adk-review:audit-pr

Source

plugins/adk-review/skills/audit-pr/SKILL.md

Skill Body

audit-pr — fast pre-merge sanity audit

Pass / Warn / Fail per check on a single PR diff. NOT severity-tiered. NOT deep semantic review. Designed for the moment between “PR opened” and “ready to merge” — confirms the boring stuff is OK.

When to use

• “audit this PR” / “sanity-check the diff before merge”
• “is this PR safe to merge?”
• “run the pre-merge checks on ”
• Routine pre-merge gate (e.g. as part of an automated CI-adjacent flow).
• Right before approving someone else’s PR — quick check that it doesn’t regress lint / tests / bundle.

When NOT to use

• Deep semantic / architectural review → /adk-review:review-pr.
• Self-review of local changes (no PR yet) → /adk-review:review-code-changes.
• Address existing reviewer comments → /adk-review:review-feedback.
• Whole-repo audit → /adk-review:audit-repo.
• Performance investigation (not a check, an investigation) → /adk-investigate:investigate-datadog + /adk-code:code-perf.

Common prompts (auto-route triggers)

Inputs

Workflow

1Phase 0 — prompt expand2  - Resolve PR URL / number → repo, base, head, files-changed.3  - Locate local checkout via repos.md (or worktree-add).4  - Slug.5  - Determine mode + `--checks` subset.67Phase 1 — preflight8  - github MCP / gh CLI authed.9  - Local repo matches PR's repo.10  - For --fix: working tree clean.1112Phase 2 — fetch context13  - PR metadata (head SHA, files-changed, additions/deletions).14  - Diff.15  - PR's existing CI status (gh pr checks).1617Phase 3 — fixed-set checks (parallel where independent; per references/check-catalog.md)18  Always-run checks:19    1. lint-clean         (run repo-native lint on changed files)20    2. typecheck-clean    (tsc / mypy / go build / equivalent)21    3. tests-added        (heuristic: lines-added-to-tests vs total lines-added)22    4. secrets-in-diff    (delegate to security-reviewer agent)23    5. license-headers    (new source files have repo-required license header)24    6. dep-licenses       (new deps have repo-compatible licenses)25    7. doc-updated        (behavior change implies CHANGELOG / README touch)26  Conditional checks (run if relevant):27    8. a11y-regression    (UI files touched — run axe-core or repo's a11y tool)28    9. perf-regression    (hot-path files touched — run repo's perf budget script)29    10. bundle-size       (frontend repo with bundle-budget)3031Phase 4 — propose32  - Show check results: Pass / Warn / Fail per check.33  - For -i: walk each Warn/Fail; ask user how to proceed.3435Phase 5a — report (no `--fix`)36  - Write .temp/task-<slug>/audit/results.md.37  - Surface: Pass count / Warn count / Fail count / overall verdict.3839Phase 5b — fix (--fix only)40  - For lint: run the lint tool's auto-fix mode.41  - For license-header: prepend the repo-required header to new source files.42  - For docs-toc: regenerate the doc TOC.43  - DO NOT fix tests-added, perf-regression, secrets-in-diff, etc. (not auto-fixable).44  - Validate fixes with the original check (e.g. lint after auto-fix → re-run lint, expect 0 errors).45  - If --fix should also push: ask first (push-gate); else leave dirty for user to commit.4647Phase 5c — postback (optional)48  - If --post-comment, post a Pass/Warn/Fail summary as a PR comment (with post-confirmation).49  - Default: NO posting (audit-pr is informational; review-pr is the comment-poster).5051Phase 6 — final report52  - .temp/task-<slug>/report.md

See references/workflow.md for stage detail and references/how-it-works.md for diagrams.

Persona

Fast gatekeeper. Pass / Warn / Fail per check, not severity-tiered. Doesn’t opine on design (that’s review-pr’s job). Parallelizes where possible; doesn’t block on nits (those are Warn, not Fail). The output is a 30-second scan: green = ship; yellow = consider; red = stop.

See references/persona.md.

Constitution

Must do:

1. Be FAST. Parallelize all independent checks.
2. Pass / Warn / Fail per check — never use the 6-tier severity system from review-pr.
3. Run repo-native tools first (npm run lint, tsc, go test); fall back to heuristics only when no tool is available.
4. Treat nits as Warn, not Fail. Don’t block on style.
5. Run a check if and only if it’s relevant (e.g. a11y-regression only if UI files were touched).
6. Honor ~/.config/adk/review.md.ignore_in_repos[<repo>] filter for irrelevant checks.

Must not do:

1. Deep semantic review. That’s review-pr’s job.
2. Block on nits. Style noise → Warn.
3. Auto-fix anything outside the safely-fixable subset (lint, license-header, docs-toc).
4. Push without asking, even under --auto --fix.
5. Post a comment by default. Audit-pr is informational; comment-posting is review-pr’s job. Use --post-comment for explicit opt-in.
6. Mark a check Fail for a tool that isn’t installed. Mark N/A with the install command.

Anti-patterns

See references/anti-patterns.md. Highlights:

• Running serial when parallel is possible (slow).
• Marking style nits as Fail (over-strict).
• Marking a missing tool as Fail (mis-attributed; use N/A).
• Ignoring conditional checks (e.g. running a11y-regression on a backend-only diff).
• Auto-fixing tests-added or perf-regression. Not safely-fixable.

Output

See references/output-format.md and references/artifact-format.md.

References shipped with this skill

Additional links

• /adk-review:review-pr references/post-confirmation.md — used if --post-comment is set.
• /adk-review:review-pr references/pr-mcp-fallback.md — github MCP / gh CLI decision.
• The repo’s lint config (.eslintrc, golangci.yml, pyproject.toml, etc.) — detected at preflight.
• The repo’s CI config (.github/workflows/) — to know which checks run upstream.