[BUG] API Error API Request Failed with Reasoning Enabled Gemini Flash latest after context condense

[nilbounce1]

Problem (one or two sentences)

Repeated API Error API Request Failed AFTER context condense with Reasoning Enabled Gemini Flash latest. If reasoning disabled then bno more such error.

Context (who is affected and when)

After context condensing

Reproduction steps

Ubuntu 24.x

Expected result

Task progress

Actual result

Repeated error

Variations tried (optional)

No response

App Version

Version: 3.36.2 (dcb04bb)

API Provider (optional)

Google Gemini

Model Used (optional)

gemini-flash-latest

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I'm looking into this issue with the Gemini Flash reasoning mode causing API errors after context condensing. Let me explore the codebase to understand the implementation and identify the root cause.

[roomote]

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

Root Cause Analysis

The issue occurs when Gemini Flash with reasoning enabled is used during context condensing. The problem appears to be that when the condensing operation uses the same Gemini API handler with reasoning enabled, it may be sending thoughtSignature blocks that are incompatible with the condensing prompt structure.

Proposed Solution

1. Disable reasoning for context condensing operations in src/core/condense/index.ts:

   • When using a condensing API handler (especially for Gemini), ensure reasoning features are disabled during the condensing operation
   • This prevents thoughtSignature blocks from interfering with the condensing prompt

2. Create a separate handler configuration for condensing that explicitly disables reasoning:

   • Modify the condensing API handler creation to override enableReasoningEffort to false
   • This ensures condensing operations don't trigger reasoning-related API features

3. Add safeguards in the Gemini handler to handle edge cases:

   • Ensure thoughtSignature handling is properly isolated from condensing operations
   • Add error handling for cases where reasoning blocks might interfere with API calls

This approach will:

• Prevent API errors during context condensing when reasoning is enabled
• Maintain reasoning functionality for normal conversation flow
• Ensure condensing operations work reliably regardless of the main model's reasoning settings

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

[nilbounce1]

@roomote Approved

[roomote]

I've seen your approval and will proceed with implementing the fix. I'll explore the codebase to address the issue where Gemini Flash with reasoning enabled causes API errors after context condensing. I'll open a PR to share progress.

[roomote]

Opened a PR with my attempt at addressing this: #9911

The fix disables reasoning features for Gemini/Vertex providers during context condensing operations to prevent the API errors you were experiencing. Feedback welcome!

[hannesrudolph]

Fixed