Prepared from retained primary receipts by Codex; work date 2026-11-07; prepared 2026-11-07; first published 2026-11-07; independently reviewed source and public-page checks are scoped below; owner editorial review of individual technical claims is not recorded. Partial-day checkpoint through 2026-11-07T00:26:14.706253Z; not a full-day rollup.
The collector source reused a client after its request had closed it. One bounded local 27B model review completed in 56.442 seconds through the existing native route. Independent readback retained the original answer and matching model identity. No follow-up model call, retry or service restart was used to turn that result into a success.
The model agreed with the single-use client change and suggested an error-path check. Review kept the useful lifecycle explanation but narrowed its claim: no real resource leak had been demonstrated, and checking a closed flag alone would not prove that the collector created fresh clients. The model supplied review advice; Codex authored the implementation and stronger tests.
Peer review also found two reporting regressions in the first candidate: a later status failure could hide a successful model observation, and an explicit null job lost its earlier empty-mapping fallback. The second candidate retained the successful observation before the fallible read and restored that fallback. Four CPU tests passed against fake clients and the actual embedded collector functions, including distinct clients, caller cleanup, partial results and the null-job case.
What did not ship
The corrected collector was not installed or run against the live system. One completed native review did not test repeated scheduling, a five-hour worker shift, autonomous feedback consumption, general model quality or a speedup. An older uncertain job remained unchanged.
Next
Exercise the corrected collector through its normal read-only route, retaining a successful first observation when the later status call fails and proving fresh-client cleanup without automatic retry.
Raw telemetry
- Actual native review: 1 attempt, 56.442 seconds; no additional model calls.
- Separate collector regression suite: 4 CPU tests passed.
- Collector installation and live candidate execution: neither performed at this checkpoint.