feat(loop): blocking review gate — adversarial findings gate the merge edge (v0.30.0)

[mabry1985]

M5 of the codified-delivery-loop plan (protoAgent docs/plans/codified-delivery-loop.md). Turns review from advisory (review_dispatch) into a gate, mirroring the CI-bounce machinery it sits next to.

The loop

1. After open_review, run the host's adversarial code-review workflow (protoAgent ADR 0077) on the PR — reached via runtime.state.STATE.workflow_run (published by the workflows plugin; no plugin import), with the configured a2a reviewer as the fallback runner.
2. Parse the findings convention (graph.review.findings, lazy import). Blocking = severity blocker/major, verdict ≠ refuted.
3. Blocking findings → stored on the bead (comment), injected into the retry prompt with the reviewed diff (the same carry-the-lesson levers as the CI bounce), label changes-requested, requeue. Bounded by review_fix_max (mirrors ci_fix_max; same-tier — findings are fixable, not a capability ceiling).
4. Budget exhausted → flag_blocked ("needs human review"). Never a silent merge — and never a silent pass: an unrunnable gate leaves review-pending set and the PR reconcile retries it next poll; a host without the findings parser sends the gate loudly inert.

Sub-states & reconcile

• Labels review-pending / changes-requested (managed by store.set_review_substate, one at a time, findings comment attached).
• _reconcile_prs: finishes an interrupted gate (pending on an in_review PR), clears the bounce budget on MERGED/CLOSED, and logs a human-override merge (merged with changes-requested still set) instead of pretending it didn't happen.

Sequencing (ADR 0064)

The gate is one call post-open_review — when execution-grounded candidate selection lands on the board face, moving it after test-passing selection is a one-line move.

Config

review_gate (default false), review_workflow (code-review), review_fix_max (2). review_dispatch stays as the advisory alternative; the gate takes precedence when both are on.

Tests

7 new (test_loop.py): bounce carries findings into _build_prompt, clean/minor/refuted pass, exhaustion blocks without requeue, unrunnable leaves pending, config parse, PR-url parse. Full suite 289 passed, ruff clean. v0.30.0 lockstep.

🤖 Generated with Claude Code

[protoquinn]

👀 Quinn is reviewing — verdict (PASS / WARN / FAIL) + findings to follow.

[protoquinn]

QA Audit — PR #70 | feat(loop): blocking review gate — adversarial findings gate the merge edge (v0.30.0)

VERDICT: WARN (CI queued — non-terminal; approve-on-green policy promotes to APPROVED once test passes and no blocking threads remain)

---

CI Status

• test: ⏳ queued

Diff Review

• Adds _review_gate to loop.py (~140 lines): runs host's adversarial code-review workflow post-open_review, parses findings, bounces on blocker/major (bounded by review_fix_max), exhausts to flag_blocked — mirrors the CI bounce pattern faithfully.
• _reconcile_prs gains resume logic for interrupted gates (in_review + review-pending) and cleanup/override-logging on MERGED/CLOSED.
• New config: review_gate (default false — safe opt-in), review_workflow (code-review), review_fix_max (2).
• _parse_pr_url regex helper extracts (pr_number, owner/name) from GitHub PR URLs — guarded for non-GitHub inputs.

Observations

• LOW: clawpatch structural review returned zero findings — the gate reuses existing CI-bounce levers (_ci_feedback, _ci_prior_diff) and the pattern is consistent with the surrounding code.
• LOW (Gap): diff truncated at 532 of 532 lines — store.py changes (labels, set_review_substate), config parsing, and test files not fully visible. PR description confirms 289 passing and 7 new tests.
• INFO: CI test still queued — the approve-on-green policy will promote this review to APPROVED automatically once all checks settle terminal-green. No action needed from the author on this pass.

— Quinn, QA Engineer

[protoquinn]

Submitted COMMENT review on protoLabsAI/projectBoard-plugin#70.