[BUG] Custom model names for Google Gemini / GCP Vertex AI are silently ignored and fallback to default model (gemini-3.1-pro)

[emplanetes]

Problem (one or two sentences)

When entering a custom model name that is not in the predefined list for Google Gemini or GCP Vertex AI, the extension silently overrides the user's input and sends the API request using the default model (gemini-3.1-pro). This is structurally highly identical to the issue reported in #12074 .

Context (who is affected and when)

1. I attempted to use the latest model (gemini-3.1-flash-lite-preview), which is currently missing from the predefined dropdown list. I manually typed it into the model field. However, the actual outgoing requests were silently replaced with gemini-3.1-pro.
2. I only discovered this bug because I hit the quota limit for gemini-3.1-pro and received an error. If the request hadn't failed due to the quota, I would have been unknowingly using a much more expensive model, resulting in significant unintended API costs.
3. Additionally, the predefined model lists for both Google Gemini and GCP Vertex AI are severely outdated (containing deprecated models while missing the latest ones). This forces users to rely on the custom model input, which is currently broken.

Reproduction steps

1. Roo Code settings and select the "Google Gemini" (or "GCP Vertex AI") API provider.
2. Provide a valid API Key.
3. In the Model field, manually enter a custom model name that is not in the predefined list (e.g., gemini-3.1-flash-lite-preview).
4. Send any prompt in the chat.
5. Check the error logs (or API billing dashboard) and observe that the request was actually made using the default model (gemini-3.1-pro).

Expected result

The extension should strictly respect the custom model name entered by the user and send the API request using exactly that model (e.g., gemini-3.1-flash-lite-preview, or gemini-2.5-flash-lite).

Actual result

The API request fails with a quota error, revealing that the request was routed to gemini-3.1-pro instead of the user-defined gemini-3.1-flash-lite-preview, or gemini-2.5-flash-lite.

Variations tried (optional)

• If I select a model that already exists in the predefined list, it works as expected.
• The exact same silent fallback bug occurs when using the GCP Vertex AI provider.

App Version

3.52.1

API Provider (optional)

Google Gemini

Model Used (optional)

gemini-3.1-flash-lite-preview,

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

Date/time: 2026-04-16T17:41:26.401Z
Extension version: 3.52.1
Provider: gemini
Model: gemini-3.1-flash-lite-preview

Gemini 生成上下文流错误：{"error":{"message":"{\n  \"error\": {\n    \"code\": 429,\n    \"message\": \"You exceeded your current quota, please check your plan and billing details. For more information on this error, head to: https://ai.google.dev/gemini-api/docs/rate-limits. To monitor your current usage, head to: https://ai.dev/rate-limit. \\n* Quota exceeded for metric: generativelanguage.googleapis.com/generate_content_free_tier_input_token_count, limit: 0, model: gemini-3.1-pro\\n* Quota exceeded for metric: generativelanguage.googleapis.com/generate_content_free_tier_requests, limit: 0, model: gemini-3.1-pro\\n* Quota exceeded for metric: generativelanguage.googleapis.com/generate_content_free_tier_requests, limit: 0, model: gemini-3.1-pro\\n* Quota exceeded for metric: generativelanguage.googleapis.com/generate_content_free_tier_input_token_count, limit: 0, model: gemini-3.1-pro\\nPlease retry in 33.912482576s.\",\n    \"status\": \"RESOURCE_EXHAUSTED\",\n    \"details\": [\n      {\n        \"@type\": \"type.googleapis.com/google.rpc.Help\",\n        \"links\": [\n          {\n            \"description\": \"Learn more about Gemini API quotas\",\n            \"url\": \"https://ai.google.dev/gemini-api/docs/rate-limits\"\n          }\n        ]\n      },\n      {\n        \"@type\": \"type.googleapis.com/google.rpc.QuotaFailure\",\n        \"violations\": [\n          {\n            \"quotaMetric\": \"generativelanguage.googleapis.com/generate_content_free_tier_input_token_count\",\n            \"quotaId\": \"GenerateContentInputTokensPerModelPerMinute-FreeTier\",\n            \"quotaDimensions\": {\n              \"location\": \"global\",\n              \"model\": \"gemini-3.1-pro\"\n            }\n          },\n          {\n            \"quotaMetric\": \"generativelanguage.googleapis.com/generate_content_free_tier_requests\",\n            \"quotaId\": \"GenerateRequestsPerMinutePerProjectPerModel-FreeTier\",\n            \"quotaDimensions\": {\n              \"location\": \"global\",\n              \"model\": \"gemini-3.1-pro\"\n            }\n          },\n          {\n            \"quotaMetric\": \"generativelanguage.googleapis.com/generate_content_free_tier_requests\",\n            \"quotaId\": \"GenerateRequestsPerDayPerProjectPerModel-FreeTier\",\n            \"quotaDimensions\": {\n              \"location\": \"global\",\n              \"model\": \"gemini-3.1-pro\"\n            }\n          },\n          {\n            \"quotaMetric\": \"generativelanguage.googleapis.com/generate_content_free_tier_input_token_count\",\n            \"quotaId\": \"GenerateContentInputTokensPerModelPerDay-FreeTier\",\n            \"quotaDimensions\": {\n              \"location\": \"global\",\n              \"model\": \"gemini-3.1-pro\"\n            }\n          }\n        ]\n      },\n      {\n        \"@type\": \"type.googleapis.com/google.rpc.RetryInfo\",\n        \"retryDelay\": \"33s\"\n      }\n    ]\n  }\n}\n","code":429,"status":"Too Many Requests"}}

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

Root cause: In src/api/providers/gemini.ts line 351, the getModel() method checks if the user-provided model ID exists in the predefined geminiModels map. If it does not, the code silently falls back to geminiDefaultModelId (gemini-3.1-pro) instead of passing through the custom model name:

let id = modelId && modelId in geminiModels ? (modelId as GeminiModelId) : geminiDefaultModelId

This is structurally identical to the bug that was already fixed for Vertex AI in commit 9724a28 (closing #12074).

Proposed fix:

1. Modify src/api/providers/gemini.ts -- Update getModel() to handle three cases instead of two:

   • If modelId is in geminiModels: use the curated ModelInfo (existing behavior).
   • If modelId is provided but not in the map: pass it through as-is with sensible default ModelInfo (new behavior for custom models).
   • If no modelId is provided: fall back to geminiDefaultModelId (existing behavior).

   This matches the exact pattern already applied to vertex.ts.

2. Update src/api/providers/__tests__/gemini.spec.ts -- Add test cases confirming that:

   • A custom/unknown model ID is passed through as-is (not replaced by the default).
   • The default model is still used when no model ID is specified.

This approach ensures custom model names entered by users are respected and sent to the Gemini API exactly as typed, while maintaining backward compatibility for known models and the no-model-specified fallback.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[emplanetes]

@roomote I approve this plan. Please proceed with the implementation. Thanks!

[roomote-v0]

I've seen the approval and will proceed with the implementation now. I'll open a PR to share progress shortly.

[roomote-v0]

Opened a PR with my attempt at addressing this: #12137. Feedback welcome!