Review Pending PR Reviews

Review Pending PR Reviews

The outbox for /review-pr. Nothing reaches GitHub until the maintainer approves it here.

This skill is where a human is guaranteed to be present. /review-pr runs headless most of the time — review-prs drives it across tmux panes, and a cron runs it overnight — so its own Step 8.5 grill is skipped on the majority of reviews. This skill is therefore the real evaluation gate: it grills the findings first (Step 3) and only then offers to post.

Configuration

• TRIAGE_DIR — where drafts live. Default: ~/.nx-pr-reviews, matching /review-pr's own default. Override with the same value if the maintainer has repointed it.

Do not read ~/.claude/triage/prs/. That is the store of the retired review-pr-deep pipeline. Its drafts predate the current review criteria (pre-PIPELINE_VERSION tiers), so posting one would publish findings the current calibrations would have dropped. If the maintainer wants those, they must be re-reviewed by /review-pr, not posted from the old store.

Each draft has YAML frontmatter (pr, verdict, head_sha, pipeline_version, posted_at, ...) and a ## Review draft section holding the body that will be posted.

A draft is pending iff posted_at: is empty. Once posted or discarded, posted_at is set and the file is kept as the historical record — it feeds /review-pr's re-review dedup, so never delete one.

1. Find pending drafts

ls $TRIAGE_DIR

Read each file's frontmatter; pending iff posted_at: is empty. Sort oldest-first by last_reviewed_at so the longest-waiting review goes first.

For each, display: PR number and title, verdict, attempt, last_reviewed_at, URL, and the first 3 lines of ## Review draft.

2. Check for drift before presenting

gh pr view PR_NUMBER --repo nrwl/nx --json headRefOid,state,isDraft -q '{head: .headRefOid, state: .state, isDraft: .isDraft}'

IMPORTANT: every Bash command must be a single, standalone command. Do not chain with &&, ||, or ;, and do not append redirects or backgrounding — these break permission matching. Use separate Bash calls.

Flag before the action prompt:

• Stale — current head differs from frontmatter head_sha. Recommend skip and re-running /review-pr PR_NUMBER.
• Closed/merged — state != OPEN. Recommend discard.
• Draft PR — isDraft == true. The maintainer may not want to post yet.
• Old pipeline — pipeline_version is missing or below the current constant for the draft's writer. The reviewer key names that writer: absent means Claude's /review-pr, whose constant is 9; codex means the Codex review-pr skill, which keeps its own. The draft was produced by weaker criteria. Recommend skip + re-review rather than posting.

3. Grill the findings before offering to post

Do this before showing the action prompt, and only for drafts that survived Step 2. Skip it entirely when the draft's frontmatter shows a ## Grill section already — /review-pr Step 8.5 grilled it interactively and re-grilling wastes the maintainer's time. Say so and go to Step 4.

Brief the maintainer before the first question — it is not optional, and it matters more here than in /review-pr. That skill's grill follows a review the maintainer just watched run; this one opens on a draft a cron produced days ago, about a PR they have likely never opened. Asking "does this hold?" cold makes the question unanswerable, and silence is defined below as stop — so a cold grill yields no evaluation at all, on the very drafts this skill calls the real gate.

PR #<N>: <title> — <author>, <age>, reviewed <date> at <head_sha>
<one or two sentences: what the PR actually changes, in plain terms>

Draft verdict: <verdict> — <what drives it>

Critical (<n>) — these gate the verdict
  1. <file:line> — <the claim in one clause>

Important (<n>)
  2. <file:line> — <the claim in one clause>

Pre-existing (<n>), Suggestions (<n>) — not grilled, listed so you know they exist

One line per finding. The detail arrives with the question that needs it, and every question restates its own finding inline — the claim, its NET-NEW line, its TRIGGER line, and one sentence on what specifically is uncertain about it. Never refer to a finding by number alone, and never make the maintainer open the draft to answer.

Invoke /grill-me, scoped to findings rather than to a plan. It works in rounds over a frontier — round 1 is every Critical finding, round 2 is Important plus whatever round 1 unblocked, round 3 is the consequences. Skip Suggestions; they never move the verdict.

For each finding, state your recommendation up front, then ask the three questions the tiers turn on:

1. Does it hold? Is the defect real as described, not merely plausible?
2. Is it this PR's? Does the NET-NEW line's base evidence support it, or is it pre-existing / widens?
3. Is the tier right? Critical = something the PR produces is wrong now. Important = wrong later, or left unguarded.

Skip any question the draft already answers. The goal is resolving genuine uncertainty, not making the maintainer re-read their own review.

Never answer your own questions. If a question goes unanswered, stop and leave the draft exactly as it is. A grill that supplies the maintainer's answers launders agent output into apparent human review — which is precisely what this skill exists to prevent.

The sandbox is long gone by now. Unlike /review-pr Step 8.5, this skill cannot re-read --ref base to settle a factual dispute — the checkout was destroyed when that review finished. If a round turns on a fact rather than a judgment, say so and recommend skip + a fresh /review-pr run rather than guessing. Do not drop a finding on an unverified hunch.

Apply each round's answers before asking the next — drop, re-tier, or keep — then recompute the verdict with /review-pr Step 7's rule: only Critical blocks, at any quantity. Dropping the last Critical moves a PR from needs-changes to lgtm.

When the frontier is empty, brief again — the same inventory the grill opened with, re-rendered against the post-grill draft, showing what moved rather than only where it landed:

Grill complete — PR #<N>

Verdict: <before> → <after>          (or "unchanged: <verdict>")

Dropped (<n>)
  <file:line> — <claim, one clause> — <the maintainer's reason, in their words>
Re-tiered (<n>)
  <file:line> — <claim, one clause> — critical→important: <reason>
Kept (<n>)
  <file:line> — <claim, one clause>

The maintainer answered one round at a time, against one finding at a time; nobody holds the cumulative effect of six answers in their head. Ask whether it matches what they decided, and fix it here if not — this is the last point before Step 4 offers to post it to a public PR.

This is a confirmation, not another round. Do not reopen a settled finding or raise a concern the grill did not. If the maintainer's reply opens something genuinely new, grill it as a new round.

Record every drop and re-tier under ## Grill in the draft file, one line each:

## Grill

- <file:line> — <finding, one clause> — DROPPED: <maintainer's reason, in their words>
- <file:line> — <finding, one clause> — RETIERED critical→important: <reason>

That section is the tuning data for /review-pr's criteria. A reason that recurs across PRs is a missing calibration — add it to that skill's "Nx-specific calibration" list instead of re-litigating it every review.

4. Present for review

Open the PR alongside the draft:

gh pr view PR_NUMBER --repo nrwl/nx --web

## Pending PR Review: #PR_NUMBER — TITLE
**Verdict:** needs-changes          ← post-grill verdict; note it if the grill changed it
**Attempt:** 1
**HEAD:** 438fa5a1 (stored) — 438fa5a1 (current)  ← or ⚠ DRIFT if different
**State:** OPEN (draft)
**Reviewed:** 2026-04-24

### Summary
<First Summary paragraph from the Review draft — 2-5 lines.>

### Counts
Critical: N · Important: N · Suggestions: N        (after the grill)
Dropped in grill: N

---
<first ~15 lines of the Review draft>
... (full body is in $TRIAGE_DIR/PR_NUMBER.md)
---

Actions: [post] [edit] [skip] [discard] [open-file]

If verdict: superseded — posting a review on a dead PR is noise. The draft already names the superseding PR in its ### Close-without-merge check section.

Actions: [close-with-pointer] [edit] [skip] [discard] [open-file]

If verdict: unnecessary — the PR shouldn't merge but there's no specific replacement to point at (no real bug / abandoned / duplicate of rejected work).

Actions: [close-with-reason] [edit] [skip] [discard] [open-file]

Wait for the maintainer to choose an action for each draft.