PR Review Orchestrator

Relies on BDK foundation (STARTUP_INSTRUCTIONS.md) for project context.

Take PR URLs, spawn one reviewer subagent per PR, and land the result where it belongs: on GitHub, as inline comments plus a templated summary ending in an explicit verdict. --verify runs the follow-up pass instead: did the author implement what the previous review asked for?

Safety Rules (MANDATORY)

• MUST NOT modify source files. disallowed-tools: Edit Write NotebookEdit removes them mechanically; the subagents are instructed to be read-only in their worktrees. Findings become GitHub comments, never edits.
• Nothing is posted to GitHub until the user confirms. Reviewer subagents compute findings, a verdict, and the data a review would be rendered from - they MUST NOT call the review-posting API or any thread-resolve mutation. The orchestrator aggregates every PR in this run, shows a full terminal report (verdict, blockers, and the complete nice-to-have list - nice-to-haves sometimes hide things that actually matter), and lets the user confirm or override each PR's verdict before any GitHub call happens.
• Once confirmed, post only through the templates. Everything posted goes to GitHub where the PR author and team see it, rendered only from references/comment-templates.md. No ad-hoc comment shapes.
• One review call per PR. All inline comments + summary + event in a single API call - per-finding top-level comments spam notifications.

Terminal Output

On start:

┌─────────────────────────────────────────────────┐
│  👁️  ORCHESTRATOR: pr-review                     │
│  📋 PRs: {N}  Mode: {review|verify}              │
│  ⚡ Model: sonnet                                │
└─────────────────────────────────────────────────┘

During execution:

[pr-review] PR #{n}: {title} ({base_ref} ← {head_ref}){ [stack: parent #{m}]}
[pr-review] Worktrees ready: {N}
[pr-review] Dispatching {N} reviewer subagents (each runs /bdk:cr --inline, no nested agents)...
[pr-review] PR #{n}: computed {verdict} ({c}C/{h}H/{m}M/{l}L) - nothing posted yet

Before asking for confirmation (Step 5), per PR:

── PR #{n}: {title} ── computed verdict: {✅ Approve | ❌ Request changes}
Blockers ({n}):
  - {path}:{line} [{SEVERITY}] {one-sentence problem}
Nice to have ({n}) - review these, real issues sometimes land here:
  - {path}:{line} [{category}] {one-sentence problem} → {one-sentence fix}

After posting (Step 6):

[pr-review] PR #{n}: ✓ posted {event} ({k} inline comments){ - verdict overridden from {computed}}
[pr-review] Done: {approved} approved, {changes_requested} changes requested

Step 0 - Parse the arguments

• Every https://github.com/{owner}/{repo}/pull/{number} token is a PR to process.
• --verify switches every PR in this invocation to verify mode.
• Remaining free text is a focus hint passed through to each reviewer.
• No URL in the arguments → ask the user for one; there is nothing to guess.

If the gh-axi skill is available, prefer it for the GitHub calls below; raw gh is the fallback.

Step 1 - Gather PR metadata

For each PR:

gh pr view {url} --json number,title,url,state,isDraft,author,baseRefName,headRefName,headRefOid,additions,deletions,changedFiles

• state not OPEN → skip it and tell the user; reviewing a merged or closed PR changes nothing.
• Draft PRs get reviewed normally, but the reviewer is told (via notes) to mention draft status in the summary.

Stack detection. Get the default branch (gh repo view {owner}/{repo} --json defaultBranchRef). When baseRefName differs from it, look for an open PR whose head is that base (gh pr list --repo {owner}/{repo} --state open --head {baseRefName} --json number,url). A hit means this PR sits in a stack (the layout stacked-PR tools such as gh-stack produce): the review scope is this PR's own diff vs its parent branch - the parent's changes get their own review in their own PR. Pass the parent PR into the prompt's {stack_context}. No auto-expansion: one URL reviews one PR; the user lists every stack entry they want reviewed.

Repo check. The PR's repo must be reachable from the current directory's git remote. When it is not, clone it once into the scratchpad (gh repo clone {owner}/{repo} {scratchpad}/{repo} -- --filter=blob:none) and run the per-PR git commands there.

Step 2 - Prepare one worktree per PR

Sequentially (concurrent fetches contend on ref locks), for each PR:

git fetch origin "pull/{number}/head" "{baseRefName}"
sha=$(git rev-parse FETCH_HEAD)   # run immediately after the fetch for THIS pr
dir=$(mktemp -d -t "bdk-pr-{number}")
git worktree add --detach "$dir" "$sha"

pull/{number}/head works for fork PRs too, where headRefName does not exist on origin. Detached worktrees keep parallel reviewers from touching each other or the user's checkout.

Step 3 - Dispatch reviewers

Read the prompt templates:

!cat ${CLAUDE_PLUGIN_ROOT}/skills/pr-review/references/reviewer-prompt.md

Fill template A (review) or B (verify) per PR - every placeholder, including {plugin_root} = ${CLAUDE_PLUGIN_ROOT} - and launch all subagents as general-purpose in a single message with multiple Agent calls. The template's no-subagent rule is the load-bearing line: each reviewer runs the whole /bdk:cr --inline engine sequentially in its own session, because a subagent cannot spawn agents and must not try. The other load-bearing line is no-posting: a reviewer computes a verdict and the data a review renders from, and returns it - it never calls the GitHub posting API or a thread-resolve mutation. Posting happens once, in Step 6, after the user has seen every PR and confirmed.

Wait for the completion notifications. Do not poll, do not schedule wake-ups. A reviewer whose finding needs clarification is continued with SendMessage, not re-spawned.

Step 4 - Parse results and clean up worktrees

1. Parse each PR_REVIEW_RESULT / PR_VERIFY_RESULT block and its JSON payload. A subagent that died or returned no block is reported as failed for that PR - never fabricate a verdict for it, and leave nothing half-posted unmentioned. It is excluded from Step 5/6 (nothing to confirm, nothing to post).
2. Remove the worktrees: git worktree remove --force "$dir" per PR, then git worktree prune. Remove the scratchpad clones too. Nothing in Step 5/6 needs worktree access, so clean up now rather than holding them through the (interactive, possibly slow) confirmation step.

Step 5 - Report and get verdict confirmation

For every parsed PR, print the per-PR report block from the Terminal Output section above: link/title, computed verdict, the full blockers list, and the complete nice-to-have list (not just a count - the whole point of surfacing them is that a real issue sometimes got classified MEDIUM/LOW, and the user can only catch that by reading it).

Then ask the user to confirm or override, one AskUserQuestion question per PR: header PR #{n}, question naming the PR and its computed verdict, two options - the computed verdict first, labeled (Recommended), and the opposite verdict second. Batch up to 4 questions in a single AskUserQuestion call (its hard limit); for more than 4 PRs in this run, issue additional calls, one batch after another. Record each PR's final_verdict from the answer - it may equal or differ from computed_verdict.

Own-PR note: still ask (the confirmed verdict still drives the summary body's verdict line), but the GitHub event sent in Step 6 is forced to COMMENT regardless of the answer, since GitHub rejects self-approval/self-request-changes.

Step 6 - Render and post confirmed reviews

For each PR, using its stored payload and final_verdict:

1. Render the summary body from references/comment-templates.md template 3 (review) or template 5 (verify) - blockers section included whenever blockers is non-empty, even when final_verdict is approve, and the override note added whenever final_verdict != computed_verdict.
2. Map to the GitHub event: approve → APPROVE, request-changes → REQUEST_CHANGES, own PR → COMMENT (see above).
3. Post the one review call per PR (inline blocker comments + summary body + event) per the "Posting mechanics" section of comment-templates.md. On a 422 anchor failure, drop that comment into the summary's Context findings section and retry once - never retry the identical payload.
4. Verify mode only: after the review call succeeds, resolve every thread in threads_to_resolve (ours, classified ✅ in Step 2 of the verify template).
5. Print the "posted" completion line for that PR.

Print the final Done: {approved} approved, {changes_requested} changes requested line once every PR in the run has been posted (or reported failed in Step 4).

Verdict Policy (single source, mirrored in the prompts)

Nice-to-haves never block by themselves - the computed verdict never turns to request-changes for MEDIUM/LOW alone. But they are never silently dropped either: the full list reaches the user in Step 5, precisely so a nice-to-have that is actually important can get its own override. Nitpicks - style pedantry, linter-territory, personal taste - are at most nice-to-haves.

Rules

• Always print the terminal block on start, the per-PR computed-verdict lines, the full report in Step 5, and the posted-completion lines in Step 6.
• Reviewer subagents compute and return; they never call the GitHub posting API or a thread-resolve mutation. Only the orchestrator posts, and only after Step 5 confirmation.
• Comments come from references/comment-templates.md only, hidden markers included - --verify depends on them.
• One reviewer subagent per PR, no nested subagents, /bdk:cr --inline inside. State this in every dispatch prompt even though the template already carries it.
• Every review and every verification ends in exactly one of: approve, request changes - as confirmed or overridden by the user in Step 5, never assumed.
• Worktrees are always removed, including on failure paths.