Agentpack Review

AgentPack Review

Use when the user invokes $agentpack-review or $agentpack-review <reviewer context>.

Do not claim correctness unless relevant checks actually ran.

Steps

1. Refresh AgentPack context for this exact review task before reading PR diff or code. Prefer MCP:

agentpack_pack_context(task="review current PR $ARGUMENTS")

If MCP is unavailable, run:

agentpack guard --agent auto --repair-stale --refresh-context

If you bypass this refresh, state the bypass reason before continuing. 2. Prepare the full review bundle. If $ARGUMENTS names a PR number or PR URL, agentpack review must bind metadata, diff, and context to that PR. If no PR target is named, use local review only with explicit --allow-local-fallback and a resolvable base (prefer --base origin/main); never run ambient gh pr view or silently compare HEAD~1.

agentpack review "$ARGUMENTS"

Use this explicit form when the PR target is known:

agentpack review --pr <number-or-url> "$ARGUMENTS"

3. Read .agentpack/review.prompt.md and follow it end to end.
4. By default, agentpack review starts a fresh run under .agentpack/reviews/<branch-or-pr>/<run_id>/ and refreshes the stable alias files in .agentpack/.
5. Do not perform the review inline from this skill. If you cannot write the required files, stop and report blocked.
6. The Anchor role starts from .agentpack/review-understanding.template.toon and writes the compatible run-scoped understanding TOON at the output path declared by agentpack review.
7. Run agentpack review --check; do not start Judge unless Anchor validates.
8. Judge must read that understanding TOON from disk, start from .agentpack/review-findings.template.toon, and write the candidate findings TOON at the declared run-scoped path.
9. Run agentpack review --check; Critic must read the canonical Anchor and Judge handoffs, start from .agentpack/review-critique.template.toon, and write exactly one accept, reject, or downgrade decision for every Judge finding at the declared path.
10. Run agentpack review --check to validate Critic and generate approved-findings.toon. --dry-run-post and --post-inline-comments consume only that approved artifact. Actor is publish-only: it never edits or pushes a PR branch. Do not produce a final review summary unless Critic validates and any intended PR-bound inline post succeeds.
11. If an older model emits valid JSON or fenced output instead of TOON, rerun agentpack review --check; AgentPack canonicalizes schema-valid output to TOON and writes a repair guide for invalid output.
12. Resume an interrupted run only with agentpack review --resume <run_id>.
13. Use the latest PR head, gh pr view, git diff, and direct code reads as source of truth.
14. Treat any non-PR portion of $ARGUMENTS only as a prioritization lens. It must not replace code evidence.
15. Report approved findings first with file evidence, then state inline-post status and validation exactly: dry-run passed, posted, failed, or not run.