audit
/audit
Przegląda plan i graf zadań, porównuje status z repozytorium, poprawia jednoznaczne rozbieżności i pyta o resztę.
- Autor
- hootener ↗
- Repozytorium
- brain ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 12.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 12.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Przegląda plan i graf zadań, porównuje status z repozytorium, poprawia jednoznaczne rozbieżności i pyta o resztę.
Wyzwalacze z opisu
Oryginalny opis (EN)
Walk a plan, spec or goal's subgraph and reconcile every task's status against what the repository actually shows — fixing what is unambiguous, asking about what is not, and surfacing work that was done with no node at all. Use when an initiative is wrapping up, before a handoff, or whenever you want to know whether the brain still matches reality.
Jak działa
- Step 1 — get the facts
- Step 2 — batch the determinable
- Step 3 — ask about the rest, and say what you checked
- Step 4 — orphan work
- Step 5 — record the audit
- Cadence
- Reporting that nothing is wrong
Podgląd SKILL.md
--- name: audit description: Walk a plan, spec or goal's subgraph and reconcile every task's status against what the repository actually shows — fixing what is unambiguous, asking about what is not, and surfacing work that was done with no node at all. Use when an initiative is wrapping up, before a handoff, or whenever you want to know whether the brain still matches reality. --- Most work arrives agentically now, and the brain hears about it only when somebody remembers to say so. This is the layer that stops the remembering from being the operator's job ([D0007](../../knowledge-catalog/decisions/D0007-status-ownership.md)). It is the **third** layer, and it only handles what the first two cannot. The hook reconciles what is mechanically checkable on every commit; the skills move status as a side effect of doing the work. What reaches here is genuine ambiguity — and if a lot reaches here, that is a signal those two are broken, not that this one is working hard. ## Step 1 — get the facts ```bash brain/bin/reconcile [initiative] ``` It compares every task's status against the history and sorts into three buckets. **It never infers**, which is exactly why its output can be trusted: - **reconciled** — status matches the evidence. Say nothing about these. - **determinable** — wrong, and the evidence is unambiguous. - **undeterminable** — honestly cannot tell. Then walk the subgraph forwards, so omission shows up as well as drift: ```bash brain/bin/ancestry <the plan, spec or goal> --forward ``` Anything marked `← outstanding` is either work remaining or a scope change nobody recorded. A goal with four specs where three shipped and one was quietly dropped is invisible any other way. ## Step 2 — batch the determinable One confirmation for all of them, not one question each: > T0005, T0009 and T0011 are all `In Progress`; their commits are merged and > their acceptance criteria are met in the code. Set all three `Complete`? ## Step 3 — ask about the rest, and say what you checked This is the part that matters, and the discipline is the same one `/grill-change` follows in citing a path. **The question must carry what was checked**, or you have handed the work back rather than done it. Not this: > I cannot tell if T0005 is complete. Is it? This: > **T0005 — Add retry to the webhook dispatcher.** Two of three acceptance > criteria are verifiably met: the retry exists, and the backoff is covered by > a test. The third — *failures surface in the operator log* — I cannot > verify: there is no test for it and no logging call I can find on that path. > Complete, or is that criterion outstanding? The first takes four seconds of yours and four minutes of theirs. The second is answerable immediately, and it is checkable — they can see what you looked at and tell you that you looked in the wrong place. **An audit never silently changes a status it inferred.** Nothing moves without an answer. If the operator walks away, write the findings as an Evidence node and leave the open questions visibly open — an unanswered question on the record is worth more than a guess that reads as a fact. ## Step 4 — orphan work ```bash brain/bin/reconcile # its last section lists commits naming no task ``` An agent told "fix the retry bug" fixes it, and no node records that it happened. **Status drift makes the brain wrong; orphan work makes it incomplete**, which is harder to notice because the graph still looks healthy. Group the commits and ask: > These seven commits map to no task. Was this T0005, or is it unrecorded work > that needs a node? Often the answer is a one-line task written retroactively, which is fine — the record is for the next reader, not for the ritual. ## Step 5 — record the audit Write an Evidence node: `brain/bin/new evidence "Audit of <initiative>, <date>"`, with `reviewed:` naming what it covered and `amends:` naming any node whose status or claims it corrected. That makes "when was this last audited" answerable, which is what lets `bin/doctor` warn that one is overdue without running one. ## Cadence **At boundaries, never on a timer.** When a plan's last task reaches `Testing`, before a handoff, or when asked. An audit that runs constantly and asks questions trains the operator to rubber-stamp, and a rubber-stamped brain is worse than a stale one because it *looks* maintained. ## Reporting that nothing is wrong Do it plainly, and expect it to be the common outcome. If the hook and the skills are doing their jobs, most audits find nothing. **An audit that always finds something means layers one and two are broken** — say that, because it is more useful than the findings.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · hootener/brain | 100 | MIT |
| marcel-bich/marcel-bich-claude-marketplace | 100 | MIT |
| jeremylongshore/tons-of-skills-marketplace | 100 | MIT |
| friendlyinternet/nuxt-crouton | 100 | MIT |
| userKael/rastro | 100 | MIT |
| Archont561/qgis-agent-harness | 100 | MIT |
| lilly-logic-labs/lll-skills | 100 | MIT |
| Come-In-Labs/agent-skills | 100 | MIT |
| minanyang/10x-agent | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki c803bdc602df (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
- 90session-reconcile
Porównuje zmiany z sesji z dokumentacją, instrukcjami agenta i pamięcią, wypisuje rozbieżności i poprawia tylko te zatwierdzone. - 100planr
Prowadzi pracę produktową jako graf zadań w Markdownie: pokazuje gotowe i zablokowane zadania oraz bezpiecznie zmienia statusy. - 100stale-proposals
Sprawdza lokalną kolejkę propozycji pod kątem wpisów, które już wdrożono, i rozstrzyga tylko jednoznaczne przypadki. - 100plan-tree
Utrzymuje drzewo dokumentów planistycznych: mapę drogową, pliki statusu, notatki decyzyjne i porządek w repozytorium. - 100ywc-spec-ready
Automatycznie powtarza planowanie i walidację specyfikacji z ostrzeżeniami, aż specyfikacja osiągnie status gotowości do zadań.