Stop load openai fast model for openapi compatible custom endpoint

[spikewang]

Summary

When using custom OPENAI compatible endpoint, gpt-4o-mini is always tried to be loaded automatically even it didn't exist.

There are always warnings in the log like this:

{"timestamp":"2026-04-19T00:48:34.916138Z","level":"WARN","fields":{"message":"Provider request failed with status: 404 Not Found. Payload: Some(Object {\"error\": Object\
 {\"message\": String(\"The model `gpt-4o-mini` does not exist.\"), \"type\": String(\"NotFoundError\"), \"param\": String(\"model\"), \"code\": Number(404)}}). Returning\
 error: RequestFailed(\"Resource not found (404): The model `gpt-4o-mini` does not exist.\")"},"target":"goose::providers::openai_compatible","span":{"gen_ai.request.mode\
l":"gpt-4o-mini","session.id":"20260419_4","name":"complete"},"spans":[{"gen_ai.request.model":"gpt-4o-mini","session.id":"20260419_4","name":"complete"}]}

{"timestamp":"2026-04-19T00:48:34.916189Z","level":"WARN","fields":{"message":"Request failed, retrying (1/3): RequestFailed(\"Resource not found (404): The model `gpt-4o\
-mini` does not exist.\")"},"target":"goose::providers::retry","span":{"gen_ai.request.model":"gpt-4o-mini","session.id":"20260419_4","name":"complete"},"spans":[{"gen_ai\
.request.model":"gpt-4o-mini","session.id":"20260419_4","name":"complete"}]}

ANALYSIS

OpenAiProvider::from_env() read OPENAI_HOST only after unconditionally calling model.with_fast("gpt-4o-mini", ...). So any user pointing the openai provider at a custom endpoint (like http://localhost:8000) always got gpt-4o-mini baked in as the fast model — which then fails with 404 at runtime whenever complete_fast() is called (MOIM summarization, context compaction, orchestrator).

FIX

The fix (openai.rs:73-86): read OPENAI_HOST first, then only set the default fast model when the host contains api.openai.com. For any other host, fast_model_config stays None, so use_fast_model() falls back to the main model instead of attempting a non-existent gpt-4o-mini.

With OPENAI_HOST: http://localhost:8000, the fast model will now silently use /models/Qwen3.5-35B-A3B-FP8 for MOIM/summarization calls instead of retrying gpt-4o-mini 3× and then falling back. If you later want a dedicated fast model on your local server, you can set it via a custom provider config with fast_model: "your-fast-model".

Testing

Tested with local VLLM hosted QWEN3.5:35B and OpenAI API Endpoints. The WARN message disappeared after the fix.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d7b14148cb

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[jh-block]

Thanks for the contribution, I agree with the need for this change but the fix is slightly too narrow: api.openai.com is not the only OpenAI API endpoint. There is at least eu.api.openai.com (EU data residency) as well, and I would not be surprised if there are other prefixes. Maybe this would be better:

    .map(|h| h == "api.openai.com" || h.ends_with(".api.openai.com"))

[spikewang]

Agreed. It will be great the API will have a way to indicate, but I don't think I could find one... Updated the fix.