[tzhouam]

The evolution loop previously needed curated gold or human stage labels, multiple model routes, and a human-approved PR before adoption. Add a default objective mode that uses one fixed API model, controller-generated fault interventions and frozen business replays, and independent executable scoring. Accepted source artifacts run as local sandbox releases with automatic canary checks and rollback; PR publication is optional audit.

Changes:

• Automatically capture review/knowledge inputs and prepare controlled engine faults plus clean controls. Keep intervention witnesses and holdout reports outside candidate access; ignore candidate scores and stamp resource measurements in the parent.
• Preserve three paired repeats, item clustering, power checks, immutable source/dependency/input verification, allowed mutation scopes, shared regressions, one candidate/week, one repair, and the shared $20 budget. Refresh free input collection during retries; use fresh independent trials after insufficient power, and retire interrupted generation without replaying paid calls.
• Route native review, knowledge drafting, lint and diagnosis through evaluated sandbox artifacts. Record deployment only after actual execution; compare the exact predecessor on eight distinct canary items, roll back on contract/isolation/budget failures or significant resource regression, and continue bounded weekly monitoring after retention. Persist activation/restore intents for restart recovery.
• Default to objective/automatic while leaving both enable switches off. Keep gold/PR mode compatible. Document Chinese setup, workflow usage, evidence limits, optional audit PRs, and operational rollback.

Validation:

• Related autonomous/legacy evolution tests: 74 passed, 1 skipped.
• Documentation links, citations, strict SPEC coverage/freshness, and knowledge linters pass.
• Full local repository suite: 3330 passed, 18 skipped; the final focused suite includes the later recovery/monitoring cases (74 passed, 1 skipped). Both push and pull-request CI runs passed on the committed final tree, including the full suite, CLI/doctor smoke, wheel runtime data, documentation, strict SPEC checks, and knowledge linters.
• A clean-source readiness smoke automatically prepared 10 development / 90 holdout scenarios (71 required) with no human/gold/judge configuration; its only missing requirement is the current host sandbox.

Limits: objective scores demonstrate controlled fault attribution, executable contracts, output preservation and resource gains, not arbitrary PR recall or knowledge semantic correctness. Unsupported mechanical production drivers cannot auto-adopt. The current host has max_user_namespaces=0; generated code and live model end-to-end execution remain deferred, without an unsandboxed fallback. Scripted worker tests verify the three workflows' protocol/state machines, not real production gains.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.