- Regenerate build/SKILL.md from template after step-transition and
  content-review gate changes
- Add --no-plan-review to integration test invocations that use
  --skip-ship and --skip-clean-check to prevent LLM calls in CI

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…lidator

Captures the follow-on work from the step-transition / living plan parse
investigation: detect structural-mirroring flat-task format before agents
are spawned, rather than after parsing returns 0 executable phases.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…rdering

Address two findings from /review:
- Add stderr assertion to --print-only exit-2 integration test so a future
  unrelated exit-2 can't pass silently (T-1)
- Add comment on ensureFeature() call after finalize() in parser.ts explaining
  the ordering: idempotent for emitted phases, load-bearing for dropped phases (M-6)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- planSynthesizer prompt now requires a `#### Test Spec` section for
  every phase: coverage target (≥80%), scenario table (ID/Scenario/
  Given/When/Then), and explicit edge cases list. Test-writer becomes a
  pure implementor — receives a spec as quality floor, MAY add cases.

- `extractCoverageTarget(phaseBody)` parses `**Coverage target: ≥N%**`
  from phase body (defaults 80 when absent — backward compatible).

- `buildGeminiTestSpecPrompt` is now spec-aware: detects `#### Test Spec`
  in phase.body and switches from generic "write failing tests" to
  "implement ALL listed cases as minimum requirement" instructions.

- `parseCoveragePercent(stdout, testCmd)` parses coverage % from test
  runner stdout for Jest/Vitest, Bun, pytest, and Go; returns null for
  unknown frameworks (advisory-only).

- `PhaseState.coverageResult?: { actual, target }` field added to types.

- `PLAN_REVIEW_PROMPT` gains criterion 6 (TEST SPEC QUALITY): CRITICAL
  for inconsistent specs across phases, IMPORTANT for all-missing (legacy
  plans), SUGGESTION for missing coverage target line.

- Test suite: 12 new tests for extractCoverageTarget + spec-aware
  buildGeminiTestSpecPrompt in cli.test.ts; 12 new tests for
  parseCoveragePercent in sub-agents.test.ts. Version assertions and
  coverage-matrix ownership map updated.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Add detectSkillFaults() with coverage for:
- CODEX_CONVERGENCE (iterations >= DEFAULT_MAX_CODEX_ITERATIONS)
- TEST_FIXER_LOOP (iterations >= DEFAULT_MAX_TEST_ITERATIONS)
- PREMATURE_COMPLETION (checked tasks for non-committed phases)
- PLAN_SYNTHESIS_INVALID (missing Origin trace: or Acceptance:)
- WORKTREE_LEAK (completed=true but worktree still exists)
- RED_SPEC_TRIVIAL (trivially-passing tests)
- PLAN_MUTATOR_MISMATCH (plan mutation failures)
- PLAN_REVIEW_STALEMATE (round>=3 with CRITICAL objections)
- FEATURE_VERIFIER_SCOPE (VERIFICATION: GAPS in stdout)

All detectors are wrapped in try/catch so bad inputs never throw.
Analytics are appended to GSTACK_HOME/analytics/skill-faults.jsonl
only when faults exist, and analytics failures are swallowed.

…rEvaluation wiring

Red-phase test spec for Phase 2.1. Tests cover:
- SKILL_FAULT_DETECTED absent from MonitorEventName/MONITOR_EXIT_CODES (guard)
- MonitorEvaluation.skillFaultEvents field exists and is always an array
- evaluateMonitorOnce populates skillFaultEvents from detectSkillFaults
- each SkillFaultDetectedEvent has required shape fields + event: 'SKILL_FAULT_DETECTED'
- skillFaultEvents is [] when detector finds no faults or state is null
- monitor exit code is unaffected by skillFaultEvents presence

11 tests fail (Red), 3 guard tests pass. No implementation code.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

The test file was placed in build/orchestrator/__tests__/ (correct for
test:build-skill) but VERIFY_RED runs bun test test/ which does not scan
that directory. Move to test/skill-monitor-fault.test.ts with adjusted
import paths so the 11 RED tests are discovered and confirmed failing.

Tests: 11 fail, 3 pass (before implementation)

- Add SkillFaultDetectedEvent type in types.ts (imports SkillFault)
- Add skillFaultEvents field to MonitorEvaluation in monitor.ts
- Add stateDir to MonitorRunSnapshot for detectSkillFaults input
- Call detectSkillFaults per snapshot in evaluateMonitorOnce with try/catch
- Print skillFaultEvents as JSON lines before terminal events in monitor mode
- SKILL_FAULT_DETECTED is not a MonitorEventName and has no exit code

- Update Step M3 monitor launch to use set -o pipefail and
  ${PIPESTATUS[0]} while teeing output to monitor-output.log
- Add Step M3.5 that scans monitor output for SKILL_FAULT_DETECTED,
  dedupes by resolved path (readlink), reads fault_investigator_model
  from configure.cm, and dispatches either GSTACK_FAULT_INVESTIGATOR_COMMAND
  or one background agent per non-duplicate fault
- Add validation tests for Step M3.5 content in skill-md.test.ts
- Fix pre-existing hardcoded model name in cli.ts comment

…_RUN_ID

The previous Step M3.5 implementation had a critical silent-failure bug:
1. `sed -n 's/.*file:////p'` is a malformed sed expression (4 slashes = bad
   flag in substitute command). `_FAULT_FILE` was always empty and the
   `[ -z "$_FAULT_FILE" ] && continue` guard silently skipped every fault.
2. The expression also assumed a `file://` URI format that the monitor never
   emits — actual SKILL_FAULT_DETECTED events are JSON lines with a
   `faults[].sourceFiles[]` array (see build/orchestrator/cli.ts:5739-5741
   and build/orchestrator/types.ts:15-23). No investigator would ever spawn.

Switch to jq-based JSON parsing that flattens each event into TSV rows
(runId<TAB>category<TAB>file) and pass FAULT_CATEGORY + FAULT_RUN_ID env
vars to the investigator alongside FAULT_FILE. Dedupe key now includes
(runId, category, resolved-path) so unrelated faults across runs aren't
collapsed. Log filename is suffixed with category to avoid collisions.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…nv passing

…atch

Adds test/skill-e2e-build-fault-investigator.test.ts (periodic tier) covering
the fault investigator E2E flow: mock gstack-build outputs SKILL_FAULT_DETECTED
JSON, Step M3.5 dispatches GSTACK_FAULT_INVESTIGATOR_COMMAND with fault env
vars, mock investigator writes report to $FAULT_PRIMARY, assertions verify
report exists with PLAN_SYNTHESIS_INVALID and no source files were edited.

Registers build-fault-investigator-e2e in touchfiles.ts — selected when
build/SKILL.md, skill-fault-detector.ts, or monitor.ts change.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Adds mock configure.cm file to prevent jq from failing in Step M3.5 mock

1. plan-selection (6 tests): `defaultActiveRunRegistryDir()` hardcoded
   `~/.gstack/build-state/active-runs` and ignored `GSTACK_BUILD_STATE_DIR`,
   causing 11 real active-run records to leak into unit tests and inflate
   candidate counts (turning expected "selected" into "ambiguous"). Fix: honour
   the env var consistently, the same way `state.ts` already does.

2. integration (3 tests): plan review subprocess called `codex` with
   `OPENAI_API_KEY` from the inherited `process.env`, triggering a real ~30s
   API call against the LLM. These tests exercise feature lifecycle, not plan
   review. Fix: add `--no-plan-review` to each CLI invocation.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…estSpec detection

Four improvements identified during code review of 3e2b8b2:

- Move `extractCoverageTarget` from cli.ts to sub-agents.ts (alongside
  parseCoveragePercent); re-export via import in cli.ts. Eliminates the
  circular-import risk when phase-runner.ts calls coverage functions.

- Fix decimal truncation in extractCoverageTarget: `(\d+)` only matched
  integers, silently returning 80 for targets like ≥90.5%. Changed to
  `([\d.]+)` + parseFloat.

- Fix `hasTestSpec` detection in buildGeminiTestSpecPrompt: was
  `phase.body.includes("#### Test Spec")` (fragile string match, false
  negative when body text differs). Now `phase.testSpecCheckboxLine !== -1`
  (parser already computes this — zero extra overhead).

- Wire coverage gate in RUN_TESTS handler: after GREEN tests pass and the
  phase has a test spec (`testSpecCheckboxLine !== -1`), call
  parseCoveragePercent(result.stdout, testCmd) and compare against
  extractCoverageTarget(phase.body). Below target → set coverageResult and
  route to test_fix_running. Unknown framework → log advisory, proceed.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Complete the coverage gate: `injectCoverageFlags(testCmd)` appends the
appropriate flag for the detected framework before the GREEN test run,
so `parseCoveragePercent` reliably finds coverage data in stdout even
when projects don't pre-configure coverage in their test script.

Framework → flag mapping:
  jest     → --coverage --coverageReporters text
  vitest   → --coverage
  bun test → --coverage
  pytest   → --cov --cov-report term-missing
  go test  → -cover
  unknown  → unchanged (advisory log, gate skips)

Injection is idempotent (no-op if flag already present) and only fires
when the phase has a test spec (testSpecCheckboxLine !== -1) — VERIFY_RED
and legacy phases use the bare test command unchanged.

11 unit tests added covering each framework, idempotency, and unknowns.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

`phase.kind !== "code" ? "" : ""` always evaluated to "" regardless
of the condition, and was silently filtered by .filter(Boolean).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…p (Bug D1)

Two failing tests document the bug:
1. After CRITICAL verdict, state.planReview must be persisted with status
   "critical_exit_pending" — currently cli.ts does not persist anything
   before process.exit(3), so planReview stays undefined on disk.
2. On resume with the sentinel set, the plan-review gate must still fire —
   the current guard (!state.planReview) is false when planReview is truthy,
   so the gate is skipped after the sentinel is introduced.

Two GREEN tests confirm baseline behavior: APPROVE verdict suppresses the
gate; undefined planReview (first run) fires the gate.

Tests MUST fail until Feature 4 implementation lands.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Before this fix, a CRITICAL plan-review verdict caused process.exit(3)
without saving any sentinel to state. On resume, !state.planReview was
true → review ran again → CRITICAL again → infinite loop.

Fix:
1. Save state.planReview = { ...verdict, status: "critical_exit_pending" }
   before releaseLock + process.exit(3) so the sentinel survives on disk.
2. Widen the plan-review gate guard from !state.planReview to
   !state.planReview || state.planReview.status === "critical_exit_pending"
   so the gate re-fires on resume when the sentinel is present.

Tests: two new tests in phase-runner.test.ts cover both the sentinel
persistence and the widened gate; 90/90 passing.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…g D2)

Introduces ExitError (errors.ts) — thrown instead of process.exit(N)
inside try/finally blocks so the finally clause runs cleanup before
the process terminates.

Changes:
- errors.ts: new ExitError class (instanceof Error, numeric code field)
- cli.ts: import ExitError; replace critical_exit process.exit(3) with
  throw new ExitError(3); update main().catch to call process.exit(err.code)
  when err instanceof ExitError
- phase-runner.test.ts: 5 new tests (ExitError shape, propagation through
  finally, default and custom messages); 95/95 passing

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ature 6)

applyResult() now populates phaseState.coverageResult when:
- action is RUN_TESTS
- tests are GREEN (status = "tests_green")
- extra.phaseBody is provided
- parseCoveragePercent() returns a non-null value for the stdout

Coverage below target emits an advisory warning but keeps status
"tests_green" — not blocking. The target defaults to 80 when no
"**Coverage target: ≥N%**" line appears in the phase body.

6 new tests in phase-runner.test.ts; 101/101 passing.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…ics + test assertions

- Add errors.ts to MODULE_TEST_OWNERS in coverage-matrix.test.ts
- Fix analytics logActivity to emit "success" for exit code 13 (FINALIZATION_REQUIRED),
  which is a success state (pending ship), not a failure
- Fix integration test assertions: --skip-ship correctly exits 13, not 0, when
  features reach origin_verified (pre-existing test/impl mismatch)

…d [Phase 1.1]

RED phase TDD: 11 tests fail because the parser does not yet stamp kind: "code"
on emitted phases, and existing Phase literal construction sites have no kind
field (undefined fails the VALID_KINDS.includes runtime assertion).

11 tests pass immediately: direct Phase construction with explicit kind values,
and PhaseKind union membership checks (both already exist in types.ts).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

… loop

Add required kind: PhaseKind field to the parser factory init and to
every Phase literal construction site in tests/fixtures. This ensures
backward-compatible default of kind: "code" for all existing phases
while the type system enforces correctness going forward.

- parser.ts: stamp kind: "code" on every emitted Phase
- state.test.ts, cli.test.ts, phase-runner.test.ts,
  feature-review.test.ts, cli-guardrails.test.ts,
  phase-kind.test.ts: add kind: "code" to all helpers and inline literals

…tations

- Fix PHASE_HEADING regex to allow optional [kind] bracket between number and colon
- Add BODY_KIND_PATTERN for <!-- kind: X --> HTML comment fallback
- Add IMPL_LABELS_BY_KIND and REVIEW_LABELS_BY_KIND maps for all 5 PhaseKind values
- Parser now stamps kind from heading bracket (primary), body comment (fallback), or defaults to "code"
- Inline kind-comment detection ensures kind is set before checkbox processing
- Add implCheckboxRe/reviewCheckboxRe for kind-specific checkbox matching
- Add 16 new parser tests covering all bracket annotations, HTML fallback, checkbox recognition

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…EW gates, ship gate

…-mutator.ts

The merge introduced exported constants at the top of the file while the
original local const declarations were still present below, causing a
"has already been declared" TypeScript error. Remove the duplicates.

Also regenerate SKILL.md files to pick up template changes from the merge.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Post-rebase gen:skill-docs sync — templates updated in Phase 1.5 commits
produced SKILL.md drift. Regenerating now from authoritative templates.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…geTarget import after rebase

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

… "paused"

T1 (integration): exit-13 --skip-ship writes "paused" record to active-run
registry. Existing test at line 685 already covers this scenario.

T3 (integration): non-zero non-13 exit writes "failed" record. Forces
validateResumeLaunch to throw via pre-written state with mismatched projectRoot.

T4 (plan-selection): paused + dead-pid (999999) record must be cleaned up by
activeRunOnlyCandidates() — verifies Feature 3 auto-clean logic.

T5 (plan-selection): paused + live-pid (process.pid) record must stay and
return a candidate — verifies existing behavior preserved.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…tions

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

…addition

Phase 1.5 added the Content Review gate to ship/SKILL.md.tmpl but the golden
baselines weren't updated. Regenerated from current ship/SKILL.md.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>