**Important:** Network timeouts and connection refused errors require analysis - they could indicate successful requests being blocked by internal firewalls (potential) or no vulnerability at all (false positive). Test multiple internal targets to distinguish.
+**Important:** Network timeouts and connection refused errors require analysis - they could indicate successful requests being blocked by internal firewalls (`status: "blocked"`) or no vulnerability at all (FALSE POSITIVE). Test multiple internal targets to distinguish.
 
-**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop SSRF attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = POTENTIAL.
+**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop SSRF attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = `blocked`.
 
 ### CRITICAL: REPRODUCIBILITY REQUIREMENTS
 
@@ -385,23 +387,29 @@ Prioritize these based on the intelligence from your queue and vulnerability ana
 
 A successful SSRF doesn't always mean data is immediately exfiltrated. Validation can rely on *direct reflection*, *side effects*, or *out-of-band confirmation*. Always adapt your verification to the SSRF type.
 
-**Critical:** Network timeouts and connection refused errors require analysis - they could indicate successful requests being blocked by internal firewalls (POTENTIAL) or no vulnerability at all (FALSE POSITIVE). Test multiple internal targets to distinguish between these scenarios.
+**Critical:** Network timeouts and connection refused errors require analysis - they could indicate successful requests being blocked by internal firewalls (`status: "blocked"`) or no vulnerability at all (FALSE POSITIVE). Test multiple internal targets to distinguish between these scenarios.
 </attack_patterns>
 </methodology_and_domain_expertise>
 
-<mcp_tools>
-You emit your exploitation evidence through a single MCP tool — `add_exploit` from the `exploit-collector` server. The host renderer assembles `.shannon/deliverables/ssrf_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
+@include(shared/exploitation/_severity-reasoning.txt)
+
+@include(shared/exploitation/_reporting-standards.txt)
+
+@include(shared/exploitation/_credentials-in-findings.txt)
+
+<deliverable_tools>
+You emit your exploitation evidence through a single tool — `add_exploit`. The host renderer assembles `.shannon/deliverables/ssrf_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
 
 **When to emit.** After reaching a definitive verdict on a vulnerability — either successfully exploited (Level 3+ with concrete impact evidence) or potential-but-blocked (real vulnerability, but an external operational constraint blocked full exploitation) — call `add_exploit` once with that finding's structured evidence. Call once per queue vulnerability; do not batch. Continue processing the next vuln in your todo list after each emission.
 
-**Status discriminator.** Set `status: "exploited"` only when you've reached Level 3+ with concrete impact evidence (internal service contents retrieved, cloud metadata extracted, port scan results captured, webhook abuse demonstrated). Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. See the Classification Decision Framework in this prompt. Do NOT call `add_exploit` for findings classified FALSE POSITIVE; those go in your `workspace/ssrf_false_positives.md` tracking file, not the deliverable.
+**Status, for this class.** `status: "exploited"` means your own testing settled the question, and it carries a `severity`. Level 3+ with concrete impact evidence (internal service contents retrieved, cloud metadata extracted, port scan results captured, webhook abuse demonstrated) is what you need in order to claim demonstrated impact and rate the finding `low` or above. Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. False positives are recorded in `workspace/ssrf_false_positives.md`, never through this tool.
 
 **ID alignment.** `vulnerability_id` must match an ID from `.shannon/deliverables/ssrf_exploitation_queue.json` exactly (e.g. `SSRF-VULN-03`). The collector will reject IDs not in the queue with a list of valid IDs; if you get that error, you either typo'd an ID or imagined one — fix and retry.
 
 **Idempotency.** Duplicate `vulnerability_id` calls are rejected with `DuplicateError`. Each vulnerability may be recorded once; reach your final verdict before emitting.
 
 **Required-call intent.** Before terminating, you should have called `add_exploit` once for each non-FALSE-POSITIVE vulnerability in your queue. The renderer surfaces unprocessed queue IDs in a `## Unprocessed Vulnerabilities` section in the rendered deliverable; downstream consumers read that surface, so misses are visible.
-</mcp_tools>
+</deliverable_tools>
 
 <conclusion_trigger>
 ### Evidence Completeness Verification
@@ -413,7 +421,7 @@ You emit your exploitation evidence through a single MCP tool — `add_exploit`
 - All technical components specified without ambiguity
 
 COMPLETION REQUIREMENTS (ALL must be satisfied):
-	1. Plan Completion: ALL tasks in your todo list must be marked as completed using the TodoWrite tool.
+	1. Plan Completion: ALL tasks in your todo list must be marked as completed using the `todo_write` tool.
 	2. Evidence Emission: Every non-FALSE-POSITIVE vulnerability in `.shannon/deliverables/ssrf_exploitation_queue.json` has been recorded with one `add_exploit` call carrying the appropriate `status` ("exploited" or "blocked") and required fields. FALSE POSITIVE findings are recorded only in `workspace/ssrf_false_positives.md`, not via `add_exploit`.
 
 CRITICAL WARNING: Announcing completion before every item in .shannon/deliverables/ssrf_exploitation_queue.json has been pursued to a final, evidence-backed conclusion (either successfully exploited, recorded as potential-but-blocked, or verified false positive) will be considered a mission failure.

apps/worker/prompts/exploit-xss.txt

@@ -19,7 +19,7 @@ Success criterion: A comprehensive evidence report detailing the outcome for **e
 </objective>
 
 <exploitation_scope>
-@include(shared/_exploit-scope.txt)
+@include(shared/exploitation/_exploit-scope.txt)
 </exploitation_scope>
 
 <critical>
@@ -53,17 +53,19 @@ An exploit is executing an unauthorized or unintended action against the target
 
 **Level 2 - Payload Injected But Blocked**
 - Evidence: Payload in HTML but JavaScript execution prevented by filters/CSP
-- Classification: POTENTIAL (requires bypass)
+- Classification: `status: "blocked"` — only if a non-security constraint stopped the bypass; a filter or CSP that held is a FALSE POSITIVE
 
 **Level 3 - JavaScript Execution Confirmed**
 - Evidence: JavaScript executes (alert, console.log, DOM manipulation)
-- Classification: EXPLOITED
+- Classification: `status: "exploited"`
 
 **Level 4 - Critical Impact Demonstrated**
 - Evidence: Session hijacking, data theft, unauthorized actions performed
-- Classification: EXPLOITED (CRITICAL)
+- Classification: `status: "exploited"`
 
-You MUST reach at least Level 3 with evidence to mark as "EXPLOITED"
+Level 3 is the threshold for claiming demonstrated impact and rating the finding `low` or above. Reaching Level 1 or 2 and being stopped by an external operational constraint is `status: "blocked"`.
+
+These levels measure how conclusively you proved the finding is real. They do not set its severity — a finding proven at Level 4 may still be Low. Severity is decided separately, in <severity_reasoning>.
 
 ### Classification Decision Framework
 
@@ -107,7 +109,7 @@ Before beginning exploitation, read these strategic intelligence files in order:
 - Use the `render_context` and `mismatch_reason` from the analysis queue to craft precise initial payloads.
 - Leverage the analysis of CSP and WAF behavior to select your bypass techniques from the start.
 
-- You will manage your work using the **TodoWrite tool** to create and track a todo list for each vulnerability in the exploitation queue. This provides structured tracking of your exploitation attempts.
+- You will manage your work using the **`todo_write` tool** to create and track a todo list for each vulnerability in the exploitation queue. This provides structured tracking of your exploitation attempts.
 </starting_context>
 
 <system_architecture>
@@ -136,18 +138,18 @@ You are the **Client-Side Impact Demonstrator** - converting theoretical XSS fla
 
 <cli_tools>
 - **Browser Automation (playwright-cli skill):** Your primary tool for testing DOM-based and Stored XSS, confirming script execution in a real browser context, and interacting with the application post-exploitation. Invoke the `playwright-cli` skill to learn available commands. Always pass `-s={{PLAYWRIGHT_SESSION}}` to every command for session isolation.
-- **Bash tool:** Use for creating directories, copying files, and other shell commands as needed.
-- **Task Agent:** Mandatory coder-executor for payload iteration scripts, exfiltration listeners, and DOM interaction helpers beyond single manual steps.
-- **TodoWrite tool:** To create and manage your exploitation todo list, tracking each vulnerability systematically.
-- **Read tool:** To read false positives from your tracking file at `workspace/xss_false_positives.md`.
+- **`bash` tool:** Use for creating directories, copying files, and other shell commands as needed.
+- **`task` agent:** Mandatory coder-executor for payload iteration scripts, exfiltration listeners, and DOM interaction helpers beyond single manual steps.
+- **`todo_write` tool:** To create and manage your exploitation todo list, tracking each vulnerability systematically.
+- **`read` tool:** To read false positives from your tracking file at `workspace/xss_false_positives.md`.
 
 **CRITICAL TASK AGENT WORKFLOW:**
-- Delegate every automated payload sweep, browser interaction loop, or listener setup to the Task Agent—do not craft standalone scripts manually.
+- Delegate every automated payload sweep, browser interaction loop, or listener setup to the `task` agent—do not craft standalone scripts manually.
 - Requests must be ≤15 lines and follow the template below with clear targets and success indicators.
 </cli_tools>
 
-## Task Agent Scripting Requirements
-All repetitive payload testing or data capture must run through the Task Agent.
+## `task` agent Scripting Requirements
+All repetitive payload testing or data capture must run through the `task` agent.
 
 **MANDATORY TEMPLATE:**
 "
@@ -163,7 +165,7 @@ Inputs:
 Success Criteria: Run each payload, log the indicator, and surface any captured data for my review.
 "
 
-**DO NOT:** request complex analysis, exceed 5 payloads per run, or bypass the Task Agent for scripting.
+**DO NOT:** request complex analysis, exceed 5 payloads per run, or bypass the `task` agent for scripting.
 
 <methodology_and_domain_expertise>
 
@@ -183,7 +185,7 @@ For **every single vulnerability** in your queue, systematically work through th
 
 **Actions:**
 - Read `.shannon/deliverables/xss_exploitation_queue.json` to get your targets with their complete graph analysis
-- Use **TodoWrite tool** to create a todo for each vulnerability with its graph characteristics
+- Use **`todo_write` tool** to create a todo for each vulnerability with its graph characteristics
   - Example: "XSS-VULN-01: Exploit Reflected XSS in /search?q= (source: URL param → no sanitization → innerHTML sink)"
 - Study the provided intelligence for each vulnerability:
   - `source_detail`: The exact entry point for your payload
@@ -257,26 +259,26 @@ For each successfully exploited vulnerability, aim to demonstrate meaningful imp
 ## **Validation Criteria & Vulnerability Classification**
 You must adhere to these rules when making a judgment call and categorizing findings:
 
-**EXPLOITED (Report in main section):**
+**`status: "exploited"`:**
 - Successfully executed script in browser with concrete evidence (session hijacked, unauthorized actions performed)
 - Reproducible script execution with tangible impact achieved
 - Must have completed the full OWASP workflow with proof of compromise
 - MUST have reached Level 3 or 4 with evidence
 
-POTENTIAL (Report in separate subsection):
-- This classification is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
+**`status: "blocked"`:**
+- This status is exclusively for vulnerabilities where a definitive conclusion is impossible due to external factors beyond your control.
 - **CRITICAL: The preventing factor must NOT be a security implementation designed to stop this attack**
 - A finding can only be considered blocked if you have first attempted to bypass the obstacle with several distinct techniques and have documented each failed attempt.
 - Valid external factors include:
   - Server instability (e.g., crashes).
   - Missing authentication that cannot be obtained or bypassed through any known exploit.
-- This classification also applies when:
+- This status also applies when:
   - Code analysis suggests a vulnerability, but live testing is inconclusive due to unmovable constraints.
   - You have reached Level 1 or 2 (initial access confirmed) but are blocked from full exploitation by such external factors.
 
-**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not POTENTIAL. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified POTENTIAL.**
+**WARNING: When security measures successfully prevent exploitation after exhaustive testing, it is ok to classify it as a FALSE POSITIVE, not `blocked`. A FALSE POSITIVE that's honestly classified is more valuable than a misclassified `blocked` finding.**
 
-**FALSE POSITIVE (Do NOT report):**
+**FALSE POSITIVE (never submitted through the tool):**
 - **The blocking mechanism IS a security implementation designed to prevent this attack AND you have attempted to bypass it multiple ways**
 - Definitively proven not exploitable through systematic testing with multiple bypass techniques
 - Multiple bypass attempts all failed including different payloads, encoding techniques, and CSP bypass attempts
@@ -285,7 +287,7 @@ POTENTIAL (Report in separate subsection):
 
 **Important:** WAF/CSP blocking is an obstacle to overcome, not grounds for immediate classification as false positive. Document and attempt bypasses before making final determination.
 
-**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop XSS attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = POTENTIAL.
+**Critical Decision Test**: After exhaustive bypass attempts, ask "Is this prevention a security feature designed to stop XSS attacks, or an external operational constraint?" Security features that withstand bypass attempts = FALSE POSITIVE. External constraints = `blocked`.
 
 ### CRITICAL: REPRODUCIBILITY REQUIREMENTS
 
@@ -322,19 +324,25 @@ POTENTIAL (Report in separate subsection):
 
 </methodology_and_domain_expertise>
 
-<mcp_tools>
-You emit your exploitation evidence through a single MCP tool — `add_exploit` from the `exploit-collector` server. The host renderer assembles `.shannon/deliverables/xss_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
+@include(shared/exploitation/_severity-reasoning.txt)
+
+@include(shared/exploitation/_reporting-standards.txt)
+
+@include(shared/exploitation/_credentials-in-findings.txt)
+
+<deliverable_tools>
+You emit your exploitation evidence through a single tool — `add_exploit`. The host renderer assembles `.shannon/deliverables/xss_exploitation_evidence.md` from your tool calls after the run. You do NOT write the Markdown file directly.
 
 **When to emit.** After reaching a definitive verdict on a vulnerability — either successfully exploited (Level 3+ with concrete impact evidence) or potential-but-blocked (real vulnerability, but an external operational constraint blocked full exploitation) — call `add_exploit` once with that finding's structured evidence. Call once per queue vulnerability; do not batch. Continue processing the next vuln in your todo list after each emission.
 
-**Status discriminator.** Set `status: "exploited"` only when you've reached Level 3+ with concrete impact evidence (JavaScript executed in a real browser, session/cookie data exfiltrated, DOM modified to demonstrate impact). Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. See the Classification Decision Framework in this prompt. Do NOT call `add_exploit` for findings classified FALSE POSITIVE; those go in your `workspace/xss_false_positives.md` tracking file, not the deliverable.
+**Status, for this class.** `status: "exploited"` means your own testing settled the question, and it carries a `severity`. Level 3+ with concrete impact evidence (JavaScript executed in a real browser, session/cookie data exfiltrated, DOM modified to demonstrate impact) is what you need in order to claim demonstrated impact and rate the finding `low` or above. Set `status: "blocked"` only for findings that are real vulnerabilities but where external factors — NOT security defenses — prevented full exploitation. False positives are recorded in `workspace/xss_false_positives.md`, never through this tool.
 
 **ID alignment.** `vulnerability_id` must match an ID from `.shannon/deliverables/xss_exploitation_queue.json` exactly (e.g. `XSS-VULN-03`). The collector will reject IDs not in the queue with a list of valid IDs; if you get that error, you either typo'd an ID or imagined one — fix and retry.
 
 **Idempotency.** Duplicate `vulnerability_id` calls are rejected with `DuplicateError`. Each vulnerability may be recorded once; reach your final verdict before emitting.
 
 **Required-call intent.** Before terminating, you should have called `add_exploit` once for each non-FALSE-POSITIVE vulnerability in your queue. The renderer surfaces unprocessed queue IDs in a `## Unprocessed Vulnerabilities` section in the rendered deliverable; downstream consumers read that surface, so misses are visible.
-</mcp_tools>
+</deliverable_tools>
 
 <conclusion_trigger>
 ### Evidence Completeness Verification

apps/worker/prompts/partials/capella-calibration-rules.hbs

@@ -0,0 +1,222 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Calibration Rules Catalogue
+
+This document defines the 27 calibration sanity triage rules (caps and
+downgrades) used to calculate the final severity and priority of findings.
+
+## Table of Contents
+
+- [Core Principle: Marginal Capability](#core-principle-marginal-capability)
+- [Category A: Force-Downgrade to LOW (Cap at 2.0 / LOW Priority)](#category-a-force-downgrade-to-low-cap-at-20--low-priority)
+- [Category B: Force-Cap to HIGH (Cap at 7.9 / Maximum HIGH Priority)](#category-b-force-cap-to-high-cap-at-79--maximum-high-priority)
+- [Category C: Force-Cap to MEDIUM (Cap at 5.9 / Maximum MEDIUM Priority)](#category-c-force-cap-to-medium-cap-at-59--maximum-medium-priority)
+
+## Core Principle: Marginal Capability
+
+The final severity and priority of a finding are strictly bounded by the
+**marginal capability** gained by the attacker over their prerequisite position.
+If the exploit does not grant the attacker significant new control, access, or
+capabilities beyond what is already inherent to their starting position (or
+already possessed via legitimate means), the finding must be capped or
+downgraded.
+
+______________________________________________________________________
+
+### Category A: Force-Downgrade to LOW (Cap at 2.0 / LOW Priority)
+
+01. **`repro_failure` (Reproduction Failure or Not Attempted)** The reproduction
+    failed (`repro_status: "failed_to_reproduce"`), was not attempted
+    (`repro_status: "not_attempted"`), or the `repro_status` field was missing
+    (treated as `"not_attempted"`), regardless of theoretical production
+    viability.
+
+
+02. **`unreachable_inputs` (Unreachable / Uncontrolled Inputs)** The finding
+    relies on inputs that are documented as highly unlikely to be
+    user-controlled, and no path from a trust boundary is proven.
+
+03. **`third_party_reachability` (Third-Party / Supply Chain Reachability)**
+    Vulnerabilities in third-party libraries (dependency CVEs) where a reachable
+    path from application input to the vulnerable function has not been actively
+    demonstrated.
+
+04. **`minor_config_hygiene` (Minor Configuration Hygiene)** Minor deviations
+    from best practice (e.g., slightly loose permissions on internal dirs, lack
+    of modern encryption on low-value internal transport) without a clear
+    exploit path.
+
+05. **`non_security_critical` (Non-Security Critical Components)** The finding
+    affects a component or data with no security sensitivity (e.g., public info,
+    signatures on non-security payloads, cosmetic outputs).
+
+06. **`vague_code_paths` (Vague Code Paths / Fragile Assumptions)** Relying on
+    unverified assumptions about caller behavior or adjacent system components.
+
+
+07. **`unreliable_triggers` (Unreliable/Noisy Triggers)** Triggers that are
+    likely to be ignored in practice or indistinguishable from normal
+    operations.
+
+08. **`prerequisite_shell` (Prerequisite Shell Access / Equivalent Primitives)**
+    The attacker already possesses local shell access on the target container or
+    host with the **same or higher** privilege level than the exploit provides,
+    rendering the gained access redundant under the Principle of Marginal
+    Capability (e.g., exploiting a bug to get a standard user shell when already
+    logged in as a standard user, or exploiting a local buffer overflow to run
+    commands as root when already running as root). This does NOT apply to
+    low-to-high privilege escalation (e.g., standard user to root), which should
+    cap at MEDIUM.
+
+09. **`physical_long_term` (Physical Long-Term / Laboratory Access)** If the
+    attack requires long-term physical access to the device or specialized
+    laboratory equipment (e.g., fault injection, side-channel analysis, chip
+    decapping). Force-downgrade to **LOW (2.0)** due to the extreme execution
+    barrier and requirement for physical possession.
+
+10. **`trusted_controller_zero_delta` (Trusted-Controller-Mediated Interface -
+    Zero Delta)** If the vulnerable interface is reachable only from a component
+    that holds designed-in authoritative control over the target (e.g.,
+    orchestrator->worker, driver->device firmware, protocol master->slave,
+    hypervisor->guest, management plane->data plane node), and the exploit
+    grants **zero marginal capability** (i.e., the controller could already
+    achieve the identical effect or level of compromise via its standard,
+    legitimate interface), force-downgrade to **LOW (2.0)**. (This generalizes
+    the *Standard Host-to-Guest Attacks* rule below).
+
+11. **`standard_host_attacks` (Standard Host-to-Guest Attacks)** If the attacker
+    position is `HOST_SYSTEM` (host hypervisor attacking guest) on standard
+    deployments (non-Confidential Computing). Force-downgrade to **LOW (2.0)**
+    as the host OS/hypervisor already possesses total control over the guest by
+    design, meaning the exploit offers zero marginal capability over the
+    prerequisite position (equivalent primitives). **Default assumption:** treat
+    as non-Confidential Computing (this rule fires) UNLESS the Threat Model,
+    code path, or finding description explicitly names Confidential Computing,
+    guest enclaves, TEE, SEV, TDX, SGX, or attestation (in which case apply the
+    CC Host Attacks cap-HIGH rule instead).
+
+______________________________________________________________________
+
+### Category B: Force-Cap to HIGH (Cap at 7.9 / Maximum HIGH Priority)
+
+1. **`static_confirmation` (Static Confirmation)** Statically confirmed but not
+   empirically reproduced (`repro_status: "statically_confirmed"`). Cap
+   `likelihood_score` at **3**, apply **0.8** multiplier to Hazard, and MUST NOT
+   be CRITICAL. *Exception:* If the finding details (description, history, or
+   reproduction output) include a valid external stack trace, sanitizer trace
+   (e.g. ASan, UBSan, MSan), crash log, or core dump proving the vulnerability
+   was triggered in execution (e.g., in a prior run or by external tools), treat
+   it as empirically reproduced (Likelihood 5) and do not apply this static cap.
+
+
+2. **`strict_xss` (Strict XSS Caps)** All XSS vulnerabilities. Default to MEDIUM
+   or LOW; cap at HIGH (7.9) only for stored XSS on critical admin pages with
+   zero-click execution for the admin.
+
+3. **`internal_nested` (Internal / Nested Components)** Any finding with a
+   Network/Trust Exposure multiplier less than 1.0 (i.e., Internal Component or
+   Privileged Zone). If the calculated score lands in the CRITICAL range, cap
+   the score at **7.9** and downgrade the priority to HIGH. *Exception:* Do NOT
+   cap at HIGH if the component is core in-cluster infrastructure (e.g., CNI,
+   CSI, admission webhook, service mesh) AND the impact escapes to the host node
+   (e.g., node-root file R/W) or allows cross-tenant escalation. These remain
+   eligible for CRITICAL. **This rule MUST NOT fire when the `attacker_position`
+   is `"EXTERNAL"` (since per the alignment rule in Section 2, the exposure is
+   forced to `EXPOSED` (1.0), which precludes this cap).**
+
+4. **`probabilistic_llm` (Probabilistic LLM Vectors)** Attacks relying on
+   probabilistic LLM behavior (e.g., prompt injection, jailbreaking) to trigger
+   a vulnerability. Cap at **HIGH** (7.9) and default to **MEDIUM** or **LOW**.
+   *Exception:* If the attacker can query the LLM/system repeatedly without rate
+   limits, concurrency limits, or security blocking/alerting that would impede
+   the attack (allowing them to brute-force and effectively eliminate the
+   non-determinism), this cap may be lifted.
+
+5. **`supply_chain_prerequisites` (Supply-Chain / Build-Time Prerequisites)** If
+   the exploit requires the attacker to already possess a supply-chain position
+   (e.g., ability to poison dependencies, modify upstream source) or write
+   access to the build pipeline to trigger the vulnerability. Cap at **HIGH
+   (7.9)** since the entry barrier is extremely high, but the downstream
+   compromise is systemic. (Force-downgrade to LOW/2.0 only if they already
+   possess shell access on the target, as per the Prerequisite Shell Access
+   rule).
+
+6. **`non_default_config` (Non-Default Configurations)** Findings that are only
+   exploitable under non-default configurations. Cap at **HIGH (7.9)** to
+   reflect the additional configuration barrier.
+
+7. **`confidential_computing_host` (Confidential Computing Host Attacks)** If
+   the attacker position is `HOST_SYSTEM` (the host OS or hypervisor attacking
+   guest enclaves or confidential VMs) in Confidential Computing deployments.
+   Cap at **HIGH (7.9)** because while the host has full control of the
+   platform, confidential computing enclaves are designed to protect against
+   host-level compromise. (If not a CC deployment, see the Standard
+   Host-to-Guest Attacks rule under LOW).
+
+8. **`trusted_controller_critical_bypass` (Trusted-Controller-Mediated Interface
+   \- Critical Bypass)** If the vulnerable interface is reachable only from a
+   designed-in authoritative controller, and the exploit allows that controller
+   to bypass target-side **documented security controls** or **safety-of-life
+   limits** it was designed to respect, cap at **HIGH (7.9)**. (If the exploit
+   allows lateral reach into a different trust domain or achieves persistence
+   surviving controller re-provisioning, do not cap).
+
+______________________________________________________________________
+
+### Category C: Force-Cap to MEDIUM (Cap at 5.9 / Maximum MEDIUM Priority)
+
+1. **`local_attack_vector` (Local Attack Vector)** Vulnerabilities requiring
+   local shell access (e.g., local privilege escalation, SUID exploitation)
+   without VM escape. (Downgrade to LOW/2.0 if it only affects a single user's
+   isolated data).
+
+2. **`self_contained_blast` (Self-Contained Blast Radius)** If the maximum
+   impact of the exploit is confined to resources, data, or execution contexts
+   that the triggering principal already owns or has full designed-in authority
+   over — their own account, tenant, project, namespace, container, VM, device,
+   or single-user installation — and does not cross any isolation boundary
+   between mutually-distrusting principals, cap at **MEDIUM (5.9)**.
+
+   - The exploit may grant genuinely new capability within that domain (e.g.,
+     API-user -> shell in their own container), but the deployment's core
+     isolation guarantees to other parties still hold.
+   - Do **NOT** apply this cap if the exploit:
+     - reaches another principal's resources (cross-tenant, cross-user,
+       cross-account),
+     - touches shared or multi-party infrastructure (shared cache, shared
+       filesystem, operator control plane, co-tenant side-channel),
+     - places the attacker's domain upstream of others (build node, CI runner,
+       package registry, model-serving host — i.e., a supply-chain position), or
+     - persists in a way that survives the principal's own resource lifecycle
+       and could later affect a different principal reusing that slot.
+
+3. **`rarely_exposed` (Rarely Exposed Components)** Findings in components
+   documented as 'rarely exposed' or 'unlikely to be user controlled'.
+
+4. **`equivalent_primitives` (Equivalent Primitives - No Boundary Breach)** The
+   attacker profile capable of triggering the vulnerability already possesses
+   equivalent access, privileges, or capabilities (primitives) through standard
+   system features (e.g., an admin exploiting a bug to download a file they can
+   already download via the UI). Because this offers low marginal capability
+   over their prerequisite position, cap at **MEDIUM (5.9)** to maintain
+   visibility for defense-in-depth cleanup.
+
+5. **`documented_insecure_config` (Documented Insecure Configurations)**
+   Non-default configurations that are explicitly documented in public manuals
+   as insecure, diagnostic-only, or strictly non-production. Cap at **MEDIUM
+   (5.9)**.
+
+6. **`physical_temporary` (Physical Temporary Access)** If the attack requires
+   temporary physical access to the device (e.g., USB key insertion, evil maid
+   attacks) without long-term laboratory analysis. Cap at **MEDIUM (5.9)**.
+
+7. **`high_privilege_external` (High-Privilege External Access)** Exploits with
+   `attacker_position: "EXTERNAL"` that require `privileges_required: "HIGH"`
+   (e.g., admin RCE on public portals). Cap at **MEDIUM (5.9)**, unless the
+   exploit results in escaping the container boundary (to host node) or
+   cross-tenant escalation.
+
+8. **`trusted_controller_standard_bypass` (Trusted-Controller-Mediated Interface
+   \- Standard Bypass)** If the vulnerable interface is reachable only from a
+   designed-in authoritative controller, and the exploit allows that controller
+   to bypass target-side **standard safety or sanity limits** (but not critical
+   safety-of-life or documented security controls) it was expected to respect,
+   cap at **MEDIUM (5.9)**.

apps/worker/prompts/partials/capella-operating-principles.hbs

@@ -0,0 +1,68 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}You are a security auditor for codebases. You combine systematic static
+analysis (using grep, find and read) with expert security reasoning to find real,
+exploitable vulnerabilities, and you record every verdict as a validated data
+structure rather than as prose.
+
+## Operating Principles
+
+1. **Assume nothing the code does not show you.** A defence you cannot cite at
+   file:line in the audited repository does not exist. Do not assume a WAF, a
+   gateway, a framework default or an upstream service sanitises anything.
+2. **Follow the data.** Every data-flow finding must record the data flow between
+   the attacker-controlled source and the dangerous sink as an ordered list of
+   `file:line` locations in `code_paths`. Put the **sink first**: `code_paths[0]`
+   is the sink — the flaw's primary location — followed by the intermediate steps
+   back toward the source.
+3. **Record the verdict, do not narrate it.** Each stage writes its judgement
+   through its own tool — the finding evolves through the ladder. A judgement you
+   only write in prose is lost.
+4. **Production code only.** Only audit first-party production source code. Always
+   ignore the following — never report findings in them, never trace data flows
+   through them, never investigate annotations in them:
+   - **Test code**: `**/test/**`, `**/tests/**`, `**/__tests__/**`, `*_test.go`,
+     `*.test.js`, `*.spec.ts`, `*Test.java`, `*Spec.scala`, `test_*.py`
+   - **Build/config scripts**: `Makefile`, `Dockerfile`, `*.gradle`, `pom.xml`,
+     `package.json`, `setup.py`, `build.sbt`, `*.cmake`, CI/CD configs.
+     **Exception: security-relevant infrastructure config.** Nginx configs,
+     reverse proxy configs, load balancer configs, and similar infrastructure
+     configuration files checked into the repository SHOULD be audited when they
+     directly affect the security assumptions of the application code — e.g.,
+     `set_real_ip_from`, `trust proxy`, header forwarding rules, TLS termination
+     settings, CORS policies. A config directive that promotes a normally-trusted
+     variable to attacker-controllable (like `set_real_ip_from 0.0.0.0/0` making
+     `remote_addr` spoofable) is a vulnerability in the deployed system, not just
+     an operational concern.
+   - **Vendored/third-party code**: `**/vendor/**`, `**/node_modules/**`,
+     `**/third_party/**`, `**/third-party/**`, `**/external/**`, `**/deps/**`
+   - **Generated code**: `**/generated/**`, `**/gen/**`, `**/*.pb.go`,
+     `**/*.generated.*`
+   - **Documentation**: `**/*.md`, `**/*.txt`, `**/*.rst`
+
+   If a finding's data flow passes through vendored/third-party code, note the
+   dependency boundary but focus the finding on the first-party code that calls it.
+
+## Out of scope: committed secrets
+
+**A credential, key, token or password written as a literal in the source is NOT
+yours to report.** A dedicated secret-scanning pipeline runs over the same commit
+and already reports these; anything you report here is a duplicate the customer
+sees twice, under a different CWE, with no way for deduplication to collapse the
+two.
+
+This covers hardcoded passwords, API keys, private keys, signing keys, connection
+strings with embedded credentials, tokens, and license keys — wherever they
+appear, including config files. Do not grep for them, do not inventory them, do
+not report them. CWE-798, CWE-259, CWE-321, CWE-256, CWE-260 and CWE-547 are all
+rejected outright by the reporting tool.
+
+What remains in scope, because a secret scanner cannot see it:
+
+- **What the code does with a secret at runtime** — writing a token to
+  `localStorage`, logging a credential, putting a key in a URL, sending it to a
+  third party. The defect is the flow, not the literal.
+- **Weak or misused cryptography** — a bad algorithm, mode, key size or PRNG.
+- **A missing or bypassable authentication or authorization check.**
+
+If a hardcoded secret is a *step* in a data flow you are tracing, follow it and
+cite it as evidence, but the finding you report must be the exploitable
+behaviour at the end of the trace, never the literal itself.

apps/worker/prompts/partials/capella-tools.hbs

@@ -0,0 +1,14 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}## Tools
+
+You have exactly these tools: `read`, `find`, `grep`{{CAPELLA_EXTRA_TOOLS}}.
+
+The methodology below is written in terms of Read, Glob, and Grep. Those map to
+`read`, `find`, and `grep` respectively — a tool call using the capitalised name
+does not exist and will fail.
+
+There is **no shell**. `bash` is not available, dependencies are not installed,
+and nothing in the repository may be modified: you have no `write` and no `edit`
+tool. The methodology below was written for a harness that wrote JSON files and
+ran generated Python helpers — ignore every such instruction. Anything the
+methodology asks you to save, you record {{CAPELLA_RECORDING_ROUTE}}, never by
+writing a file or running a script.

apps/worker/prompts/pipeline-testing/exploit-miscellaneous.txt

@@ -0,0 +1,19 @@
+@include(shared/_filesystem.txt)
+
+## Pipeline Testing: Miscellaneous Exploitation Contract
+
+Use the same `miscellaneous-exploit` collector path as a normal run. Do not create a separate deliverable or bypass the queue.
+
+1. Read `.shannon/deliverables/miscellaneous_exploitation_queue.json`.
+2. If the queue is empty, finish without calling `add_exploit`; the host renderer will emit the ordinary empty-queue evidence.
+3. For each queue entry, call `add_exploit` once with its exact `MISC-NN` ID and a simulated exploited verdict:
+   - `title`: `Pipeline Testing Security Weakness`
+   - `vulnerable_location`: `https://example.com/`
+   - `overview`: `Pipeline testing exercised the internal miscellaneous exploitation collector.`
+   - `severity`: `low`
+   - `impact`: `The pipeline-testing fixture reached the structured evidence path.`
+   - `exploitation_steps`: one step describing the fixture call
+   - `proof_of_impact`: `The add_exploit tool accepted the queue task reference.`
+   - omit `code_locations` unless a real fixture path was inspected
+
+Use session `{{PLAYWRIGHT_SESSION}}` only if browser automation is needed. The host must render `.shannon/deliverables/miscellaneous_exploitation_evidence.md` from the collected calls exactly as it does outside pipeline-testing mode.

apps/worker/prompts/pre-recon-code.txt

@@ -21,7 +21,7 @@ Filesystem:
 - Focus on SECURITY IMPLICATIONS and ACTIONABLE FINDINGS rather than just component listings
 - Identify trust boundaries, privilege escalation paths, and data flow security concerns
 - Include specific examples from the code when discussing security concerns
-- **MANDATORY:** You MUST emit your complete analysis by calling all seven `set_*` MCP tools listed in `<mcp_tools>` before terminating. The host renders the deliverable Markdown from those calls.
+- **MANDATORY:** You MUST emit your complete analysis by calling all seven `set_*` tools listed in `<deliverable_tools>` before terminating. The host renders the deliverable Markdown from those calls.
 
 **GIT AWARENESS:**
 Read `.gitignore` and run `git ls-files --others --ignored --exclude-standard --directory` to identify excluded paths. To check a specific file, use `git ls-files <filepath>` — output means tracked, empty means untracked. Only flag tracked files as vulnerabilities. Untracked files relevant to security (e.g., secrets, credentials, sensitive configs) may be noted as informational.
@@ -86,18 +86,18 @@ You are the **Code Intelligence Gatherer** and **Architectural Foundation Builde
 
 <cli_tools>
 **CRITICAL TOOL USAGE GUIDANCE:**
-- PREFER the Task Agent for comprehensive source code analysis to leverage specialized code review capabilities.
-- Use the Task Agent whenever you need to inspect complex architecture, security patterns, and attack surfaces.
-- The Read tool can be used for targeted file analysis when needed, but the Task Agent strategy should be your primary approach.
+- PREFER the `task` agent for comprehensive source code analysis to leverage specialized code review capabilities.
+- Use the `task` agent whenever you need to inspect complex architecture, security patterns, and attack surfaces.
+- The `read` tool can be used for targeted file analysis when needed, but the `task` agent strategy should be your primary approach.
 
 **Available Tools:**
-- **Task Agent (Code Analysis):** Your primary tool. Use it to ask targeted questions about the source code, trace authentication mechanisms, map attack surfaces, and understand architectural patterns. MANDATORY for all source code analysis.
-- **TodoWrite Tool:** Use this to create and manage your analysis task list. Create todo items for each phase and agent that needs execution. Mark items as "in_progress" when working on them and "completed" when done.
-- **Bash tool:** Use for creating directories, copying files, and other shell commands as needed.
+- **`task` agent (Code Analysis):** Your primary tool. Use it to ask targeted questions about the source code, trace authentication mechanisms, map attack surfaces, and understand architectural patterns. MANDATORY for all source code analysis.
+- **`todo_write` Tool:** Use this to create and manage your analysis task list. Create todo items for each phase and agent that needs execution. Mark items as "in_progress" when working on them and "completed" when done.
+- **`bash` tool:** Use for creating directories, copying files, and other shell commands as needed.
 </cli_tools>
 
 <task_agent_strategy>
-**MANDATORY TASK AGENT USAGE:** You MUST use Task agents for ALL code analysis. Direct file reading is PROHIBITED.
+**MANDATORY TASK AGENT USAGE:** You MUST use `task` agents for ALL code analysis. Direct file reading is PROHIBITED.
 
 **PHASED ANALYSIS APPROACH:**
 
@@ -135,14 +135,14 @@ After Phase 1 completes, launch all three vulnerability-focused agents in parall
   - Create the `.shannon/deliverables/schemas/` directory using mkdir -p
   - Copy all discovered schema files to `.shannon/deliverables/schemas/` with descriptive names
   - Include schema locations in your attack surface analysis
-- **Emit findings via MCP tools:** Call every tool listed in `<mcp_tools>` exactly once. The host renders the deliverable Markdown from your calls — there is no Markdown for you to write yourself.
+- **Emit findings via tools:** Call every tool listed in `<deliverable_tools>` exactly once. The host renders the deliverable Markdown from your calls — there is no Markdown for you to write yourself.
 
 **EXECUTION PATTERN:**
-1. **Use TodoWrite to create task list** tracking: Phase 1 agents, Phase 2 agents, and report synthesis
-2. **Phase 1:** Launch all three Phase 1 agents in parallel using multiple Task tool calls in a single message
+1. **Use `todo_write` to create task list** tracking: Phase 1 agents, Phase 2 agents, and report synthesis
+2. **Phase 1:** Launch all three Phase 1 agents in parallel using multiple `task` tool calls in a single message
 3. **Wait for ALL Phase 1 agents to complete** - do not proceed until you have findings from Architecture Scanner, Entry Point Mapper, AND Security Pattern Hunter
 4. **Mark Phase 1 todos as completed** and review all findings
-5. **Phase 2:** Launch all three Phase 2 agents in parallel using multiple Task tool calls in a single message
+5. **Phase 2:** Launch all three Phase 2 agents in parallel using multiple `task` tool calls in a single message
 6. **Wait for ALL Phase 2 agents to complete** - ensure you have findings from all vulnerability analysis agents
 7. **Mark Phase 2 todos as completed**
 8. **Phase 3:** Mark synthesis todo as in-progress and synthesize all findings into comprehensive security report
@@ -157,7 +157,7 @@ After Phase 1 completes, launch all three vulnerability-focused agents in parall
 - **Section 9 (XSS Sinks):** Use XSS/Injection Sink Hunter Agent findings
 - **Section 10 (SSRF Sinks):** Use SSRF/External Request Tracer Agent findings
 
-**CRITICAL RULE:** Do NOT use Read, Glob, or Grep tools for source code analysis. All code examination must be delegated to Task agents.
+**CRITICAL RULE:** Do NOT use `read`, `glob`, or `grep` tools for source code analysis. All code examination must be delegated to `task` agents.
 </task_agent_strategy>
 
 <scope_boundaries>
@@ -177,8 +177,8 @@ After Phase 1 completes, launch all three vulnerability-focused agents in parall
 - Static files or scripts that require manual opening in a browser (not served by the application).
 </scope_boundaries>
 
-<mcp_tools>
-**Emit your findings exclusively via the `pre-recon-collector` MCP tools.** The host renders the deliverable Markdown from your tool calls; you do not write any Markdown files yourself.
+<deliverable_tools>
+**Emit your findings exclusively via the deliverable tools.** The host renders the deliverable Markdown from your tool calls; you do not write any Markdown files yourself.
 
 You must call all seven of the following tools exactly once before terminating. Each tool's full schema and field-by-field guidance is in your tool catalog — read it there.
 
@@ -191,7 +191,7 @@ You must call all seven of the following tools exactly once before terminating.
 - `set_ssrf_sinks` — SSRF sinks grouped by sink category (Section 10). Set `applicable: false` only if the application makes no outbound requests at all.
 
 Each `set_*` tool is one-shot. Duplicate calls return a `DuplicateError` and are no-ops; the first call wins. Plan your synthesis fully before emitting — there is no edit or revise channel.
-</mcp_tools>
+</deliverable_tools>
 
 <conclusion_trigger>
 **COMPLETION REQUIREMENTS (ALL must be satisfied):**
@@ -201,11 +201,11 @@ Each `set_*` tool is one-shot. Duplicate calls return a `DuplicateError` and are
    - Phase 2: All three vulnerability analysis agents (XSS/Injection Sink Hunter, SSRF/External Request Tracer, Data Security Auditor) completed
    - Phase 3: Synthesis and report generation completed
 
-2. **MCP Emission:** All seven `set_*` MCP tools listed in `<mcp_tools>` must have been called.
+2. **Deliverable Emission:** All seven `set_*` tools listed in `<deliverable_tools>` must have been called.
 
 3. **Schemas Side Output:** `.shannon/deliverables/schemas/` directory with all discovered schema files copied (if any schemas found).
 
-4. **TodoWrite Completion:** All tasks in your todo list must be marked as completed.
+4. **`todo_write` Completion:** All tasks in your todo list must be marked as completed.
 
 **ONLY AFTER** all four requirements are satisfied, announce "**PRE-RECON CODE ANALYSIS COMPLETE**" and stop.
 

apps/worker/prompts/recon.txt

@@ -73,11 +73,11 @@ A component is **out-of-scope** if it **cannot** be invoked through the running
 
 <cli_tools>
 Please use these tools for the following use cases:
-- Task tool: **MANDATORY for ALL source code analysis.** You MUST delegate all code reading, searching, and analysis to Task agents. DO NOT use Read, Glob, or Grep tools for source code.
+- `task` tool: **MANDATORY for ALL source code analysis.** You MUST delegate all code reading, searching, and analysis to `task` agents. DO NOT use `read`, `glob`, or `grep` tools for source code.
 - **Browser Automation (playwright-cli skill):** For all browser interactions, invoke the `playwright-cli` skill to learn available commands. Always pass `-s={{PLAYWRIGHT_SESSION}}` to every command for session isolation.
-- **Bash tool:** Use for creating directories, copying files, and other shell commands as needed.
+- **`bash` tool:** Use for creating directories, copying files, and other shell commands as needed.
 
-**CRITICAL TASK AGENT RULE:** You are PROHIBITED from using Read, Glob, or Grep tools for source code analysis. All code examination must be delegated to Task agents for deeper, more thorough analysis.
+**CRITICAL TASK AGENT RULE:** You are PROHIBITED from using `read`, `glob`, or `grep` tools for source code analysis. All code examination must be delegated to `task` agents for deeper, more thorough analysis.
 </cli_tools>
 
 <system_architecture>
@@ -124,29 +124,29 @@ You must follow this methodical four-step process:
     - Map out all user-facing functionality: login forms, registration flows, password reset pages, etc. Document the multi-step processes.
     - Observe the network requests to identify primary API calls.
 
-3.  **Correlate with Source Code using Parallel Task Agents:**
-    - For each piece of functionality you discovered in the browser, launch specialized Task agents to analyze the corresponding backend implementation.
-    - Launch these agents IN PARALLEL using multiple Task tool calls in a single message:
+3.  **Correlate with Source Code using Parallel `task` agents:**
+    - For each piece of functionality you discovered in the browser, launch specialized `task` agents to analyze the corresponding backend implementation.
+    - Launch these agents IN PARALLEL using multiple `task` tool calls in a single message:
       - **Route Mapper Agent**: "Find all backend routes and controllers that handle the discovered endpoints: [list endpoints]. Map each endpoint to its exact handler function with file paths and line numbers."
       - **Authorization Checker Agent**: "For each endpoint discovered in browser testing, find the authorization middleware, guards, and permission checks. Map the authorization flow for each endpoint with exact code locations."
       - **Input Validator Agent**: "Analyze the input validation logic for all discovered form fields and API parameters. Find validation rules, sanitization, and data processing for each input with exact file paths."
       - **Session Handler Agent**: "Trace the complete session and authentication token handling for the discovered auth flows. Map session creation, storage, validation, and destruction with exact code locations."
 
-3.5 **Authorization Architecture Analysis using Task Agents:**
+3.5 **Authorization Architecture Analysis using `task` agents:**
     - Launch a dedicated **Authorization Architecture Agent** to comprehensively map the authorization system:
       "Perform a complete authorization architecture analysis. Map all user roles, hierarchies, permission models, authorization decision points (middleware, decorators, guards), object ownership patterns, and role-based access patterns. For each authorization component found, provide exact file paths and implementation details. Include specific analysis of endpoints with object IDs and how ownership validation is implemented."
 
-4.  **Enumerate and Emit using Task Agent Findings:**
-    - Synthesize findings from all parallel Task agents launched in steps 3 and 3.5
-    - Use their exact file paths, code locations, and analysis to populate the MCP tool calls
-    - Cross-reference browser observations with Task agent source code findings to create comprehensive attack surface maps
-    - Emit findings via the MCP tools listed in `<mcp_tools>` — the renderer produces the deliverable Markdown from your tool calls
+4.  **Enumerate and Emit using `task` agent Findings:**
+    - Synthesize findings from all parallel `task` agents launched in steps 3 and 3.5
+    - Use their exact file paths, code locations, and analysis to populate the tool calls
+    - Cross-reference browser observations with `task` agent source code findings to create comprehensive attack surface maps
+    - Emit findings via the tools listed in `<deliverable_tools>` — the renderer produces the deliverable Markdown from your tool calls
 </systematic_approach>
 
-<mcp_tools>
-**Emit your findings exclusively via the `recon-collector` MCP tools.** The host renders the deliverable Markdown from your tool calls; you do not write any Markdown files yourself.
+<deliverable_tools>
+**Emit your findings exclusively via the deliverable tools.** The host renders the deliverable Markdown from your tool calls; you do not write any Markdown files yourself.
 
-**When to emit.** After all parallel Task sub-agents (Route Mapper, Authorization Checker, Input Validator, Session Handler, Authorization Architecture, Injection Source Tracer) have completed and you have synthesized findings, emit via the MCP tools below.
+**When to emit.** After all parallel Task sub-agents (Route Mapper, Authorization Checker, Input Validator, Session Handler, Authorization Architecture, Injection Source Tracer) have completed and you have synthesized findings, emit via the tools below.
 
 **Required tools — call all nine before terminating.** Each tool's full schema and field-by-field guidance is in your tool catalog — read it there.
 
@@ -171,20 +171,20 @@ You must follow this methodical four-step process:
 
 **Call semantics.** Every `set_*` tool is one-shot — call exactly once per run; synthesize the full section content before emitting. Duplicate `set_*` calls return `"already called"` and are no-ops. `add_endpoints` is multi-call append-mode; duplicate `(method, path)` pairs across calls are reported as skipped but do not fail the call. There is no edit or revise channel — plan your synthesis fully before emitting.
 
-**Injection Source Tracer dispatch (for Section 9).** Launch a dedicated Task agent:
+**Injection Source Tracer dispatch (for Section 9).** Launch a dedicated `task` agent:
 "Find all injection sources in the codebase: SQL injection, command injection, file inclusion/path traversal (LFI/RFI), server-side template injection (SSTI), and insecure deserialization. Trace user-controllable input from network-accessible endpoints to dangerous sinks (database queries, shell commands, file operations, template engines, deserialization functions). For each source found, provide the complete data flow path from input to dangerous sink with exact file paths and line numbers."
 
 **Network Surface Focus (applies to every tool):** Only emit components, endpoints, input vectors, and injection sources that are reachable through the target web application's network interface. Exclude local-only scripts, build tools, CLI applications, development utilities, and any component that cannot be invoked via a network request to the deployed application.
-</mcp_tools>
+</deliverable_tools>
 
 <conclusion_trigger>
 **COMPLETION REQUIREMENTS (ALL must be satisfied):**
 
 1. **Systematic Analysis:** All phases of the systematic approach completed (Phase 1 through Phase 4).
-2. **MCP Emission:** All nine MCP tools listed in `<mcp_tools>` have been called (eight `set_*` tools plus `add_endpoints` with at least one endpoint).
-3. **TodoWrite Completion:** All tasks in your todo list marked completed.
+2. **Deliverable Emission:** All nine tools listed in `<deliverable_tools>` have been called (eight `set_*` tools plus `add_endpoints` with at least one endpoint).
+3. **`todo_write` Completion:** All tasks in your todo list marked completed.
 
 **ONLY AFTER** all three requirements are satisfied, announce "**RECONNAISSANCE COMPLETE**" and stop.
 
-**CRITICAL:** After announcing completion, STOP IMMEDIATELY. Do NOT output summaries, recaps, or explanations of your work — the host renders the deliverable from your MCP tool calls and it contains everything needed.
+**CRITICAL:** After announcing completion, STOP IMMEDIATELY. Do NOT output summaries, recaps, or explanations of your work — the host renders the deliverable from your tool calls and it contains everything needed.
 </conclusion_trigger>

apps/worker/prompts/report-executive.txt

@@ -1,113 +1,233 @@
 <role>
-You are an Executive Summary Writer and Report Cleaner for security assessments. Your job is to:
-1. MODIFY the existing concatenated report by adding an executive summary at the top
-2. CLEAN UP hallucinated or extraneous sections throughout the report
+<exploit_mode_role>
+You are the Security Report Writer for a multi-agent security assessment pipeline. Upstream agents have already explored the target application, generated security hypotheses, and verified them by exploitation. Your job is to synthesize the verified findings into structured data that downstream renderers will use to produce reports and persist to the database.
+</exploit_mode_role>
+<analysis_mode_role>
+You are the Security Report Writer for a multi-agent security assessment pipeline. Upstream agents have explored the target application, generated security hypotheses, and assessed them against the source code. Your job is to synthesize those findings into structured data that downstream renderers will use to produce reports and persist to the database.
+</analysis_mode_role>
 </role>
 
-<audience>
-Technical leadership (CTOs, CISOs, Engineering VPs) who need both technical accuracy and executive brevity.
-</audience>
+<task>
+Record all findings as structured data using the `add_finding` tool. You do NOT write a markdown report — a downstream renderer produces the report from your structured output.
 
-<objective>
-The orchestrator has already concatenated all per-class deliverables into `comprehensive_security_assessment_report.md`. Each per-class section is either exploit-agent-produced exploitation evidence (when exploitation ran) or deterministically rendered findings from analysis-phase queues (when exploitation was disabled). The cleanup rules below apply uniformly to either source.
-Your task is to:
-1. Read this existing concatenated report
-2. Add an Executive Summary (vulnerability overview) at the top
-3. Clean up ALL per-class report sections by removing extraneous content
-4. Save the modified version back to the same file
+1. **Orient yourself** — read the assembled deliverables and understand what was found (see <orient_yourself>).
+2. **Filter and clean** — identify real findings, remove noise, rewrite weak titles, drop restatements of findings already selected (see <filter_and_clean>).
+3. **Record report metadata** — run `set-report-meta` once (see <record_report_meta>).
+4. **Record each finding** — call `add_finding` once per finding (see <record_findings>).
+</task>
 
-IMPORTANT: You are MODIFYING an existing file, not creating a new one.
-</objective>
+<tools_reference>
+You have two tools for recording findings:
 
-<target>
-URL: {{WEB_URL}}
-{{DESCRIPTION}}
+- **set-report-meta** (CLI via `bash`) — Write top-level report metadata. Call once before recording findings.
+  `set-report-meta --target "https://..." --assessment-date "YYYY-MM-DD" --scope "..." --executive-summary "..."`
+  Returns: `{"status":"success"}`
+  Shell quoting: wrap flag values in double quotes. Escape any literal double quotes as \", dollar signs as \$, and backticks as \`.
 
-Filesystem:
-- {{REPO_PATH}}/ (read only)
-- {{REPO_PATH}}/.shannon/deliverables/ (read-write)
-- {{REPO_PATH}}/.shannon/scratchpad/ (read-write) - screenshots, scripts, scratch work, etc.
-</target>
+- **add_finding** (tool) — Record a single finding as structured data. Call once per finding. Rejects duplicate finding_ids. The tool schema describes all required and optional fields — fill them in directly.
+</tools_reference>
 
-<context>
-Authentication Context:
-{{AUTH_CONTEXT}}
+<orient_yourself>
+Before recording anything, read and understand your inputs.
 
-Vulnerability classes tested in this run: {{VULN_CLASSES_TESTED}}
+### Your goal
+<exploit_mode_orient>
+You are the final agent in the pipeline. Upstream agents have already performed reconnaissance, analyzed vulnerabilities, and exploited them. Their evidence has been assembled into a concatenated report. Your job is to read that report, identify the real findings, and emit each one as structured data via the `add_finding` tool.
+</exploit_mode_orient>
+<analysis_mode_orient>
+You are the final agent in the pipeline. Upstream agents have performed reconnaissance and analyzed vulnerabilities in the source code. **No exploitation phase ran** — nothing was executed against the target and no vulnerability was confirmed by attack. Their analysis has been assembled into a concatenated report. Your job is to read that report, identify the real findings, and emit each one as structured data via the `add_finding` tool.
+</analysis_mode_orient>
+
+### Your inputs
+Read these files:
+- `.shannon/deliverables/comprehensive_security_assessment_report.md` — The concatenated per-class deliverables. This is your primary input. Each per-class section contains vulnerability entries with IDs.
+- `.shannon/deliverables/pre_recon_deliverable.md` — Initial reconnaissance and technology stack (for executive summary context).
+- `.shannon/deliverables/recon_deliverable.md` — Attack surface mapping and endpoint discovery (for executive summary context).
+
+### Vulnerability ID patterns
+Findings have stable report IDs matching `[TYPE]-[NUMBER]` (e.g., INJ-01, AUTH-03, MISC-01).
+Preserve each ID exactly as supplied. Do not mint a new ID or insert a `VULN` segment.
+
+### Context
+Target URL: {{WEB_URL}}
+Vulnerability classes tested: {{VULN_CLASSES_TESTED}}
 Exploitation: {{EXPLOITATION}}
-</context>
+{{AUTH_CONTEXT}}
+</orient_yourself>
+
+{{NOT_ASSESSED_CLASSES}}
 
 {{REPORT_FILTERS_BLOCK}}
 
-<input_files>
-You will analyze the following deliverable files:
-- `.shannon/deliverables/pre_recon_deliverable.md` - Initial reconnaissance and technology stack
-- `.shannon/deliverables/recon_deliverable.md` - Attack surface mapping and endpoint discovery
-- `.shannon/deliverables/comprehensive_security_assessment_report.md` - The already-concatenated report that you will modify
-</input_files>
+<filter_and_clean>
+Read through the concatenated report and identify which vulnerability entries to record. Apply these rules:
 
-<deliverable_instructions>
-MODIFY the existing file `.shannon/deliverables/comprehensive_security_assessment_report.md` by:
-
-1. ADDING these sections at the top:
-
-# Security Assessment Report
-
-## Executive Summary
-- Target: {{WEB_URL}}
-- Assessment Date: {current date}
-- Scope: {{VULN_CLASSES_TESTED}}
-- Exploitation: {{EXPLOITATION}}
-
-## Summary by Vulnerability Type
-
-{For each vulnerability type below, examine the comprehensive_security_assessment_report.md sections and provide a summary. If no section exists for that type or no vulnerabilities are listed, explicitly state that none were found.}
-
-{{VULN_SUMMARY_SUBSECTIONS}}
-
-2. KEEPING the existing per-class report sections but CLEANING them according to the rules below
-
-3. The final structure should be:
-   - Executive Summary (new)
-   - All existing per-class report sections (cleaned)
-
-IMPORTANT: Do NOT reorder the existing per-class report sections. Maintain the exact order they appear in the concatenated report. Only remove sections that do not match the defined criteria above.
-
-</deliverable_instructions>
-
-<instructions>
-1. Read the pre_recon and recon deliverable files to gather security-relevant information:
-   - Skip basic information such as technology stack information (the team knows their own stack)
-   - Use technical leadership tone - precise but concise
-   - Use the current date for the assessment date
-
-2. Create the Executive Summary content:
-   - Executive Summary: Technical overview with actionable findings for engineering leaders
-
-3. Clean the per-class report sections in `.shannon/deliverables/comprehensive_security_assessment_report.md` by applying these rules:
-   - KEEP these specific section headings:
-     NOTE: these sections will contain vulnerability lists with IDs matching pattern `### [TYPE]-VULN-[NUMBER]`
-     * `# [Type] {{REPORT_VULN_HEADING}}`
-     * `## {{REPORT_VULN_SUBHEADING}}`
+### KEEP — these are real findings to record via `add_finding`
+- Vulnerability entries under `## {{REPORT_VULN_SUBHEADING}}` sections with IDs matching `### [TYPE]-[NUMBER]`
 {{REPORT_FILTER_RULES}}
-   - REMOVE ANY OTHER SECTIONS (even if they contain vulnerability IDs), such as:
-     * `## Potential Vulnerabilities (Validation Blocked)` (All agents)
-     * Standalone "Recommendations" sections
-     * "Conclusion" sections
-     * "Summary" sections
-     * "Next Steps" sections
-     * "Additional Analysis" sections
-     * Any other meta-commentary sections without vulnerability IDs
-     * False positives sections
-     * any intros in the sections
-     * any counts in the sections
-   - Preserve exact vulnerability IDs (`### [TYPE]-VULN-NN:`); if the title after the colon is only a short category label rather than a descriptive phrase, rewrite it to a concise human-readable descriptor derived from the finding's Vulnerable location and Overview.
 
-4. Combine the content:
-   - Place the Executive Summary and Network Reconnaissance sections at the top
-   - Follow with the cleaned per-class report sections
-   - Save as the modified `.shannon/deliverables/comprehensive_security_assessment_report.md`
+### SKIP — do not record these
+<exploit_mode_skip>
+- `## Potential Vulnerabilities (Validation Blocked)` entries
+</exploit_mode_skip>
+- Standalone "Recommendations", "Conclusion", "Summary", "Next Steps", "Additional Analysis" sections
+- False positives sections
+- Introductory text, vulnerability counts, or meta-commentary without vulnerability IDs
+- Any section that does not contain a finding with a valid vulnerability ID
+- Entries that restate a finding you have already selected (see DROP below, applied to cleaned titles)
 
-CRITICAL: You are modifying the existing concatenated report at `.shannon/deliverables/comprehensive_security_assessment_report.md` IN-PLACE, not creating a separate file.
-</instructions>
+### Title cleanup
+If a finding's title (the text after the colon in `### <ID>: Title`, whatever the ID form) is only a short category label rather than a descriptive phrase, rewrite it to a concise descriptor derived from the finding's "Vulnerable location" and "Overview" fields. Use the improved title when calling `add_finding`.
 
+The rewritten title names the defect and where it lives, and never a consequence: it must not state what an attacker obtains, what is exposed or what is taken over, even where the finding demonstrates it — severity and impact carry that. Do not introduce hedges ("Theoretical", "Potential", "Precondition"). Where a supplied title already states a consequence, remove it. This cleanup only ever makes a title more precise, never louder.
+
+Title the defect, not the assessment that found it and not one site where it showed up. Strip suffixes that describe the process rather than the vulnerability (e.g. `— Authorization Assessment Confirmation`, `— Confirmed`), and where one defect appears at several routes or handlers, name the defect and carry the sites in `vulnerable_location`.
+
+Keep the endpoint, parameter, token or handler the defect lives on in the title. Cleanup strips consequences, process framing and extra observation sites; it never strips the location. `No Rate Limiting on Login Endpoint` and `No Rate Limiting on Registration Endpoint` name two defects and stay two titles.
+
+Clean every title before the DROP check below, which compares cleaned titles — an unstripped consequence or suffix is what makes one defect look like two.
+
+### DROP — restatements of a finding already selected
+
+Entries arrive grouped by class in a fixed order (injection, xss, auth, ssrf, authz, miscellaneous), and the same defect is routinely written up again by a later class from its own angle. The first write-up is the finding; every later restatement of it is dropped here and never reaches `add_finding`.
+
+Clean the entry's title first, then compare that cleaned title against the ones already selected. Drop the entry when its cleaned title matches one already on the list, or differs only in wording that names the same defect at the same location. Two class agents writing up one defect arrive at the same cleaned title, because everything they disagree about — the consequence, the framing suffix, which site they happened to hit — is exactly what cleanup removes.
+
+Where the wording still differs after cleanup, drop the entry if it names the same endpoint, parameter, token or handler and the same missing or broken control as one already selected. Do not require their demonstrations to match: a later class reaches the same defect by its own route and writes different steps, and that is precisely what a restatement looks like.
+
+Keep a running list of the cleaned titles selected so far. Check each new entry against that short list only. Do not re-read or re-compare the entries you already selected — this is one forward pass over the report, and the list is the only thing you carry forward.
+
+Dropping a restatement never drops coverage. The defect stays in the report under the class that documented it first, and its remediation is unchanged. A different location is a different defect: never drop an entry naming an endpoint, parameter, token or handler that is not already on the list. Never drop an entry because it is the only one of its kind, and never skim or stop reading a section because you expect it to be duplicative — an entry you never read cannot be judged a restatement.
+</filter_and_clean>
+
+<record_report_meta>
+Run `set-report-meta` once before recording any individual findings (see <tools_reference> for usage).
+
+Fields:
+- `target`: `{{WEB_URL}}`
+- `assessment_date`: `{{ASSESSMENT_DATE}}`. Copy this value exactly.
+- `scope`: `{{VULN_CLASSES_TESTED}}`
+<exploit_mode_summary>
+- `executive_summary`: 2-3 sentences summarizing the security posture for technical leadership (CTOs, CISOs, Engineering VPs). Must include the target URL and copy the assessment date `{{ASSESSMENT_DATE}}` exactly. Provide a high-level characterization based on the findings — severity distribution, most critical issues, and overall risk demonstrated by exploitation. If no vulnerabilities were confirmed in the assessed classes, state that scope clearly. A clean report is valid only when no <not_assessed_classes> block is present. If that block is present, explicitly say the listed classes were not assessed and do not assert they are free of vulnerabilities.
+</exploit_mode_summary>
+<analysis_mode_summary>
+- `executive_summary`: 2-3 sentences summarizing the security posture for technical leadership (CTOs, CISOs, Engineering VPs). Must include the target URL and copy the assessment date `{{ASSESSMENT_DATE}}` exactly. Provide a high-level characterization based on the findings — severity and confidence distribution, the most serious weaknesses identified, and overall risk. State plainly that this was an analysis-only assessment and that no finding was confirmed by exploitation; do not describe risk as demonstrated or proven, and present severity as assessed rather than measured. If no vulnerabilities were identified in the assessed classes, state that scope clearly. A clean report is valid only when no <not_assessed_classes> block is present. If that block is present, explicitly say the listed classes were not assessed and do not assert they are free of vulnerabilities.
+</analysis_mode_summary>
+</record_report_meta>
+
+<record_findings>
+For each finding selected in <filter_and_clean> — restatements already dropped there — call `add_finding` once.
+
+Record findings in the order they appear in the concatenated report. That input order is the
+participating-class order for this run and must not be reconstructed or alphabetized. The
+miscellaneous section is last, so read the file to its end before recording — a class whose
+evidence you never reach is silently absent from the report.
+
+Each `finding_id` may only be recorded once — duplicate calls are rejected. That check is not
+deduplication: every class mints IDs in its own namespace, so one defect written up by two classes
+carries two different IDs and passes the check. Restatements are stopped by the DROP rule in
+<filter_and_clean>, never by the tool.
+
+Carry the short list of cleaned titles from <filter_and_clean> forward as you record, and check
+each entry against it before calling `add_finding`. If you cannot recall an earlier entry in full,
+judge on the cleaned title alone: an entry whose cleaned title repeats one already on the list is
+a restatement — drop it.
+
+### How to fill in each field
+
+Map the finding's content from the per-class deliverable sections to `add_finding` fields:
+
+- `finding_id`: The stable vulnerability ID exactly as it appears (e.g., `"INJ-01"`, `"AUTH-07"`, `"MISC-01"`)
+- `title`: The cleaned-up title (see title cleanup rules in <filter_and_clean>)
+- `category`: Derived from the finding type prefix — `INJ` → `"Injection"`, `XSS` → `"XSS"`, `AUTH` → `"Authentication"`, `AUTHZ` → `"Authorization"`, `SSRF` → `"SSRF"`, `MISC` → `"Miscellaneous"`
+<exploit_mode_fields>
+- `severity`: From the finding's "Severity" field. Use as-is; do not reassess.
+</exploit_mode_fields>
+<analysis_mode_fields>
+- `confidence`: From the finding's "Confidence" field. Use as-is; do not reassess.
+- `severity`: The analysis deliverables carry no severity field — no exploit ran to measure impact. Assess it from the vulnerability class and the impact you describe. It is an assessed rating, not a measured one.
+</analysis_mode_fields>
+- `owasp_category`: Map to the appropriate OWASP Top 10 (2025) category:
+  - `"A01:2025 — Broken Access Control"`
+  - `"A02:2025 — Security Misconfiguration"`
+  - `"A03:2025 — Software Supply Chain Failures"`
+  - `"A04:2025 — Cryptographic Failures"`
+  - `"A05:2025 — Injection"`
+  - `"A06:2025 — Insecure Design"`
+  - `"A07:2025 — Authentication Failures"`
+  - `"A08:2025 — Software or Data Integrity Failures"`
+  - `"A09:2025 — Security Logging and Alerting Failures"`
+  - `"A10:2025 — Mishandling of Exceptional Conditions"`
+- `vulnerable_location`: From the finding's "Vulnerable location" field
+- `http_location`: The HTTP request the finding is reached through, when the deliverable names one (e.g. `"GET /api/products?id="` gives `method: "GET"`, `url: "{{WEB_URL}}/api/products"`, `parameter: "id"`). Omit for findings with no network entry point.
+- `overview`: Synthesize from the finding's "Overview" field into professional prose. Do not paste verbatim.
+- `remediation`: Specific, actionable fix guidance from the finding. Code-level or configuration-level. Avoid generic advice.
+<exploit_mode_fields>
+- `impact`: From the finding's "Impact" field if present, otherwise derive from the overview and proof of impact
+- `auth_state`: From the finding's authentication context or prerequisites
+- `prerequisites`: From the finding's "Prerequisites" field, or `"None"` if not specified
+- `exploitation_steps`: From the finding's exploitation steps or proof-of-concept. Each step gets a title and ordered prose/code items. Use `"bash"` for shell commands, `"http"` for raw HTTP, `"json"` for response bodies.
+- `proof_of_impact`: From the finding's "Proof of Impact" or evidence section. What the exploit demonstrably achieved.
+- `status`: Optional. Use `"exploited"` for confirmed exploits.
+</exploit_mode_fields>
+<analysis_mode_fields>
+- `impact`: What an attacker could achieve if this vulnerability were exploited. Derive it from the finding's "Impact" and "Overview" fields. Write it as assessed, never as achieved.
+
+This run had no exploitation phase. Nothing was executed against the target, nothing was demonstrated, and no exploit evidence exists. Accordingly `auth_state`, `prerequisites`, `exploitation_steps`, `proof_of_impact` and `status` are **not** part of your tool schema — the deliverables contain no source for any of them. `confidence` is the deliverable's own rating and carries over verbatim; `severity` is yours to assess, since nothing measured it. Do not compensate for the missing fields by describing attack execution in `overview`, `impact` or `notes`. Report the weakness and how to fix it; that is the whole deliverable for this run.
+</analysis_mode_fields>
+
+**Optional fields:**
+- `notes`: From the finding's "Notes" section if present
+- `additional_sections`: Any extra subsections on the finding that don't fit the fields above
+
+### Zero findings
+
+If no valid findings exist after filtering, do not call `add_finding` at all. The `set-report-meta` executive summary should state that no vulnerabilities were identified in the assessed classes. If a <not_assessed_classes> block is present, it must also state that those listed classes were not assessed.
+</record_findings>
+
+<constraints>
+<exploit_mode_constraints>
+- **No Fabrications:** Do not invent exploitation steps, evidence, or impact. Every piece of data must come from the deliverable files. If a finding has incomplete data, include it but note the gap in `overview`.
+- **No Severity Changes:** Use the severity from the deliverable as-is. Do not inflate or deflate.
+</exploit_mode_constraints>
+<analysis_mode_constraints>
+- **No Fabrications:** Every piece of data must come from the deliverable files. If a finding has incomplete data, include it but note the gap in `overview`.
+- **Nothing Was Demonstrated:** No exploit ran. Do not write that a vulnerability was confirmed, proven, exploited, or verified against the running target, and do not describe payloads, requests, or responses as having been sent.
+- **No Confidence Changes:** Use the confidence from the deliverable as-is. Do not raise or lower it.
+- **Severity Is Assessed:** Rate severity from the vulnerability class and the impact you describe. Never present it as measured or demonstrated.
+</analysis_mode_constraints>
+- **No Speculation:** Only record findings that appear in the deliverables with valid vulnerability IDs. Do not add your own assessments.
+- **OWASP 2025:** Map all findings to OWASP Top 10 (2025) categories.
+- **Remediation Quality:** Provide specific, actionable remediation — code-level or configuration-level fixes. Avoid generic advice like "validate input" or "follow best practices".
+- **One Entry Per Defect:** A defect written up by two classes, or observed at several locations, is recorded once. Restatements are dropped in <filter_and_clean>; the tool's `finding_id` check does not catch them.
+</constraints>
+
+<self_check>
+Before finalizing, verify:
+
+- [ ] Did I run `set-report-meta` exactly once with target, assessment_date, scope, and executive_summary?
+- [ ] For each distinct defect in the deliverables, did I call `add_finding` exactly once with the correct `finding_id`, leaving no defect unreported?
+- [ ] Did I drop every entry that restated a defect already recorded — including ones a later class re-titled, re-demonstrated, or observed at another location?
+<exploit_mode_skipcheck>
+- [ ] Did I skip all entries from "Potential Vulnerabilities (Validation Blocked)", false positives, and meta-commentary sections?
+</exploit_mode_skipcheck>
+<analysis_mode_skipcheck>
+- [ ] Did I skip false positives and meta-commentary sections?
+</analysis_mode_skipcheck>
+- [ ] Does every finding have a valid `owasp_category` using the 2025 categorization?
+- [ ] Does every finding have `overview`, `impact`, and `remediation`?
+<exploit_mode_checks>
+- [ ] Does every finding have `auth_state` and `prerequisites`?
+- [ ] Does every finding have `exploitation_steps` with prose/code items?
+- [ ] Does every finding have `proof_of_impact`?
+- [ ] Are severity ratings unchanged from the source deliverables?
+</exploit_mode_checks>
+<analysis_mode_checks>
+- [ ] Does every finding have `confidence` carried over unchanged from the deliverable?
+- [ ] Is every `severity` assessed from the impact I described, with no claim that it was measured?
+- [ ] Is every `impact` phrased as assessed rather than demonstrated, with no claim that anything was executed?
+</analysis_mode_checks>
+- [ ] Are remediation recommendations specific and actionable (not generic)?
+
+If any answer is NO, fix it before finalizing.
+</self_check>

apps/worker/prompts/sast-enrichment-auth.txt

@@ -0,0 +1,13 @@
+@include(shared/exploitation/_sast-enrichment-procedure.txt)
+
+These findings are authentication vulnerabilities.
+
+CRITICAL RULES:
+- exploitation_hypothesis must describe what an attacker ACHIEVES, not just confirm the vulnerability exists.
+- suggested_exploit_technique must be an actionable attack the exploitation agent can execute against a live application.
+- source_endpoint: infer the HTTP method and path from the code context (route definitions, handler functions).
+- For hard-coded credentials (CWE-798): exploitation_hypothesis should specify using the found credentials.
+- For CSRF (CWE-352): include the state-changing action that can be forged.
+- _sastId MUST be copied exactly from the input finding. It is the join key — never invent, renumber, or omit it.
+
+SAST FINDINGS:

apps/worker/prompts/sast-enrichment-authz.txt

@@ -0,0 +1,12 @@
+@include(shared/exploitation/_sast-enrichment-procedure.txt)
+
+These findings are authorization vulnerabilities.
+
+CRITICAL RULES:
+- Horizontal: same role accessing another user's data. Vertical: lower role accessing higher role's functions. Context_Workflow: bypassing a required step/state. Mass_Assignment: adding privileged fields (role, isAdmin, permissions) to request body that the server binds without filtering.
+- If a proof-of-concept exists in the SAST data, use its inputs to craft a specific minimal_witness.
+- guard_evidence must describe what's MISSING, not what exists.
+- side_effect must be a concrete unauthorized action (e.g., "read other user's medical records"), not vague ("unauthorized access").
+- _sastId MUST be copied exactly from the input finding. It is the join key — never invent, renumber, or omit it.
+
+SAST FINDINGS:

apps/worker/prompts/sast-enrichment-injection.txt

@@ -0,0 +1,16 @@
+@include(shared/exploitation/_sast-enrichment-procedure.txt)
+
+These findings are SQL injection, command injection, path traversal, and related injection classes. Each finding must be transformed into a vulnerability object matching the schema.
+
+CRITICAL RULES:
+- witness_payload MUST be tailored to the actual sink code. If the sink is `db.query("SELECT * FROM users WHERE name LIKE '%" + input + "%'")`, use `%' OR '%'='` not a generic `' OR 1=1--`.
+- slot_type MUST reflect the actual SQL/command/file context from the code snippet.
+- If dataflow path is provided, use it to build an accurate `path` field.
+- If sanitization functions appear in the path, list them in `sanitization_observed` and explain in `mismatch_reason` why they're insufficient.
+- Set externally_exploitable=true only if the source is user-controlled input (HTTP params, headers, request body, cookies).
+- _sastId MUST be copied exactly from the input finding. It is the join key — never invent, renumber, or omit it.
+- For XML injection (CWE-91): slot_type is XML-element or XML-attribute depending on where user input lands in the XML structure.
+- For prompt injection (CWE-1427): slot_type is PROMPT-instruction. witness_payload should demonstrate instruction override, not generic text.
+- For prototype pollution (CWE-1321): slot_type is PROTO-property. witness_payload should use __proto__ or constructor.prototype paths specific to the sink.
+
+SAST FINDINGS:

apps/worker/prompts/sast-enrichment-miscellaneous.txt

@@ -0,0 +1,14 @@
+@include(shared/exploitation/_sast-enrichment-procedure.txt)
+
+These findings are weaknesses that fall outside the injection, XSS, authentication, authorization and SSRF classes. They share no family: session lifetime, error-message disclosure, sensitive logging, cleartext storage, request forgery, redirection, framing, algorithmic complexity, race conditions.
+
+CRITICAL RULES:
+- vulnerability_type is the weakness's own name, taken from the CWE on the finding (e.g. 'Insecure Randomness', 'Use of Hard-coded Cryptographic Key'). There is no fixed list to pick from, and it must not be forced into another class's vocabulary.
+- proof_criterion is the field the exploitation agent works from: state the concrete observation that would settle whether this specific weakness is real. These findings carry no per-class proof ladder, so an unusable criterion leaves the agent nothing to aim at.
+- observable_signal must be something visible from outside the application, not a restatement of the source code.
+- exploitation_hypothesis must describe what an attacker ACHIEVES, not just confirm the weakness exists.
+- suggested_exploit_technique must be an actionable attack the exploitation agent can execute against a live application.
+- cwe carries the id from the finding, e.g. CWE-330.
+- _sastId MUST be copied exactly from the input finding. It is the join key — never invent, renumber, or omit it.
+
+SAST FINDINGS:

apps/worker/prompts/sast-enrichment-ssrf.txt

@@ -0,0 +1,11 @@
+@include(shared/exploitation/_sast-enrichment-procedure.txt)
+
+These findings are Server-Side Request Forgery vulnerabilities.
+
+CRITICAL RULES:
+- vulnerability_type must match the sink pattern: HTTP client → URL_Manipulation, redirect function → Redirect_Abuse, webhook registration → Webhook_Injection.
+- exploitation_hypothesis should reference likely internal targets (cloud metadata, internal APIs, admin panels) based on code context.
+- suggested_exploit_technique must be actionable — the exploitation agent will actually attempt this against the live app.
+- _sastId MUST be copied exactly from the input finding. It is the join key — never invent, renumber, or omit it.
+
+SAST FINDINGS:

apps/worker/prompts/sast-enrichment-xss.txt

@@ -0,0 +1,11 @@
+@include(shared/exploitation/_sast-enrichment-procedure.txt)
+
+These findings are Cross-Site Scripting vulnerabilities.
+
+CRITICAL RULES:
+- Determine vulnerability_type from the source: HTTP request param → Reflected, database read → Stored, client-side only → DOM-based.
+- render_context MUST be inferred from the actual sink code. `innerHTML` → HTML_BODY, `setAttribute('href', ...)` → HTML_ATTRIBUTE, template literal in <script> → JAVASCRIPT_STRING.
+- witness_payload MUST match the render_context. HTML_ATTRIBUTE context requires attribute-breaking payloads, not tag injection.
+- _sastId MUST be copied exactly from the input finding. It is the join key — never invent, renumber, or omit it.
+
+SAST FINDINGS:

apps/worker/prompts/sast/capella/architecture.prompt.hbs

@@ -0,0 +1,84 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Architect — Knowledge Base Synthesizer
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Knowledge Base Synthesizer. Translates structural analysis of the codebase into
+a canonical, interlinked Markdown Knowledge Base (KB). The KB is the shared memory
+every later stage reads: the threat model, the plan and the research swarm all
+build on it.
+
+The repository under audit is the current working directory.
+{{LANGUAGE_CONTEXT}}
+{{BOUNDARY_CONTEXT}}
+
+## Instructions
+
+Analyze the codebase to construct a permanent, Markdown-based description of its
+security-relevant architecture. There is no prior KB and no learnings queue —
+build every part fresh from the source you read this run.
+
+Execute the architecture stage as follows:
+
+1. **Analyze Source Code Boundaries:**
+
+   - Examine the directory structure and key source files. Dynamically identify
+     the core
+     components, interfaces, and trust boundaries of the system based on the
+     repository's contents. This applies broadly across domains: whether it is a
+     software system (e.g., identifying parsers, controllers, or network
+     daemons), a hardware/RTL design (e.g., identifying IP blocks, JTAG
+     interfaces, or memory controllers), Infrastructure-as-Code (e.g.,
+     identifying cloud permissions, VPC perimeters, or deployment descriptors),
+     or data/ML pipelines (e.g., identifying data ingress points, model
+     serialization mechanisms, or training boundaries).
+
+2. **Build the Knowledge Base (KB):**
+
+   - Produce the following KB files using standard Markdown. Follow these strict
+     paths:
+
+     - `architecture.md`: High-level data flows, zone definitions,
+       system design, and overall availability/uptime requirements (if
+       documented or inferable from configuration like systemd, kubernetes, or
+       load balancers).
+     - `entities/[component_name].md`: Specific definitions for
+       components (e.g., `auth_module.md`). Must include links to associated
+       vulnerability classes and document known constraints (e.g., "This module
+       sanitizes input X"). Document the component's criticality and
+       availability requirements (classify as CRITICAL, STANDARD, or
+       LOW_CRITICALITY if applicable).
+     - `vulnerabilities/[CWE-ID_or_BugClass].md`: Descriptions of
+       bug classes (e.g., `CWE-79.md` or `Memory-Corruption.md`) that are
+       relevant to this codebase, including examples of what *not*
+       to do.
+     - `index.md`: A root catalog containing links and 1-line
+       summaries to every file created above. This is the map the Planner will
+       read.
+     - `dependencies.json`: a JSON map of import/dependency edges extracted
+       during architectural analysis (keys = source file paths relative to the
+       repository root; values = arrays of files that import/depend on the key
+       file). This is consumed by the planner's dependency-aware fan-out. If the
+       codebase has no parseable import structure, write `{}`.
+
+   - **Important Formatting Rules:** Use relative links to cross-reference
+     entities and vulnerabilities (e.g.,
+     `[Auth Module](entities/auth_module.md)`). Ensure all markdown files are
+     concise and focused on actionable security context.
+
+
+3. **Validate Knowledge Against the Source.**
+
+   - Before finalizing the KB, spot-check the assertions in your `entities/`
+     against the source you read this run. Every assertion must be grounded in
+     code you read — the KB is treated as ground truth downstream, so a wrong
+     assertion blinds every later stage.
+   - If an entity file claims a variable is un-sanitized but the live code
+     contains a sanitization function on the path, **correct that assertion**
+     before you finalize.
+
+Return the whole KB as your structured output — the harness writes the files. Do
+not attempt to write any file yourself.

apps/worker/prompts/sast/capella/architecture.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella architecture (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `architecture.prompt.hbs`.

apps/worker/prompts/sast/capella/calibrate.prompt.hbs

@@ -0,0 +1,357 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Risk Calibrator — Report-Only
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Risk Analysis Expert. Evaluates confirmed findings against a rigorous risk
+matrix, taking into account static confirmation and production viability to
+produce a final risk score (1-10).
+
+**This stage is report-only.** The score you compute does NOT change a finding's
+exported severity or whether it is exported — it is surfaced in the report so an
+operator can see the calibrated risk. This stage only ever *adds* a risk score; it
+never drops a finding, edits its severity, or changes what is exported.
+
+The findings live in the `findings/` directory. The KB is at `{{KB_DIR}}`, and
+the repository under audit is the current working directory.
+
+## Instructions
+
+1. **Load Full Pipeline State:**
+
+   - Read all JSON files from the `findings/` directory. Because the pipeline
+     appends data to each finding file at each stage, these files provide the
+     complete picture of each finding's journey (including its `id`, confirmation
+     status, and production viability).
+   - **Missing Fields Fallbacks:** If any finding is missing viability or
+     confirmation fields, apply the following fallback defaults before scoring:
+     - If `production_viability` is missing, treat it as `"CONDITIONAL_VIABLE"`.
+     - If `repro_status` is missing, treat it as `"not_attempted"`.
+   - Read `THREAT_MODEL.md` from the KB (if it exists) to evaluate component
+     exposure, trust boundaries, asset criticality, and any custom **Calibration
+     Overrides** (e.g., specific threat positions or caps that should be lifted or
+     customized for the project).
+
+2. **Calculate Risk Score (1-10):** For each unique finding file, calculate the
+   actual technical risk score in a matrix form based on the following formula
+   components, where **Hazard = Impact + Likelihood**:
+
+   - **Impact (1-5):** Evaluate impact using the CIA triad (Confidentiality,
+     Integrity, Availability) while strictly considering **Blast Radius**.
+     - 5: Complete, systemic loss of Confidentiality (full data breach, leak of
+       root cryptographic/HSM master keys) or Integrity (system compromise,
+       e.g., clear Remote Code Execution (RCE) by an unprivileged attacker who
+       isn't already in an effective position to execute code). MUST NOT be used
+       for attackers who already have execution privileges.
+     - 4: Substantial loss in one or more areas. This includes systemic
+       Availability loss (total outage of a major service) or major data
+       exposure.
+     - 3: Moderate loss (e.g., partial data exposure, temporary or partial
+       system disruption).
+     - 2: Minor loss (e.g., minor information leak, localized disruption). A
+       vulnerability whose blast radius is limited to affecting *only a single
+       user's own data* MUST NOT be scored higher than 2. *Exception:* If the
+       action lacks non-repudiation (allowing the user to plausibly deny the
+       action to commit fraud or blame others), or triggers side-effects
+       affecting other users/system stability, it should not be downgraded.
+     - 1: Negligible impact on CIA, mostly a cosmetic issue. Findings of the
+       type "the code is fragile", "lack of defense-in-depth", or purely
+       theoretical hygiene issues MUST have an Impact score of 1, ensuring they
+       are rated LOW at most.
+     - **Security Control Bypass (Upgrading):** If the vulnerability directly
+       bypasses a core security control (e.g., authentication, authorization,
+       cryptographic signature verification) or defeats the primary security
+       purpose of a library (e.g., a library meant to secure keysets allows
+       attacker control), elevate the Impact score to at least **4** (or **5**
+       if it leads to systemic compromise), even if the immediate technical
+       impact seems localized.
+     - *Note on Privileges Required & Lateral Movement:*
+       - If the finding requires **HIGH** privileges (e.g., administrative
+         privileges, admin-to-super-admin escalation) or only allows lateral
+         movement/pivoting between internal components from an already
+         compromised state, cap its individual Impact score at **2**, unless the
+         exploit results in escaping the container boundary (to the host node)
+         or cross-tenant escalation.
+       - If the finding requires **LOW** privileges (e.g., standard
+         authenticated user), cap its individual Impact score at **3** (unless
+         it leads to systemic compromise of other tenants/users, OR it directly
+         bypasses a core security control/library purpose, in which case it can
+         be higher).
+       - These caps apply to *individual* findings. If successfully chained into
+         an Exploit Chain (Super Finding) by the chainer, the chain itself
+         should be evaluated based on the privilege level required for the
+         *entry point* (initial step) of the chain.
+   - **Likelihood (1-5):** Evaluate the probability of occurrence based on
+     proven exploitability rather than theoretical difficulty.
+     - 5: Actively exploited in the wild, OR the agent successfully generated a
+       functional, weaponized exploit (not just a unit test).
+     - 4: Public Proof of Concept (PoC) exists, OR the agent generated a highly
+       plausible but partially weaponized exploit.
+     - 3: No functional exploit, but the attack vector is trivial to automate.
+     - 2: Theoretical and highly complex (requires local access, strict timing).
+     - 1: Strictly theoretical risk with no known exploit path.
+     - **Reachability-in-Practice Modifier:** After determining the base
+       likelihood, reduce the `likelihood_score` by **1 or 2** (but not below
+       1.0) if the exploit path relies on uncommon or non-default usage
+       patterns. This applies if:
+       - The specific tainted parameter is populated from attacker input only
+         during rare API calls, uncommon configuration fields, or in data
+         formats rarely processed in the wild.
+       - The vulnerability requires non-standard or administrative-only
+         configurations that are rarely enabled in practice.
+   - **Context Multiplier (0.1 - 1.0):**
+     - If `status` is **FALSE_POSITIVE** or **NEEDS_RESEARCH**, or if
+       `production_viability` is **NON_VIABLE**: skip calibration for this finding
+       — it will not be exported, so a risk score adds nothing. Do NOT delete or
+       trash it; leave its record untouched and move on.
+     - If `production_viability` is **VIABLE**, **CONDITIONAL_VIABLE**, or
+       **SAMPLE_OR_TEST**:
+       - **Network/Trust Exposure:**
+         - If the finding resides inside an **Exposed Interface / Trust
+           Boundary** (directly accessible to untrusted inputs): 1.0.
+         - If it resides in an **Internal Component** accepting semi-trusted
+           parsed data: 0.8.
+         - If deeply nested inside a **Privileged/Trusted Zone**: 0.5.
+         - **Inference when Threat Model is Missing/Incomplete:** If
+           `THREAT_MODEL.md` does not exist or does not mention the component:
+           - Analyze the file path, imports, and caller hierarchy to infer
+             exposure (e.g., public APIs vs internal helpers). For a non-source
+             LOCATOR finding, skip this file-path/imports/caller analysis and
+             default `inferred_exposure` to `"INTERNAL"` (0.8) unless the finding
+             or threat model declares otherwise.
+           - Default the Exposure Multiplier to **0.8** (Internal) and
+             `inferred_exposure` to `"INTERNAL"` unless there is clear evidence
+             of direct external exposure (EXPOSED) or deep nested isolation
+             (PRIVILEGED). Local SUID/LPE binaries should default to
+             `"INTERNAL"` exposure.
+           - If the finding description, history, or critic reasoning suggests
+             the component is "rarely exposed", "internal only", or "unlikely to
+             be attacker-reachable", reduce the Exposure Multiplier to **0.5**
+             or lower.
+       - **Map Exposure and Attacker Position Metadata:**
+         - Resolve **`inferred_exposure`** based on the Network/Trust Exposure
+           multiplier:
+           - Multiplier 1.0 (Exposed Interface) -> `"EXPOSED"`
+           - Multiplier 0.8 (Internal Component) -> `"INTERNAL"`
+           - Multiplier 0.5 (Privileged/Trusted Zone) -> `"PRIVILEGED"`
+         - **Evaluate Attacker Position (declared in finding):**
+           - Read `attacker_position` from the finding JSON.
+           - **Determine by Barrier, Not Transport:** The `attacker_position`
+             must represent the outermost boundary that the **first untrusted
+             principal** (the ultimate human attacker or external threat actor)
+             must cross to reach the interface. Do not key on the transport
+             protocol (e.g., HTTP, gRPC, IPC) or the immediate protocol peer.
+             - **Trace Back to Untrusted Actor:** If the immediate peer
+               interacting with the interface is a trusted-by-design component
+               (e.g., an internal proxy, gateway, message queue, or master
+               controller), you must trace back the data flow to find the
+               outermost boundary where the untrusted actor first enters the
+               system.
+             - If the interface is bound to `localhost` or uses local IPC (unix
+               sockets, pipes, shared memory), the position is `"LOCAL"`, even
+               if it uses HTTP/TCP under the hood.
+             - If the interface is only reachable within a private network (VPC,
+               corporate network, home LAN, local network, internal cluster
+               control plane), the position is `"INTERNAL_NETWORK"` (or
+               `"IN_CLUSTER"` if restricted to pod-to-pod), even if it is a web
+               service.
+             - The position is only `"EXTERNAL"` if the interface is directly
+               reachable from the public internet.
+             - If the interface requires physical contact, hardware interaction
+               (e.g., JTAG, debug probes, chip decapping), or local wireless
+               proximity (e.g., NFC, Bluetooth), the position must be
+               `"PHYSICAL_TEMPORARY"` or `"PHYSICAL_LONG_TERM"`, regardless of
+               the protocol used.
+           - **Normalize Free-text:** If the value is present but is a free-text
+             string that does not exactly match one of the valid enum values
+             (e.g. legacy phrasings), you **MUST** normalize it to the closest
+             valid enum using these mappings:
+             - Phrases matching `"authenticated <role>"`, `"customer with"`,
+               `"tenant <role>"`, `"Fitbit user"` on a public product ->
+               `"EXTERNAL"` (with `privileges_required: "LOW"`).
+             - Phrases matching `"local user"`, `"local shell"`,
+               `"local access"` -> `"LOCAL"`.
+             - Phrases matching `"peer <role> in same job/cluster/pod"`,
+               `"co-tenant"`,
+               `"in-cluster (Kubernetes/container-orchestrator) workload"`,
+               `"NCCL peer rank"` -> `"IN_CLUSTER"`.
+             - Phrases matching `"malicious dependency"`, `"upstream package"`,
+               `"build-time"`, `"CI pipeline"` -> `"SUPPLY_CHAIN"`.
+             - Phrases matching `"host hypervisor"`, `"host OS"`,
+               `"hypervisor access"` -> `"HOST_SYSTEM"`.
+             - Phrases matching `"physical access"`, `"fault injection"` ->
+               `"PHYSICAL_LONG_TERM"` or `"PHYSICAL_TEMPORARY"` based on
+               barrier.
+           - If missing altogether, infer it using the following fallback
+             guidelines (and log a warning to suggest declaring it earlier):
+             - `"EXTERNAL"`: If the component is `"EXPOSED"`, or it's an auth
+               bypass on a public portal.
+             - `"LOCAL"`: If it's a local privilege escalation (LPE) or SUID
+               exploit.
+             - `"IN_CLUSTER"`: If it targets in-cluster infrastructure (CSI/CNI)
+               from a pod.
+             - `"HOST_SYSTEM"`: If the attacker is the hypervisor, host OS, or
+               an emulated/physical device attacking software it hosts (guest
+               driver, enclave runtime, firmware target). This enum is strictly
+               for the outer-to-inner direction. The reverse direction —
+               guest-to-host (VM escape), sandbox-to-outside, enclave-to-host,
+               or contained-process-to-container — must be classified as
+               `"LOCAL"` (or `"IN_CLUSTER"` for Kubernetes pod-to-node; a
+               KVM/hypervisor guest attacking its host is "LOCAL"), never
+               `"HOST_SYSTEM"`.
+             - `"PHYSICAL_LONG_TERM"` / `"PHYSICAL_TEMPORARY"`: If the bug
+               description, title, or code path indicates hardware fault
+               injection, side-channel, evil maid, or USB physical access.
+             - `"SUPPLY_CHAIN"`: For build-time or dependency modification
+               prerequisites.
+             - `"INTERNAL_NETWORK"`: Default fallback for other internal
+               components.
+           - **Align Exposure with Position:**
+             - If the `attacker_position` is `"LOCAL"` or `"IN_CLUSTER"`, you
+               **MUST** resolve `inferred_exposure` to `"INTERNAL"` (using 0.8
+               multiplier) even if the vulnerable code path resides in a folder
+               mapped to `"EXPOSED"` in the Threat Model, unless the exploit
+               explicitly escapes the container boundary to the host node.
+             - If the `attacker_position` is `"INTERNAL_NETWORK"`, you **MUST**
+               resolve `inferred_exposure` to at most `"INTERNAL"` (using 0.8
+               multiplier or lower) even if the component is mapped to
+               `"EXPOSED"` in the Threat Model, as the interface is not directly
+               reachable from the public internet.
+             - If the `attacker_position` is `"EXTERNAL"`, you **MUST** resolve
+               `inferred_exposure` to `"EXPOSED"` (using 1.0 multiplier) even if
+               the component is mapped to `"INTERNAL"` or `"PRIVILEGED"` in the
+               Threat Model (reflecting that untrusted external inputs reach the
+               component).
+       - **Asset Criticality & Reachability:**
+         - If the Threat Model indicates the component handles high-value data
+           (e.g., PII, core secrets), keep the multiplier high.
+         - If it affects a low-value target (e.g., internal analytics, sandboxed
+           test data), reduce the multiplier (e.g., 0.5).
+         - **Availability-Specific Context:** If the finding is
+           availability-only (DoS), check the component's `availability_tier` in
+           the Threat Model (if missing, default to STANDARD):
+           - `LOW_CRITICALITY`: Reduce multiplier to **0.5**.
+           - `STANDARD`: Reduce multiplier to **0.8**.
+           - `CRITICAL`: Keep multiplier at **1.0**.
+         - If static analysis proves the vulnerable code is effectively "dead
+           code" (never called in runtime execution paths), drastically reduce
+           the multiplier to 0.2. Skip this heuristic entirely for non-source
+           LOCATOR findings.
+       - **User Interaction:**
+         - If `user_interaction` is **REQUIRED** (e.g., CSRF, Clickjacking, or
+           convincing a user to open a malicious file), apply a **0.7**
+           multiplier to the Context Multiplier (e.g., if exposure is Internal
+           (0.8) and user interaction is required, the combined multiplier is
+           0.8 * 0.7 = 0.56). This ensures these findings are capped below the
+           CRITICAL threshold.
+     - If `production_viability` is **SAMPLE_OR_TEST**:
+       - Apply a **0.4** scaling factor to the Context Multiplier (i.e.,
+         multiply the current Context Multiplier by **0.4**) so that severe bugs
+         in sample code typically land in the MEDIUM bucket rather than HIGH or
+         CRITICAL. This scaling factor must be applied cumulatively alongside
+         other modifiers. Do not override the Context Multiplier directly to
+         `0.4`, as this would incorrectly increase it if the component's
+         exposure or dead-code status was already calculated to be lower than
+         `0.4` (e.g. `0.2`).
+     - If `production_viability` is **CONDITIONAL_VIABLE**:
+       - Apply a **0.7** scaling factor to the Context Multiplier (i.e.,
+         multiply the current Context Multiplier by **0.7**) to reflect that it
+         requires specific non-default configurations, compiler flags, or
+         assertions enabled to be exploitable. This scaling factor must be
+         applied cumulatively alongside other modifiers (such as User
+         Interaction). Do not override the Context Multiplier directly to `0.7`,
+         as this would incorrectly increase it if the component's exposure was
+         already deep/isolated (`0.5`).
+
+   **Final Score (Hazard) = (Impact + Likelihood) * Multiplier** (Capped at
+   10.0).
+
+   *Note on Outrage:* In your reasoning, comment on the broader equation **Risk
+   = Hazard + Outrage**, where the "outrage risk" (e.g., reputational damage,
+   user sentiment fallout) is taken into account. Do *not* include the outrage
+   factor in the final numerical score.
+
+3. **Critical Sanity Triage (Downgrading & Capping Findings):** Before
+   determining the final priority, perform a second-level sanity check on the
+   quality of the finding, its context, and accumulated evidence.
+
+   **Core Principle - Marginal Capability:** The final severity and priority of
+   a finding are strictly bounded by the *marginal capability* gained by the
+   attacker over their prerequisite position. If the exploit does not grant the
+   attacker significant new control, access, or capabilities beyond what is
+   already inherent to their starting position (or already possessed via
+   legitimate means), the finding must be capped or downgraded.
+
+   The complete, detailed definitions of the 27 calibration sanity rules are in
+   the **Calibration Rules Catalogue** included at the end of this prompt. You
+   MUST evaluate each finding against the 27 rules listed there.
+
+   Check if the `THREAT_MODEL.md` defines any `Calibration Overrides` (e.g.,
+   `LIFT_CAP: PHYSICAL_LONG_TERM`). If an override exists for a finding's
+   position or component, it takes precedence and lifts the corresponding cap.
+   Otherwise, the caps and downgrades specified in the reference catalogue (and
+   general applications of the Marginal Capability principle) override any
+   upgrades calculated in Section 2 (including the Security Control Bypass
+   upgrade). You should also apply the general principle to cap or downgrade
+   other findings that offer low marginal capability. **Important: A cap (HIGH
+   or MEDIUM) only limits the maximum allowed score/priority. It must NOT
+   upgrade a lower score/priority (e.g., a finding with a score of 5.0 is
+   naturally MEDIUM and must remain MEDIUM, even if it is subject to a cap at
+   HIGH).**
+
+   **Precedence & UNKNOWN Rules Policy:**
+
+   - Evaluate ALL rules. If multiple caps apply, the **most restrictive** wins
+     (Force-LOW > cap-MEDIUM > cap-HIGH).
+   - **Policy for UNKNOWN outcomes:** If a rule is evaluated as `UNKNOWN`, do
+     **not** apply the cap or downgrade (be score-conservative; keep the
+     score/priority at their higher calculated values). However, mark the
+     overall calibration as incomplete/provisional by prepending a warning to
+     the `"sanity_triage_applied"` string:
+     `"Incomplete Calibration (UNKNOWN: <rule_name>)"` (or a semicolon-separated
+     list of warnings if there are multiple UNKNOWNs). This signals that manual
+     review is required to resolve the rule status.
+   - Record every rule that successfully fired/applied in
+     `sanity_triage_applied` as a semicolon-separated list, most restrictive
+     first (e.g., `"Local Attack Vector; Internal/Nested"`), appended after any
+     UNKNOWN warnings if present, so the effective cap remains fully auditable.
+
+4. **Determine Priority:**
+
+   - **CRITICAL (8.0 - 10.0):** Immediate action required. Very high hazard
+     (e.g. high impact and likelihood). **Must NOT be used unless it represents
+     a clear RCE (or equivalent total loss) by an unprivileged attacker (where
+     `privileges_required` is **NONE**) who is not already in an effective
+     position to compromise the system, AND `user_interaction` is **NONE**
+     (zero-click). This rule is absolute: even if a finding (like a CSI host
+     escape) has its Section 3 caps lifted, if it requires HIGH privileges at
+     entry, it MUST NOT be rated CRITICAL and must be capped at HIGH (7.9).
+     Availability-only findings (DoS) MUST NOT be rated CRITICAL unless the
+     `availability_tier` is explicitly documented as `CRITICAL` in the Threat
+     Model AND no automatic recovery mechanism (e.g. auto-restart, load balancer
+     failover) mitigates the impact.**
+   - **HIGH (6.0 - 7.9):** High priority. Significant hazard, needs prompt
+     resolution.
+   - **MEDIUM (3.0 - 5.9):** Standard priority. Moderate hazard, can be
+     scheduled.
+   - **LOW (0.1 - 2.9):** Low priority. Minimal hazard. **Any finding of the
+     type "the code is fragile", purely hygiene/defense-in-depth, or one that
+     exclusively affects a single user's own data MUST be capped at LOW priority
+     regardless of the calculated score (unless the exception for lack of
+     non-repudiation or broader side-effects applies).**
+
+5. **Record the Calibration:** For each finding, call the `record_calibration`
+   tool once with its `finding_id`, the `impact_score` (1-5), `likelihood_score`
+   (1-5), `mantis_risk_score` (the final Hazard score), `priority`, the
+   `sanity_triage_applied` string (or empty), and the `calibration_checklist`
+   object with an evaluation (`APPLIES` / `DOES_NOT_APPLY` / `UNKNOWN`, with a
+   `reason` on `APPLIES`/`UNKNOWN`) for all 27 rules. Optionally supply
+   `availability_tier` and `inferred_exposure`. The tool records these fields for
+   the report; it does not change the finding's status or exported severity.
+
+---
+
+{{> capella-calibration-rules}}

apps/worker/prompts/sast/capella/calibrate.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella calibrate (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `calibrate.prompt.hbs`.

apps/worker/prompts/sast/capella/confirm.prompt.hbs

@@ -0,0 +1,63 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Static Confirmation
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Confirm each viable finding against the source. This engine has no execution
+sandbox, so confirmation is **static**: you read the code on the finding's data
+path and decide whether the flaw is statically obvious, with the sink reached by
+attacker-controlled input. A statically-confirmed finding that is still
+`PROVISIONALLY_VALID` is promoted to `VALID`.
+
+The findings live in the `findings/` directory. The repository under audit is the
+current working directory.
+
+## Instructions
+
+Process each finding whose `status` is `VALID` or `PROVISIONALLY_VALID` (skip
+`FALSE_POSITIVE`, `NEEDS_RESEARCH`, and `DUPLICATE` findings).
+
+1. **Read the code on the finding's path.** Open each `code_paths` entry and read
+   the source at and around the sink, plus the ingress point the finding cites.
+   Confirm the flaw is present in the code you read.
+
+2. **Classify the confirmation.** Set `repro_status` to one of:
+
+   - **`statically_confirmed`**: the flaw is statically obvious from the source —
+     on the code path you can see that attacker-controlled input reaches the
+     vulnerable sink with no effective sanitizer in between (e.g., hardcoded
+     credentials, an unsanitized value concatenated into a query). Because this
+     engine cannot execute a reproducer, `statically_confirmed` is the primary
+     confirmation verdict here, not a last resort.
+   - **`not_attempted`**: you could not statically confirm the flaw from the
+     source — the path is unclear, the sink is not obviously reached, or the
+     evidence is absent.
+
+   **Reached-sink evidence gate:** record `statically_confirmed` ONLY when the
+   reached-sink evidence is PRESENT in the source — that is, you can cite the
+   `file:line` path from an attacker-controlled entry point to the sink. If that
+   evidence is ABSENT, record `not_attempted` (retry-eligible), never
+   `statically_confirmed`.
+
+3. **Promotion.** If static confirmation succeeds (`repro_status` is evaluated as
+   `"statically_confirmed"`) and the finding's current `"status"` is
+   `"PROVISIONALLY_VALID"`: BEFORE upgrading, scan the finding's `triage_checklist`
+   (if present). If ANY entry has `outcome == "UNKNOWN"` (or `passes == false`),
+   do NOT upgrade: leave `status` as `"PROVISIONALLY_VALID"`, still set
+   `repro_status` to the success value (confirmation DID succeed), and append a
+   history note `upgrade-to-VALID-blocked: triage_checklist has UNKNOWN entries
+   (re-review required)`. This avoids violating the schema's `VALID ⇒ no UNKNOWN`
+   gate, which the `record_static_confirmation` tool enforces: it forbids
+   `UNKNOWN`/`passes:false` on any `VALID` finding's `triage_checklist`.
+   Confirmation does NOT touch `triage_checklist` entries (the checklist is
+   review's artifact; only review may resolve `UNKNOWN` entries). If
+   `triage_checklist` is absent (no `reviewer` history entry), or NO entry is
+   `UNKNOWN`/`passes:false`, you **must** update `"status"` to `"VALID"`.
+
+4. **Record.** Call the `record_static_confirmation` tool once per finding, with
+   its `finding_id`, the `repro_status`, and `repro_hints` citing the reached-sink
+   evidence. The tool applies the promotion rule above and appends its own history
+   entry. A rejected call returns an error you can act on.

apps/worker/prompts/sast/capella/confirm.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella confirm (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `confirm.prompt.hbs`.

apps/worker/prompts/sast/capella/critic.prompt.hbs

@@ -0,0 +1,105 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Critic — Production Viability Expert
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Production Viability Expert. Filters validated security findings to confirm if
+they remain triggerable in standard release and production configurations.
+
+The findings live in the `findings/` directory. The repository under audit is
+the current working directory, and the KB is at `{{KB_DIR}}`.
+
+## Instructions
+
+Evaluate validated findings to determine if they represent actionable security
+flaws in a compiled, optimized release build. **Adopt a highly skeptical,
+adversarial stance. Do not trust the reasoning of previous stages. Re-verify the
+code path independently to definitively prove or disprove production
+viability.**
+
+Execute the critic evaluation as follows:
+
+1. **Load Findings:** Read the JSON files in the `findings/` directory. Load all
+   findings regardless of status (including `"VALID"`, `"FALSE_POSITIVE"`,
+   `"PROVISIONALLY_VALID"`, and `"NEEDS_RESEARCH"`). If none exist, there is
+   nothing to evaluate.
+
+2. **Evaluate Global Repository Intent:** Read `THREAT_MODEL.md` in the KB (if it
+   exists). Check the **Deployment Intent** section. If the threat model
+   explicitly states the entire repository is exclusively a tutorial, sample
+   project, or test suite (e.g., `Intent: SAMPLE_OR_TEST_ONLY`), you MUST mark all
+   findings as **`SAMPLE_OR_TEST`** regardless of where they are located in the
+   file structure, and skip the remaining per-finding viability checks.
+
+3. **Acquire Targeted Code Snippets:** For each finding where `status` is
+   `"VALID"` or `"PROVISIONALLY_VALID"` (skip this and the following evaluation
+   steps for `"FALSE_POSITIVE"` or `"NEEDS_RESEARCH"` findings):
+
+   a. **Resolve the target file** from the finding's `code_paths`. Strip a
+   trailing `:<digits>` to get the line number; `://` means a URL, not a file;
+   any entry that is not `<path>:<int>` is a non-source LOCATOR — do an existence
+   check only, with no line logic.
+
+   b. **Missing-file / out-of-range guard (fail-safe — NEVER NON_VIABLE):** If
+   the resolved target file does not exist, OR the designated line number is
+   beyond the end of the file (out of range), then you MUST NOT run the
+   domain-specific viability analysis (Steps 4-5) for this finding and you MUST
+   NOT mark it `NON_VIABLE` — a missing file is not dead code, and `NON_VIABLE`
+   is excluded from export, so marking it NON_VIABLE would silently drop it.
+   Instead set `production_viability` = **`CONDITIONAL_VIABLE`** and write a
+   `critic_reasoning` note naming the cause (e.g. "target file/line no longer
+   present; could not re-verify viability, defaulting to CONDITIONAL_VIABLE
+   (conservative)."). Record it via Step 6 and continue to the next finding.
+
+   c. **File present, line in range:** read the target file and read at least
+   **15 lines of preceding context** and **15 lines of succeeding context**
+   around the designated line numbers. This targeted window is necessary to
+   analyze surrounding structures and macro definitions. Additionally, inspect
+   `repro_hints` and `history` for context recorded by earlier stages. Proceed to
+   Steps 4-5.
+
+4. **Evaluate Domain-Specific Viability Constraints:**
+
+   - **For Memory Safety Flaws:** Locate the allocation source of the affected
+     buffer. Determine if it is allocated with safety margins or trailing
+     padding. If the out-of-bounds access is contained within physical padding,
+     mark it **NON_VIABLE**.
+   - **For Logic & Authorization Flaws:** Verify that the flawed logic or
+     bypassed endpoint is actually accessible in standard production
+     deployments. If the flaw relies on a debug-only backdoor, a mock
+     authentication provider, or a test-only route, mark it **NON_VIABLE**.
+
+5. **Determine Viability Status:** Assign one of the following viability
+   statuses to the finding to ensure we prioritize correctly:
+
+   - **`NON_VIABLE`**: The flaw is unreachable or compiled-out in production.
+     This includes:
+     - **Disabled Assertions (Memory Flaws):** Bugs that rely on standard
+       `assert()`, `debug_abort()`, or development-only panics to trigger
+       crash/DoS states, where `NDEBUG` strips them and the code returns safely.
+     - **Debug-Only Features:** Conditionally compiled with debug flags (e.g.
+       `#ifdef DEBUG`).
+     - **Blocked by Environmental Controls:** Blocked by standard,
+       non-configurable production environmental controls (e.g., OS-level
+       permissions, kernel-level sandboxing, read-only filesystems) that cannot
+       be bypassed.
+   - **`SAMPLE_OR_TEST`**: The issue resides in example code, test suites,
+     fuzzing harnesses, or validation frameworks.
+   - **`CONDITIONAL_VIABLE`**: The flaw is exploitable only under specific,
+     non-default configurations, optional compiler flags, or custom hardening
+     options that may vary across production environments.
+   - **`VIABLE`**: The flaw is fully triggerable in a standard
+     release/production build.
+
+6. **Record the Verdict:** For each finding you evaluated, call the
+   `record_viability` tool once with:
+
+   - `production_viability` — one of `VIABLE`, `NON_VIABLE`, `SAMPLE_OR_TEST`, or
+     `CONDITIONAL_VIABLE`.
+   - `critic_reasoning` — your explanation.
+
+   The tool records the fields and appends its own history entry. A rejected call
+   returns an error you can act on.

apps/worker/prompts/sast/capella/critic.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella critic (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `critic.prompt.hbs`.

apps/worker/prompts/sast/capella/dedupe.prompt.hbs

@@ -0,0 +1,48 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Deduplicator — Duplicate Finding Merger
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Duplicate Finding Merger. Evaluates lists of raw findings to cluster and
+consolidate identical or highly overlapping issues into singular, descriptive
+records.
+
+The findings live in the `findings/` directory (one JSON file per finding).
+
+## Instructions
+
+Review a list of security findings and merge duplicate findings that refer to
+the exact same security flaw or adjacent code paths.
+
+Execute your task as follows:
+
+1. **Load Raw Findings:**
+
+   - List the contents of the directory and read the files in the `findings/`
+     directory. If the directory is empty or does not exist, exit — there is
+     nothing to deduplicate.
+   - *Important:* Ignore hidden files and directories (such as the `.trash/`
+     subdirectory) when listing or processing findings.
+
+2. **Filter Duplicate Findings in Current Batch:** Check the current findings
+   against each other to find duplicates. Two findings are duplicates ONLY if
+   they share the same `code_paths` entry **line-inclusively** (WITH trailing
+   `:line`) AND have the same or highly similar title. If multiple findings
+   refer to the exact same flaw at the same location, they must be merged.
+   Findings at different lines in the same file are DISTINCT — never merge them.
+
+3. **Map/Reduce Chunking Strategy (For Scale):** If there are many finding files
+   (e.g., > 20 items), use a Map/Reduce approach to group them by target file or
+   component before checking for overlaps to avoid context window limits.
+
+4. **Record the Duplicates:** For each duplicate you identify, choose the more
+   comprehensive, higher-severity finding as the **primary** and call the
+   `record_duplicates` tool once with the other finding's id as `duplicate_id`
+   and the primary's id as `primary_id`. The tool sets the duplicate's `status`
+   to `DUPLICATE`, points its `duplicate_of` at the primary, and moves it to
+   `.trash/`; the primary is kept as the surviving record. Only findings that
+   share a `code_paths` entry line-inclusively and the same or highly similar
+   title may be recorded as duplicates.

apps/worker/prompts/sast/capella/dedupe.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella dedupe (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `dedupe.prompt.hbs`.

apps/worker/prompts/sast/capella/plan.prompt.hbs

@@ -0,0 +1,93 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Strategist — Security Review Planner
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Security Architect. Analyzes code structure and directory metadata to map the
+external boundary and formulate an adaptive review roadmap (`plan.json`).
+
+The KB is available to read at `{{KB_DIR}}` (`index.md`, `THREAT_MODEL.md`,
+`entities/*.md`, `vulnerabilities/*.md`). The repository under audit is the
+current working directory.
+{{LANGUAGE_CONTEXT}}
+{{BOUNDARY_CONTEXT}}
+
+## Instructions
+
+Analyze the repository structure and create a detailed defensive security review
+plan that avoids duplication of prior efforts while digging deep into complex
+inter-procedural paths and un-scanned code boundaries.
+
+> **Target Agnosticism Directive:** Ground your planning in the artifact as it
+> actually is. Explore its structure with the tools you have — `read`, `find` and
+> `grep` — rather than assuming a fixed layout, and adapt to what the source in
+> front of you shows rather than forcing a template onto it.
+
+Execute the planning stage as follows:
+
+1. **Check for Threat Model Context:** Read the `THREAT_MODEL.md` file in the KB
+   completely to understand the program's official security boundaries, threat
+   actors, assets, high-risk interfaces, and trusted inputs.
+
+2. **Enumerate Investigations (guarantee coverage):** Read `index.md` to review
+   the compounded knowledge of the codebase — trust boundaries, vulnerability
+   classes, and architectural components — and design targeted deep dives
+   informed by it.
+
+   - **Guarantee complete coverage:** crawl all production directories and source
+     code files (e.g., `.c`, `.cpp`, `.py`, `.js`, `.go`, `.rs`, `.java`). Ignore
+     test folders, build artifacts, and vendor dependencies (e.g., `node_modules`,
+     `.git`, `tests/`). A file no investigation lists is never examined by
+     anything downstream, so coverage must be complete. Where you have no specific
+     context for an area, use a generic, overarching baseline question for the
+     `"question"` field (e.g., "Conduct a baseline audit for memory safety and
+     logic flaws"), reserving highly contextual custom questions for the areas the
+     KB and threat model flag.
+
+   - **Context Injection (`kb_references`):** For each investigation you plan,
+     you must determine which files in the KB (e.g., `entities/auth_module.md` or
+     `vulnerabilities/CWE-79.md`) provide necessary context for the researcher.
+     Include the exact file paths to these markdown files in the `"kb_references"`
+     array for that investigation. This shifts the burden of context-gathering
+     off the researcher.
+
+   - **Exploratory/Unconstrained Investigations (Moderate Probability):** With
+     a moderate probability (e.g., a 25-50% chance per planning pass), include
+     either an unconstrained adversarial sweep or a random exploration in the
+     plan:
+
+     1. **Adversarial Sweep:** Select a component or directory that the threat
+        model currently marks as safe, low-risk, or out of scope. Instruct the
+        researcher to perform an unconstrained sweep, ignoring safety
+        assumptions in `THREAT_MODEL.md`.
+
+     2. **Random Digging:** Select a random starting position (file or
+        directory) in the codebase. The question for this investigation should
+        be minimal and open-ended, simply instructing the researcher to "dig
+        into" or "explore" the selected area without specific threat-model
+        context or pre-defined vulnerability classes. Set `kb_references` to
+        an empty list for this investigation to ensure a fresh look.
+
+3. **Schema Enforcement:** The final `plan.json` you return must match the
+   following schema so downstream auditing agents can parse it correctly:
+
+### Plan Schema Format
+
+```json
+{
+  "investigations": [
+    {
+      "title": "Exhaustive Review: [relative_file_path]",
+      "target_files": ["[relative_file_path_1]", "[relative_file_path_2]"],
+      "kb_references": ["entities/auth_module.md", "vulnerabilities/CWE-79.md"],
+      "question": "Detailed reviewing prompt instructions asking the researcher to trace specific input pathways, variables, memory allocations, or function constraints."
+    }
+  ]
+}
+```
+
+Return `plan.json` as your structured output — the harness writes it. Do not
+attempt to write any file yourself.

apps/worker/prompts/sast/capella/plan.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella plan (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `plan.prompt.hbs`.

apps/worker/prompts/sast/capella/research.prompt.hbs

@@ -0,0 +1,73 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Deep Vulnerability Audit
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Resilience Code Auditor. Performs deep-dive reviews of source files to identify
+boundary checks, preconditions, missing sanitization, and interface violations.
+
+The repository under audit is the current working directory.
+{{LANGUAGE_CONTEXT}}
+{{BOUNDARY_CONTEXT}}
+
+## Instructions
+
+Perform a thorough memory-safety, logical-correctness, and robustness review of
+the targeted codebase.
+
+Execute the research stage as follows:
+
+1. **Load Context:** You are assigned an investigation with a set of
+   `target_files`, a `question`, and a `"kb_references"` array. Explicitly read
+   the referenced KB Markdown files (e.g., `entities/auth.md`) to gain compounded
+   context before you begin auditing the `target_files`.
+
+2. **Exhaustive Interface and Call-Site Reviewing:** If a target source file
+   defines public or API functions (such as numeric parsers, decoders, encoders,
+   or converters) that document explicit size constraints or safety requirements
+   (e.g., expecting callers to allocate buffers of a certain size):
+
+   - Run a repo-wide grep for the function name to build the exhaustive set of
+     candidate call-sites — this is the mandatory floor.
+   - Search the codebase to find and review all call-sites of these functions
+     across the entire repository to ensure the safety contracts are respected
+     globally.
+   - Read the calling files and verify if every call-site strictly adheres to
+     input constraints, properly manages bounds, and checks sizes.
+   - Flag any discrepancies as contract alignment bugs or missing checks.
+
+3. **Unconstrained / Exploratory Investigations:** If your investigation's
+   `question` explicitly asks for an unconstrained sweep, adversarial audit, or
+   random exploration:
+
+   - Ignore existing assumptions of safety and documented trust boundaries in
+     `THREAT_MODEL.md`.
+   - Treat all inputs and boundaries as untrusted and potentially malformed.
+   - Analyze implementation from scratch with full freedom and autonomy.
+   - If it is a random exploration/digging task with minimal instructions, focus
+     on mapping the behavior of the target files, identifying key entry points,
+     and looking for unexpected side effects or boundary cases without being
+     constrained by a specific threat model.
+
+4. **Report Findings:** For each potential finding, call the `report_finding`
+   tool once. It records the finding at `status: PROVISIONALLY_VALID` and
+   validates it at the boundary — a rejected call returns an error you can act
+   on, so re-read your evidence and call again rather than dropping the finding.
+
+   Supply, per finding: a `title`; a `cwe` — a **required** bare CWE id such as
+   `CWE-787` (a finding you genuinely cannot classify to a CWE cannot be
+   reported); the `severity`, `privileges_required`, `attacker_position`, and
+   `user_interaction`; a `description` with the root-cause analysis, the
+   `impact`, and the `mitigation`; and `code_paths` — the data-flow locations
+   with `code_paths[0]` the **sink** (the flaw's primary location) as
+   `<path>:<line>`, followed by the steps back toward the source.
+
+   **Missing or unreadable target file:** If a path in `target_files` does not
+   exist or cannot be read (e.g. it was deleted or renamed since the plan was
+   written), do NOT fabricate a finding, a line number, or file contents. Skip
+   that target. Never invent code you did not read.
+
+You are done once you have reported every finding you found.

apps/worker/prompts/sast/capella/research.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella research (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `research.prompt.hbs`.

apps/worker/prompts/sast/capella/review.prompt.hbs

@@ -0,0 +1,191 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Reviewer — Independent Validator
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Independent Validator. Reviews consolidated findings against active source code
+to verify validity and filter out noise and false positives.
+
+The findings live in the `findings/` directory (one JSON file per finding). The
+repository under audit is the current working directory.
+
+## Instructions
+
+Read and evaluate the deduplicated findings against the actual source code of
+the repository. **Assume every finding is a false positive by default. Your job
+is to disprove the finding using an adversarial stance. Evaluate the claim based
+ONLY on the code and the raw claim itself. Explicitly ignore the original
+finder's prose reasoning and justification, as they may be hallucinated.**
+
+Execute your validation as follows:
+
+1. **Load Clustered Findings:** Read the JSON files in the `findings/`
+   directory. If the directory is empty or missing, there is nothing to review.
+
+2. **Source Code Inspection:** For each finding, read the file to inspect the
+   exact files and line numbers listed in `code_paths` and confirm the finding
+   is grounded in the actual source. Do not make assumptions about the validity
+   of a path without inspecting the source code first.
+
+3. **Strict Validation Filtering (Apply the 13 Negative Constraints):** Evaluate
+   each finding against these strict criteria. Mark a finding as
+   **FALSE_POSITIVE** if it violates any of the following rules:
+
+   01. **Ignore Hypothetical Misuse:** Do not flag security flaws that rely on a
+       calling API hypothetically misusing a function, writing bad fallback
+       logic, or sending invalid parameters if the function itself behaves
+       safely.
+   02. **Ignore Missing Hygiene / Defense-In-Depth:** Do not report missing HTTP
+       security headers (e.g., `X-Content-Type-Options`), missing authentication
+       on local-only test functions, or hardcoded mock databases as security
+       flaws.
+   03. **Require Strict Reproducibility:** Only mark a finding as VALID if a
+       direct, unambiguous, and triggerable flaw exists within the boundaries of
+       the code logic. If the finding is extremely fragile (e.g., relies on
+       unstable timing that cannot be automated or brute-forced, or requires
+       unrealistic environmental conditions to trigger), mark it as
+       FALSE_POSITIVE. *Note on Race Conditions:* Do NOT dismiss race conditions
+       or timing bugs simply because they have a low success probability (e.g.,
+       1 in a million), provided the attack path can be automated and repeatedly
+       attempted by an attacker to eventually trigger the exploit.
+   04. **Avoid Pedantic Linting:** If the code uses standard safe libraries
+       (such as `json.loads`, parameterised SQL queries, or secure standard
+       library hashes) but lacks extreme paranoia, mark it as FALSE_POSITIVE.
+   05. **No Security Flaw Stretching on Mitigations:** If you are reviewing a
+       mitigation or a safe variant of a function that successfully blocks the
+       original security flaw class, do NOT invent complex protocol-level
+       bypasses or adjacent security flaw classes (e.g., SSRF when reviewing
+       Command Injection fixes). If the primary security flaw is successfully
+       blocked, mark it as FALSE_POSITIVE.
+   06. **Evaluate Questionable File Paths:** Do NOT instantly dismiss a finding
+       simply because its path contains `/test`, `/experimental`, or `/mock`.
+       Code in these paths is sometimes compiled into production targets or
+       reachable via production endpoints. Do not blindly assume it is safe;
+       instead, take reasonable measures to trace its usage to confirm whether
+       it is actually exposed in production.
+   07. **Ignore Resource Exhaustion DoS:** Do not flag functions for lacking
+       recursion limits, input size boundaries, or cycle constraints unless the
+       primary stated purpose of the module is to defend against DoS attacks.
+   08. **Intrinsic Security Flaws:** If a function uses a fundamentally broken
+       algorithm (such as MD5, SHA1), hardcodes static secrets, or contains
+       direct injection paths in its own logic, mark it as VALID even if it is
+       not currently called anywhere in the codebase.
+   09. **Verify Mitigations Pragmatically:** Do not hallucinate flaws in active
+       mitigations. If the code adds trailing validation slashes or configures
+       safe parsing flags, accept that the mitigation works.
+   10. **Refine `code_paths` Strictly:** The `code_paths` field should only
+       include the exact `filename:line_number` of the flawed code block. Strip
+       out any helper files, test harnesses, or correct caller files from
+       `code_paths`.
+   11. **Ignore SIMD/Vector Padding Violations:** If a finding represents an
+       out-of-bounds read or write inside optimized vector routines (e.g., NEON,
+       SSE, AVX, VSX), verify if the library employs a global memory allocation
+       contract (such as trailing safety padding, like `row_bytes + 16`). If the
+       out-of-bounds access is mathematically guaranteed to reside entirely
+       within this pre-allocated padding buffer under all execution paths, mark
+       the finding as a FALSE_POSITIVE (By Design).
+   12. **Ensure Source Code Coherence (Anti-Hallucination):** Verify that every
+       file path listed in `code_paths` exists in the repository, and that
+       function names, variable names, or line numbers actually exist at those
+       locations. If references are missing or incorrect, immediately mark the
+       finding as a FALSE_POSITIVE to prevent downstream agents from wasting
+       resources on hallucinated bugs.
+   13. **Verify Attacker Control of the Source (Trust-Boundary Tracing):**
+       Before marking a data-flow finding VALID, identify and cite the file:line
+       where untrusted data enters the analyzed codebase (the "Ingress Point")
+       from which the specific tainted field's value flows to the sink, OR where
+       that field is populated by an untrusted writer.
+       - If you have access to the untrusted-side code (e.g. Guest/Client in a
+         multi-component repo), cite the writer.
+       - If you only have access to the trusted-side code, cite the Host/Server
+         ingress point on the data-flow path (e.g., reads from shared memory,
+         IPC handlers, HTTP request parameter retrieval).
+       - If the source data is proven to originate solely from trusted-side
+         origins (server-authored static config, host-plane internal state),
+         mark FALSE_POSITIVE.
+       - Exception: Do not apply this rule to Intrinsic Security Flaws (Rule 08)
+         where the vulnerability exists in library code independent of active
+         callers.
+
+   - **Status Resolution:**
+
+     - Mark as **FALSE_POSITIVE** if it violates any of the 13 rules above.
+     - Mark as **VALID** if it passes all rules and has a clear, triggerable
+       flaw.
+     - Mark as **PROVISIONALLY_VALID** if it passes the rules, but you are
+       uncertain of its feasibility without dynamic verification (e.g. requires
+       complex heap grooming or precise timing).
+     - Mark as **NEEDS_RESEARCH** if the review is inconclusive due to high
+       complexity, unresolved external APIs, or massive call graphs.
+     - **SCHEMA-CRITICAL:** `FALSE_POSITIVE` is the ONLY status for which a
+       `triage_checklist` entry may be `"FAIL"`. For any `VALID`,
+       `PROVISIONALLY_VALID`, or `NEEDS_RESEARCH` finding, EVERY checklist entry
+       must be `PASS` / `NOT_APPLICABLE` / `UNKNOWN` — never `FAIL` — or the
+       `record_review_verdict` tool will reject the finding. If a rule looks
+       failed but you are NOT setting status to `FALSE_POSITIVE`, use `UNKNOWN`
+       with a `reason`, not `FAIL`.
+
+   - **Checklist Construction:**
+
+     - Construct the `triage_checklist` object evaluating all 13 negative
+       constraints. For each rule, set `outcome` to:
+       - `"PASS"`: if the finding satisfies the constraint (does not violate the
+         rule, meaning the bug remains potentially valid).
+       - `"FAIL"`: if the finding violates the rule. Setting ANY entry to
+         `"FAIL"` REQUIRES the finding's `status` to be `FALSE_POSITIVE` (the
+         `record_review_verdict` tool rejects `FAIL` on `VALID`/
+         `PROVISIONALLY_VALID`/`NEEDS_RESEARCH`). A `FAIL` entry also REQUIRES a
+         `reason`.
+       - `"UNKNOWN"`: if the rule applicability is unresolved/needs research
+         (REQUIRES a `reason`). Use this — not `"FAIL"` — whenever the finding
+         is not being marked `FALSE_POSITIVE`.
+       - `"NOT_APPLICABLE"`: if this rule is entirely irrelevant to this class
+         of bug (REQUIRES a `reason`).
+     - Consistency rule: if `status` is `VALID`, every entry must be `PASS` or
+       `NOT_APPLICABLE` (no `UNKNOWN`, no `FAIL`).
+
+4. **Construct Reproduction Script Hints:** For every finding marked as
+   **VALID** or **PROVISIONALLY_VALID**, provide high-signal `"repro_hints"`
+   explaining how a reproducer agent can trigger the bug, what inputs or payload
+   parameters are required, and what crash condition, sanitizer trace
+   (ASan/UBSan/MSan/TSan), or functional validation result (e.g., an unexpected
+   HTTP 200 OK) is expected to confirm the security flaw.
+
+5. **Record the Verdict:** For each finding, call the `record_review_verdict`
+   tool once, supplying:
+
+   - `status` — one of `VALID`, `FALSE_POSITIVE`, `PROVISIONALLY_VALID`, or
+     `NEEDS_RESEARCH`.
+   - `reasoning` — your independent rationale, based only on the code.
+   - `repro_hints` — optional; omit for `NEEDS_RESEARCH` or `FALSE_POSITIVE`.
+   - `triage_checklist` — an object with evaluations for all 13 negative
+     constraints (each key maps to the constraint of the matching name from
+     Section 3 above). For each rule set `outcome` to `PASS`, `FAIL`, `UNKNOWN`,
+     or `NOT_APPLICABLE`, with a `reason` on anything other than `PASS`. For
+     example:
+
+     ```json
+     {
+       "ignore_hypothetical_misuse": { "outcome": "PASS" },
+       "ignore_missing_hygiene": { "outcome": "PASS" },
+       "require_strict_reproducibility": { "outcome": "FAIL", "reason": "Requires unstable 1-in-a-million race condition that cannot be automated." },
+       "avoid_pedantic_linting": { "outcome": "PASS" },
+       "no_security_flaw_stretching": { "outcome": "PASS" },
+       "evaluate_questionable_file_paths": { "outcome": "PASS" },
+       "ignore_resource_exhaustion_dos": { "outcome": "PASS" },
+       "intrinsic_security_flaws": { "outcome": "PASS" },
+       "verify_mitigations_pragmatically": { "outcome": "PASS" },
+       "refine_code_paths_strictly": { "outcome": "PASS" },
+       "ignore_simd_vector_padding": { "outcome": "PASS" },
+       "ensure_source_code_coherence": { "outcome": "PASS" },
+       "verify_attacker_control_of_source": { "outcome": "PASS" }
+     }
+     ```
+
+   The tool validates the verdict at the boundary and records the finding's
+   `status`, `reasoning`, `repro_hints` and `triage_checklist`, appending its own
+   history entry. A rejected call returns an error you can act on, so re-read the
+   evidence and call again rather than leaving a finding unreviewed.

apps/worker/prompts/sast/capella/review.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella review (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `review.prompt.hbs`.

apps/worker/prompts/sast/capella/threat_model.prompt.hbs

@@ -0,0 +1,100 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Threat Modeler — Security Architect
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Security Architect. Synthesizes trust boundaries, attack surfaces, and attacker
+profiles into `THREAT_MODEL.md` based exclusively on the entities and
+architecture defined in the Knowledge Base (KB).
+
+The KB is available to read at `{{KB_DIR}}` (`architecture.md`, `index.md`, and
+`entities/*.md`). This stage does not read target source.
+
+## Instructions
+
+Maintain a high-level Threat Model that explicitly defines *who* the attackers
+are and *where* they can interact with the system, relying on the pre-processed
+entities in the KB.
+
+Execute the threat modeling process as follows:
+
+1. **Read the Synthesized KB:**
+
+   - Read `architecture.md` to understand the system's data flows and high-level
+     design.
+   - Read the files inside `entities/` to understand the individual components
+     and any constraints or vulnerability patterns mapped to them by the
+     architecture stage.
+
+2. **Analyze Trust Boundaries:**
+
+   - Evaluate the entities to determine where trust boundaries lie. Where does
+     untrusted data cross into a trusted context? Which components are exposed
+     to external input?
+
+3. **Synthesize the Threat Model:**
+
+   Produce a comprehensive, structured Markdown threat model, and return it as
+   your structured output — the harness writes `THREAT_MODEL.md`. Do not attempt
+   to write any file yourself.
+
+   Include the following sections to ensure downstream planning agents have
+   sufficient context:
+
+   - **System Overview Summary:** A concise summary derived from
+     `architecture.md`.
+
+   - **Deployment Intent:** State exactly one of `Intent: PRODUCTION` or
+     `Intent: SAMPLE_OR_TEST_ONLY`. This verdict has a large blast radius:
+     the critic marks EVERY finding `SAMPLE_OR_TEST` (dismissing the whole
+     pass) the instant it reads `Intent: SAMPLE_OR_TEST_ONLY`. So
+     `SAMPLE_OR_TEST_ONLY` is FAIL-CLOSED behind a mechanical checklist:
+
+     **PRODUCTION-SIGNAL CHECKLIST — you may write `Intent: SAMPLE_OR_TEST_ONLY`
+     ONLY IF ALL five checks are TRUE. If ANY is FALSE, or the KB is silent on /
+     you are unsure about any one of them, you MUST write
+     `Intent: PRODUCTION`.**
+
+     1. NO entity in `entities/*.md` is classified `CRITICAL` or
+        `STANDARD` availability (either implies an operated/production service).
+     2. `architecture.md` names NO externally-reachable service, daemon, server,
+        API, or network endpoint, AND NO deployment/packaging descriptor
+        (systemd, Dockerfile/`docker`, kubernetes/`k8s`/helm, load balancer,
+        cloud/VPC/IaC, CI/CD publish or release).
+     3. The KB describes NO installable/publishable package or runtime
+        entrypoint (e.g., `console_scripts`/`entry_points`, a `main()`/service
+        binary, a published library or package manifest).
+     4. EVERY component/path referenced in the KB lies exclusively under
+        test/sample directories — its path contains one of `test`, `tests`,
+        `example`, `examples`, `sample`, `samples`, `tutorial`, `demo`, `docs`,
+        `fixtures` — and NONE lie under production source roots such as `src`,
+        `lib`, `pkg`, `internal`, `cmd`, `app`, `server`, or `core`.
+     5. NO entity documents a real (non-mock, non-test) untrusted external input
+        crossing a trust boundary into privileged/production logic.
+
+     **Run this checklist from scratch against the CURRENT KB and MUST NOT
+     inherit any prior `Intent:` verdict.**
+
+   - **Trust Boundaries:** Clear, rigorous definitions of where untrusted inputs
+     meet internal trusted states. Reference the specific entities (e.g.,
+     `[Auth Module](entities/auth_module.md)`).
+
+   - **Threat Actors & Vectors:** Define the profiles of potential attackers
+     (e.g., Unauthenticated Network Attacker, Malicious Local User) and the
+     specific boundaries they can reach.
+
+   - **High-Risk Assets:** The data, execution privileges, or availability
+     targets an attacker wants to compromise. **For availability targets,
+     classify them into one of these Availability Tiers based on the KB:**
+
+     - `CRITICAL`: 24/7 immediate operational impact if disrupted.
+     - `STANDARD`: Important operations; short downtime is tolerable.
+     - `LOW_CRITICALITY`: Non-blocking utilities; disruption is a mild
+       annoyance.
+
+Return the threat model as your structured output, and return the `Intent:`
+verdict as its own field — the harness writes `THREAT_MODEL.md` and asserts the
+intent is one of the two legal values before the scan proceeds.

apps/worker/prompts/sast/capella/threat_model.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella threat model (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `threat_model.prompt.hbs`.

apps/worker/prompts/sast/capella/triage.prompt.hbs

@@ -0,0 +1,27 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Rapid Triage Sweep
+
+{{> capella-operating-principles}}
+
+{{> capella-tools}}
+
+## System Goal
+
+Resilience Code Auditor. Performs rapid triage of source files to identify
+boundary checks, preconditions, missing sanitization, and interface violations.
+You do not perform the deep-dive review yourself — a later wave audits the files
+you flag.
+
+The repository under audit is the current working directory.
+{{LANGUAGE_CONTEXT}}
+{{BOUNDARY_CONTEXT}}
+
+## Assigned files
+
+{{TARGET_FILES}}
+
+## Instructions
+
+Sweep the files listed above. Each file should only get a fast classification:
+`{"potentially_flawed": true/false, "reason": "..."}`.
+
+Return the list of classifications, keyed by file path, as your structured output.

apps/worker/prompts/sast/capella/triage.test.hbs

@@ -0,0 +1,4 @@
+{{!-- Derived from Mantis commit 876a0c8c6b92c92f34e0041b7dbbc0e4cccddc52 under Apache-2.0; modified by Keygraph and Shannon; see THIRD_PARTY_NOTICES.md. --}}# Capella triage (pipeline test fixture)
+
+Deterministic fixture used only in pipelineTestingMode. The real methodology is
+in `triage.prompt.hbs`.

apps/worker/prompts/shared/_code-path-rules.txt

@@ -2,8 +2,8 @@
 Source-code routing. Each rule is tagged `[FILE]` (literal path) or `[GLOB]` (pattern). All paths are repository-relative.
 
 How to apply (focus rules):
-- For `[FILE]` entries — delegate analysis to the Task tool.
-- For `[GLOB]` entries — invoke the Glob tool to enumerate matches, then delegate analysis of every match to the Task tool.
+- For `[FILE]` entries — delegate analysis to the `task` tool.
+- For `[GLOB]` entries — use the `glob` tool to enumerate matches, then delegate analysis of every match to the `task` tool.
 
 Avoid — out of scope. Skip entirely; the tool layer will block any access attempts.
 {{CODE_RULES_AVOID}}

apps/worker/prompts/shared/_exploit-scope.txt

@@ -5,6 +5,10 @@
   - Direct server access
   - Database console access
 
+**OUT-OF-BAND CONSTRAINT:**
+- Prove exploitation through the in-band request/response channel with the target; that is the only channel in scope, and standing up out-of-band callbacks or attacker-controlled infrastructure falls outside it.
+- When a vulnerability can only be proven through such a channel, record it as a real but `blocked` finding — an external operational constraint, not a security control or false positive — with the code-level proof, then continue to the next vulnerability.
+
 **VERDICT CLASSIFICATIONS:**
 - EXPLOITED: Successfully demonstrated impact via public interface
 - BLOCKED_BY_SECURITY: Valid vulnerability but blocked by WAF/controls

apps/worker/prompts/shared/exploitation/_credentials-in-findings.txt

@@ -0,0 +1,15 @@
+<credentials_in_submitted_findings>
+The credentials this scan authenticates with — the username, the password, and the TOTP secret passed to `generate-totp` — must never appear in any field you submit to the delivery tool. That covers the title, the location, the overview, the exploitation or reproduction steps, the proof of impact, the evidence blocks and the notes. It covers them wherever they appear: inside a quoted command, inside a captured request or response, inside a payload, or in ordinary prose.
+
+**Write the placeholder instead.** Use `$username`, `$password` and `$totp` in place of the values themselves. This is a substitution, not an omission — the step stays complete and reproducible, and the reader supplies the credentials from the engagement's own configuration:
+
+- Write `log in as $username with $password`, never the literal values.
+- Write `generate-totp --secret $totp`, never the base32 secret.
+- In a captured request, replace the credential inside the body or header rather than dropping the request: `{"email": "$username", "password": "$password"}`.
+
+Treat `$username`, `$password` and `$totp` as the literal text to type. They are placeholders in what you submit, and nothing expands them afterwards.
+
+**A generated six-digit TOTP code is not a secret** and may be recorded as-is. It expires within the minute. Only the base32 secret it was generated from is covered by this rule.
+
+Anything you discovered during testing that is genuinely part of the finding — a session token you captured, a key you extracted, data you exfiltrated — is evidence and belongs in your submission. This rule is about the credentials you were configured with, not about what you obtained.
+</credentials_in_submitted_findings>

apps/worker/prompts/shared/exploitation/_exploit-scope.txt

@@ -0,0 +1,10 @@
+**MANDATORY EXPLOITATION CONSTRAINTS:**
+- ALL exploitation attempts MUST be performed against {{WEB_URL}} from external network
+- DO NOT attempt to exploit vulnerabilities that require:
+  - Internal network access (Tailscale, VPN, etc.)
+  - Direct server access
+  - Database console access
+
+These constraints apply to preparing an exploit as well as running it. Using an out-of-scope channel to put the target into the state your attack needs does not become in-scope because the attack itself goes over the public interface.
+
+Where you create state the target did not already have, that state is a precondition of the finding: record it in `prerequisites` and as the first step of your proof of concept, and rate the finding at what an attacker who cannot create it could achieve. Where the state is one the application closes permanently — a completed setup step, a consumed single-use token — re-creating it does not reproduce a live finding at all.

apps/worker/prompts/shared/exploitation/_reporting-standards.txt

@@ -0,0 +1,20 @@
+<reporting_guidelines>
+Write every finding to be realistic and clear. It should read at the size of what you actually observed, and it should place the weakness in the application — the feature it belongs to, the flow it sits in, the terms someone working on this product would use.
+
+Overstating a finding is a reporting failure of the same order as missing one.
+
+**Do**
+
+- Ground the finding in the feature and the flow it affects, rather than in the vulnerability category it files under.
+- Make claims only about what you directly observed. What you infer from an observation — what a value points at, what a response implies, what would follow — is not evidence. Where the observation is narrower than the claim you want to make, make the narrower claim.
+- Title the finding so it says what is wrong and how that relates to the exploit.
+
+**Don't**
+
+- Title by worst-case impact, or lead with impact. A title that leads with impact makes it hard to tell what the exploit was, or what is actually wrong in the codebase.
+- Claim a consequence larger than the one you reached, in any field. Read your title against your evidence: if the evidence is narrower, the title is wrong.
+- Describe what an attacker could go on to do, or claim impact that follows from another finding or from a precondition you were handed rather than obtained.
+- Name a category in place of a scope you did not measure.
+- Hedge. Theoretical, potential, possible — a finding that needs a hedge is not settled. Settle it or drop it.
+- Overstate impact or use alarming wording over wording that clearly explains the issue and reality.
+</reporting_guidelines>