Summary
Selecting Provider: Google (Gemini) with Model: Gemini 2.0 Flash in Pickaxe results in calls to Google’s OpenAI-compat endpoint /v1beta/openai/chat/completions, producing 400 Client Error: Bad Request. The request appears to use OpenAI Chat Completions schema instead of Gemini native schema.
Environment
- App: Pickaxe (Form-style tool with Connected Action disabled for test)
- Model: Google Gemini 2.0 Flash
- Provider key: Google API key
- Date/Time:
- Region:
- Project/Tool ID:
Expected
When Provider = Google (Gemini), Pickaxe should call Gemini native endpoint and payload format (either models/gemini-...:generateContent or the current GA equivalent), not the OpenAI-compat route.
Actual
Network error shown in UI:
400 Client Error: Bad Request for url: https://generativelanguage.googleapis.com/v1beta/openai/chat/completions
Symptoms indicate Pickaxe routed via OpenAI-compat and sent an OpenAI-style chat.completions body.
Repro Steps
- Create a new Pickaxe (any tool).
- In Model settings, choose Provider: Google (Gemini).
- Select Model: Gemini 2.0 Flash.
- Save and run a simple prompt (e.g., “hello world”).
- Observe 400 error with URL ending in
/v1beta/openai/chat/completions.
Evidence / Logs
- Screenshot of error (attached).
- (If available) Network tab shows request to
.../v1beta/openai/chat/completionswith OpenAI-style fields (model,messages, etc.).
Impact
- Blocks use of Gemini models in Pickaxe even with correct Google API key.
- Affects any user selecting Gemini 2.0 Flash via native provider flow.
Hypothesis
Provider/model switcher isn’t switching the transport + serializer; it continues using the OpenAI Chat Completions client and payload while swapping only the model id, causing a 400 at Google’s OpenAI-compat endpoint.
