review-pr

Review PR #$ARGUMENTS by invoking @ai-sdlc/orchestrator's executeReview() for the three review perspectives. The orchestrator already drives the LLM-based ReviewAgentRunner and applies the meta-review pass that filters medium-confidence findings; this skill's job is to fetch context, fan out to the three review types, and present the structured verdicts.

Step 1 — Fetch PR context

Don't hardcode --repo — let the cwd's git remote drive gh.

PR=$ARGUMENTS

# Diff + metadata for the review agents
gh pr diff "$PR" > /tmp/pr-diff.txt
gh pr view "$PR" --json number,title,body,headRefName,changedFiles > /tmp/pr.json

# Linked issue (if any) — feeds acceptance-criteria extraction
LINKED=$(gh pr view "$PR" --json body --jq '
  (.body | scan("(?i)(?:closes|fixes|resolves)\\s+#([0-9]+)"))[0][0] // empty
')
if [ -n "$LINKED" ]; then
  gh issue view "$LINKED" --json number,title,body > /tmp/issue.json
fi

If there's no linked issue, omit --issue-file from the calls below — cli-review falls back to the PR title/body.

Step 2 — Run the three review types

for TYPE in testing critic security; do
  pnpm --filter @ai-sdlc/dogfood review \
    --pr "$PR" \
    --diff-file /tmp/pr-diff.txt \
    --type "$TYPE" \
    ${LINKED:+--issue-file /tmp/issue.json} \
    > "/tmp/review-$TYPE.json" 2>"/tmp/review-$TYPE.stderr"
done

Each call writes a structured ReviewVerdict JSON to its own file:

{
  "approved": true | false,
  "findings": [
    { "severity": "critical"|"major"|"minor"|"suggestion",
      "file": "path",
      "line": 42,
      "message": "string" }
  ],
  "summary": "string"
}

If any of the three calls writes to stderr, surface it — typically a config issue, not a true review failure.

Step 3 — Present verdicts

For each review type in order (testing, critic, security):

1. Header line — Testing: APPROVED with 2 suggestions or Critic: CHANGES REQUESTED — 1 critical, 3 major
2. Summary — the orchestrator's summary string
3. Findings — only critical and major; minor and suggestion go in a collapsed list