gstack/AGENTS.md

gstack — AI Engineering Workflow

gstack is a collection of SKILL.md files that give AI agents structured roles for software development. Each skill is a specialist: CEO reviewer, eng manager, designer, QA lead, release engineer, debugger, and more.

Available skills

Skills live in .agents/skills/ (or ~/.claude/skills/gstack/ on Claude Code). Invoke them by name (e.g., /office-hours). This index lists names only; what each skill does, with examples, is in docs/skills.md. The iOS companion CLIs (gstack-ios-qa-daemon, gstack-ios-qa-mint, gstack-ios-qa-regen) are in docs/howto-ios-testing-with-gstack.md.

Skill	Area
/office-hours	plan
/plan-ceo-review	plan
/plan-eng-review	plan
/plan-design-review	plan
/plan-devex-review	plan
/plan-tune	plan
/autoplan	plan
/design-consultation	plan
/spec	plan
/review	build + review
/deslop-shared-libs	build + review
/write-skill	build + review
/test-audit	build + review
/codex	build + review
/claude-code	build + review
/investigate	build + review
/verify	build + review
/design-review	build + review
/design-shotgun	build + review
/design-html	build + review
/design-motion	build + review
/devex-review	build + review
/qa	build + review
/qa-only	build + review
/scrape	build + review
/skillify	build + review
/ship	release + deploy
/land-and-deploy	release + deploy
/modernize	release + deploy
/canary	release + deploy
/landing-report	release + deploy
/document-release	release + deploy
/document-generate	release + deploy
/setup-deploy	release + deploy
/gstack-upgrade	release + deploy
/context-save	operational + memory
/context-restore	operational + memory
/learn	operational + memory
/retro	operational + memory
/health	operational + memory
/captain	operational + memory
/benchmark	operational + memory
/ubicloud	operational + memory
/runaway-stopper	operational + memory
/benchmark-models	operational + memory
/eval-plan	operational + memory
/cso	operational + memory
/setup-gbrain	operational + memory
/sync-gbrain	operational + memory
/browse	browser
/open-gstack-browser	browser
/setup-browser-cookies	browser
/pair-agent	browser
/ios-qa	iOS QA
/ios-fix	iOS QA
/ios-design-review	iOS QA
/ios-clean	iOS QA
/ios-sync	iOS QA
/careful	safety + scoping
/freeze	safety + scoping
/guard	safety + scoping
/unfreeze	safety + scoping
/make-pdf	safety + scoping
/diagram	safety + scoping

Validation discipline

When fixing failures or preparing /ship, follow this order:

1. List the known failing cases, their logs and source revision, the demonstrated cause, and the smallest check that can prove each repair. Keep one current list in .context/; update it instead of starting overlapping repair plans. Reconcile the runner's failure total with named failures and unhandled or module-load errors; the named-test footer alone is not the complete inventory.
2. Resolve base-branch integration and assign one owner per shared file before editing. Keep repairs within the observed failures and the user's scope. Before a fixture writes through a link, resolve its target and verify it stays inside that fixture's temporary root; live skill registrations can point back into this checkout. When upstream replaces a helper API, inventory every direct caller, mock adapter, source snapshot, generated golden, and selection edge before choosing focused checks. Verify extracted test adapters supply the current imports and result schema; an adapter failure is not evidence that production failed. Schedule independent checks independently. Gate a check only on inputs or prerequisites it actually needs; an unrelated failure must not serialize the whole validation plan. Keep source fixed while tests live-link its files.
3. Diagnose before changing code. Distinguish a product defect, an invalid test expectation, a detector/fixture defect, and a launch/environment failure. Preserve the original failure. Do not call it pre-existing without evidence. Verify pinned runtime tool schemas and defaults before treating omitted fields as model noncompliance. Check that a bounded evaluation’s fixture scope and automated answers support its metric. Do not let the driver approve unrelated expansion, then blame the skill for the extra work; preserve required findings and evidence limits.
4. Reproduce with the smallest relevant test. For agent tests, reuse captured public events in free regressions, including negative controls, before paying for another agent run. Check behavior and acknowledgments; match exact prose only when that prose is the contract. Do not lower thresholds, increase model budgets, skip cases, or rejudge a failure to manufacture a pass. A pre-registered fixed panel is not rejudging. For any measured speedup, regression or eval delta, name the limiter that moved and the run-to-run variation; a gap smaller than that variation is no difference. For policy or validation repairs, exercise the actual registered callback with representative native input and assert that it uses the helper’s result. When renderer or parser failures recur at the same boundary, verify the supported input class against the pinned runtime. Keep adversarial controls; do not add one spelling or glyph per paid failure. For workflow clarity failures, read the complete evaluated excerpt and its referenced source. Resolve all demonstrated ambiguities together: order, definitions, ownership and approval. Consolidate dense instructions into executable steps instead of appending more clauses. Review the resulting workflow as a whole; prose snapshots alone do not prove it is clear. For each gate, identify when its inputs exist and trace normal, skipped/unavailable and late-change paths to catch circular prerequisites or bypassed checks.
5. Run required cheap CI checks, including credential scanning, before paid work. Also run adjacent cheap checks: generated-content freshness, prompt-size/parity limits, source assertions, fixture checks, and dependency selection as applicable. A changed prompt must clear these before its eval. For skill edits, include bun test test/parity-suite.test.ts: its historical union-size cap is separate from the other prompt-size and context budgets. When workflow wording changes, search the entire test tree for removed clauses, including always-loaded prompt guards. Test fixtures containing subprocess examples must pass test/spawnsync-timeout-tripwire.test.ts; its scanner also checks quoted code. Run its selected quality judge before long behavioral evaluations that read the same changed prompt. If a repair supersedes an active run's inputs, cancel that run, preserve completed outcomes, and label unfinished cases as cancelled. Check each edit or setup command’s result before running dependent checks. A failed edit is not a reason to test the unchanged input again.
6. Declare a fixture actor’s supported interactions before the model starts. Keep its answers and permission handling within that declared interface. Bind artifact ownership to the same isolated state passed to the child; ambient environment paths do not establish ownership. Check whole-file and CI supervision against every case and configured retry, not just one attempt. Preflight the actual launcher: required binaries, isolated state, display when needed, explicit test tier, selection, and expected executed-case counts. Match the runtime versions pinned by the workflow and its container image. Keep socket-bearing temporary paths short after the runner adds its nested directories; exercise that exact layout in the smoke check. Store long-lived logs separately from socket directories. Verify required tool execution with a no-cost smoke check under that launch environment; versions and authentication alone do not prove it works. Set private artifact modes explicitly and preserve normal fixture permissions. Prove a diagnostic snapshot survives fixture cleanup in the final artifact directory before paid work; native snapshots require EVALS_RUN_ID or GSTACK_EVAL_DIR. Bind complete spool filenames and classify Bun's out-of-tier describe.skip placeholders separately, with zero selected-case credit. Put standalone Git fixtures outside another checkout; verify their resolved project slug and state root before interpreting a failure. Prove a seed commit succeeds there: repository-local author configuration does not establish the identity available to a fresh fixture repository. Reject missing explicit test files before invoking Bun; it can silently ignore a nonexistent file selector and pass the remaining files. Preserve exit status through logging. Use the documented detached runner and eval lock. Review the final launcher after edits; preparation and --list modes must not start monitors, retainers, or test processes. Verify this with a before/after process check. During long runs, inspect the last public tool result and pending permission state; diagnose a blocked actor before waiting through its deadline. Preserve cancellation separately from a test verdict. Skipped or unstarted cases do not satisfy coverage; preserve every attempt. Paid evals never retry, with one exception: the gate reruns a red rule, behavior or judge case verdict once at the same head (Garry, 2026-10-19: "you can just run the failed test again and note that it is flaky"); a pass is green with a flaky annotation and a flake record, a second failure is a real red, and harness, INFRA, INCOMPLETE or unattributed reds never rerun (docs/TESTING_INTERNALS.md#flaky-reruns). Each case's kind (E2E_KINDS) fixes its trials before the run: rule one trial; behavior a panel of 3 independent trials, PASS at >= 2 with no contract violation; judge 3 samples on one output, each dimension gated on its median (at least 2 of 3 samples) against the unchanged threshold. Never add trials, samples or dispatches after seeing a result, never change a kind to change a verdict without pass-rate evidence, and report every trial. Quarantine follows CASE_QUARANTINE's entry and exit rules only (EVAL_POLICY, docs/TESTING_INTERNALS.md). A census whose every red is machine-classified INFRA or INCOMPLETE may be re-dispatched once as a new run; report both runs.
7. Prove all known repairs with focused tests, including affected paid cases. Rerun a failed case only after a concrete repair or a demonstrated launch correction. Run the remaining required selected evaluations on the integrated code. Do not use the full free suite to discover predictable adjacent failures. Reuse a passing check when its consumed inputs and relevant environment are unchanged. For model judges, compare the expanded prompt, rubric, parameters and dependencies; a different commit alone does not invalidate the result. Do not resample an unchanged passing judge to simplify launcher configuration. Preserve its original source and label the result as reused evidence. Use actual prompt builders and compare complete bytes when proving model-input identity; preserve literal text in excerpts and record the consumed inputs.
8. Finish review fixes, generation, release metadata, and build before final acceptance. Freeze the code, then run bun run test once at the end. During repair, focused checks replace a full-suite run before every commit. If final acceptance unexpectedly fails, retain the failure, diagnose it narrowly, and report the changed validation plan before another full run; never retry it blindly or claim a pass from an older revision.
9. Publish only with passing required checks, unless the user explicitly grants an exception for identified failures. Report revision, actual pass/fail/skip counts, and incomplete coverage. A passing subset is not release acceptance.

Build commands

bun install              # install dependencies
bun run typecheck        # strict tsc over product code; must report zero errors
bun run typecheck:test   # test-code type-debt ratchet (new diagnostics fail; --write-baseline locks in fixes)
bun run format:cso       # format lib/cso/*.ts (format:cso:check is the CI gate)
bun run test:quick       # fast measured free subset for edit feedback (not acceptance)
bun run test             # complete free suite via the strict shard runner (no API spend)
bun run eval:bg:pr       # changed live probes + selected judges; dispatches CI when HEAD is clean and pushed, else runs locally
bun run eval:bg:release  # fresh complete gate + periodic live coverage (same backend choice)
bun run test:health      # audit success metrics and weekly health from CI history (free; needs gh)
bun run build            # generate docs + compile binaries
bun run gen:skill-docs   # regenerate SKILL.md files from templates
bun run skill:check      # health dashboard for all skills

Every other test and eval command (Ubicloud, the Windows subset, one paid tier or case, branch validation in CI, pass rates, plan previews), with its cost and prerequisites, is in Which command do I run?. Agents poll eval:bg:* logs for the ### gstack-detach EXIT=<code> ### sentinel (CLAUDE.md).

Platform support

Support tiers (full: macOS, Linux; minimal: Windows, Gemini models) are in docs/support-tiers.md.

• macOS + Linux: full test suite supported.
• Windows: minimal support; docs/support-tiers.md says what is promised. Every PR must pass the windows-smoke check (./setup, the core skills start, the fastest Windows-safe tests); the curated Windows-safe subset runs nightly on main in windows-free-tests and never blocks a PR. Setup script (./setup) requires Git Bash or MSYS today; native PowerShell support is a future expansion. The bin/gstack-paths helper resolves state roots through CLAUDE_PLUGIN_DATA / GSTACK_HOME so plugin installs work on every platform.
• CSO native helper: /cso additionally needs Bun's four --no-compile-autoload-* flags and a native toolchain (static-capable C on Linux, Xcode command-line tools on macOS, or VS 2022 C++ Build Tools on Windows). Setup installs everything else and leaves /cso explicitly unavailable when that optional toolchain is absent.
• Browser and renderer: the browser skills, /make-pdf, and /diagram drive the Aside browser first, which is macOS 15+ only. On Linux and Windows (or a Mac with Aside closed) the readiness check says so once and the same skills use gstack's own bundled browser, built by ./setup.

Key conventions

• SKILL.md files are generated from .tmpl templates. Edit the template, not the output.
• Run bun run gen:skill-docs --host codex to regenerate Codex-specific output.
• Browser steps in skills are aside repl scripts per scripts/resolvers/aside.ts (Aside first), each with a $B equivalent for the fallback engine — $B <command> is the browse binary and is a legitimate tool when the Aside probe does not print READY. Local HTML renders through bin/gstack-render.ts, which picks the same way.
• Safety skills (careful, freeze, guard) use inline advisory prose — always confirm before destructive operations.
• State paths resolve through one chain owned by lib/state-root.ts and its sourced bash twin bin/gstack-state-root.sh (GSTACK_STATE_ROOT → GSTACK_HOME → GSTACK_STATE_DIR → gstack's CLAUDE_PLUGIN_DATA → ~/.gstack; see docs/state-root.md). Skill prose uses GSTACK_STATE_ROOT=$(bin/gstack-paths --get GSTACK_STATE_ROOT) with its ${GSTACK_STATE_ROOT:?…} guard (worktree-isolated Claude Code sessions refuse eval); test/state-root-ratchet.test.ts rejects hand-rolled chains.
• Browse daemon HTTP routes are entries in browse/src/routes/table.ts (its header shows how to add one); never dispatch on url.pathname in server.ts.
• Both test lanes run shards through scripts/lib/shard-engine.ts; the free and paid runners hold lane policy only. PTY harness code lives in test/helpers/pty/* behind the claude-pty-runner.ts barrel.
• Outside-voice failure prose (auth, timeout, empty, fallback) comes only from outsideVoiceFailurePolicy() in scripts/resolvers/outside-voice.ts.
• Every env-var-host fence starts with the shared prelude from scripts/resolvers/runtime-root.ts (inserted by its post-render pass); never resolve gstack's root by hand in a template.
• Every Codex run's verdict goes through lib/outside-review-result.ts; its reason codes live in lib/gate-outcomes.ts, each with a docs/troubleshooting.md anchor.
• test/module-size-ratchet.test.ts keeps refactored owner modules at or under 800 lines (150 per function) and residual files from growing.
• The claude CLI binary resolves via lib/claude-bin.ts (re-exported from browse/src/claude-bin.ts for browse internals; Bun.which() + GSTACK_CLAUDE_BIN override). Set GSTACK_CLAUDE_BIN=wsl plus GSTACK_CLAUDE_BIN_ARGS='["claude"]' to run Claude through WSL on Windows.
• Every Bun gstack starts ignores the working directory's .env files and bunfig.toml: bun shebangs are #!/usr/bin/env -S bun --no-env-file --config=/dev/null, skill prose passes bun --no-env-file --config=/dev/null, shipped shell files define GSTACK_BUN_CONFIG with the line in test/bun-entry-isolation.test.ts (NUL when OSTYPE is msys or cygwin and cygpath is on PATH, where native Bun cannot open /dev/null, #3415) and pass bun --no-env-file --config="${GSTACK_BUN_CONFIG:-/dev/null}", TypeScript spawns use bunIsolationArgs() from lib/bun-isolation.ts, and bun build --compile adds --no-compile-autoload-dotenv --no-compile-autoload-bunfig. test/bun-entry-isolation.test.ts enforces all four.