soa-pr-review

name	soa-pr-review
description	Detect and triage AI-generated pull request review comments for the delivery orchestrator flow. Use when poll-review finds AI review feedback or when you need to inspect recent AI review comments on the current PR.

AI Code Review

Supported external review agents: coderabbit, qodo, greptile, sonarqube. Other vendors are unsupported unless repo policy adds them.

Boundary

The orchestrator owns polling cadence, state transitions, artifact persistence, and auto-recording clean on the final check.

This skill owns: fetching review data with gh, detection logic, normalizing comments into structured artifacts, and triage judgment.

Contract:

• fetcher outputs: detected, agents, reviewed_head_sha, vendors, comments
• triager outputs: outcome (clean|needs_patch|patched), note, action_summary, non_action_summary, vendors

When triager returns needs_patch, follow-up must conclude as patched or operator_input_needed — not stop permanently at needs_patch.

Workflow

1. Resolve PR number with gh pr view if not provided.
2. Fetch with .agents/skills/pr-review/scripts/fetch_pr_review_comments.sh <pr-number>.
3. If orchestrator already saved review.fetch.json, use that as the source of truth for vendor attribution and comment shape.
4. Detection policy: supported vendor identities, explicit vendor wording in comment body, check-run annotations. Human drive-by comments do not count. Preserve head SHA, inline resolution/outdated state, and native thread identity.
5. Return fetcher contract when inside poll-review: detected=false → keep polling or auto-clean; detected=true → orchestrator inspects agent state and decides.
6. Triage each detected comment: actionable, stale, wrong, over-scoped, or out of scope.
7. Treat AI review as advisory, not authoritative.
8. Push back on stale, over-scoped, unnecessary, or policy-conflicting suggestions.
9. When running as triager hook, return the triager contract and let the agent environment decide whether to patch or stop.
10. If user approves a patch: apply it, run the smallest relevant verification, commit, push. If the finding came from a native GitHub inline thread still resolvable, mark it resolved.

Output Expectations

• State whether AI review comments were found.
• Filter summary noise; focus on unresolved review items.
• Group by unresolved issue, not raw API payload.
• Include file and line when available.
• Distinguish current-head review from stale-history review when reviewed SHA no longer matches branch head.
• State which comments are actionable and which should be rejected.
• Call out stale or already-addressed comments rather than blindly implementing them.
• Stop for operator input only when the right action is genuinely ambiguous.