Codex Desktop: historical user messages treated as current instructions during cross-thread delegation

Please investigate Codex Desktop treating historical messages as current instructions during cross-thread delegation. Windows app package: OpenAI.Codex 26.915.4065.0. Observed 20 September 2026, Australia/Sydney.

Observed sequence

1. I explicitly approved starting an end-to-end test in a long-running project with multiple Work threads.
2. An implementation thread treated a complaint from approximately a week earlier as my latest instruction and paused preparation without an explicit stop request.
3. After correction, another report again inferred a stop command from criticism that did not request stopping.
4. The agent later said it had mistaken inherited conversation history for a current instruction. This is the agent's explanation, not an independently verified platform root cause.

Expected

Historical context must remain distinguishable from new user instructions. Criticism must not be converted into a stop command. Cross-thread reports should preserve instruction provenance and task revision.

Impact and limitations

Repeated interruptions, unnecessary coordination, wasted time and Codex usage. Exact wasted token usage is unknown. This has not yet been reduced to a deterministic minimal reproduction; no claim is made that compaction or backend transcript corruption is the established cause. Exact model/reasoning configuration for the affected turn has not been verified. No private project content, full transcript, or credentials are included.

Please investigate historical-message provenance and recency handling in cross-thread workflows. I can provide relevant redacted excerpts privately through an appropriate support channel.

User-observed timing relative to Astra

I clearly began noticing this class of problem after Astra became available. That change is noticeable in my day-to-day use, not merely a hypothetical concern. This is my reported experience and a temporal association; it is not a controlled before/after benchmark or proof that Astra alone caused the issue. Please include the model rollout and associated context/delegation handling in the investigation.

No file, test, or entry point is identified; first establish a deterministic minimal reproduction for cross-thread delegation using a redacted sequence that separates historical messages from new instructions. Done should demonstrate that historical context remains distinguishable, criticism is not treated as a stop command, and instruction provenance and task revision are preserved.