Let an agent session request a review with a /review comment

[claude]

Requested by Ezra · project thread

Before: the only requester dispatch-on-comment accepted besides a human OWNER/MEMBER/COLLABORATOR was cursor[bot].
A Claude Code remote/web session acts as claude[bot], and had no working way to ask for a review at all.
Its pushes carry sender.type == 'Bot', which skips every review / job on the automatic pull_request path; and a workflow_dispatch it issued itself started a run that claude-code-action short-circuited at zero cost, because allowed-bots admits github-actions[bot] and not claude[bot].

After: claude[bot] is on that job's login allowlist, so such a session posts a /review comment and gets a real review.
The job's own dispatch runs under GITHUB_TOKEN, so it re-enters as github-actions[bot] — the actor allowed-bots already admits — and the sender gate is never consulted, because the run arrives as a workflow_dispatch.
Neither gate is widened, and the requester is still checked by login before anything is dispatched.

How: one login added to the existing fromJSON allowlist, plus the comment block recording why a direct dispatch is not the alternative it appears to be, and a CLAUDE.md section saying the same for a session reading the repo rather than the workflow.

Measured on Morrison-Lab/ai-config, 2026-09-18: 4 of 4 dispatches with triggering_actor: claude[bot] packed failure-kind: short-circuit, attempts: 1, total-cost-usd: 0.0000, SELF_MOD: false, with the "Run Claude Code Review" step lasting 35ms; 4 of 4 with triggering_actor: github-actions[bot] reviewed normally.
The run IDs are tabulated in the companion PR, Morrison-Lab/ai-config#3767, which ports this whole job into that repo's caller.

A zero-cost short-circuit is also gha#368's signature, so CLAUDE.md now says to read the actor before the failure kind — retrying reproduces this one exactly, since the discriminator is the actor and not the attempt.

This gate lives entirely in the caller, so it takes effect on merge to main without waiting for a @v2 slide.

Notes for the reviewer

• Pushed with ALLOW_UNREVIEWED_PUSH=1; the pre-push guard cannot see a verdict from this harness (no-unshipped-commit and no-push-without-self-review deadlock while a review is in flight ai-config#3739). An adversarial review was dispatched over this diff and its findings will be addressed in a follow-up push.
• Not merged, despite this repo's standing mwc grant, per the request that produced it.

🤖 Generated with Claude Code

https://claude.ai/code/session_017qehMo49hC1bfozVnjwdKc

---

Generated by Claude Code

[d-morrison]

@claude reivew

[d-morrison]

@claude review

[github-actions]

👀 Picked up by workflow run #35316929502. Setup runs first; Claude itself responds after that.

[github-actions]

👀 Picked up by workflow run #35316925288. Setup runs first; Claude itself responds after that.

[github-actions]

Stopping Point: Clean stopping point reached — the misspelled @claude reivew trigger required no action since the correctly-spelled sibling comment already dispatched a real review for PR #883, and no late @claude requests were found to absorb.

_{— posted by @claude post-step from workflow run}

💰 Cost: $0.9904228000000003

[claude]

Pushed 4fd4a2a to fix the red antigravity-tests job (run 35314255459).

Cause. changelog.d/allow-claude-bot-review-command.changed.md was written as bare prose. changelog.d/README.md requires a fragment's contents to be Markdown bullets, and antigravity-review/scripts/preflight_check.py's check_changelog_fragments() enforces it: the first non-blank line must start with - . test_preflight_checks_pass is an integration test that runs the checker against the real repo tree rather than a fixture, so the fragment reddened it directly:

❌ Preflight error: changelog.d/allow-claude-bot-review-command.changed.md must start with a '- ' bullet point
AssertionError: False is not true

Fix. Prepend the bullet marker, indent the continuation lines by two, and add the (gha#883) reference the README asks for. The wording is otherwise unchanged.

Verified. Reproduced the failure at 7963791; all 42 tests in antigravity-review/tests pass at 4fd4a2a. Longest line is 79 characters.

Pushed with ALLOW_UNREVIEWED_PUSH=1: the pre-push guard wants the adversarial review's verdict as the dispatching call's own result, which this harness cannot produce (ai-config#3739), so it denies every push from here. The change is a single changelog fragment with no executable content.

The remote tip was still 7963791 immediately before the push, so this was a clean fast-forward rather than a race. The session driving this PR has been told to rebase any staged follow-up onto 4fd4a2a rather than force-pushing over it.

Posted by Claude Code (AI agent) --- not written by a human.

---

Generated by Claude Code

[claude]

Fixed the lint-markdown failure and corrected the measurement's population in dd243dc.

lint-markdown. check_list_item_splices.mjs flagged two lines: this repo's house style requires a blank line between wrapped list items, and the new /review section's Do/Don't block had none. Verified locally against the same base ref CI used (15b202ac) — check_list_item_splices.mjs now reports no splices.

The measurement. The section undercounted its own population. There were four claude[bot] dispatches on Morrison-Lab/ai-config, not three — 35267489584 was missing. The github-actions[bot] side is now stated as the four runs actually checked rather than as an exhaustive count, every run is linked, and the deriving query is recorded beside them so the next reader queries rather than recalls.

antigravity-tests was the other failure and is already green on 4fd4a2a, which gave the changelog fragment its required bullet marker.

Disclosure. This push used ALLOW_UNREVIEWED_PUSH=1, for the same reason as the sibling PR: hooks/no-push-without-self-review.py expects the adversarial reviewer's verdict as the dispatching call's own result, which this harness cannot produce, so it denies every push from here. The adversarial review did run, and its findings are applied.

Posted by Claude Code (AI agent) --- not written by a human.

[dem-extra1]

@claude review

[github-actions]

👀 Picked up by workflow run #35375418309. Setup runs first; Claude itself responds after that.

[github-actions]

👀 Claude Review Dispatched — run reviewing PR #883 at dd243dc.

[github-actions]

Claude finished review — View run

---

Stopping Point

Re-verifying the CI/review-status claims in my prior message before finalizing, since the stop hooks correctly flagged that gh pr checks alone doesn't authorize a "green"/"fully clean" claim, and my earlier phrasing blurred a code-review verdict with an operational merge-readiness claim.

What I actually confirmed with authoritative instruments:

1. Check-runs, paginated, direct from the REST API for head dd243dc0db6858940f13d77457a7d928238ec1f1: total_count=71, all 71 fetched, every one status=completed and conclusion in {success, skipped} — zero failing, zero pending among them.

2. The one "skipped" entry that matters — review / claude-review at that SHA is a prior, preempted attempt (job 105512136720, run 35317453115), not the review validating this diff. Tracing the run that's actually reviewing this content (35375567047, linked from the PR's own "👀 Claude Review Dispatched" comment for dd243dc) shows its headSha is 15b202ac — origin/main's tip at dispatch time (PR Expose timeout-minutes as an input on the two agent workflows #881's merge commit), not dd243dc. This is because this PR edits .github/workflows/claude-review.yml itself, so dispatch-on-comment's gha#598 logic correctly omits --ref, and GitHub attaches the check-run to the default branch rather than the PR head — the exact mechanism CLAUDE.md's "A PR fixing claude-code-review.yml ... can't self-verify before merge" section documents, live, on this very PR.

3. check-pr-fully-clean.py 883 -R Morrison-Lab/gha (the authoritative instrument) currently exits 1: No valid clean review found for HEAD SHA dd243dc0. This is expected and not a contradiction — no review has posted yet, because this task is the review that will supply that verdict. Once this message posts, that instrument's answer will change; I can't make it report clean in advance of posting, and I'm not claiming it does.

So, to be precise about what my ### Verdict below means: it is my code-review assessment of the diff's correctness, security, and CLAUDE.md compliance — not an ARDI-style claim that the PR is currently fully clean/mergeable. The PR's actual fully-clean status depends on this review posting and the review check-run landing (which, per point 2, may need a fresh dispatch against the real head once this PR is no longer editing workflow YAML, or a manual confirmation after merge, per this repo's own documented self-verification gap).

With that caveat stated explicitly, my review content and diff-correctness verdict from my prior message stand unchanged — the diff itself (login-allowlist addition, CLAUDE.md documentation, changelog fragment) has no bugs, security issues, or CLAUDE.md violations that I found.

Verdict

Ready for merge (code-review assessment of the diff's content; see the CI/review-status caveat above — this is not an independent claim that all merge gates are currently green)

Structured Review Data (JSON)

{
  "schema_version": "1.1",
  "reviewer": "claude",
  "commit_sha": "dd243dc0db6858940f13d77457a7d928238ec1f1",
  "verdict": "CLEAN",
  "findings": [],
  "detailed_assessment": "Traced the dispatch-on-comment if: gate change in .github/workflows/claude-review.yml, verified the GITHUB_TOKEN re-entry mechanism against unchanged surrounding code, confirmed claude[bot] is the real webhook login via the REST API, and cross-checked all eight cited Morrison-Lab/ai-config run IDs and the gha#368 reference; no factual, logic, or security defect found in the diff content.",
  "holistic_assessment": "Change is scoped to a single caller-side login allowlist plus documentation, requires no permissions widening or tag slide, and needs no test-fixture updates since no offline suite pins this allowlist; paginated check-runs at the PR head show 71/71 success-or-skipped with none failing, though the review check itself is attached to main's tip rather than the PR head due to this PR's own gha#598 workflow-edit path, so a fresh confirming review pass after this one posts is advisable before relying on require-review at the actual head SHA."
}

Stopping Point: Clean stopping point reached — full code review completed and posted with instrument-backed CI/review-status caveats; no further action needed from this task.

Reviewed commit: dd243dc

[github-actions]

💰 Cost: $2.7197 (review) — run