Codex Desktop: active threads change effective model without a user action

Summary

Two active Codex Desktop threads appeared to change model without a user-initiated model selection. This is a report of unexpected per-thread model drift, not an intentional user model switch or a model-routing feature request.

Observed

• Relay / coordination thread
• Thread ID: 019fbf2d-6cdf-74d3-9ba9-b8649ce6931c
• Observed change: 5.6 → 5.5

• Implementation thread
• Thread ID: 019fca93-ad62-70c3-a45b-388d652cbe2a
• Observed change: Sol → Terra

• Submission/message reference: 019fd38b-f728-7562-b453-fac2b23590f0

The two thread IDs are distinct. The reported issue is that the effective model for an existing thread changed on its own, rather than remaining pinned to the model selected by the user for that thread.

Expected behavior

• A thread's selected/effective model should remain stable unless the user explicitly changes it.
• If a backend fallback, migration, or policy requires a model change, Codex should make that change visible before the next turn, with the reason and the effective model.
• Thread metadata, the model selector, and the model actually used for each turn should agree.
• A silent change must not occur during long-running, delegated, or recovery work, where model continuity affects execution quality and predictability.

Actual behavior

The relay thread was observed as changing from 5.6 to 5.5, and the implementation thread from Sol to Terra, without a corresponding user-directed model change.

Impact

This undermines confidence in long-running threaded work. Model differences can change reasoning behavior, tool use, and delegation outcomes. It is especially difficult to diagnose when one thread coordinates work and another performs it.

Requested diagnostics

Please make the per-turn effective model and the reason for any change inspectable in the desktop UI and durable thread/session telemetry. A sanitized event trail should include:

• thread ID
• selected model before/after
• effective model before/after
• trigger/reason (user action, fallback, capacity policy, migration, etc.)
• timestamp
• submission/turn/message reference when applicable

I searched existing open openai/codex issues for automatic thread model changes and Sol/Terra model changes before filing. I found adjacent issues about model-change-triggered compaction and account switching, but not this silent per-thread model drift.