Codex App with gpt-5.6-sol aborts with unsupported prompt_cache_retention

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

ChatGPT/Codex Desktop 26.814.41407, build 6720. Bundled app-server reports codex-cli 0.148.0-alpha.15.

What subscription do you have?

ChatGPT/Codex Desktop account. The local diagnostics do not expose the exact subscription tier. This is using the official OpenAI provider, not a custom/Azure provider.

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

Codex Desktop repeatedly aborts GPT-5.6 Sol turns with:

{
  "error": {
    "message": "prompt_cache_retention is not supported on this model",
    "type": "invalid_request_error",
    "param": "prompt_cache_retention",
    "code": "invalid_parameter"
  }
}

This is happening with model gpt-5.6-sol in Codex Desktop using the official OpenAI provider. My local ~/.codex/config.toml does not set prompt_cache_retention or any prompt-cache override; it only sets the default model to gpt-5.6-sol.

The issue became frequent on 2026-08-19 after the Desktop app updated from build 6662 to build 6720 around 09:20 Asia/Shanghai time. The current bundled app-server is 0.148.0-alpha.15. I also observed one occurrence shortly before the app update while the old bundled runtime was still 0.148.0-alpha.9, so this may be a backend/client compatibility rollout issue rather than a purely local config problem.

Local evidence, sanitized:

• App version: 26.814.41407, build 6720
• Bundled app-server: codex-cli 0.148.0-alpha.15
• Previous app-server observed before update: 0.148.0-alpha.9
• Sparkle update event: 2026-08-19T01:20:02Z, current build before update was 6662
• First same-day failure observed at 2026-08-19T01:04:13Z, before the app relaunched into 0.148.0-alpha.15
• Additional failures observed after update at approximately 2026-08-19T01:53Z, 02:15Z, 02:32Z, 02:44Z, 02:52Z
• Same error reproduced at both xhigh and high reasoning effort, so changing reasoning effort does not appear to fix it.

This looks like a request-shape mismatch for GPT-5.6 models. The official OpenAI prompt caching docs say GPT-5.6 and later use prompt_cache_options.ttl, while earlier models use prompt_cache_retention:

https://developers.openai.com/api/docs/guides/prompt-caching

Related but not identical issues:

• #31882 is about Azure/custom providers and bundled model metadata. My repro is the official OpenAI provider in Codex Desktop.
• #32037 contains older logs showing prompt_cache_retention: "24h" in a GPT-5.6 Sol response path, but that issue primarily tracks updater/app-server lifecycle behavior. This report is specifically about the unsupported prompt_cache_retention invalid request error on official OpenAI-backed Codex Desktop.

What steps can reproduce the bug?

1. Use Codex Desktop on macOS.
2. Set/use model gpt-5.6-sol.
3. Run a normal Codex task with a large existing context, especially after the 2026-08-19 Desktop update to build 6720.
4. The turn can abort with invalid_request_error for param: "prompt_cache_retention".

I do not currently have a minimal deterministic reproducer, but this has occurred repeatedly today across multiple local session rollouts.

What is the expected behavior?

Codex Desktop should not send, trigger, or route a GPT-5.6 Sol request using prompt_cache_retention if that field is unsupported for this model. It should use the GPT-5.6-compatible prompt-cache option shape, or disable the unsupported retention field for GPT-5.6-family models.

Additional information

This interrupted multiple paid/interactive Codex sessions and forced retries. Please investigate and ship a client or backend-side fix. I am also separately contacting support to request a usage/credit review for the failed runs.

The issue names no source file or test; start by tracing where model metadata and prompt-cache parameters are assembled for the official provider, then compare the request for gpt-5.6-sol with the documented supported shape. Done means repeated GPT-5.6 Sol turns no longer send or trigger prompt_cache_retention and a regression check covers the failure.