[BUG] Frequent first API call fails when switching modes with Gemini 3

[S-PAC-Digital]

Problem (one or two sentences)

When all (or multiple?) Roo modes are set to use Gemini 3, most of the time when Roo switches modes the first API call fails with the error:

"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."

However this failed request usually still has a "Thinking" icon that can be expanded to see the model's thinking output, so the API call was at least partially successful.

It's unclear if this is a model response issue, or an issue with Roo's interpretation of the model response.

In these cases the first retry attempt is usually successful.

Context (who is affected and when)

Users of Gemini 3 models in multiple Roo modes.

Reproduction steps

1. Use gemini-3-pro-preview for all Roo modes
2. Allow auto approve for everything
3. Create task in Orchestrator that will will require multiple mode changes

Watch the first request after each mode change. I have experienced failures of 80+% of first request after mode change.

Expected result

Roo mode processes request that was provided by Orchestrator

Actual result

First API request fails and requires retry

Variations tried (optional)

No response

App Version

v3.34.0

API Provider (optional)

Google Gemini

Model Used (optional)

gemini-3-pro-preview

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

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.

[roomote]

I'm exploring a potential fix for the Gemini 3 mode switching issue. I'll investigate why the first API call after mode switches fails to parse assistant messages correctly.

[roomote]

I've opened PR #9521 to fix this issue: #9521

The fix modifies the response validation logic to properly handle reasoning-only responses from Gemini 3 models after mode switches. Instead of treating these as errors, the system now:

• Accepts reasoning-only responses as valid content
• Doesn't count them as mistakes
• Provides gentle prompts to continue with the task

This should resolve the frequent first API call failures you were experiencing with Gemini 3 models.