bug(runtime): a per-model output limit makes every OpenAI Codex subscription turn fail with HTTP 400

[liugddx]

What happened

On an openai-codex (ChatGPT subscription) connection, setting a per-model output limit (modelOverrides[model].maxOutputTokens, exposed in the connection's model settings) makes every main turn on that model fail. The model adapter forwards the limit as max_output_tokens (packages/runtime/src/model-adapter.ts:715-745 → ai-sdk-turn.ts:1580-1622). The Codex backend rejects that field:

failureClass: request_rejected
Codex OAuth request failed: HTTP 400 {"detail":"Unsupported parameter: max_output_tokens"}

The setting is accepted and saved without complaint. The user only learns about the problem when every turn on that model fails.

How to reproduce

1. Configure an openai-codex connection (ChatGPT sign-in).
2. In the connection's model settings, set an output limit for e.g. gpt-6-astra (verified with 4096).
3. Start a turn on that model. It fails with the 400 above. Remove the limit and the same turn completes.

I reproduced this headlessly against the real backend on main at 87fc9f69c, using a Runtime Host on a copy of a Desktop workspace.

Environment

• Maka commit: 87fc9f69c
• OS: Windows 10 x64
• Connection: openai-codex via ChatGPT OAuth

Logs, screenshots, or additional context

Found while addressing review on #5723. That PR keeps streamed turns byte-identical to main, so it neither causes nor fixes this.

Possible directions, for maintainers to choose:

• The settings UI and codec refuse an output limit on openai-codex models, with a message saying the subscription backend does not accept one.
• Or the adapter drops it for openai-codex and the UI marks it as ignored there.

AI disclosure: Claude Code (Claude Opus) reproduced the failure against the real backend and drafted this issue; liugddx reviewed it and is the human contributor of record.

[liugddx]

take