fix(providers): strip models/ prefix for Gemini OpenAI-compat endpoint (#175)

packages/providers/src/gemini-compat.test.ts

@@ -0,0 +1,41 @@
+import { describe, expect, it } from 'vitest';
+import { isGeminiOpenAICompat, normalizeGeminiModelId } from './gemini-compat';
+
+describe('isGeminiOpenAICompat', () => {
+  it('detects the official Gemini OpenAI-compat endpoint', () => {
+    expect(isGeminiOpenAICompat('https://generativelanguage.googleapis.com/v1beta/openai/')).toBe(
+      true,
+    );
+  });
+
+  it('returns false for non-Gemini bases', () => {
+    expect(isGeminiOpenAICompat('https://api.openai.com/v1')).toBe(false);
+  });
+
+  it('returns false when baseUrl is undefined', () => {
+    expect(isGeminiOpenAICompat(undefined)).toBe(false);
+  });
+});
+
+describe('normalizeGeminiModelId', () => {
+  it('strips the models/ prefix for Gemini hosts', () => {
+    expect(
+      normalizeGeminiModelId(
+        'models/gemini-3.1-pro-preview',
+        'https://generativelanguage.googleapis.com/v1beta/openai/',
+      ),
+    ).toBe('gemini-3.1-pro-preview');
+  });
+
+  it('leaves non-Gemini model ids untouched', () => {
+    expect(normalizeGeminiModelId('gpt-4', 'https://api.openai.com/v1')).toBe('gpt-4');
+  });
+
+  it('does not strip models/ prefix when baseUrl is not a Gemini host', () => {
+    expect(normalizeGeminiModelId('models/foo', 'https://api.openai.com/v1')).toBe('models/foo');
+  });
+
+  it('is a no-op when baseUrl is undefined', () => {
+    expect(normalizeGeminiModelId('models/gemini-2-pro', undefined)).toBe('models/gemini-2-pro');
+  });
+});

packages/providers/src/gemini-compat.ts

@@ -0,0 +1,18 @@
+/**
+ * Google's OpenAI-compatible endpoint
+ * (https://generativelanguage.googleapis.com/v1beta/openai/) accepts the same
+ * request shape as OpenAI Chat Completions but rejects model ids carrying the
+ * `models/` prefix that its own /models listing returns. Settings UI keeps the
+ * prefixed id (so it matches the /models response), and we strip it only on
+ * the wire. See issue #175.
+ */
+
+export function isGeminiOpenAICompat(baseUrl: string | undefined): boolean {
+  if (!baseUrl) return false;
+  return baseUrl.includes('generativelanguage.googleapis.com');
+}
+
+export function normalizeGeminiModelId(modelId: string, baseUrl: string | undefined): string {
+  if (!isGeminiOpenAICompat(baseUrl)) return modelId;
+  return modelId.replace(/^models\//, '');
+}

packages/providers/src/index.test.ts

@@ -303,4 +303,40 @@ describe('complete', () => {
       ),
     ).rejects.toMatchObject({ code: 'ATTACHMENT_TOO_LARGE' });
   });
+
+  it('strips models/ prefix from modelId when routing through Gemini OpenAI-compat endpoint', async () => {
+    getModelMock.mockReturnValue(undefined);
+    completeSimpleMock.mockImplementationOnce(async (piModel) => {
+      expect(piModel.id).toBe('gemini-2-pro');
+      return {
+        role: 'assistant',
+        content: [{ type: 'text', text: 'hi' }],
+        api: 'openai-completions',
+        provider: 'custom-gemini',
+        model: 'gemini-2-pro',
+        usage: {
+          input: 0,
+          output: 0,
+          cacheRead: 0,
+          cacheWrite: 0,
+          totalTokens: 0,
+          cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+        },
+        stopReason: 'stop',
+        timestamp: Date.now(),
+      };
+    });
+
+    await complete(
+      { provider: 'custom-gemini', modelId: 'models/gemini-2-pro' },
+      [{ role: 'user', content: 'hello' }],
+      {
+        apiKey: 'token',
+        wire: 'openai-chat',
+        baseUrl: 'https://generativelanguage.googleapis.com/v1beta/openai/',
+      },
+    );
+
+    expect(getModelMock).toHaveBeenCalledWith('custom-gemini', 'gemini-2-pro');
+  });
 });

packages/providers/src/index.ts

@@ -18,6 +18,7 @@ import {
   looksLikeClaudeOAuthToken,
   shouldForceClaudeCodeIdentity,
 } from './claude-code-compat';
+import { normalizeGeminiModelId } from './gemini-compat';
 
 /** Subset of pi-ai's `ThinkingLevel` we expose. Maps directly to its `reasoning`
  * field, which Anthropic adapters translate to extended-thinking effort/budget
@@ -221,6 +222,11 @@ export async function complete(
   }
   const apiKey = opts.apiKey || 'open-codesign-keyless';
 
+  // Gemini's OpenAI-compat endpoint rejects the `models/` prefix that its own
+  // /models listing returns (issue #175). Normalize on the wire only; Settings
+  // keeps the prefixed form so provider/model UX stays in sync with /models.
+  const effectiveModelId = normalizeGeminiModelId(model.modelId, opts.baseUrl);
+
   const pi = (await import('@mariozechner/pi-ai')) as unknown as {
     getModel: (provider: string, modelId: string) => PiModel | undefined;
     completeSimple: (
@@ -237,12 +243,12 @@ export async function complete(
     ) => Promise<PiAssistantMessage>;
   };
 
-  let piModel = pi.getModel(model.provider, model.modelId);
+  let piModel = pi.getModel(model.provider, effectiveModelId);
   if (!piModel) {
     if (opts.wire !== undefined) {
-      piModel = synthesizeWireModel(model.provider, model.modelId, opts.wire, opts.baseUrl);
+      piModel = synthesizeWireModel(model.provider, effectiveModelId, opts.wire, opts.baseUrl);
     } else if (model.provider === 'openrouter') {
-      piModel = synthesizeOpenRouterModel(model.modelId);
+      piModel = synthesizeOpenRouterModel(effectiveModelId);
     } else {
       throw new CodesignError(
         `Unknown model ${model.provider}:${model.modelId}`,