Before anything else: read $HOME/.agentic-workflow/toolkit/skills/_preamble.md and follow it (skill index, provider capability map, bootstrap check, session context). Run its Session Close section when this skill finishes.

PR Review Orchestrator

Runs parallel domain-specific reviewers and saves findings locally. Does not post to GitHub — run /postReview when ready.

Step 1: Resolve the PR

If an argument was provided, use it directly:

gh pr view <argument> --json number,title,headRefName,baseRefName,url,headRepository

If no argument was provided, auto-detect from the current branch:

gh pr list --head $(git branch --show-current) --json number,title,url

If multiple PRs are returned, Ask the user to pick one before proceeding.

If no PRs are found: "No open PR found for the current branch. Use /review <number> to specify one."

Step 2: Fetch PR Context

Run in parallel:

gh pr diff {number}
gh pr view {number} --json number,title,body,additions,deletions,headRefName,baseRefName,headRepository
gh pr view {number} --json files --jq '[.files[].path]'
gh pr view {number} --json commits --jq '.commits[-1].oid'

Step 3: Ensure output directory exists and gather evidence

SHARED_DIR="$HOME/.agentic-workflow/toolkit/skills/_shared"
source "$SHARED_DIR/repo-slug.sh"
mkdir -p "$AW_DIR/reviews"
ls -1dt "$AW_DIR/verification/"*/ 2>/dev/null | head -1   # newest pack.json dir
ls -1dt "$AW_DIR/design/verify/"*/ 2>/dev/null | head -1  # newest comparison-report.json dir

Read pack.json and comparison-report.json from those newest dirs (per _shared/evidence-pack.md). Their contents become {evidence} for reviewers — the literal none when neither exists.

Step 4: Triage

Spawn a subagent (general-purpose) with the triage prompt from triage-prompt.md.

Inject:

• {title}, {body} — PR metadata
• {file_list} — JSON array of changed file paths
• {diff} — full diff output

Parse the returned JSON array of reviewer assignments. If the user passed --lens <persona,...>, append one assignment per named lens/persona (dedupe by agent name) — these run in addition to triage output.

Step 5: Spawn Parallel Reviewers

Dispatch in parallel — all reviewers at once, per _shared/parallel-dispatch.md. Each reviewer receives the prompt from reviewer-prompt.md with these values injected:

• {agent}, {focus}, {number}, {title}, {diff}, {evidence} (from Step 3)

Each reviewer returns a JSON object as its final output (not GitHub comments). Collect all responses.

Sub-skill Dispatch

After collecting all reviewer JSON outputs:

1. Check for any reviewer output with investigation_needed: true
2. If found: rank candidates deterministically — severity first (blocking before issue), then longer error trace, then lowest issue id (lexicographic) — and take up to 3
3. Ask the user:

   "Found {N} blocking bug(s) with error traces. Run rootCause on the top {min(N,3)}? (yes/no)"

4. If yes: Dispatch in parallel up to 3 rootCause runs — one Invoke skill rootCause (args: "<investigation_error value>") per candidate, launched together
   • Attach each investigation file path to its issue in the state file under "investigation"
   • If a rootCause returns scope-breach, note it in the state file and continue — do not block the review
5. If no, or no reviewer flagged investigation_needed: skip and proceed to writing the state file

Step 6: Write State File

Combine all reviewer outputs into ~/.agentic-workflow/<repo-slug>/reviews/{number}.json:

{
  "pr": {
    "number": 123,
    "title": "Add auth middleware",
    "branch": "feature/auth",
    "owner": "myorg",
    "repo": "myrepo",
    "url": "https://github.com/myorg/myrepo/pull/123"
  },
  "commit_sha": "<latest commit oid>",
  "review_run": 1,
  "previous_run_finding_count": null,
  "lenses": ["security-sentinel", "kieran-typescript-reviewer"],
  "reviewed_at": "<ISO timestamp>",
  "posted": false,
  "posted_at": null,
  "reviewers": [
    {
      "agent": "security-sentinel",
      "focus": "auth, input validation",
      "summary": "Found 2 blocking issues...",
      "issues": [
        {
          "id": "sec-0",
          "severity": "blocking",
          "path": "src/auth.ts",
          "diff_position": 42,
          "summary": "JWT not verified before use",
          "body": "**[blocking] JWT not verified before use**\n\nFull comment text...",
          "type": "inline",
          "addressed": false,
          "posted_comment_id": null
        }
      ]
    }
  ]
}

Save this file (write it in full). lenses lists the agent names actually spawned in Step 5. On re-review (file already exists): set review_run to prior value + 1 and previous_run_finding_count to the prior run's total issue count before overwriting.

Step 7: Report to User

Review complete for PR #{number}: "{title}"
Findings saved to ~/.agentic-workflow/<repo-slug>/reviews/{number}.json
Run #{review_run} (previous run findings: {previous_run_finding_count})
Lenses: {lenses, comma-joined}

Reviewers:
  • security-sentinel (focus: auth, input validation) — 2 blocking, 1 issue
  • kieran-typescript-reviewer (focus: type safety) — 0 issues

Run /postReview to publish comments to GitHub.
Run /addressReview to start implementing fixes.

Next steps

• /postReview — publish the findings to GitHub as batched comments
• /addressReview — start fixing the issues in parallel
• /rootCause — investigate a specific blocking bug from the review