Intent

Author a firstmate-loaded internal lavish-repair skill that safely diagnoses and repairs wedged Lavish decision surfaces. The workflow must prove the failing layer in cheapest-first order: HTTP serving, the known MaxListeners live-channel leak, a missing per-board poll listener, and only then the browser. It must prohibit fleet-wide pkill -f patterns with the concrete 2026-07-28 loss of 15 of 16 crews, require listing and argv inspection before explicit-PID signaling, prescribe least-destructive per-layer repair, preserve session URLs when re-serving the same source file, and explain that reports, worktrees, commits, and pushed PRs survive a crew kill while only the live agent session is lost. Keep it under .agents/skills with internal metadata, add the minimum section-13 load trigger and safety regression coverage, open a PR without merging, and do not conflate or fix the unrelated Behavior-test failure already red on main.

What Changed

• Add an internal lavish-repair skill with cheapest-first diagnosis and least-destructive recovery procedures for wedged Lavish surfaces.
• Document explicit-PID process safety, stable session URLs, and which crew artifacts survive process termination.
• Register the skill’s conditional load trigger and add regression coverage for its routing, diagnosis order, safety guardrails, and recovery contracts.

Risk Assessment

✅ Low: The change is a well-bounded internal skill and static regression contract that satisfies the required diagnosis order, explicit-PID safety gate, least-destructive repairs, stable session URLs, and crew-kill durability guidance without touching the unrelated existing failure.

Testing

Completed 1 recorded test check.

• Outcome: ⚠️ 1 error across 1 run (1h9m39s)

Pipeline

Updates from git push no-mistakes