fix: handle reasoning-only responses from Gemini 3 Pro Preview

[roomote]

This PR attempts to address Issue #10422 - Gemini Pro failing on both Gemini and OpenRouter.

Problem

When using google/gemini-3-pro-preview with high reasoning effort and native tool calling enabled, the model sometimes returns ONLY reasoning content (via reasoning_details) without any text content or tool uses. The current validation in Task.ts only checks for hasTextContent and hasToolUses, so reasoning-only responses fall through to the error handling path, causing the "Unexpected API Response: The language model did not provide any assistant messages" error.

This happens on the second message (and subsequent messages) every time, as confirmed by the reporter.

Solution

1. Added reasoningOnlyResponse() helper in src/core/prompts/responses.ts:

   • Prompts the model to continue with actionable output after a reasoning-only response
   • Instructs the model to use tools or provide text content

2. Added hasReasoning check in src/core/task/Task.ts:

   • New variable that checks if reasoningMessage.length > 0
   • Added else if (hasReasoning) block between the successful case and error case
   • When reasoning-only response is detected:
     • Reset error counters (valid response received)
     • Save assistant message with reasoning to API history
     • Add prompting message to continue with actionable output
     • Continue the task loop instead of failing

Testing

• TypeScript compilation passes
• All 122 existing Task tests pass
• No regressions introduced

Notes

This fix allows the task loop to continue instead of failing when the model produces a reasoning-only response. The model is prompted to provide actionable output (tools or text), which should resolve the issue for users with high reasoning effort enabled.

Feedback and guidance are welcome!

[roomote]

Rooviewer   See task on Roo Cloud

Review complete. No issues found.

The implementation correctly handles reasoning-only responses from Gemini 3 Pro Preview by detecting when the model returns only reasoning content (without text or tool uses), saving the reasoning to API history, and prompting the model to continue with actionable output.

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}