babysit-prs: batch confirmed findings into one commit per cycle; severity-route SUGGESTION findings to deferred work items

[kyle-sexton]

Context

Review-churn audit (2026-07-19 session): across the 6 noisiest PRs of the last 24h (claude-code-plugins #356/#359/#355/#341/#372, standards#184) there were ~92 fix rounds, but only ~15-20 were driven by genuinely pre-existing MAJOR issues. 30% of all 222 bot findings were NEW-SCOPE cascade — real bugs, but in code a previous fix commit had just added — manufactured by the current one-finding-at-a-time fix/push loop: each small push re-triggers both reviewers (Codex on every push; claude-review on synchronize until scoped). #355's root refactor landed ~8 rounds late after a ladder of patch-shaped fixes.

The skill currently has no rule coalescing findings: each finding runs its own D1-D7 cycle and "batch" refers only to worker concurrency (reference/orchestration.md worker lifecycle ~:459-483, :232-233). Nits currently always qualify as "small + related → fix now" (reference/review-discipline.md fix-now-or-ticket rule), so every cosmetic finding becomes a pre-merge round-trip.

Proposed work

• Batch discipline: per cycle, collect ALL confirmed findings on a PR, diagnose root-cause classes, fix the whole class in one commit, push once. One push per PR per cycle, not per finding.
• Severity routing default: CRITICAL/IMPORTANT → fix in the batch commit pre-merge. SUGGESTION/cosmetic → VALID (defer) by default: one batched follow-up work item per PR (all nits, with file:line pointers + PR link), fixed post-merge by background drain loops. Uses the existing fix-now-or-ticket mechanism in review-discipline.md — this changes only the default for nit-class findings ("small + related" no longer forces fix-now when severity is SUGGESTION).
• Reconcile with babysit-prs: codify the decide-and-defer non-convergence resolution playbook (own-once, no duplication of safety.md) #375 (decide-and-defer playbook, deferred disposition in reference/feedback.md) — the deferred-nits work item is the same disposition class; don't create a second mechanism.
• Reconcile with the fix-round-cap ledger (reference/feedback.md :66-86): at cap, remaining findings route to the deferred work item and the PR proceeds to gate-proven merge — autonomous, no human interrupt (per babysit-prs: codify the decide-and-defer non-convergence resolution playbook (own-once, no duplication of safety.md) #375's autopilot-only thread-resolution basis).

Acceptance criteria

• Skill docs mandate one batched commit+push per PR per cycle covering all confirmed findings.
• SUGGESTION/cosmetic findings default to VALID (defer) with a single batched work item per PR; nothing is silently dropped.
• Round-cap behavior at limit: defer remainder + merge when gates green, no human escalation in the loop.
• No duplication of babysit-prs: codify the decide-and-defer non-convergence resolution playbook (own-once, no duplication of safety.md) #375's mechanics — pointers only.

References

• plugins/source-control/skills/babysit-prs/ — SKILL.md loop :308-335, reference/orchestration.md, reference/feedback.md, reference/cadence.md
• plugins/source-control/reference/review-discipline.md (VALID (defer), "never merge past a confirmed-valid finding with neither a fix nor a tracked issue")
• plugins/source-control/skills/pull-request/reference/monitor.md :304-319
• babysit-prs: codify the decide-and-defer non-convergence resolution playbook (own-once, no duplication of safety.md) #375 (decide-and-defer non-convergence playbook)
• Companion trigger change: melodic-software/ci-workflows (claude-review drop synchronize)

[kyle-sexton]

This was generated by AI (triage lane — gate-audit claim). Claiming this item for one gate-audit cycle: an independent panel re-classifies whether the needs-human / status: needs-decision gate is still warranted, or whether the item is C2/C3 and liftable to agent-ready. Outcome comment follows in this cycle.

[kyle-sexton]

This was generated by AI (triage lane — gate audit). Independent panel: two fresh-context agents (one fable, one opus), each given only the issue content and the repo's own plugins/autonomy/reference/guardrails/work-classes.md, plus the triage lane's own read. This lane never executes work items, authors fixes, or merges.

Gate audit: GATE UPHELD — merge-policy change, not yet operator-ratified

Panel: 2-of-2 STAYS-GATED (classed C3 and C4 respectively). Lane read concurs.

The batching half is a bounded, well-briefed skill change. What holds the gate is that two of the acceptance criteria change autonomous merge policy: SUGGESTION findings flip from fix-now to defer-by-default, and at the fix-round cap the PR "proceeds to gate-proven merge — autonomous, no human interrupt." That is the same merge authority #1630 is currently re-litigating, and no comment records an operator ratifying it.

One leg also found a hard dependency problem: AC #4 requires "no duplication of #375's mechanics", but #375 is still OPEN and labelled wayfind: design, so the deferred disposition it points at does not exist yet — a grep for "defer" in plugins/source-control/skills/babysit-prs/reference/feedback.md returns nothing, and the issue's cited line ranges there have drifted.

Decision required

Should SUGGESTION-class findings default to post-merge deferral, and should a PR at the fix-round cap merge autonomously with the remainder deferred?

RECOMMENDED — split it, and land only the uncontroversial half now. Implement "batch confirmed findings into one commit per cycle" as a standalone scoped change (the churn evidence for it is strong and it alters no merge authority). Hold the severity-routing default until #375 decides the deferred-disposition mechanism it must reuse, and hold the round-cap merge change as a separate operator decision alongside #1630 — deciding it here would grant merge authority in a side channel.

Alternative — ratify both as written. Nothing is dropped (each PR gets a tracked deferred item) and it composes with #375's disposition model. Costs: it decides #1630's question implicitly, and it builds on a mechanism #375 has not designed yet.

[kyle-sexton]

This was generated by AI (triage lane — claim release). Gate-audit cycle complete for this item; the triage claim posted earlier this cycle is released. No lease is held — another instance or lane may pick this up immediately. The outcome comment above this one carries the classification and either the lift rationale or the decision summary.

[kyle-sexton]

Frontier disposition (drain session)

Ruling: Close as not planned.

Batch-commit / severity-routing enhancement is backlog nice-to-have without a blocking acceptance brief.