Bug: Claude requests fail when thinking tokens enabled

Your OS version: Sequoia 15.5

Your BoltAI app version. 2.8.0 (build 55)

Are you a Setapp user. No.

The AI provider & model you’re using. Claude 4.6 Sonnet on Google Vertex and Amazon Bedrock endpoints via OpenRouter.

Steps to reproduce the issue. Create a chat and specify the Thinking tokens budget to be non-zero. The first request will work; subsequent responses will fail most of the time, but not always. Resetting the thinking token budget to zero always makes it work. Note that this occurs even when the total input tokens + thinking budget is very low.

The error message. Provider returned error.