0078 — Fleet PR review as a protoAgent tier: harvest Quinn, keep the panel

• Status: Accepted (2026-07-06 — Phases A/B/B2/C shipped and accepted live; D built, shadow data collection is the standing posture; stage-2/3 promotion stays a per-repo operator decision earned by the three-way eval)
• Date: 2026-07-06
• Builds on: ADR 0077 (the adversarial review workflow + findings convention — the engine this operationalizes), ADR 0025 (delegates spine), ADR 0042 (bundle archetypes — how the QA engineer presents), ADR 0006 (flywheel — the eval/learning posture), projectBoard v0.30.0 (the board-tier review gate that shares the engine).
• Harvest source: protoWorkstacean's Quinn × protoPatch review system (production: 617 reviews/7d, 99.4% completion; sources quinn.yaml, pr-inspector.ts, github.ts read 2026-07-06).
• Plan: docs/plans/qa-review-takeover.md.

Context

The fleet's PR review runs on Quinn — a single-agent DeepAgent in protoWorkstacean whose reliability comes from a specific discipline: every objective decision (structural trigger, CI-terminal gating, approve-on-green, dedup, thread guards) was moved out of the prompt into deterministic code, each move retiring a named failure mode (#863, #748, #891, #858/#903). Meanwhile protoAgent grew a structurally stronger review engine (ADR 0077): a four-angle adversarial panel with dedup and independent verification, findings as a parseable convention, and a board gate that closes the loop by bouncing findings into the authoring coder's retry prompt.

The two systems are complementary, not competing: Quinn has the operational guards and the verdict/merge surface; we have the better reviewer and the fix loop. Taking over her role means porting her guards around our engine — not rebuilding either.

Decision

D1 — Four-layer placement, matching the fleet's existing tiers. The review engine stays in core (ADR 0077 — domain-neutral). Generic GitHub verdict capability goes in github-plugin: formal Review API tools (github_review_approve / github_review_comment / github_review_request_changes) plus github_path_exists. Role-specific machinery goes in a new pr-reviewer-plugin: webhook ingress, the dispatch chokepoint, structural-trigger recipe sizing, the approve-on-green policy layer + sweep, prior-review recall, the protoPatch client, and the review eval. The agent itself is a new qaEngineer bundle (pins + archetype + persona + defaults), the leadEngineer pattern applied to QA.

D2 — Guards live below the model, verbatim from Quinn's ledger. The CI-terminal guard is enforced INSIDE the verdict tools (refuse APPROVE/REQUEST_CHANGES while any check is non-terminal or unverifiable — the model receives a comment-instead error, so the #863 busy-wait class cannot exist). Approve-on-green is a pure function used by both the webhook edge and a 3-minute level sweep: promote COMMENTED→APPROVE only when all required checks are terminal-green AND zero unresolved review threads AND not already promoted for this head SHA; every unknown falls through fail-closed. The structural trigger (>3 files, >120 changed lines, or a sensitive path — auth/session/token/crypto/payment/billing/migrations/CI-CD/docker) is computed from authoritative PR JSON + full-diff paths, never from a truncated diff, and selects the recipe (full panel vs lite).

D3 — Fail closed at every judgement boundary. A review run with ANY failed panel step is not a review: the caller (board gate, reviewer dispatch) must treat it as unreviewed — retry or escalate to the operator — never synthesize a verdict from a partial panel (a promotable verdict from a starved run is how an unreviewed PR auto-merges; observed in our own stack 2026-07-06). A run that cannot verify an external reference records a Gap, never a severity. The reviewer never reviews its own PRs (author/branch guard in the bundle policy).

D4 — The panel stays; noise discipline comes from the prompt ledger. We keep the ADR 0077 multi-angle panel + verifier (Quinn's own roadmap wished for this — ws-2uy) and port her signal-to-noise rules INTO the role prompts as a deliberate negative-prompting exception: the out-of-scope ledger (linter-owned style, theoretical risks behind impossible preconditions, subjective preference, speculation about unread code, already-resolved threads), the ≥80% confidence bar, and consolidation of duplicate findings. Findings remain the ADR 0077 parseable convention — which makes per-finding resolution rate (Quinn's open ws-91a) directly measurable.

D5 — GitHub-native review memory. The posted verdict is the store: unchanged head ⇒ reaffirm; advanced head ⇒ delta review carrying forward still-open findings via a prior_findings recipe input. No local review store to drift.

D6 — Shadow mode before cutover. The qaEngineer bundle ships shadow_mode: true (COMMENT-only, no promotion), running on the same PRs Quinn reviews. Cutover is per-repo, judged against her own eval metrics (completion, verdict mix, latency, catch-rate vs merge outcomes). protoPatch remains a finding source either way — called, not rebuilt.

Consequences

• Four PR surfaces evolve in step (core prompts, github-plugin tools, reviewer plugin, bundle); the findings convention (ADR 0077) is now load-bearing across all of them — schema changes need a compat sweep.
• The verdict tools are write-scoped and identity-bearing: which GitHub App/token the QA agent posts as (inherit @protoquinn vs a new identity) is an operational decision deferred to Phase D — shadow mode works under any identity.
• The multi-step panel costs minutes where Quinn's single pass costs ~45s median. The lite recipe + structural trigger is the cost lever; shadow mode measures where the panel earns its latency. If it doesn't on small PRs, the lite path must win there.
• Two new repos to operate (pr-reviewer-plugin, qaEngineer bundle) — accepted: it mirrors the proven leadEngineer/pm-stack shape and keeps github-plugin lean for its other consumers.
• Revisit triggers: shadow-mode report (Phase D accept); or the engine gaining data-driven fan-out (ADR 0077's parked per-finding verify), which would strengthen the panel's precision claim.