harness-audit-engine
/harness-audit-engine
Wykrywa dryf konfiguracji harnessu agenta: pliki konwencji, ONBOARDING, nieaktualne migawki i regresje ograniczeń.
- Autor
- Habitat-Thinking ↗
- Repozytorium
- ai-literacy-superpowers ↗
- Licencja
- Apache-2.0
- Ostatnia zmiana repozytorium
- 19.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 19.10.2026, bez odwołań do wycofanych modeli i API.
Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.
Co robi i kiedy
Wykrywa dryf konfiguracji harnessu agenta: pliki konwencji, ONBOARDING, nieaktualne migawki i regresje ograniczeń.
Oryginalny opis (EN)
Use when running the shared drift-detection logic that backs /harness-audit and /harness-sync — produces a structured drift report covering convention files, ONBOARDING.md, snapshot staleness, constraint regressions, recurring reflection patterns, and HARNESS.md Status section accuracy.
Jak działa
- Overview
- What the engine scans
- Drift-report shape
- State vocabulary
- Auto-fixable classification rule
- Caller contract
- Adding a new surface to the engine
Podgląd SKILL.md
--- name: harness-audit-engine description: Use when running the shared drift-detection logic that backs /harness-audit and /harness-sync — produces a structured drift report covering convention files, ONBOARDING.md, snapshot staleness, constraint regressions, recurring reflection patterns, and HARNESS.md Status section accuracy. --- # Harness Audit Engine ## Overview The harness audit-engine is the shared drift-detection layer that both `/harness-audit` and `/harness-sync` invoke. Audit is read-only; sync uses the same engine to drive its multi-select prompt. This skill defines the contract — what surfaces are scanned, the drift-report shape, and the per-finding `[auto]` / `[manual]` classification rule. `/harness-audit` and `/harness-sync` are the two callers. Their UX differs (audit prints, sync prompts), but the underlying findings are the same. ## What the engine scans The engine evaluates every surface or fact that could be out of sync with `HARNESS.md`. The current scan covers: | Category | Finding | Auto-fixable? | | --- | --- | --- | | Convention files | `.cursor/rules/` matches HARNESS.md | yes — `/convention-sync` | | Convention files | `.github/copilot-instructions.md` matches HARNESS.md | yes — `/convention-sync` | | Convention files | `.windsurf/rules/` matches HARNESS.md | yes — `/convention-sync` | | Onboarding | `ONBOARDING.md` matches HARNESS.md + AGENTS.md + REFLECTION_LOG.md | manual — `/harness-onboarding` | | Observability | Most recent snapshot in `observability/snapshots/` is < 30 days old | yes — `/harness-health` | | Status section | `HARNESS.md` Status block matches actual constraint enforcement counts | yes — `/harness-audit` (audit updates Status as a side-effect) | | Constraint regression | Any constraint marked `deterministic` whose tool no longer succeeds | manual — `/harness-constrain` | | Reflection pattern | Recurring failure pattern in REFLECTION_LOG.md (2+ similar entries, not yet a constraint) | manual — `/harness-constrain` | | CI / CD | Constraint scope handled at runtime by harness-enforcer | informational — handled at runtime | New surfaces added to the engine appear in both commands automatically. ## Drift-report shape The engine returns a list of findings. Each finding has: - `surface` — short label (e.g. `.cursor/rules/`, `Snapshot staleness`) - `status` — one of `drifted`, `missing`, `in sync`, `managed`, `candidate` - `details` — short string with the specific evidence (e.g. `last: 2026-04-15`, `HARNESS: 0.31, plugin: 0.34`) - `action_command` — the slash command that remediates this finding (e.g. `/convention-sync`, `/harness-upgrade`) - `auto_fixable` — `true` if the action runs without further user judgement; `false` if the action requires user input ## State vocabulary - `drifted` — file or fact exists but does not match HARNESS.md / reality. - `missing` — file expected but not present. - `in sync` — matches HARNESS.md / reality. - `managed` — handled at runtime by other layers (CI/CD); informational only. - `candidate` — finding that is not strict drift but warrants review (recurring reflection patterns; deferred constraints). ## Auto-fixable classification rule A finding is `auto_fixable: true` only when: 1. There is a deterministic remediation (a specific command that, given the same HARNESS.md state, would always produce the same fix). 2. The remediation writes to allow-listed paths (the four push-direction surfaces) OR mutates HARNESS.md only in defined ways (Status section regen via `/harness-audit`, snapshot creation via `/harness-health`). 3. No user judgement is required between detection and remediation. Findings that need user input — choosing which constraint to add, deciding whether enforcement should be promoted, deciding when to take an upstream template upgrade — are `auto_fixable: false`. Sync surfaces them as `[manual]` rows; selecting them prints the suggested command without writing. ## Caller contract Each caller is responsible for its own UX: - `/harness-audit` prints the findings in its existing format and updates `HARNESS.md` Status section. - `/harness-sync` builds its drift table from the findings, prompts for selection, and dispatches the action commands. The engine itself does not print or prompt. It returns the findings and exits. ## Adding a new surface to the engine When a new check is added to the engine: 1. Document the surface in the Categories table above. 2. Decide its `auto_fixable` classification using the rule. 3. Define the `action_command` it maps to. 4. Both `/harness-audit` and `/harness-sync` pick up the new finding without further code changes. This is the value of factoring the engine: surface coverage evolves in one place.
Wersje
W bazie nie ma innych skilli o nazwie harness-audit-engine.
Migawka bazy z 13.11.2026 · wersja paczki e136a5c24c92 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100constraint-design
Pomaga projektować ograniczenia w sekcji Constraints pliku HARNESS.md: pisze reguły harnessu, dobiera typ egzekwowania, promuje ograniczenia i konfiguruje sloty weryfikacji. - 95harness-doctor
Skanuje strukturę .claude, CLAUDE.md i reguły projektu, wykrywa dryf i brakujące pliki, a potem proponuje poprawki. - 100harness
Buduje harness jako meta-skill, który definiuje wyspecjalizowanych agentów i generuje używane przez nich skille, a także przebudowuje, audytuje i synchronizuje istniejący harness. - 90harness-observability
Sprawdza zdrowie harnessu agenta, konfiguruje cykliczne przeglądy, eksport telemetrii i format migawek. - 90vc:audit-vc
Audytuje zdrowie konfiguracji agentów: zgodność Claude i Codex, rejestr skilli, README i pliki protokołu. - 100harness
Wyznacza etapy i zasady budowy oraz ulepszania harnessu Claude Code: skille, agenci, hooki i CLAUDE.md, z ewaluacją i kosztami.