Keep verification state independent of context refreshes

Proposed recovery design: use conversation history to recover reasoning, but keep exact operational state in durable records. Preserve the contract revision, candidate commit, outstanding blockers, captured checks, review artifacts, and pending side effects. Recompute readiness from those records and authoritative evidence rather than trusting the model's recollection.

This prevents a context refresh from silently turning an unfinished test into a pass or an already-created pull request into a second creation attempt. It does not require a competing conversation summarizer when the harness already supplies notes and searchable history.

Test recovery while a check is running, between review and repair, and after an external action but before its receipt is saved. Recover or reconcile the action before repeating it, preserve blockers, and reject approvals tied to stale inputs.

Editorial synthesis of the September 5, 2026 submission based on public Codex context-management documentation. This is an architecture and test proposal, not a verified guarantee for any experimental hook path. Source context: https://learn.chatgpt.com/docs/models#experimental-context-management .