agent-os v0.6.2
Independent review gate
deliver-worknow requires independent review for material changes.- Review may be skipped only when every small-fix condition holds: one localized supported defect, no new capability or public/sensitive boundary, and direct regression evidence.
- Reviewer identity must come from a real launch result in the current run. Empty waits, invented reviewer labels, self-review, and unreturned reviewers do not count.
- Supported findings require a fix, targeted re-review, and fresh final verification.
- Integrated
batch-workcandidates use the same material review gate.
Evidence
- Repository validator and all red cases passed.
- Batch manifest tests passed.
- VitePress production build passed.
- Fresh installed-plugin trigger matrix: 26/26 measured cases passed.
- Forward tests passed for a localized fix, a material feature with a real reviewer, and an MCP auth/public-schema/external-write change. The MCP reviewer found a missing/empty-token bypass; the implementer fixed it and a targeted re-review returned no findings.
Known limitation
A raw CLI host that exposed wait but no reviewer launch tool did not obey the required pre-mutation stop and invented a reviewer identity. The contract now rejects that behavior, and the published Evals page records the case as FAIL. Material delivery is not claimed as independently reviewed in that host.