Files
shannon/apps/worker/src/temporal/workflows.ts
T
ezl-keygraph 5ff40f8c6f feat(worker): migrate agent runtime from Claude Agent SDK to pi harness (#389)
* feat(worker): migrate agent runtime from Claude Agent SDK to pi harness

* feat: remove Google Vertex AI provider support

* fix(worker): route Bedrock and custom-base-URL providers from env

* feat(prompts): instruct agents to call submit_exploitation_queue and submit_auth_result

* fix(worker): count sub-agent cost and surface compaction failures

* refactor(worker): rename claude-executor to pi-executor

* feat(worker): pi-event-driven output formatting

* fix(worker): gate adaptive thinking to Opus models, drop CLAUDE_THINKING_LEVEL

* fix(worker): restore minLength/minItems on vuln-collector schemas

* feat(worker): give task sub-agent write+bash, align tool descriptions

* feat(worker): add glob custom tool and route code_path globs to it

* refactor(prompts): use pi tool names (task, todo_write, read, bash, glob)

* refactor(prompts): drop stale MCP terminology for collector tools

* refactor(prompts): drop collector server names from deliverable instructions

* fix(worker): restore minLength/minItems on pre-recon and exploit collector schemas

* feat(worker): load playwright-cli skill via pi resource loader

* refactor(cli): remove CLAUDE_CODE_MAX_OUTPUT_TOKENS config

* build: drop @anthropic-ai/claude-code from worker image

* docs: remove vertex references from llms context

* docs(worker): update stale sdk comments

* refactor(worker): unify provider precedence between preflight and executor

* feat(worker): enforce bounded bash timeouts via pi extension

* ci: bump the beta release line to 2.0.0 (#356)

* fix(cli): pin npx command hints to beta tag

* fix: render agent deliverables before the success commit so resume preserves them (#377)

* feat(cli): restructure run folder and improve terminal UX (#383)

* feat: surface report at run root and nest run internals under .shannon

* feat: use plain-language wording in user-facing terminal messages

* feat(cli): guide users to watch scan progress and surface report path on start

* docs: sync run-folder layout and CLI wording across docs and comments

* feat(cli): add version command reporting package version or git SHA

* feat(cli): detect TTY for interactive prompts, color, and progress output

* docs: document --yes flag, version command, and tty module

* fix(cli): FORCE_COLOR precedence and plain uninstall --yes output

* fix(cli): respect empty NO_COLOR

* fix(cli): let NO_COLOR take precedence over FORCE_COLOR

* docs: mark claude-code-router integration as removed

* refactor(worker): converge shared core with shannon-oss (#388)

* fix(worker): port keygraph shared-core correctness fixes

* refactor(worker): adopt collectors/ and ai/pi/ layout; add task budget cap and cancellation

* refactor(worker): drop inconsistent Collector "Server" suffix

* refactor(worker): drop unused providerConfig/apiKey seams, resolve credentials from env only

* refactor(worker): port oss code_path pattern expansion + external_directory allow

* fix(worker): preserve dotfile paths in code_path avoid patterns (.env no longer stripped to env)

* feat(worker): render Unprocessed Vulnerabilities section in exploit deliverable (align with oss)

* feat(worker): request set_blind_spots for all vuln classes (align auth/ssrf with production prompts)

* refactor(worker): adopt unified permissionSystem* naming and helper layout

* refactor(worker): inline blind_spots into vuln deliverable section array

* chore(worker): drop unused zod dependency (tree is typebox-native)

* fix(worker): normalize base32 TOTP secret to accept padding and whitespace

* refactor(worker): adopt shared toolResult helper and flatSchema naming in collectors

* refactor(worker): use undefined over null in queue-schema builders

* docs(worker): converge renderer/collector doc comments to current pi terminology

* refactor(worker): adopt schema.ts cleanInput/stringEnum helpers in collectors

* feat(worker): converge exploit-collector/renderer with vendored; capture and render overview for blocked findings

* refactor(worker): converge session-tools/pipeline/exploitation-checker with vendored

* refactor(worker): converge task-tool usage reporting with vendored onUsage callback

* refactor(worker): converge structured output onto a submitTool executor channel

* docs(worker): expand exploit-renderer docstring to match shannon-oss

* docs(worker): adopt richer vuln-renderer docstring from shannon-oss

* docs(worker): neutralize billing-detection wording for shannon-oss parity

* fix(worker): verify checkpoint hash in the deliverables clone being reset

* fix(worker): fail fast on malformed exploitation queue JSON

* fix(worker): honor retryable flag when classifying exploitation-queue check failures

* fix(worker): fail fast on corrupted session.json in run-scope validation

* feat(worker): propagate Temporal cancellation signal into agent and auth pi sessions

* fix(worker): mark exploit agent complete when exploitation is skipped so resume skips it

* prompts: drop scan description from executive report prompt

* refactor(worker): add createGenericSubmitTool for raw JSON-schema submit tools

* refactor(worker): gate playwright-cli skill to browser agents via skillsOverride (adopt shannon-oss mechanism)

* docs(worker): correct formatLogTime comment to UTC to match toISOString

* refactor(worker): converge queue-schemas with shannon-oss (guarded count, decl order)

* refactor(worker): converge task-tool with shannon-oss (byte-identical; modelRegistry optional)

* fix(worker): use replaceLiteral for all prompt value insertions to prevent $-mangling

* fix(worker): classify agent execution failures by error type instead of hardcoding validation

* fix(worker): cap auth-failure detail at 250 chars to match shannon-oss

* style(worker): apply biome formatting

* refactor(worker): remove per-session task delegation cap from task tool

* style(cli): collapse usage hint now that the beta tag is gone

* chore: mark the pi harness migration as a breaking change

BREAKING CHANGE: Google Vertex AI is no longer a supported provider. The
CLAUDE_CODE_USE_VERTEX, ANTHROPIC_VERTEX_PROJECT, CLOUD_ML_REGION, and
GOOGLE_APPLICATION_CREDENTIALS environment variables, along with the
use_vertex, vertex_project, and cloud_ml_region config.toml keys, are
removed. Vertex users must switch to Anthropic, AWS Bedrock, or a custom
Anthropic-compatible base URL.

The CLAUDE_CODE_MAX_OUTPUT_TOKENS environment variable and the
max_output_tokens config.toml key are also removed.
2026-07-16 19:13:13 +05:30

734 lines
28 KiB
TypeScript
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
// Copyright (C) 2025 Keygraph, Inc.
//
// This program is free software: you can redistribute it and/or modify
// it under the terms of the GNU Affero General Public License version 3
// as published by the Free Software Foundation.
/**
* Temporal workflow for Shannon pentest pipeline.
*
* Orchestrates the penetration testing workflow:
* 1. Pre-Reconnaissance (sequential)
* 2. Reconnaissance (sequential)
* 3-4. Vulnerability + Exploitation (5 pipelined pairs in parallel)
* Each pair: vuln agent → queue check → conditional exploit
* No synchronization barrier - exploits start when their vuln finishes
* 5. Reporting (sequential)
*
* Features:
* - Queryable state via getProgress
* - Automatic retry with backoff for transient/billing errors
* - Non-retryable classification for permanent errors
* - Audit correlation via workflowId
* - Graceful failure handling: pipelines continue if one fails
*/
import {
ApplicationFailure,
CancellationScope,
isCancellation,
log,
proxyActivities,
setHandler,
workflowInfo,
} from '@temporalio/workflow';
import type { AgentName, VulnType } from '../types/agents.js';
import { ALL_AGENTS } from '../types/agents.js';
import { ALL_VULN_CLASSES, type VulnClass } from '../types/config.js';
import type * as activities from './activities.js';
import type { ActivityInput } from './activities.js';
import {
type AgentMetrics,
getProgress,
PipelineExecutionError,
type PipelineInput,
type PipelineProgress,
type PipelineState,
type PipelineSummary,
type ResumeState,
type VulnExploitPipelineResult,
} from './shared.js';
import { toWorkflowSummary } from './summary-mapper.js';
import { classifyErrorCode, formatWorkflowError } from './workflow-errors.js';
/** Agents this run is expected to produce — drives the resume short-circuit. */
function computeExpectedAgents(vulnClasses: readonly VulnClass[], exploit: boolean): string[] {
const expected: string[] = ['pre-recon', 'recon'];
for (const cls of vulnClasses) {
expected.push(`${cls}-vuln`);
if (exploit) {
expected.push(`${cls}-exploit`);
}
}
expected.push('report');
return expected;
}
// Retry configuration for production (long intervals for billing recovery)
const PRODUCTION_RETRY = {
initialInterval: '5 minutes',
maximumInterval: '30 minutes',
backoffCoefficient: 2,
maximumAttempts: 50,
nonRetryableErrorTypes: [
'AuthenticationError',
'PermissionError',
'InvalidRequestError',
'RequestTooLargeError',
'ConfigurationError',
'InvalidTargetError',
'ExecutionLimitError',
'AuthLoginFailedError',
],
};
// Retry configuration for pipeline testing (fast iteration)
const TESTING_RETRY = {
initialInterval: '10 seconds',
maximumInterval: '30 seconds',
backoffCoefficient: 2,
maximumAttempts: 5,
nonRetryableErrorTypes: PRODUCTION_RETRY.nonRetryableErrorTypes,
};
// Activity proxy with production retry configuration (default)
const acts = proxyActivities<typeof activities>({
startToCloseTimeout: '2 hours',
heartbeatTimeout: '60 minutes', // Extended for nested pi task execution
retry: PRODUCTION_RETRY,
});
// Activity proxy with testing retry configuration (fast)
const testActs = proxyActivities<typeof activities>({
startToCloseTimeout: '30 minutes',
heartbeatTimeout: '30 minutes', // Extended for sub-agent execution in testing
retry: TESTING_RETRY,
});
// Retry configuration for subscription plans (5h+ rolling rate limit windows)
const SUBSCRIPTION_RETRY = {
initialInterval: '5 minutes',
maximumInterval: '6 hours',
backoffCoefficient: 2,
maximumAttempts: 100,
nonRetryableErrorTypes: PRODUCTION_RETRY.nonRetryableErrorTypes,
};
// Activity proxy for subscription plan recovery (extended timeouts)
const subscriptionActs = proxyActivities<typeof activities>({
startToCloseTimeout: '8 hours',
heartbeatTimeout: '2 hours',
retry: SUBSCRIPTION_RETRY,
});
// Retry configuration for preflight validation (short timeout, few retries)
const PREFLIGHT_RETRY = {
initialInterval: '10 seconds',
maximumInterval: '1 minute',
backoffCoefficient: 2,
maximumAttempts: 3,
nonRetryableErrorTypes: PRODUCTION_RETRY.nonRetryableErrorTypes,
};
// Activity proxy for preflight validation (short timeout)
const preflightActs = proxyActivities<typeof activities>({
startToCloseTimeout: '2 minutes',
heartbeatTimeout: '2 minutes',
retry: PREFLIGHT_RETRY,
});
// Credential rejection is not retryable; transient provider errors get 3 attempts.
const AUTH_VALIDATION_RETRY = {
initialInterval: '10 seconds',
maximumInterval: '1 minute',
backoffCoefficient: 2,
maximumAttempts: 3,
nonRetryableErrorTypes: PRODUCTION_RETRY.nonRetryableErrorTypes,
};
// Browser-driving validation measured at 60180s; 10 min start-to-close leaves headroom for slow SSO/MFA flows.
const authValidationActs = proxyActivities<typeof activities>({
startToCloseTimeout: '10 minutes',
heartbeatTimeout: '10 minutes',
retry: AUTH_VALIDATION_RETRY,
});
/**
* Compute aggregated metrics from the current pipeline state.
* Called on both success and failure to provide partial metrics.
*/
function computeSummary(state: PipelineState): PipelineSummary {
const metrics = Object.values(state.agentMetrics);
return {
totalCostUsd: metrics.reduce((sum, m) => sum + (m.costUsd ?? 0), 0),
totalDurationMs: Date.now() - state.startTime,
totalTurns: metrics.reduce((sum, m) => sum + (m.numTurns ?? 0), 0),
agentCount: state.completedAgents.length,
};
}
const MAX_PIPELINE_ERROR_MESSAGE_LENGTH = 2000;
function truncatePipelineErrorMessage(message: string): string {
if (message.length <= MAX_PIPELINE_ERROR_MESSAGE_LENGTH) {
return message;
}
return `${message.slice(0, MAX_PIPELINE_ERROR_MESSAGE_LENGTH - 20)}\n[truncated]`;
}
/**
* Core pipeline orchestration. Coordinates the pentest pipeline stages.
*
* IMPORTANT: This function uses Temporal workflow APIs internally (proxyActivities,
* queries). It can ONLY be called from within a Temporal workflow execution.
* Do not call from standalone scripts or activity code.
*/
export async function pentestPipeline(input: PipelineInput): Promise<PipelineState> {
// Validate repoPath: reject traversal attempts and require absolute path
if (!input.repoPath || input.repoPath.includes('..')) {
throw ApplicationFailure.nonRetryable(
`Invalid repoPath: path traversal not allowed (received: ${input.repoPath ?? '<empty>'})`,
'ConfigurationError',
);
}
if (!input.repoPath.startsWith('/')) {
throw ApplicationFailure.nonRetryable(
`Invalid repoPath: absolute path required (received: ${input.repoPath})`,
'ConfigurationError',
);
}
const { workflowId } = workflowInfo();
// Select activity proxy based on mode: testing (fast), subscription (extended), or default
function selectActivityProxy(pipelineInput: PipelineInput) {
if (pipelineInput.pipelineTestingMode) return testActs;
if (pipelineInput.pipelineConfig?.retry_preset === 'subscription') return subscriptionActs;
return acts;
}
const a = selectActivityProxy(input);
const state: PipelineState = {
status: 'running',
currentPhase: null,
currentAgent: null,
completedAgents: [],
failedPipelines: [],
failedAgent: null,
error: null,
startTime: Date.now(),
agentMetrics: {},
summary: null,
};
setHandler(
getProgress,
(): PipelineProgress => ({
...state,
workflowId,
elapsedMs: Date.now() - state.startTime,
}),
);
// Build ActivityInput with required workflowId for audit correlation
// Activities require workflowId (non-optional), PipelineInput has it optional
// Use spread to conditionally include optional properties (exactOptionalPropertyTypes)
// sessionId is workspace name for resume, or workflowId for new runs
const sessionId = input.sessionId || input.resumeFromWorkspace || workflowId;
const activityInput: ActivityInput = {
webUrl: input.webUrl,
repoPath: input.repoPath,
workflowId,
sessionId,
...(input.configPath !== undefined && { configPath: input.configPath }),
...(input.outputPath !== undefined && { outputPath: input.outputPath }),
...(input.pipelineTestingMode !== undefined && {
pipelineTestingMode: input.pipelineTestingMode,
}),
// Config fields — flow through to getOrCreateContainer()
...(input.configYAML !== undefined && { configYAML: input.configYAML }),
...(input.deliverablesSubdir !== undefined && { deliverablesSubdir: input.deliverablesSubdir }),
...(input.auditDir !== undefined && { auditDir: input.auditDir }),
...(input.promptDir !== undefined && { promptDir: input.promptDir }),
...(input.sastSarifPath !== undefined && { sastSarifPath: input.sastSarifPath }),
};
const selectedVulnClasses: readonly VulnClass[] =
input.vulnClasses && input.vulnClasses.length > 0 ? input.vulnClasses : ALL_VULN_CLASSES;
const selectedClassSet = new Set<VulnClass>(selectedVulnClasses);
const exploit: boolean = input.exploit ?? true;
const expectedAgents = computeExpectedAgents(selectedVulnClasses, exploit);
await a.persistOrValidateRunScope(activityInput, [...selectedVulnClasses], exploit);
let resumeState: ResumeState | null = null;
if (input.resumeFromWorkspace) {
// 1. Load resume state (validates workspace, cross-checks deliverables)
resumeState = await a.loadResumeState(
input.resumeFromWorkspace,
input.webUrl,
input.repoPath,
input.deliverablesSubdir,
);
// 2. Restore git workspace and clean up incomplete deliverables
const incompleteAgents = ALL_AGENTS.filter(
(agentName) => !resumeState?.completedAgents.includes(agentName),
) as AgentName[];
await a.restoreGitCheckpoint(
input.repoPath,
resumeState.checkpointHash,
incompleteAgents,
input.deliverablesSubdir,
);
// 3. Short-circuit when every agent expected by this run is done.
// Uses dynamic expectedAgents (not ALL_AGENTS) so a class-scoped run completes sooner.
const allExpectedDone = expectedAgents.every((a) => resumeState?.completedAgents.includes(a));
if (allExpectedDone) {
log.info(`All ${expectedAgents.length} expected agents already completed. Nothing to resume.`);
state.status = 'completed';
state.completedAgents = [...resumeState.completedAgents];
state.summary = computeSummary(state);
return state;
}
// 4. Record this resume attempt in session.json and workflow.log
await a.recordResumeAttempt(
activityInput,
input.terminatedWorkflows || [],
resumeState.checkpointHash,
resumeState.originalWorkflowId,
resumeState.completedAgents,
);
log.info('Resume state loaded and workspace restored');
}
const shouldSkip = (agentName: string): boolean => {
return resumeState?.completedAgents.includes(agentName) ?? false;
};
// Run a sequential agent phase (pre-recon, recon)
async function runSequentialPhase(
phaseName: string,
agentName: AgentName,
runAgent: (input: ActivityInput) => Promise<AgentMetrics>,
): Promise<void> {
if (!shouldSkip(agentName)) {
state.currentPhase = phaseName;
state.currentAgent = agentName;
await a.logPhaseTransition(activityInput, phaseName, 'start');
state.agentMetrics[agentName] = await runAgent(activityInput);
state.completedAgents.push(agentName);
if (input.checkpointsEnabled) {
await a.saveCheckpoint(activityInput, agentName, phaseName, state);
}
await a.logPhaseTransition(activityInput, phaseName, 'complete');
} else {
log.info(`Skipping ${agentName} (already complete)`);
state.completedAgents.push(agentName);
}
}
// Build pipeline configs for the 5 vuln→exploit pairs
function buildPipelineConfigs(): Array<{
vulnType: VulnType;
vulnAgent: string;
exploitAgent: string;
runVuln: () => Promise<AgentMetrics>;
runExploit: () => Promise<AgentMetrics>;
}> {
return [
{
vulnType: 'injection',
vulnAgent: 'injection-vuln',
exploitAgent: 'injection-exploit',
runVuln: () => a.runInjectionVulnAgent(activityInput),
runExploit: () => a.runInjectionExploitAgent(activityInput),
},
{
vulnType: 'xss',
vulnAgent: 'xss-vuln',
exploitAgent: 'xss-exploit',
runVuln: () => a.runXssVulnAgent(activityInput),
runExploit: () => a.runXssExploitAgent(activityInput),
},
{
vulnType: 'auth',
vulnAgent: 'auth-vuln',
exploitAgent: 'auth-exploit',
runVuln: () => a.runAuthVulnAgent(activityInput),
runExploit: () => a.runAuthExploitAgent(activityInput),
},
{
vulnType: 'ssrf',
vulnAgent: 'ssrf-vuln',
exploitAgent: 'ssrf-exploit',
runVuln: () => a.runSsrfVulnAgent(activityInput),
runExploit: () => a.runSsrfExploitAgent(activityInput),
},
{
vulnType: 'authz',
vulnAgent: 'authz-vuln',
exploitAgent: 'authz-exploit',
runVuln: () => a.runAuthzVulnAgent(activityInput),
runExploit: () => a.runAuthzExploitAgent(activityInput),
},
];
}
// A rejected settle can be a genuine Temporal cancellation (runVulnExploitPipeline rethrows in
// its isCancellation branch). Cancellation must win over failed/partial classification — there is
// no report worth shipping once the user has cancelled, and rethrowing lets the workflow's outer
// isCancellation handler produce a real cancelled state instead of a hard failure.
function throwIfPipelineCancelled(results: PromiseSettledResult<VulnExploitPipelineResult>[]): void {
const cancelled = results.find(
(r): r is PromiseRejectedResult => r.status === 'rejected' && isCancellation(r.reason),
);
if (cancelled) {
throw cancelled.reason;
}
}
// Classify the settled pipeline results into clean / partial / fail-hard.
// Metrics and completedAgents are updated incrementally inside runVulnExploitPipeline
// so that getProgress queries reflect real-time status during execution.
function aggregatePipelineResults(
results: PromiseSettledResult<VulnExploitPipelineResult>[],
alreadyCompletedPipelineCount: number,
): void {
throwIfPipelineCancelled(results);
const failed: { vulnType: VulnClass; error: string }[] = [];
// A rejected settle is now unexpected (runVulnExploitPipeline catches and returns its error in
// the value). Without a value we cannot attribute the failure to a class, so we treat it as a
// hard failure rather than risk an under-qualified report.
const unattributable: string[] = [];
for (const result of results) {
if (result.status === 'fulfilled') {
if (result.value.error !== null) {
failed.push({ vulnType: result.value.vulnType, error: result.value.error });
}
} else {
const rawMessage = result.reason instanceof Error ? result.reason.message : String(result.reason);
unattributable.push(truncatePipelineErrorMessage(rawMessage));
}
}
const failedCount = failed.length + unattributable.length;
const totalPipelineCount = results.length + alreadyCompletedPipelineCount;
if (failedCount === 0) {
return;
}
// All run pipelines failed, or a failure we cannot attribute to a class → fail-hard. There is
// no report worth shipping, and we must never render an un-assessed class as if it passed.
if (failedCount === totalPipelineCount || unattributable.length > 0) {
const allErrors = [...failed.map((f) => `${f.vulnType}: ${f.error}`), ...unattributable];
const message = `${failedCount} vulnerability/exploitation pipeline(s) failed`;
state.status = 'failed';
state.failedAgent = 'pipelines';
state.error = `${message}: ${allErrors.join('; ')}`;
log.warn(message, { failures: allErrors });
throw ApplicationFailure.nonRetryable(state.error, 'PipelineFailedError', [{ failures: allErrors }]);
}
// Partial: at least one class succeeded and at least one failed. Record the failed classes and
// set the partial terminal status; do NOT throw — the successful pipelines still ship.
state.failedPipelines = failed;
state.status = 'partial';
log.warn(`${failed.length} of ${totalPipelineCount} pipeline(s) failed — continuing with partial results`, {
failures: failed.map((f) => `${f.vulnType}: ${f.error}`),
});
}
// Run thunks with a concurrency limit, returning PromiseSettledResult for each.
// When limit >= thunks.length (default), all launch concurrently — identical to Promise.allSettled.
// NOTE: Results are in completion order, not input order. Callers must key on value fields, not index.
async function runWithConcurrencyLimit(
thunks: Array<() => Promise<VulnExploitPipelineResult>>,
limit: number,
): Promise<PromiseSettledResult<VulnExploitPipelineResult>[]> {
const results: PromiseSettledResult<VulnExploitPipelineResult>[] = [];
const inFlight = new Set<Promise<void>>();
for (const thunk of thunks) {
const slot = thunk()
.then(
(value) => {
results.push({ status: 'fulfilled', value });
},
(reason: unknown) => {
results.push({ status: 'rejected', reason });
},
)
.finally(() => {
inFlight.delete(slot);
});
inFlight.add(slot);
if (inFlight.size >= limit) {
await Promise.race(inFlight);
}
}
await Promise.allSettled(inFlight);
return results;
}
try {
// === Preflight Validation ===
// Quick sanity checks before committing to expensive agent runs.
// NOT using runSequentialPhase — preflight doesn't produce AgentMetrics.
state.currentPhase = 'preflight';
state.currentAgent = null;
await preflightActs.runPreflightValidation(activityInput);
log.info('Preflight validation passed');
// === Playwright stealth config ===
// Write the playwright-cli config before any browser session opens so the
// validator and downstream agents inherit anti-detection defaults.
await preflightActs.syncPlaywrightStealthConfig(activityInput);
// === Authentication Validation ===
state.currentPhase = 'auth-validation';
state.currentAgent = 'validate-authentication';
await authValidationActs.runAuthenticationValidation(activityInput);
state.currentAgent = null;
log.info('Authentication validation passed');
// === Initialize Deliverables Git ===
await a.initDeliverableGit(activityInput);
// === Sync code_path deny rules ===
await a.syncCodePathDenyRules(activityInput);
log.info(`Run scope: vuln_classes=[${selectedVulnClasses.join(', ')}] exploit=${exploit}`);
// === Phase 1: Pre-Reconnaissance ===
await runSequentialPhase('pre-recon', 'pre-recon', a.runPreReconAgent);
// === Phase 2: Reconnaissance ===
await runSequentialPhase('recon', 'recon', a.runReconAgent);
// === Phases 3-4: Vulnerability Analysis + Exploitation (Pipelined) ===
// Each vuln type runs as an independent pipeline:
// vuln agent → queue check → conditional exploit agent
// Exploits start immediately when their vuln finishes, not waiting for all.
state.currentPhase = 'vulnerability-exploitation';
state.currentAgent = 'pipelines';
await a.logPhaseTransition(activityInput, 'vulnerability-exploitation', 'start');
// Closure over shouldSkip and activityInput by design (Temporal replay safety)
async function runVulnExploitPipeline(
vulnType: VulnType,
runVulnAgent: () => Promise<AgentMetrics>,
runExploitAgent: () => Promise<AgentMetrics>,
): Promise<VulnExploitPipelineResult> {
const vulnAgentName = `${vulnType}-vuln`;
const exploitAgentName = `${vulnType}-exploit`;
// A class failure must not reject the pipeline set — that would lose the class identity
// (results are completion-ordered) and force fail-hard. Catch here and return the error in
// the result's `error` field so aggregatePipelineResults can attribute it to `vulnType`.
try {
// 1. Run vulnerability analysis (or skip if resumed)
let vulnMetrics: AgentMetrics | null = null;
if (!shouldSkip(vulnAgentName)) {
vulnMetrics = await runVulnAgent();
state.agentMetrics[vulnAgentName] = vulnMetrics;
state.completedAgents.push(vulnAgentName);
if (input.checkpointsEnabled) {
await a.saveCheckpoint(activityInput, vulnAgentName, 'vulnerability-analysis', state);
}
} else {
log.info(`Skipping ${vulnAgentName} (already complete)`);
state.completedAgents.push(vulnAgentName);
}
// 1.5. Merge external findings from consumer provider into exploitation queue
await a.mergeFindingsIntoQueue(activityInput, vulnType);
// 2. Check exploitation queue for actionable findings
const decision = await a.checkExploitationQueue(activityInput, vulnType);
// 3. Previously-completed exploits are preserved regardless of mode; new exploits gated by mode.
let exploitMetrics: AgentMetrics | null = null;
if (shouldSkip(exploitAgentName)) {
log.info(`Skipping ${exploitAgentName} (already complete)`);
state.completedAgents.push(exploitAgentName);
} else if (decision.shouldExploit && exploit) {
exploitMetrics = await runExploitAgent();
state.agentMetrics[exploitAgentName] = exploitMetrics;
state.completedAgents.push(exploitAgentName);
if (input.checkpointsEnabled) {
await a.saveCheckpoint(activityInput, exploitAgentName, 'exploitation', state);
}
} else {
// Exploitation did not run (exploit mode off, or no actionable findings) — still
// mark the agent complete so a resume does not treat it as unfinished work.
log.info(
`Marking ${exploitAgentName} complete (${decision.shouldExploit ? 'exploit mode disabled' : 'no actionable findings'})`,
);
state.completedAgents.push(exploitAgentName);
if (input.checkpointsEnabled) {
await a.saveCheckpoint(activityInput, exploitAgentName, 'exploitation', state);
}
}
return {
vulnType,
vulnMetrics,
exploitMetrics,
exploitDecision: {
shouldExploit: decision.shouldExploit,
vulnerabilityCount: decision.vulnerabilityCount,
},
error: null,
};
} catch (error) {
// Let cancellation propagate to the workflow-level handler.
if (isCancellation(error)) {
throw error;
}
const rawMessage = error instanceof Error ? error.message : String(error);
const message = truncatePipelineErrorMessage(rawMessage);
log.warn(`Pipeline ${vulnType} failed`, { error: message });
return {
vulnType,
vulnMetrics: state.agentMetrics[vulnAgentName] ?? null,
exploitMetrics: state.agentMetrics[exploitAgentName] ?? null,
exploitDecision: null,
error: message,
};
}
}
const maxConcurrent = input.pipelineConfig?.max_concurrent_pipelines ?? 5;
const pipelineConfigs = buildPipelineConfigs();
const pipelineThunks: Array<() => Promise<VulnExploitPipelineResult>> = [];
let alreadyCompletedPipelineCount = 0;
for (const config of pipelineConfigs) {
// Excluded classes drop entirely; any prior deliverables stay on disk but don't count this run.
if (!selectedClassSet.has(config.vulnType)) {
log.info(`Skipping ${config.vulnType} pipeline (class not selected this run)`);
continue;
}
if (!shouldSkip(config.vulnAgent) || !shouldSkip(config.exploitAgent)) {
pipelineThunks.push(() => runVulnExploitPipeline(config.vulnType, config.runVuln, config.runExploit));
} else {
log.info(`Skipping entire ${config.vulnType} pipeline (both agents complete)`);
state.completedAgents.push(config.vulnAgent, config.exploitAgent);
alreadyCompletedPipelineCount++;
}
}
const pipelineResults = await runWithConcurrencyLimit(pipelineThunks, maxConcurrent);
aggregatePipelineResults(pipelineResults, alreadyCompletedPipelineCount);
state.currentPhase = 'exploitation';
state.currentAgent = null;
await a.logPhaseTransition(activityInput, 'vulnerability-exploitation', 'complete');
// === Phase 5: Reporting ===
if (!shouldSkip('report')) {
state.currentPhase = 'reporting';
state.currentAgent = 'report';
await a.logPhaseTransition(activityInput, 'reporting', 'start');
// First, assemble the concatenated report from per-class deliverables
await a.assembleReportActivity(activityInput, exploit);
// Then run the report agent to add executive summary and clean up
state.agentMetrics.report = await a.runReportAgent(activityInput);
state.completedAgents.push('report');
if (input.checkpointsEnabled) {
await a.saveCheckpoint(activityInput, 'report', 'reporting', state);
}
// Inject model metadata into the final report
await a.injectReportMetadataActivity(activityInput);
await a.logPhaseTransition(activityInput, 'reporting', 'complete');
} else {
log.info('Skipping report (already complete)');
state.completedAgents.push('report');
}
// Runs after the skip gate so consumer providers still execute on resume.
await a.generateReportOutputActivity(activityInput);
if (input.checkpointsEnabled) {
await a.saveCheckpoint(activityInput, 'report-output', 'reporting', state);
}
// Preserve a partial verdict (set by aggregatePipelineResults) — a clean run is 'completed',
// a run where some classes were not assessed is 'partial'.
const terminalStatus: 'completed' | 'partial' = state.failedPipelines.length > 0 ? 'partial' : 'completed';
state.status = terminalStatus;
state.currentPhase = null;
state.currentAgent = null;
state.summary = computeSummary(state);
// Log workflow completion summary
await a.logWorkflowComplete(activityInput, toWorkflowSummary(state, terminalStatus));
return state;
} catch (error) {
// Cancellation: return structured state instead of throwing
if (isCancellation(error)) {
state.status = 'cancelled';
state.error = `Cancelled during phase: ${state.currentPhase ?? 'unknown'}`;
state.summary = computeSummary(state);
// Finalization runs I/O activities; shield them from the cancellation so the
// cancelled state is still logged rather than aborted mid-write.
await CancellationScope.nonCancellable(async () => {
try {
await a.logWorkflowComplete(activityInput, toWorkflowSummary(state, 'cancelled'));
} catch (completionError) {
log.warn('Failed to finalize cancelled workflow', {
error: completionError instanceof Error ? completionError.message : String(completionError),
});
}
});
return state;
}
state.status = 'failed';
state.failedAgent = state.currentAgent;
state.error = formatWorkflowError(error, state.currentPhase, state.currentAgent);
const errorCode = classifyErrorCode(error);
if (errorCode) {
state.errorCode = errorCode;
}
state.summary = computeSummary(state);
// Log workflow failure summary
try {
await a.logWorkflowComplete(activityInput, toWorkflowSummary(state, 'failed'));
} catch (completionError) {
log.warn('Failed to finalize failed workflow', {
error: completionError instanceof Error ? completionError.message : String(completionError),
});
}
// Carry the populated state so a consumer can report real spend instead of a zeroed
// failed state. The original error rides as `cause` for classification/reporting.
throw new PipelineExecutionError(state.error ?? 'Pipeline failed', state, { cause: error });
}
}
/** OSS workflow entry point — thin shell around the extracted pipeline function. */
export async function pentestPipelineWorkflow(input: PipelineInput): Promise<PipelineState> {
return pentestPipeline(input);
}