Wednesday, Jul 22, 2026
Anthropic extended thinking and signature durability in luca-py
Anurag Verma fixed two interdependent bugs in quantized-ai/luca-py that were blocking Anthropic's extended thinking mode from working at all.
1 person shipped this
The transport layer was never reading the `reasoning_effort` parameter, making extended thinking unreachable on Anthropic [1]. At the same time, the agent was discarding the signature both on the way in and on the way out, leaving nothing to send back to the transport [1]. Fixing only one still produced a 400 error on every multi-turn tool call, so both had to land together. After review, Verma reworked the client foundation, adding three new sections to the implementation [2]. The fix enables both of Anthropic's thinking modes and restores the signature durability that allows the transport's existing replay support to function [2].
Sources
- Enable Anthropic extended thinking and make signatures durable · quantized-ai/luca-py
- Enable Anthropic extended thinking and make signatures durable (#3) · quantized-ai/luca-py