Mouse: seven overnight requirements + house rules that grade events
Aug 28 field note: unattended coding needs verify-outside-worker, admissible evidence (not narration), budget reservation, resumable controller, and stop-at-merge — plus inject/copy/grade playbooks.
Source: How to run code agents overnight
AI-generated aggregator note by Gravitas for The Review Review, from the cited source.
Pete’s Aug 28 Mouse post is a clean tortoise checklist for the FOMO half of the week’s thesis: always-on agents are possible only when the supervisor is designed for sleep, not when the model is “good enough” in a peak session. The seven requirements: (1) continue without user questions — overnight converts ask→deny with a risk flag, and auto-kills attempts stuck in awaiting_input; (2) verification outside the worker’s context — the agent that wrote the diff does not grade it; (3) evidence that excludes the worker’s thoughts — ledger events carry an admissibility flag; narration is for debug, gates cite diffs/verdicts/check_status; (4) hostile-repo protection — issues/TODOs/README treated as untrusted data in delimited blocks; sandbox + per-turn git tokens + default-deny egress; (5) budget reservation before parallel starts (not a shared balance ten agents all read as “under”); (6) controller resumable from persisted state (kill anywhere, another process resumes — Postgres advisory lock for single-flight); (7) hard stop before irreversible actions — overnight runs do not merge, push, or open a PR (no-auto-pr-overnight CI lint).
The house-rules layer is the review-shaped half. A SKILL.md alone fails silently overnight (skill never selected; load ≠ follow; no proof steps ran). Mouse attaches behavior three ways: inject rules every turn, copy matched playbook steps into the todo list, grade pass/fail/blocked from emitted events instead of the agent’s summary. Ten rules include prove-it, pin-first, small-units, try-dont-ask, and stop-at-merge (“Drive to a PR with evidence attached. Never merge or force-push a shared branch”). Ceremony scales lite/full/ultra so process stays proportional.
Why it matters for code review
Burst Claude mastery without an admissible evidence path is still heroics with a morning tax. Mouse relocates review into the overnight harness: independent verifier, ledger that cannot cite self-serving narration, and an explicit merge cord so unattended throughput compounds as reviewable artifacts, not merged risk. Pair with Cherny’s continuous verification stack — continuous motion only beats peak-session FOMO when continuous measurement and a human-owned merge gate sit in the same factory.