project-audit
/project-audit
Koordynuje audyt istniejącego projektu pod kątem zachowania kodu, testów, bezpieczeństwa, ścieżek użytkownika i architektury, z jednym raportem.
- Autor
- JessePomeroy ↗
- Repozytorium
- iced-coffee-suite ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 06.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 06.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Koordynuje audyt istniejącego projektu pod kątem zachowania kodu, testów, bezpieczeństwa, ścieżek użytkownika i architektury, z jednym raportem.
Oryginalny opis (EN)
Coordinate an evidence-backed audit of an existing project across code behavior, tests, security, user journeys, and architecture, using applicable installed skills and one consolidated report. Use for a full project audit or a scoped multi-discipline review; no repairs or delivery by default.
Jak działa
- 1. Scope and inventory
- 2. Safe baseline
- 3. Code, tests, and security
- 4. User-facing and runtime verification
- 5. Architecture and optional design review
- 6. Reconcile once and report
- Coordination and next steps
Podgląd SKILL.md
--- name: project-audit description: Coordinate an evidence-backed audit of an existing project across code behavior, tests, security, user journeys, and architecture, using applicable installed skills and one consolidated report. Use for a full project audit or a scoped multi-discipline review; no repairs or delivery by default. --- # Project audit Audit an existing project through complementary skills, not a stack of duplicate reports. Resolve the repository/path from the request or current project. Ask for the target only when it cannot be determined unambiguously; never infer a project from this skill's own location. Do not start auditing an unrelated checkout. The default is report-only. No source/config/skill edits, test deletion, dependency installation, automatic fixes, commits, PRs, deployment, publishing, or messages. User-approved local verification may create task-owned scratch/results, but must preserve existing tracked and untracked work and shared state. Production mutations, active attacks, paid calls, and sending data to external services need appropriate explicit authority. A complete audit is not a production certification. ## 1. Scope and inventory Read project and nested instructions, Git state/revision, package/toolchain config, supported entrypoints, application surfaces, data stores, integrations, CI/release configuration, and existing tests/verification skills. Inspect secrets handling without reading or printing secret values. Identify safe test data and whether local runtime work can be isolated. Start no server just to inspect source. Use `recon` for a bounded ownership/entrypoint map when available. Its Svelte lens applies to Svelte; do not manufacture rune concerns for other stacks. Map omitted packages and external dependencies as limits, not implicitly audited code. Choose full or focused scope from the user request: - **Full:** all applicable disciplines below across a declared domain/feature inventory, with risk-based deep checks and explicit sampling limits. - **Focused:** the named subsystem, task, or diff plus necessary caller context. Write a compact coverage plan: domain/feature, intended check, owner skill, expected evidence, and known prerequisites. Full does not mean every possible state or every line is proved correct. Do not silently reduce a full request to one homepage or a file inventory. Process planned domains in bounded passes; stop at blockers, explicit budget, cancellation, or scope completion, and show unfinished coverage. Do not repeat hunts until no new findings appear. ## 2. Safe baseline Inspect runner scripts, imports, fixtures, network access, and environment loading before executing checks. Collection can have side effects. Run the relevant safe existing type/lint/build/test checks, recording exact commands and revision. Use the project's environment and package manager. Do not install dependencies, run migrations, seed shared databases, or switch environments to force a baseline. Record existing failures, missing dependencies, and unavailable checks separately. Locate the project `verify-*` skill and feature map. Reuse its health/isolation recipes. If none exists, use the supported harness that is already available. Recommend `create-verification-skill` as a separate setup task; do not create or repair a project harness during a report-only audit. A missing harness need not block source review, but limits runtime claims. ## 3. Code, tests, and security Load each selected skill's instructions when its phase starts. Reuse context and evidence without claiming a second independent review for the same agent. - **Code behavior — `bug-hunter`:** review each assigned domain and necessary callers for reachable failures. Use its Hunter/Skeptic/Referee reasoning and distinguish local passes from independent reviewers. Map full component behavior for Svelte; use general runtime/request-boundary analysis for other stacks. A triage/index entry is not a reviewed file. Do not enable fix or commit modes. - **Tests — `testprune` in audit mode:** assess invariants, misleading tests, unsafe external calls, collection gaps, fixture costs, and test command/CI meaning. No removal, gate rewrite, or new test infrastructure in this phase. - **Security — `security-review`:** inspect trust boundaries, authorization, sensitive data, and relevant dependency reachability. Reuse a verified threat model or build the required context within the audit's authorized output scope. Apply `vulnerability-validation` to serious candidates; benign local proof is preferred, and unavailable exploitability evidence remains unresolved. Security workflow artifact paths refer to the run's isolated `.bug-hunter/`, not permission to write into the source repository. Follow Bug Hunter's runtime workflow for structured outputs and canonical filenames/schemas. Use an explicit source target and isolated output root. If a helper cannot separate these or requires unapproved access, report the capability gap rather than bypassing it. Resolve helper paths against the installed skill bundle, not the project cwd. Check available dependency scanner support for the actual lockfile/language. An unsupported scanner or unavailable/current advisory lookup is not a clean result. Do not upload source, lockfiles, or dependency inventories to a service without appropriate authority. Public documentation/advisory research must follow the host's allowed retrieval tools and privacy rules. ## 4. User-facing and runtime verification For an app with a UI, use `ux-audit` with the matching declared scope and its project feature map. Use `web-debug` to investigate specific failing interactions, not to duplicate the entire walkthrough. Cover relevant keyboard/focus, responsive layout, errors/recovery, and persistence with actual interaction evidence. Measure performance only through supported tooling; unmeasured is not zero. For APIs, CLIs, libraries, or workers, exercise supported request/command/consumer boundaries with existing safe fixtures. UI-only checks are not applicable, not a reason to force a browser or mark the entire project failed. Deployment smoke checks must target an authorized environment and remain within its permissions. Maintain one driver per shared instance and clean up only task-owned resources. If auth, browser capability, or safe fixtures are missing, retain completed source findings and mark the affected runtime coverage incomplete. ## 5. Architecture and optional design review Use `codebase-design` and `improve-codebase-architecture` in review/design-only mode to assess observed ownership, coupling, caller complexity, and testability friction. Prioritize actual hotspots and evidence from earlier phases; do not produce speculative redesigns for every module. Keep improvement proposals separate from verified defects. Use `blast-radius` when there is an actual proposed change, pending diff, or named migration whose downstream effects need review. Do not invent a diff to make it fit an audit of the current system. Visual hierarchy, usability, and accessibility observations belong in the live UX phase. A separate Impeccable critique is optional when design judgment is part of the request; follow its actual read-only/permission protocol and disclose any unavailable required detector or independent assessment. Do not silently run its setup/repair operations or force its full critique for every technical audit. **Punchlist is opt-in only:** invoke it when the user specifically requests its named-defect inspection. Preserve its current critic, coverage, privacy, and verification protocol; this wrapper does not implement the proposed bounded mode. If its required independence cannot run under the user's delegation permissions, report that phase blocked or obtain the required choice. Never simulate its panel. ## 6. Reconcile once and report Join phase outputs by root cause and affected behavior, preserving evidence and all affected paths. Recheck apparent contradictions against current source. Use `vulnerability-validation` for security reachability and `diagnosing-bugs` for a specific unresolved failure when the needed probe remains in audit scope. Do not reinterpret an uncertain claim as confirmed just because several tools reported it. A source change mid-audit invalidates affected evidence; record the revision mismatch and recheck that portion or report it incomplete. Return one report: - Target/revision, scope, applicable disciplines, and declared exclusions. - Confirmed findings ordered by demonstrated impact, with trigger, expected and observed result, source/evidence, and a scoped correction direction. - Architecture/design/test-maintenance recommendations, clearly separate. - Unresolved hypotheses and the exact evidence needed to decide them. - Coverage by domain and check: exercised, not applicable, or not assessed. - Actual commands/interactions, baseline failures, and verification limits. - Prioritized next actions and dependencies, without implementing them. Report completion within the declared scope separately from issues found. Incomplete coverage cannot become a clean bill of health. Do not issue a global numerical quality grade or pad a fixed finding count. Use in-chat output by default; save a report when requested and keep structured run artifacts in task-owned scratch. Repository-owned report files need that write scope. ## Coordination and next steps Multiple skills do not imply multiple agents. Run sequentially by default and reuse the domain map. Delegate only when the user or applicable project authority permits it, with disjoint scopes and explicit output ownership. Read each required phase's skill; never claim it ran because its name appeared in the plan. If a skill is absent, report the gap and use an honest equivalent review when possible, without installing a plugin or pretending that skill's protocol completed. Repairs, verification-harness creation, lint installation, and maintenance edits are separate follow-ups. Use `fixer` for explicitly approved confirmed repairs, `babysit` only for authorized delivery through merge, and `reflect` only when a retrospective is requested. A project audit never auto-ships its findings.
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · JessePomeroy/iced-coffee-suite | 90 | MIT |
| drayline/rootnode-skills | 100 | Apache-2.0 |
| engida32/project-audit-skill | 100 | MIT |
| sd0xdev/sd0x-harness | 100 | MIT |
| static-var/keystone | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 95c05764febc (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
- 100comprehensive-review-full-review
Koordynuje przeglądy jakości kodu, architektury, bezpieczeństwa, wydajności, testów i dokumentacji, gdy zmiana lub baza kodu potrzebuje skonsolidowanej oceny. - 90audit-flow
Przeprowadza audyt projektu, zmiany lub planu pod kątem ryzyka, jakości i gotowości, tylko do odczytu, bez narzucania poprawek. - 100error-debugging-multi-agent-review
Koordynuje równoległe przeglądy kodu pod kątem bezpieczeństwa, architektury i wydajności oraz zbiera ich wyniki. - 90project-delivery-orchestrator
Koordynuje wieloetapowe dostarczenie oprogramowania od zakresu, przez implementację i przegląd, po weryfikację i wydanie, łącząc dostępne narzędzia i ludzi. - 100security-audit
Przegląda kod, diffy i zmiany zależności pod kątem problemów bezpieczeństwa potwierdzonych dowodami. - 100architecture-audit
Przeprowadza audyt architektury istniejącego kodu tylko do odczytu: sprzężenia, cykle zależności, bezpieczeństwo i plan naprawy z priorytetami.