Record the two-campaign night's lessons in ship-campaign

[KyleMit]

The self-heal pass that closes campaign #2529. It writes four lessons from that night's run into the ship-campaign sources (.ruler/, then regenerated). Left open for the owner: a skill change is outside the campaign's queue, so the orchestrator does not merge it.

What happened, and where each lesson now lives

1. Two campaigns on one queue. Leftovers campaign from #2500: 9 decisions and 10 follow-ups (2026-09-29 night) #2529 (Claude) and Ship campaign: #2500 leftover decisions and follow-ups #2530 (Codex) started 94 seconds apart on the same Code-smell burndown campaign, 2026-09-29 evening (4 hours) #2500 leftovers comment. Free-form units carry no in-progress label, so neither preflight could see the other. They collided on decisions 1–3 before either noticed.
   • Preflight bullet "No second campaign on the same queue" (SKILL.md step 1): search for another open issue or PR that references the same source, and for another runner's branches.
   • Step 2 re-check, repeated before every launch, because a peer that starts after this preflight is invisible to it. It covers both cases: splitting with a peer that is running, and leaving a stopped peer's drafts to that peer.
2. The same outside change, read two ways. A newly published brace-expansion advisory failed pnpm audit on main's own lockfile.
   • Ship campaign: #2500 leftover decisions and follow-ups #2530 treated it as "main red for a reason no campaign merge caused" and stopped.
   • Leftovers campaign from #2500: 9 decisions and 10 follow-ups (2026-09-29 night) #2529 treated it as "a check proven broken on its own base" and shipped gate repair Lift brace-expansion past two new high audit advisories (gate repair) #2538.
   • Step 3 now says an outside change failing a check on main's own tree is a gate repair when the gate's own conventional remedy fixes it. For pnpm audit, that remedy is a narrow override. The queue-wide blocker list is narrowed to match.
3. Mechanical coupled verdicts (references/parallel.md). After a merge changes pnpm-lock.yaml or eslint.config.js, every later catch-up surveys coupled. That means a quick full reconcile pass, not a reason to leave a PR behind. The rival review of this campaign's takeover decision caught exactly that wrong assumption.
4. Harness facts for lanes (references/parallel.md):
   • The auto-mode classifier refused the orchestrator's end-of-campaign batch removal of its own clean worktrees, so the reference now prefers cleanup after each merge.
   • Worktree-isolated lanes cannot run a brokered command that invokes git in the rival's worktree. The reference says how to answer the rival instead.

The skill note (.ruler/skill-notes/ship-campaign.md.template) records the incident chronology.

Not healed, as not durable: npm run show:free-port -- --help prints a Node stack. The error line already names the unknown option, and the script takes no flags worth documenting.

Checks

• npm run ruler:apply, then npm run ruler:check: in sync.
• npm run format:check, check:skill-refs, check:doc-refs: all exit 0.
• npm run test:tools: 279 files, 6391 tests passed.

Update (2026-09-30 afternoon)

This branch caught up with main after #2537 merged. #2537 edited other lines of the same ship-campaign sources; the merge was clean, with no overlapping hunks. Step 2's stopped-peer rule no longer says a peer's drafts "stay its own". It now says adopting them needs the user's explicit hand-over, plus a survey, a fresh rival review and a new CI round. That afternoon's adoption and merge of #2532 and #2537 is the evidence, and the note records it. ruler:check, format:check, check:doc-refs and check:skill-refs all exit 0.

Refs #2529 · Refs #2500

🤖 Generated with Claude Code

[KyleMit]

Independent review by the codex rival agent, round 1, of pull request 2548.

Reviewed the full diff, checking peer-campaign coordination, gate-repair rules, cleanup guidance, and broker handling against the existing workflow and tools. Generated copies match their sources; Markdown and reference checks pass, along with 193 targeted tests. Direct checks confirmed the documented catch-up classifications. No substantiated defects found.

[KyleMit]

Independent review by the codex rival agent, round 2, of pull request 2548.

Round 1 reported no findings, so none required resolution. Reviewed the new hand-over requirement and fresh survey, rival-review, and CI requirements for adopted drafts. Generated copies match their sources, and Markdown and reference checks pass. No new defects found.