[BUG] Gemini 3 pro causes provider errors

[TsunamiWind]

Problem (one or two sentences)

Gemini 3 Pro Preview is causing errors.

Context (who is affected and when)

Roughly ever 2-3 requests this error occurs with Gemini 3 pro, regardless if its through Google's api or through an AI Gateway such as LiteLMM.

Date/time: 2026-02-17T05:47:34.685Z
Extension version: 3.47.3
Provider: gemini
Model: gemini-3-pro-preview

Unexpected API Response: The language model did not provide any assistant messages. This may indicate an issue with the API or the model's output.

Reproduction steps

1. Use Gemini API
2. Select the gemini-3-pro-preview

Expected result

Should work smoothly without any errors

Actual result

Works once in 3 requests

Variations tried (optional)

No response

App Version

3.47.3

API Provider (optional)

Google Gemini

Model Used (optional)

gemini-3-pro-preview

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I've analyzed this issue and here's what I found:

Root Cause Analysis

The error "The language model did not provide any assistant messages" is triggered in src/core/task/Task.ts (line ~3421) when the API stream completes but yields no text content and no tool_use blocks. For gemini-3-pro-preview, there are two likely scenarios causing this:

1. Reasoning-only responses: gemini-3-pro-preview supports thinking/reasoning (efforts: ["low", "high"]). Sometimes the model returns only reasoning/thought parts with no actual text or tool calls. The Gemini handler (src/api/providers/gemini.ts) correctly yields reasoning chunks, but Task.ts only checks for text and tool_use content when determining if the response was valid. Reasoning-only responses are treated as empty.

2. Silent empty responses with unhandled finish reasons: The Gemini handler tracks finishReason (e.g., SAFETY, RECITATION, MAX_TOKENS) but never surfaces it. If the API returns a non-STOP finish reason with no content, the stream ends silently with no chunks, and no diagnostic information reaches the user or retry logic.

The existing grace-retry mechanism (silent retry on first failure, error after 2 consecutive) helps, but with ~33% failure rate the user experience is significantly degraded.

Proposed Implementation Plan

1. Surface finishReason from the Gemini handler (src/api/providers/gemini.ts): After the stream loop ends, if there was no content (!hasContent) and the finishReason is not STOP, yield an informational event or throw a descriptive error that includes the finish reason (e.g., "Response blocked by safety filter"). This gives Task.ts better information for retry decisions.

2. Handle reasoning-only responses in Task.ts (src/core/task/Task.ts, around line 3415): Expand the content check to also consider reasoningMessage.length > 0. When a response contains reasoning but no actionable content, auto-retry (similar to existing empty-response retry) since the model clearly attempted to respond but failed to produce actionable output.

3. Add a Gemini-specific automatic retry in the handler (src/api/providers/gemini.ts): When the stream ends with no content and the finish reason is STOP (suggesting the model just didn't produce output rather than being blocked), automatically retry once within the handler before yielding back to Task.ts. This reduces the latency of the retry cycle.

4. Add/update tests: Add test cases in the existing grace-retry and Gemini provider test files to cover reasoning-only responses and non-STOP finish reasons.

This approach targets the root causes while working within the existing retry infrastructure.

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

[TsunamiWind]

@roomote approve

[roomote]

I've seen the approval and will proceed with the implementation plan. I'll explore the codebase, implement the fixes for the Gemini 3 Pro empty response handling, and open a PR to share progress.

[roomote]

Opened PR #11497 with my attempt at addressing this. The fix handles reasoning-only responses from Gemini 3 Pro (auto-retry without counting as failure) and surfaces non-STOP finish reasons as descriptive errors. Feedback welcome!