[Bug] ChatGPT model override is discarded: chatgpt-engine.js ignores its engine argument and hardcodes model: 'auto'

[sjgold]

Summary

ChatGPT model selection is plumbed all the way from the CLI to the injected browser call, then discarded. chatgpt-engine.js accepts an engine argument and never reads it, hardcoding model: 'auto' in the request to /backend-api/conversation.

Gemini's equivalent works, so the routing layer is fine. Only the ChatGPT engine is missing its half of the implementation.

Version: v5.0.0, Session (Browser) mode, Windows 11.

Reproduction

proxima ask chatgpt "Reply with your model name, then OK. Nothing else."
# -> GPT-5.5-mini

proxima ask -m chatgpt:gpt-4o "Reply with your model name, then OK. Nothing else."
# -> GPT-5.5-mini  (override ignored)

Selecting a different model in the ChatGPT tab's own model picker inside the Agent Hub changes nothing either, since the engine posts to /backend-api/conversation directly and bypasses whatever the web UI has selected.

Trace

The override survives the entire chain and dies at the last consumer:

1. cli/proxima-cli.cjs:705 reads --model and posts it as { model, message } (:335)
2. electron/api/rest-api.cjs:269-283 _parseResolved splits on : generically, so chatgpt:gpt-4o resolves to { providers: ['chatgpt'], gemini: 'gpt-4o' } (the field name says gemini but it functions as a provider-agnostic engine override)
3. electron/api/rest-api.cjs:401 passes it to queryProvider
4. electron/main-v2.cjs:525-527 builds targetProvider = provider + ':' + engineOverride, generic across providers
5. electron/providers/api.cjs:129-131,159 splits that and injects window.__proximaChatGPT.send(message, "gpt-4o", attachments, conversationId)
6. electron/providers/engines/chatgpt-engine.js:374 receives "gpt-4o" as engine and never references it. Line 453 is model: 'auto'.

For comparison, gemini-engine.js:507-539 parses its engine argument into a workspace ID plus a {modelId, customIndex11, inner79} profile and bakes it into request headers (:627) and innerReq[79] (:614).

Suggested fix

Minimal pass-through at chatgpt-engine.js:445-453:

var requestedModel = 'auto';
if (typeof engine === 'string' && engine.trim() && engine.trim() !== 'auto') {
    requestedModel = engine.trim();
    console.log('[Proxima ChatGPT] Using requested model:', requestedModel);
}

var payload = {
    action: 'next',
    messages: [ /* ... */ ],
    model: requestedModel,
    // ...
};

That keeps auto as the default and forwards anything explicit. A follow-up would be validating against the slugs /backend-api/models actually returns, so an unknown slug fails with a clear message instead of a raw backend error.

Related note

MODEL_ALIASES in rest-api.cjs:25 maps gpt-4o and gpt-4 to the provider chatgpt, so a bare -m gpt-4o is read as a provider selection with no engine. The colon form is the only way to express an engine today, which is worth documenting in proxima --help whichever way the fix lands.