[Bug] 400 Bad Request error when using Google Gemini API via Custom Provider (OpenAI Chat protocol)

[CaioGS06]

Description

When attempting to use a Google Gemini API key through the "Custom provider" option using the OpenAI Chat wire protocol, the initial connection and model fetching succeed, but generating a design fails immediately with a 400 status code (no body) error.

Note for triage: Based on the logs, there are two likely culprits regarding Gemini's strict OpenAI wrapper:

1. Model Name Prefix: The connection test successfully fetches models, but they are returned with the models/ prefix (e.g., models/gemini-3.1-pro-preview). Passing this prefix back in the generation payload might be causing the API to reject the request.
2. Unsupported Agent Parameters: Gemini's OpenAI compatibility wrapper is very strict. If Open CoDesign's agent runtime is passing specific OpenAI parameters in the payload (like presence_penalty, frequency_penalty, or strict JSON schema enforcement) that Google hasn't mapped, it will reject the entire request with a 400 error regardless of prompt size.

Steps to reproduce

1. Go to Settings -> Models -> + Add provider.
2. Select Custom provider.
3. Set Wire protocol to OpenAI Chat.
4. Set Base URL to https://generativelanguage.googleapis.com/v1beta/openai/.
5. Enter a valid Google Gemini API key.
6. Click Test connection (this succeeds and populates the model dropdown).
7. Select a model from the dropdown (e.g., models/gemini-3.1-pro-preview).
8. Return to the chat and attempt to generate a design prompt.

Expected behavior

The app should successfully pass the prompt to the Gemini API via the OpenAI compatibility layer and return a generated design.

Actual behavior

The generation fails immediately. The UI displays an error dialog stating: Error invoking remote method 'codesign:v1:generate': CodesignError: 400 status code (no body).

open-codesign version

0.1.3

Platform

Windows

OS version

Windows 10

Model provider (if relevant)

Other

Error code

PROVIDER_ERROR

Diagnostics bundle

open-codesign-diagnostics-2026-04-22T21-30-40.zip

Logs

[2026-04-22 18:02:54.037] [info]   (main:ipc) [generate] step=load_config.ok {
  generationId: 'moajhfkr-0krhrt',
  ms: 0,
  hasApiKey: true,
  baseUrl: 'https://generativelanguage.googleapis.com/v1beta/openai/'
}
...
[2026-04-22 18:02:54.419] [error]  (main:ipc) [title] step=send_request.fail { ms: 424, errorClass: 'CodesignError' }
[2026-04-22 18:02:54.419] [error]  (main:ipc) [title] generate-title.fail {
  provider: 'custom-gemini-api-dnpm',
  modelId: 'models/gemini-3.1-pro-preview',
  baseUrl: 'https://generativelanguage.googleapis.com/v1beta/openai/',
  message: '400 status code (no body)',
  code: 'PROVIDER_ERROR'
}

[github-actions]

Thanks for the detailed report and logs — this is very helpful.

I checked the current code path and can confirm a few concrete things:

• The model ID returned by /models is passed through as-is (no normalization/stripping of models/ prefix before use).
• The selected custom provider model is persisted and then used for both title generation and main generation calls.
• In this repository, we do not explicitly set presence_penalty, frequency_penalty, JSON schema response_format, or similar OpenAI-only fields in our own request-building layer.

So based on current code, your models/... prefix hypothesis is consistent with what we do today; and if unsupported params are being sent, that would likely be inside the underlying provider adapter path, not explicit app-side fields in this repo.

Relevant code:

• apps/desktop/src/main/connection-ipc.ts:248 — model IDs from provider response are returned directly via id/name
• apps/desktop/src/main/connection-ipc.ts:274 — extractModelIds accepts OpenAI-compatible { data: [{ id }] } shape
• apps/desktop/src/main/onboarding-ipc.ts:596 — custom defaultModel is accepted as a non-empty string and stored
• apps/desktop/src/main/provider-settings.ts:235 — active model resolution uses selected model ID
• apps/desktop/src/main/index.ts:579 — generation resolves active model and passes it to core
• apps/desktop/src/main/index.ts:895 — title generation uses the same active provider/model path
• packages/core/src/index.ts:353 — core sends limited options (wire/baseUrl/headers/maxTokens/reasoning)
• packages/providers/src/index.ts:254 — provider bridge passes apiKey/baseUrl/signal/maxTokens/reasoning/headers

Need more info:

• Can you share the provider-side HTTP response body (if any) for the failing POST /chat/completions call? Right now the app log only shows 400 status code (no body).

Also, quick context: Open CoDesign is still in pre-alpha (v0.1.x), so some behaviors are still being hardened and roadmap items are not fully implemented yet.

---

open-codesign Bot

[rward27]

I confirm this issue persist in latest version. I am facing same issue with google gemini models.
What's the solution to this?