Deferred: replace claude-review Action with loop-driven /code-review (trigger: PR-review background loop exists)

[kyle-sexton]

Context

Deferred decision, recorded with trigger per convention. Alternative to the claude-review GitHub Action: a background review loop (software-factory shape) that finds PRs lacking review and runs the code-review plugins / built-in /code-review from a headless Claude Code session. Running BOTH the Action and a loop would double-review every PR — the same duplication churn the 2026-07-19 audit measured (claude[bot] contributed ~2 unique majors vs Codex across the 6 noisiest PRs).

Trigger

A PR-review background loop / software factory exists and reviews PRs reliably. Until then: no action; the scoped Action (open + on-demand) stays.

Proposed work (when triggered)

• Decide: disable the claude-review Action in caller repos, or keep it for a subset (e.g. public repos where a seat-driven loop isn't running).
• Dedup mechanism: label-marker protocol (e.g. a needs-review / reviewed-by-loop marker, analogous to agent-ready pickup) so loop and Action can never both fire on one PR if a hybrid period is needed.
• Compare cost: Action = subscription OAuth seat per run; loop = local/headless session capacity under operator control, plus scheduling freedom (review once per ready-state, not per event).

Acceptance criteria

• Explicit decision recorded (Action off / hybrid with marker dedup), not both blindly enabled.
• Exactly one automated Claude review path per PR.

References

• ci-workflows/.github/workflows/claude-review.yml + caller pattern
• Trigger-scoping item (this repo) — predecessor change
• melodic-software/claude-code-plugins review plugins (/code-review, review:* agents) as the loop's review engine
• 2026-07-19 churn audit: duplication + seat-burn evidence

[kyle-sexton]

This was generated by AI during triage.

Status check: trigger still unmet — correctly parked

The deferred trigger for this issue is "a PR-review background loop / software factory exists and reviews PRs reliably" as a replacement for the claude-review Action. Based on the current state of melodic-software/claude-code-plugins and this repo, that condition is not met — and the issue's own framing needs a correction before it can be actioned.

What the repo does today

claude-review is an active, event-driven reusable workflow (.github/workflows/claude-review.yml), invoked per pull_request event via thin callers using anthropics/claude-code-action. It is the sole automated Claude review path.

Why the trigger is not met (and the framing is off)

The one background loop that exists — source-control/skills/babysit-prs — does not review PRs itself; it summons an external reviewer and depends on that reviewer existing. Primary-source evidence: plugins/source-control/skills/babysit-prs/reference/review-trigger.md describes a bot-agnostic module that posts a trigger-phrase comment to summon an AI reviewer (e.g. the claude-review bot) and reads its engagement gate. It is dormant by default (all four config slots unset). So the existing loop is a complement to the Action (a scheduler/summoner), not a replacement review engine.

The issue title — "replace claude-review Action with loop-driven /code-review" — presupposes a headless /code-review review path the loop could run instead of the Action. No such path exists: the review plugin ships review skills (fanout, quality-gate, setup) and review agents, but there is no self-running /code-review command that a headless loop drives against a PR. Removing the Action today would break the summon mechanism the current loop relies on, with nothing to replace it.

Bounded limitation (does not change the verdict)

Whether any loop is actually deployed and running on a schedule is a runtime fact not verifiable from repo state. It is moot here: no independent headless /code-review review engine exists regardless of what is scheduled, so the replacement premise fails either way.

Recommendation

Keep parked. No action. The scoped Action (open + on-demand) stays. Re-evaluate when either (a) the loop gains a headless /code-review engine that reviews PRs itself rather than summoning the Action, or (b) a deployed software-factory review loop is demonstrably reviewing PRs reliably. Until then the acceptance criterion "exactly one automated Claude review path per PR" is already satisfied by the Action alone.

[kyle-sexton]

Do-not-break check: the claude-review cadence cut does NOT change this issue's trigger

Record-only note from the claude-review-lanes Phase 5 close-out sweep. This issue stays open and deferred; nothing here changes its state or disposition.

What changed upstream

Phase 2c cut the claude-review lane's cadence to a single pass — the canonical caller shape ci-workflows documents is now [opened, ready_for_review, reopened] (synchronize removed), with a job-level draft guard and a new max-reviews-per-pr input (default 5, claude-review.yml:160-172) capping automatic reviews per PR. The Brief records the motivation as cutting roughly 85% of review volume and the seat-contention failures it causes (PLAN.md:10).

Deployment status — not yet fleet-wide. As of this comment the new cadence runs only in ci-workflows' own dogfood caller (claude-review-self.yml). Consumer repos have not been re-pinned: medley's caller still declares types: [opened, synchronize, ready_for_review] and pins the reusable at v0.6.1 (90f1c549), and the caller-component rollout (melodic-software/standards#286) is still open and unmerged. Treat the single-pass cadence as shipped-in-the-reusable, not as observed consumer behavior.

Effect on this issue's trigger: NONE

This issue's trigger, verbatim from its body:

A PR-review background loop / software factory exists and reviews PRs reliably. Until then: no action; the scoped Action (open + on-demand) stays.

The trigger is keyed on the existence of a background review loop, which the cadence cut neither creates nor removes. The trigger condition is unchanged and remains unmet. No action.

One substantive nuance worth recording

The cadence cut erodes one of this issue's stated rationales, without touching the trigger. The body lists among the loop's advantages:

plus scheduling freedom (review once per ready-state, not per event).

Once the cadence change reaches consumers, the Action will offer essentially that property itself. So when this issue is eventually triggered and evaluated, the cost/benefit comparison should be re-derived against whatever cadence the Action is actually running in consumer repos at that time — not against the per-event Action the comparison was originally written for, and not against the reusable's documented shape if the fleet has not been re-pinned to it. The duplication concern (running both paths double-reviews every PR) is unaffected and still governs.

Not re-verified here

The 2026-07-19 churn-audit numbers cited in the body were not re-measured. Post-cadence-cut volume data does not exist yet at fleet scale, since consumers still run the old cadence. If that comparison becomes decision-relevant at trigger time, it needs fresh data gathered after the rollout.

---

From claude-review-lanes Phase 5 do-not-break sweep. Corrected after review: an earlier revision of this comment described the single-pass cadence as already in effect; it is shipped in the reusable but not yet rolled out to consumers.

[kyle-sexton]

Status (2026-08-12 drain): closing as deferred carve-out — trigger still unmet.

Trigger required a PR-review background loop / software-factory that itself reviews PRs as a replacement for claude-review. That loop does not exist; babysit-prs summons an external reviewer and does not replace the Action. Scoped Action remains the sole automated Claude review path.

Reopen when a reliable loop-driven /code-review path exists and a Action-off vs hybrid decision is needed.

[kyle-sexton]

Closed: deferred until PR-review background loop exists (trigger unmet as of 2026-08-12).