autonomous-gate-operations

When to use. Any multi-day autonomous run with LLM reviewers as the only check. Why: the trail is only as trustworthy as its weakest execution claim; all three of our gate rounds caught real defects precisely because the reviewer re-ran everything instead of reading the driver's prose.

Procedure.

1. Reviewers run the harness/trace/probes themselves and quote real output; a verdict citing an un-run result is invalid. Fresh context per round; findings routed with severities; verdict blocks recorded append-only.
2. Never mask exit codes (cmd | tail in a && chain ate a red test through two commits); check $? explicitly. Never run the test suite concurrently with pipeline jobs sharing output dirs.
3. Decide-and-log instead of blocking on absent humans: every judgment call gets a rationale in the append-only log; ratified decisions are extended or corrected by entry, never silently re-decided.
4. Keep honest-red visible: failing baselines recorded before fixes; Draft/Blocked items never claimed; disclosures (e.g. judge noise floor) promoted to the surface the human signs.
5. Root-fix flakes an unattended loop would trip on (ours: Windows os.replace sharing violation → bounded winerror-scoped retry + two tests), and maintain a per-turn resume note ending with the explicit terminal state ("no pending agent action — do not reopen gates").
6. Done when: an independent reviewer can APPROVE from executions and the written trail alone, with zero claims they had to take on faith.

Knowledge: FIELD-KNOWLEDGE-NOTHOMEWRECKER.md §F4 (+ §F1/§F2 for the loop and evidence rules).