gpt-5.3-codex-spark is incorrectly advertised as supporting reasoning summaries, causing persistent HTTP 400 errors

Start by tracing the model/list catalog refresh and shared models_cache.json handling, then follow how supports_reasoning_summaries becomes the reasoning.summary request parameter. Verify that Spark is advertised as unsupported and that an unsupported_parameter response for reasoning.summary is handled as described, including behavior for existing threads.

What version of the Codex App are you using (From “About Codex” dialog)?

26.715.31925 (build 5551); bundled Codex Core: codex-cli 0.145.0-alpha.18

What subscription do you have?

ChatGPT Pro

What platform is your computer?

Darwin 25.3.0 arm64 arm

What issue are you seeing?

The Codex remote model catalog advertises gpt-5.3-codex-spark with:

"supports_reasoning_summaries": true

However, the Spark serving endpoint rejects reasoning.summary. A minimal prompt such as test fails with:

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "code": "unsupported_parameter",
    "message": "Unsupported parameter: 'reasoning.summary' is not supported with the 'gpt-5.3-codex-spark' model.",
    "param": "reasoning.summary"
  },
  "status": 400
}

This is a capability mismatch: the catalog tells Codex that Spark supports reasoning summaries, so the client sends the parameter, but the model endpoint rejects it. It is unrelated to project source code and reproduces in projectless threads.

What steps can reproduce the bug?

Feedback ID: 019f7cfd-f023-73d3-9a44-2d124613a7fc

What is the expected behavior?

The remote catalog should advertise gpt-5.3-codex-spark with:

"supports_reasoning_summaries": false

Codex should omit reasoning.summary for Spark. As a defensive fallback, when the endpoint returns unsupported_parameter for reasoning.summary, the client should remove the parameter and retry automatically.

Additional information

Evidence

The incorrect Spark capability was observed with at least two catalog ETags:

• W/"db2a6dc50b1d003969cdc236274e488a"
• W/"82064eb631c228baaf368627eaa0d23f"

Model-manager logs show complete-catalog refreshes initiated by both Codex Desktop model/list and the internal codex-auto-review process. Because all processes share ~/.codex/models_cache.json, manually changing Spark to false is later overwritten by the remotely supplied catalog.

Configurations model_reasoning_summary = "none" and model_supports_reasoning_summaries = false were attempted. Existing threads may retain summary=auto or reasoning_summary=detailed.

Versions observed

• Current Codex App: 26.715.31925 (build 5551)
• Current bundled Core: codex-cli 0.145.0-alpha.18
• Earlier affected Core versions: 0.144.0 and 0.144.1

Temporary workaround

Use a local model_catalog_json in which Spark has supports_reasoning_summaries: false, stop every existing Codex App/CLI process, and create a new thread. Editing models_cache.json directly is not durable because it is automatically refreshed.

The feedback session and diagnostic logs are attached through the Feedback ID above.