메뉴
Use after build has produced build-report.md. Validates the build via four parallel sub-auditors (eval-replay, lint, regression, build-quality) and produces ALL-PASS verdict. Adversarial mode default-on per CLAUDE.md.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | audit |
| description | Use after build has produced build-report.md. Validates the build via four parallel sub-auditors (eval-replay, lint, regression, build-quality) and produces ALL-PASS verdict. Adversarial mode default-on per CLAUDE.md. |
Sprint 1.2 fan-out + Sprint 3 composable skill (v8.16+). Sub-auditors run in parallel via
subagent-run.sh dispatch-parallel auditor.
build produces build-report.mdbuild phase per cycle-stateaudit-eval-replay sub-auditor)| Step | Action | Exit criteria |
|---|---|---|
| 1 | Verify <workspace>/build-report.md exists, fresh, status ≠ FAIL | Build verified |
| 2 | Dispatch 4 sub-auditors in parallel | 4 worker artifacts |
| 3 | Aggregator applies ALL-PASS rule | <workspace>/audit-report.md first line is Verdict: <X> |
| 4 | Phase gate gate_audit_to_ship enforces PASS | Gate passes only on PASS |
| Verdict | Trigger | Phase-gate behavior |
|---|---|---|
PASS | Every sub-auditor reports PASS | Allow ship |
FAIL | Any sub-auditor reports FAIL | Block ship; orchestrator → retrospective |
WARN | Any sub-auditor reports WARN (no FAIL) | Block ship; review case-by-case |
Default ON: each sub-auditor's prompt prepends "ADVERSARIAL AUDIT MODE — require positive evidence for PASS". Disable only via ADVERSARIAL_AUDIT=0 for deliberately permissive sweeps. Auditor model defaults to Opus while Builder defaults to Sonnet — different family breaks same-model-judge sycophancy.
For any cycle that changes a scored metric — a flag-reduction cycle, a
registry/gate/marker/allowlist edit, or any cycle claiming a count reduction —
the auditor MUST apply the goal-integrity rubric in
skills/adversarial-testing/SKILL.md §10.1
as a mandatory BLOCK (co-equal with the deterministic gates, not a backstop). A
claimed reduction must cite the reader that was deleted and confirm no
surviving reader on any surface; "the row is gone" is not evidence. FAIL on
writer-fabrication, off-namespace/reflection rename, contract under-delivery, or
any --class cycle edit of a guards.IsProtectedSurface control-plane file.
| Fact | Value |
|---|---|
| Phase | audit (evaluate archetype, mandatory) |
| Persona | agents/evolve-auditor.md |
| Profile | .evolve/profiles/auditor.json — CLI claude-tmux, tier deep, single-writer |
| Inputs | build-report.md · tester-report.md |
| Artifact | audit-report.md (cycle workspace) |
audit-report.md must declare:
## Verdict (also accepted: Verdict:)Verdict tokens: PASS | FAIL | WARN | SKIPPED.
Invoked by:
/evo:auditloop macro after /evo:buildFan-out prompts live in .evolve/profiles/auditor.json:parallel_subtasks (count projected into Phase facts above).
.evolve/profiles/auditor.jsonlegacy/scripts/dispatch/aggregator.sh (phase=audit)legacy/scripts/lifecycle/phase-gate.sh:gate_audit_to_shipskills/loop/phase4-audit.md (legacy detailed workflow)Use when the user invokes /evo:loop or asks to run autonomous improvement cycles, self-evolving development, compound discovery, or multi-cycle code improvement with research, build, audit, and learning phases
Use when the user asks to explain how a feature/subsystem works and how the loops built it, or what the concurrent loops/lanes/cycles did and why. Default output = per-FEATURE HTML explainer pages (two movements — how it works today, then how it got built, with verified file:line claims and drift callouts) under docs/explain/; per-batch/per-lane story pages are the fallback framing. Derived from geoffreylitt/explain-diff-html, customized for evolve-loop.
Use when writing or changing any production code or tests — new features, bug fixes, refactors, or test authoring, in any language, by any model on any CLI.
Use when running any LLM agent below Fable 5 — Claude (Opus/Sonnet/Haiku), GPT via codex, Gemini via agy, or local models via ollama — to adopt Fable 5's operating discipline — evidence-first investigation, premise verification, root-cause-only fixes, adversarial self-review, calibrated autonomy, and honest failure reporting. Load at session start or as a persona overlay for phase agents on any CLI.
Use after ship completes. Three sub-reflectors (instinct, gene, failure) run in parallel to extract lessons, update gene pool, and analyze any failures. Off the latency-critical path.
Use when the user invokes /evo:publish or asks to release a new version, ship a release, or publish a tag. Wraps the go-native self-healing release pipeline (`evolve release`) — pre-flight checks, auto-changelog, atomic ship, marketplace propagation poll, auto-rollback on failure.