Responses-Lite model compatibility error can loop through auto-compaction and waste tokens #31607

Summary

When Codex Desktop is used with a model/channel combination that is not supported by X-OpenAI-Internal-Codex-Responses-Lite, the app can repeatedly emit the same compatibility error while also auto-compacting context. This can consume a large amount of token budget without making progress.

Observed error:

This model is not supported when using X-OpenAI-Internal-Codex-Responses-Lite.

Observed Behavior

A Codex Desktop thread repeatedly showed:

• Context automatically compacted
• the compatibility error above
• another compaction/retry cycle

In the reported case, the user saw 30%+ of the available token usage consumed before manually stopping the run.

Expected Behavior

Codex should detect model/channel incompatibility before starting or resuming an automatic Goal/continuation run. Once the incompatibility is known, it should fail fast with a single blocking UI error and stop further auto-retry or auto-compaction for that run.

Impact

• Significant avoidable token consumption
• No useful task progress
• Context degradation from repeated compaction
• Poor user trust because the error is deterministic and should be preflightable

Suggested Fix

Add a preflight compatibility check between the selected model and the Responses-Lite mode/header before Goal execution or continuation. If unsupported:

1. Show one actionable UI error.
2. Disable auto-resume/auto-retry for that run.
3. Avoid triggering context compaction as part of the failed retry path.

Environment

• Product: Codex Desktop
• Model shown in UI: 5.5 Medium
• Mode/header implicated by error: X-OpenAI-Internal-Codex-Responses-Lite
• Scenario: Goal/automatic continuation with repeated context compaction

I am filing this on behalf of a user who hit the loop and lost substantial token budget to the compatibility error.