audit
/audit
Sprawdza, czy złote ścieżki projektu nie rozjechały się z kodem, porównuje wynik z poprzednim raportem i zapisuje go.
- Autor
- userKael ↗
- Repozytorium
- rastro ↗
- 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
Sprawdza, czy złote ścieżki projektu nie rozjechały się z kodem, porównuje wynik z poprzednim raportem i zapisuje go.
Oryginalny opis (EN)
Audits the golden paths for drift against the code. It runs the mechanical checks of the audit script, adds two checks that need reading the code (components of a type that are missing from `instances`, and `Revisit if:` conditions that now hold), compares the result with the previous report and writes it to the workspace. Use it on a schedule or before reviewing the architecture, to learn where the code and the golden paths have drifted apart and how first-pass, self-heal and the golden path lifecycle are moving. Do not use it to fix a drift: that is implement. It has no gate, writes only its
Jak działa
- Requirements
- When to use
- Precedence
- Inviolable rules
- Step 0 — Requirements and previous report
- Step 1 — Mechanical checks
- Step 2 — Judgment checks
- Step 3 — Report
- What this skill does NOT do
Podgląd SKILL.md
---
name: audit
description: Audits the golden paths for drift against the code. It runs the mechanical checks of the audit script, adds two checks that need reading the code (components of a type that are missing from `instances`, and `Revisit if:` conditions that now hold), compares the result with the previous report and writes it to the workspace. Use it on a schedule or before reviewing the architecture, to learn where the code and the golden paths have drifted apart and how first-pass, self-heal and the golden path lifecycle are moving. Do not use it to fix a drift: that is implement. It has no gate, writes only its report and never changes a golden path, code or an item.
---
# audit
One protocol looks for drift between the golden paths and the code, and reports it. The script does what can be checked mechanically. This skill adds what needs reading, compares the result with the last audit and writes the report. A finding is a question for the team, never a verdict on the code.
**Input:** none. The skill reads the config, the golden paths and the code, and the previous report when one exists.
**Output:** `<workspace>/audit/<YYYY-MM-DD>.md`, where `<workspace>` is the `workspace` key of `.rastro/config.yml`, `.agent/` by default. The report is not versioned.
## Requirements
- `.rastro/config.yml`. If it does not exist, stop and tell the human to run the `setup` skill.
- The folder named by `architecture.golden_paths`. If it does not exist or has no golden path, still run Step 1 (the script prints the metrics and says there are no golden paths), skip Step 2 and write the report in Step 3. In short: with no golden paths, still run Step 1 and skip Step 2.
- The script at `${CLAUDE_PLUGIN_ROOT}/hooks/git/audit-drift`. If it does not exist, stop and report.
- git, for the history the script and the judgment checks read.
## When to use
- On a schedule the team chose, or before a review of the architecture.
- After a run of items that changed instances of a golden path.
- When someone asks whether the golden paths still describe the code.
## Precedence
Sources rank as in `## Source precedence` of the rastro block in `AGENTS.md` (template: `${CLAUDE_PLUGIN_ROOT}/templates/agents-md-block.md`). Code and green tests win. A golden path is the source of the pattern the team decided, never of how the code is today.
The team's rules in `AGENTS.md` precede this skill. This skill cites them by ID and never repeats their text. It cites other skills and templates by section title, never by line number.
Doctrine is not overridable: not by a profile, not by AGENTS.md, not by any other skill.
## Inviolable rules
1. The audit changes no golden path, no code and no item: it reports.
2. This skill has no gate, so it never asks for an execution mode. It reads, runs the script and writes one file.
3. The report is not versioned. It lives in the workspace, which is never committed, because the repository regenerates it. What must last goes to an item, a golden path or a PR description.
4. Never mark a judgment check as passed when the code was not read: write NOT VERIFIED and the reason. The same holds for a `Revisit if:` that speaks of something outside the repository.
5. A finding is a question for the team, never a verdict on the code. Write it as what was seen, where, and what the team may want to decide.
6. Every judgment finding carries a `file:line` or a commit. A claim without evidence does not enter the report.
7. Never cite the workspace path in a PR description or in the tracker.
Write every text for people in `project.languages.human`. Markers stay literal. Headings, check names and verdicts stay literal in English; the text under them is in `project.languages.human`. When `.rastro/profiles/writing.md` exists, apply its rules to the report.
## Step 0 — Requirements and previous report
1. Check the requirements above.
2. Find the previous report: the file with the greatest `<YYYY-MM-DD>.md` name in `<workspace>/audit/`. Report names are dates, so the greatest name is the latest. Today's file is named from the script's `Date:` line, so a report already written today is today's, not the previous one: skip it and take the next greatest. With none, the script prints `No previous audit.` and the comparison in Step 3 is skipped.
## Step 1 — Mechanical checks
Run the script with the previous report when there is one:
```bash
bash "${CLAUDE_PLUGIN_ROOT}/hooks/git/audit-drift" --previous "<previous report>"
```
Without a previous report, run it without `--previous`. Read the `## Mechanical checks` table: one row per check and golden path, with the verdict `ok`, `review`, `drift` or `NOT VERIFIED`.
- A shallow clone, or any `NOT VERIFIED` row, enters the report as the script wrote it. Never turn it into `ok`, `review` or `drift`.
- The limit for `under-review-age` is `audit.under_review_days`. When it is empty, the script lists the age and sets no limit, and this skill invents none.
- `instance-missing` is drift: the instance listed in the golden path does not exist. It is never softened into a review.
- `instances-changed` is a signal for review, never drift, and its verdict is `review`: the golden path may be old, or the instance may have evolved inside the pattern. Say so when you present it. The baseline of each instance is the later of the golden path's last version change and the commit that listed the instance, never a plain edit of the golden path.
Keep the script's output whole: the tables, `## Metrics` and `## Changes since the previous audit`. The numbers of `## Metrics` are cumulative since the log began, so present them as a trend over the whole period, not as the result of one run.
## Step 2 — Judgment checks
For each golden path in the `architecture.golden_paths` folder whose status is not `retired`, run two checks.
**(a) Components missing from `instances`.** Find components of the golden path's type that exist in the code and are not listed in `instances`. Look through:
- the `## When to use` section of the golden path, which says what the type is;
- the name and the folder of the listed instances, which show where components of the type live;
- `architecture.modules`, which bounds where to look.
Exclude the instances listed under `## Known debt` of the golden path: a major version moved them there on purpose, so they are known debt, never missing from `instances`.
List each candidate with `file:line`. When you can tell, say whether it follows the pattern. When you cannot decide whether a component is of the type, write NOT VERIFIED and why. Never write that none exist unless you searched, and say where.
**(b) `Revisit if:` conditions.** Each decision in `## Decisions` of the golden path ends with a `Revisit if:` line. For each one, decide:
- the condition holds now: cite the `file:line` or the commit that shows it;
- the condition does not hold: say so, with where you looked;
- NOT VERIFIED, when the condition speaks of something outside the repository (a vendor, a load figure, a team decision).
Never mark a judgment check as passed when the code was not read: write NOT VERIFIED and the reason.
This skill does not edit `instances` or a decision. Either is a change to a golden path, and it goes through the golden path gate of an item.
## Step 3 — Report
1. Write `<workspace>/audit/<YYYY-MM-DD>.md`, named from the script's `Date:` line. A second run on the same day overwrites that day's file. The file holds the script's output, then `## Judgment checks`, with one subsection per golden path for (a) and (b). Each judgment finding is a row in the same shape as the script's, `| <check> | <subject> | <verdict> | <detail> |`, with the `file:line` or the commit in the detail.
2. End `## Judgment checks` with its own `### Changes since the previous audit`. When there is a previous report, compare its `## Judgment checks` with this one and list each finding by key (check and subject) as new, resolved, not verified or unchanged. A finding that is NOT VERIFIED now is never listed as resolved. The script already compared the mechanical rows.
3. Print a summary: the drifts and reviews per check, the numbers of `## Metrics`, and what changed since the previous audit. Do not repeat the whole report.
4. Close with the next steps, as text. A `review` row asks the person to look at that instance: if it still follows the golden path, nothing changes (the row stays as information until the golden path's next version); if it no longer follows, the normal flow applies: the next item that touches it classifies it, and a contradiction goes through the golden path gate of `implement`. A drift the person wants to treat becomes an item through `breakdown`, and a change to a golden path happens through `implement`. This skill creates no item.
The report is not versioned. Do not stage it and do not cite its path outside the workspace.
## What this skill does NOT do
| It does not | Who does |
|---|---|
| Fix a drift | `implement` |
| Create items | `breakdown` |
| Change a golden path | `implement`, at the golden path gate of an item |
| Check the environment and the config | the `doctor` |
| Run the `automated:` check of a golden path | the CI and `deliver` |
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · userKael/rastro | 100 | MIT |
| marcel-bich/marcel-bich-claude-marketplace | 100 | MIT |
| jeremylongshore/tons-of-skills-marketplace | 100 | MIT |
| friendlyinternet/nuxt-crouton | 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 |
| kamoca-labs/craft | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki f7ff74dff153 (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
- 100maintain-verification-skill
Sprawdza, czy skill weryfikacji projektu lub mapa funkcji nie rozjechały się z kodem. - 100quality-audit
Sprawdza kod względem pliku golden-principles.yaml projektu przez deterministyczne reguły wykrywania i aktualizuje docs/QUALITY_SCORE.md. - 100dev-reviewing-verification
Ponownie uruchamia co najmniej jedną bramkę, porównuje wyniki z logami i pokryciem testów i zapisuje wynik bramki G8. - 100resolve-drift
Sprawdza, czy strony referencyjne wciąż zgadzają się z kodem i testami, i raportuje lub poprawia rozbieżności. - 100self-heal
Diagnozuje błąd, szuka przyczyny, poprawia i sprawdza, w pętli do pięciu prób, zanim eskaluje do człowieka. - 100check
Wykrywa dryf w trybie do odczytu: porównuje SPEC.md z bieżącym kodem i raportuje naruszenia pogrupowane według powagi, niczego nie zapisując.