work-items(work): implementation subagents flip PRs to ready without confirming review findings were actually resolved

[kyle-sexton]

Lane: work · Category: operational (subagent behavior gap, observed twice in one session)

Triggering examples

1. PR feat(claude-ops): lanes skill — scripted background loop-lane launcher (#508) #616 (issue ops: scripted lane launcher — start/restart loop lanes headlessly from canonical prompt files #508, claude-ops lanes skill): after CI went green, the review bot flagged 3 MEDIUM findings. The subagent's first resume pass acknowledged and claimed to fix them, but the orchestrator's own diff/CI re-check found the same 3 findings still present — required a second explicit resume with a more forceful instruction ("verify by re-reading the diff yourself before pushing").
2. PR fix(work-items): exclude open-linked-PR issues from work frontier (#463) #643 (issue work-items(work): status: ready issues with an open PR stay pickable — no in-review state (re-pick risk) #463, work-items open-linked-PR filter): after fixing an unrelated CI-blocking typos failure, the subagent flipped the PR to ready-for-review — but 2 REAL findings from an earlier review round (a regex missing a word-boundary lookbehind; an inaccurate "truncation-safe" claim in a doc) were still unaddressed in HEAD. The subagent's fix-and-flip pass apparently didn't re-check prior review rounds, only the CI failure it was explicitly asked to fix.

Observed vs expected

• Observed: a subagent asked to "fix the CI failure, then flip to ready" (or similar scoped instruction) sometimes flips to ready without independently re-confirming that ALL outstanding review-bot findings — including ones from earlier rounds, not just the one it was explicitly pointed at — are actually resolved in the pushed HEAD.
• Expected: before flipping any PR from draft to ready, a subagent should re-fetch the full review-bot comment history (not just the most recent one) and self-verify every finding is addressed in the current diff, not just the specific issue it was asked to fix in that resume.

Category

behavioral (subagent self-verification gap before a state-changing action)

Suggested fix direction (not investigated in depth)

This may already be partially covered by #513 (role-separated mini-SDLC — implementer ≠ verifier) — a dedicated verifier role re-checking the full review history before flip would structurally prevent this class of gap, rather than relying on the implementer subagent's own diligence. Could also be a narrower fix: the orchestrator's dispatch/resume brief template could explicitly instruct "before flipping to ready, re-fetch ALL review-bot comments (not just the latest), and confirm each finding is resolved in the current diff" as a standing step, independent of #513's larger architecture change.

Related

#616 (PR), #643 (PR), #513 (mini-SDLC role separation — likely supersedes a narrow fix here)

[kyle-sexton]

This was generated by AI during triage.

Triage (umbrella-fold) — reporter's own body flags this as likely superseded by #513 (role-separated mini-SDLC: implementer ≠ reviewer ≠ verifier). The observed gap (implementer subagent flips to ready without re-confirming ALL prior review-round findings, not just the one it was explicitly asked to fix) is exactly the same-agent-implements-and-verifies failure mode #513 exists to structurally eliminate via a dedicated verifier role. Folding rather than routing standalone per the umbrella-fold rule.

Blocked by #513's decision. Not routed to needs-human individually.

[kyle-sexton]

Deferral-discipline audit note (control tower): this is the behavioral twin of #642 — a second independent path for an unresolved finding to reach merge-ready. Recommended fix direction is the structural one: route to #513 (implementer≠verifier role separation) rather than a narrow per-lane patch; interim mitigation is a dispatch-brief step to re-fetch ALL review comments before any ready flip. Merge-safety class — should ride the same pre-gate-off window as #642.