audit
/audit
Prowadzi audyty jakości kodu w przestrzeni roboczej QGIS SDK i zapisuje zwięzły raport z zadaniami.
- Autor
- Archont561 ↗
- Repozytorium
- qgis-agent-harness ↗
- 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
Prowadzi audyty jakości kodu w przestrzeni roboczej QGIS SDK i zapisuje zwięzły raport z zadaniami.
Wyzwalacze z opisu
Oryginalny opis (EN)
Conduct evidence-based code-quality audits of this QGIS SDK workspace: source, a diff, one package, or the whole repository. Use whenever the user asks to audit, scan, assess, or review code for maintainability, architecture, readability, duplication, complexity, code smells, error handling, or refactor opportunities, including a plain "audit this" in a coding context. Produces a compact findings report and files one backlog task per validated actionable follow-up. Classifies behavior-preserving refactors separately from bugs, missing features and contract changes, and pauses in chat when inte
Jak działa
- 1. Establish scope and constraints
- 2. Load the review criteria
- 3. Scan and verify findings
- 4. Clarify contradictions before acting
- 5. File focused backlog task(s)
- 6. Report
- 7. Evaluation prompts
Podgląd SKILL.md
--- name: audit description: >- Conduct evidence-based code-quality audits of this QGIS SDK workspace: source, a diff, one package, or the whole repository. Use whenever the user asks to audit, scan, assess, or review code for maintainability, architecture, readability, duplication, complexity, code smells, error handling, or refactor opportunities, including a plain "audit this" in a coding context. Produces a compact findings report and files one backlog task per validated actionable follow-up. Classifies behavior-preserving refactors separately from bugs, missing features and contract changes, and pauses in chat when intent, contracts or decisions conflict. Not a substitute for security, compliance or performance review. --- # Code-quality audit An audit discovers and triages maintainability problems; it does not change production code. The deliverable is a concise evidence-backed report and, by default, a backlog task for every validated actionable follow-up. Track behavior-preserving cleanup as `refactor` work and track bugs, missing features or approved contract changes under their correct task types. ## 1. Establish scope and constraints - Follow the user's requested target: named files, a diff, a package, or the repository. If the request is simply "audit this" and the target is not clear from context, ask one scope question instead of silently scanning everything. State any reasonable scope assumption. - Read the controlling context before judging design: [AGENTS.md](../../../AGENTS.md), the [workspace skill](../qgis-sdk/SKILL.md), and the accepted decisions under `backlog/decisions/decision-NN - *.md` that touch the code in scope. A rule in a decision record outweighs a general principle; if the code disagrees with an accepted decision, that is a contract finding, not a style opinion. - Inspect `git status --short` and the relevant diff first so uncommitted work is not mistaken for an audit finding. - Keep the scan read-only: no source edits, no auto-format, no "while I am here" fixes, no starting the fix during the audit. Honor an explicit request to report only, or to skip paths. - Exclude generated and vendored output unless the generator is the target: `**/src/generated/`, `packages/qgis-sdk-typescript/tests/contracts/generated/`, `packages/qgis-sdk-typescript/tests/contracts/pydantic/generated/`, `packages/**/dist/`, `node_modules/`, `.pixi/`. Generated bytes are the generator's responsibility: audit `packages/qgis-sdk-runtime/src/qgis_sdk/bridge/generate.py` and the scripts under `packages/qgis-sdk-python/scripts/` instead. Say what was and was not inspected. ## 2. Load the review criteria - [refactor](../refactor/SKILL.md) for behavior-preserving restructuring, small steps, and avoiding speculative change. - [clean-code-principles](../clean-code-principles/SKILL.md) for principle selection, then read the specific rules that match the evidence found: `rules/core-fail-fast.md`, `rules/core-dry-single-source.md`, `rules/core-kiss-simplicity.md`, `rules/core-separation-concerns.md`, `rules/solid-srp-function.md`, `rules/solid-lsp-contracts.md`, `rules/core-yagni-abstractions.md` and so on. Cite the rule by its file id (`core-fail-fast`); do not cite a rule that is not in that directory as if it were implemented. - [tdd](../tdd/SKILL.md) and [backlog](../backlog/SKILL.md) when a finding will become a task. The repository's own conventions are the controlling context. A principle is a lens, not proof: line count, a design-pattern opportunity, or personal taste alone is not an actionable finding. Prefer the smallest fix; never recommend an abstraction without a concrete problem it solves. ## 3. Scan and verify findings Read the implementation together with its callers, public surface, tests, and the task or decision that introduced it. The two Python/TypeScript halves of the bridge are one contract: a finding on one side must be checked against the other. Run the scoped checks that already exist instead of inventing a gate: | Target | Command | | --- | --- | | Python lint and format | `pixi run lint-python`, `pixi run format-python-check` | | Focused Python tests | `pixi run -e default python -m pytest packages/qgis-sdk-python/tests/<file> -q` | | Workspace tests | `pixi run -e bun bun run test`, or `pixi run -e bun turbo-test` | | Coverage gates | `pixi run -e bun bun run coverage` | | TypeScript lint and types | `pixi run -e bun biome-check`, `pixi run -e bun typecheck` | | Generated client drift | `pixi run -e build python packages/qgis-sdk-python/scripts/generate_bridge_clients.py --check` | In a sandbox without Pixi or Bun, the documented fallback is a lower-fidelity check, not an equivalent one: `pip install --user --break-system-packages -e packages/qgis-sdk-runtime -e packages/qgis-sdk-tools -e "packages/qgis-sdk-python[dev]"` plus `PYTHONPATH=packages/qgis-sdk-runtime/src:packages/qgis-sdk-tools/src python3 -m pytest -c packages/qgis-sdk-python/pyproject.toml packages/qgis-sdk-python/tests -q` for Python, and a node toolchain under `/tmp` for `bun test`, `tsc` and `biome`. Say that the gate ran through the fallback, never that the Pixi gate passed. Two known sandbox-only artifacts: `test_pure_imports.py` fails under a plain interpreter (it asserts the SDK imports without native modules), and CI remains the authority for anything the fallback cannot cover. Keep a finding only when the evidence supports it. For each candidate: 1. Reproduce it or quote the exact tool output. A probe is worth more than a reading: a call, a generated artifact, a dispatcher round trip, a test seam. 2. Cite `path:line` and name the clean-code rule id. 3. State the concrete cost in this codebase: what a caller, page, plugin author, or maintainer loses. "Feels wrong" and "not idiomatic" are not costs. 4. Check whether a test, doc, accepted decision, public export, wire shape, error code, grant, thread affinity or other explicit contract changes the interpretation. 5. Separate verified evidence from a hypothesis, and say which follow-up is the smallest fix. Do not label a behavior bug, missing feature or contract change as a refactor. Report it separately and, once its expected behavior and scope are clear, track it under the matching task type. ## 4. Clarify contradictions before acting Pause when a finding depends on unclear intent, or when the code conflicts with a test, an accepted decision, a documented contract, or the user's constraints. This repository has no `grill-me` skill: ask the same focused question in chat, state the options and the consequence of each, and wait for the answer before deciding the finding or writing acceptance criteria. Never encode one side of an unresolved dispute in a task. A clear objective defect with no behavioral ambiguity does not need a clarification round. ## 5. File focused backlog task(s) Unless the user asked for a report only, create or update a task for every validated, actionable follow-up once its scope and expected behavior are clear. Do not ask for a second confirmation to create a task after an audit was requested. - Search first: `pixi run -e bun bunx backlog.md search "<query>" --plain`. Reuse or update an existing task rather than duplicating it; if two tasks overlap, ask which owner keeps the work. - Use the right type: `--type bug` for wrong behavior, `--type feature` for missing behavior, `--type refactor` for behavior-preserving structure, `--type chore` for tooling. A defect is never filed as a refactor. - File against the milestone that owns the surface (`-m m-2` for the bridge/0.3.0 track) and set `--priority` from practical impact, not from how interesting the finding is. - Write criteria a reader can verify: the input or call that proves the defect today, the intended behavior, the seam the test attaches to, and the non-goals that keep the task bounded. Follow [tdd](../tdd/SKILL.md) for that wording. - Create with the CLI only — `pixi run -e bun bunx backlog.md task create ... --plain`. Never hand-edit task metadata, ids, filenames or status. If Pixi and the pinned CLI are unavailable, say exactly which command failed, hand back a ready-to-enter `backlog.md task create` draft, and do not claim the task exists. - Leave every created task open. The audit does not implement its fixes and never marks a task Done. If no actionable follow-up is confirmed, say so and create no task. ## 6. Report Use this structure, omitting empty sections: ```markdown ## Audit **Scope:** what was inspected, at which commit, on which branch. **Summary:** one or two sentences: what holds up, what does not. ### Findings - **[severity] [rule-id]** `path:line` — evidence, impact, and the smallest fix. ### Follow-up tasks - `TASK-NN` — title; or why nothing was filed, or the exact blocker and the draft. ### Checks - `command` — result, or the exact reason it could not run. ### Scope limits - Paths, packages or concerns not examined. ``` Severity describes practical impact only. Distinguish verified evidence from uncertainty, and never imply that a check, probe, or task creation ran when it did not. ## 7. Evaluation prompts - "Audit `packages/qgis-sdk-runtime/src/qgis_sdk/bridge/dispatch.py` for maintainability problems" — expects a read-only scan, `path:line` evidence, rule ids, and tasks filed only for validated follow-ups. - "Audit the last commit before I merge it" — expects the diff to be checked against `git status`, the accepted decisions, and both halves of the bridge contract. - "Audit this, report only" — expects no backlog writes at all and no source edits. - "The audit found a bug that contradicts decision-22" — expects a clarification question and a pause, not a task whose criteria pick a side.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · Archont561/qgis-agent-harness | 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 |
| 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 ddced747a4bc (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
- 100software-code-review
Ocenia kod źródłowy pod kątem błędów, bezpieczeństwa, stylu i utrzymywalności, tworząc raport z ustaleniami i zaleceniami. - 100workflow-refactor
Wykonuje refaktoryzację o ograniczonym zakresie, zachowującą działanie kodu: mapuje zależności, zmienia strukturę i uruchamia testy dotkniętych miejsc. - 100code-quality-audit
Audytuje kod pod kątem code smells, luk w obsłudze błędów i problemów z utrzymywalnością, z konkretnymi poprawkami, w ogólnym przeglądzie niezwiązanym z regułami projektu. - 100code-quality
Uruchamia analizę jakości kodu: złożoność cyklomatyczna, code smells, naruszenia SOLID, martwy kod, duplikacja i ocena utrzymywalności pliku lub katalogu. - 100checking-code-quality
Sprawdza metryki jakości kodu: złożoność, duplikację, konwencje nazw i długość funkcji, przy bramkach jakości, przeglądzie code smells i regułach lintera. - 100slop-scan
Uruchamia analizę jakości kodu przez usługę Slop Bot, wykrywając zapachy kodu i problemy złożoności.