bug(runtime): every auxiliary model call fails on OpenAI Codex subscription connections (titles, goal evaluation, recap, Daily Review, prompt suggestions) #5711

What happened

On an openai-codex (ChatGPT/Codex subscription OAuth) connection, main turns work, but every Host auxiliary model call fails. These are the calls that go through runHostAuxiliaryModelCall → generateToolFreeModelCall / generateProviderPrefixModelCall, which use non-streaming generateText:

• Session titles
• Goal evaluation
• Session recap
• Memory extraction
• Daily Review
• WorkHub routing
• feat(desktop): add opt-in next prompt suggestions #5629's next-prompt suggestions

For example, sessions on a Codex subscription never get an auto-generated title.

This is from a real Desktop workspace's usage_llm_calls. Every auxiliary row for the Codex connection is a failure, and none succeeded:

I captured the backend's response bodies and found three separate causes, each exposed once the one before it is fixed:

1. 400 {"detail":"Stream must be set to true"}. The Codex backend only accepts stream: true. Main turns stream, but generateText sends a non-streaming Responses request. buildOpenAiCodexFetch (packages/runtime/src/subscription-model-fetch.ts:127-178) rewrites instructions, store and text.verbosity, but not stream.
2. 400 {"detail":"Unsupported parameter: max_output_tokens"}. Every auxiliary call sets maxOutputTokens, which the Codex backend rejects. Main turns don't send it.
3. Empty text after a successful stream. When forced to stream, the terminal response.completed event carried no output items. The usage row was success with 11 output tokens, but the assembled text was "". The message item only arrives in the earlier response.output_item.done events.

How to reproduce

1. Configure an openai-codex connection (ChatGPT sign-in) and pick e.g. gpt-6-astra, gpt-5.6-sol or gpt-5.5.
2. Start a new Session and finish a turn.
3. The Session never gets a generated title. The Usage store records a session_title row with status: 'error' and errorClass: 'OpenAiCodexHttpError', and the error message carries HTTP 400 {"detail":"Stream must be set to true"}.

For a headless check, I ran a Runtime Host from source (upstream/main) on a copy of that Desktop workspace:

• Unmodified: every session.prompt-suggestion.generate call recorded OpenAiCodexHttpError.
• With all three causes worked around in the Codex fetch layer: suggestions came back from gpt-6-astra, e.g. Let's go with parseISODate. and Yes, give me a checklist., and the usage rows were success.

Environment

• Maka commit: f1f259b67 (also present at 99cfeb7e9)
• OS: Windows 10 x64
• Surface: Runtime Host auxiliary model calls, visible in Desktop (Session titles, Goal, Daily Review, prompt suggestions)
• Connection: openai-codex via ChatGPT OAuth

Logs, screenshots, or additional context

Proposed fix, contained in buildOpenAiCodexFetch:

• When the caller's body is not streaming, send stream: true.
• Drop max_output_tokens.
• Read the SSE stream to its terminal event (response.completed / response.incomplete / response.failed).
• Rebuild output from response.output_item.done items when the terminal response omits them.
• Return that response as the JSON body the non-streaming Responses parser expects.

This keeps the provider quirk in the provider's fetch adapter. Auxiliary callers stay unchanged, and streaming callers are untouched.

Side effect: Codex auxiliary calls lose their output cap. They are bounded only by their prompts. The next-prompt path also rejects output over 80 characters.

A related tuning point, not part of this bug: measured gpt-6-astra suggestion latency was 3.2–8.3 s, while #5629's deadline is 5 s. So on Codex some suggestions will still time out once this is fixed.

I'm happy to send the fix.

AI disclosure: Claude Code (Claude Opus) found the failure while verifying #5690 against a real Codex connection, captured the backend responses, and drafted this issue; liugddx reviewed it and is the human contributor of record.