Module thinking_prune 

Strip extended-thinking blocks from older messages.

Modern reasoning models (Claude extended thinking, DeepSeek R1, GPT-5 reasoning, Gemini thought summaries) emit Thinking content parts that can be 10-100× larger than the assistant’s actual reply. These blocks help the current turn’s decision but carry almost no value once the turn has produced its tool calls and the loop has moved on — the final answer/action already reflects them.

This module removes ContentPart::Thinking from every message older than KEEP_LAST_MESSAGES. Recent thinking is preserved so the model can still reference its own recent chain-of-thought.

Safety

• Providers that inject thinking for correctness (cache-coherent thought signatures on Gemini ToolCall) are unaffected — those signatures live on ContentPart::ToolCall::thought_signature, not on Thinking blocks.
• An assistant message whose only content was a thinking block becomes empty; such messages are removed entirely to keep the buffer a valid provider-consumable shape.

Always-on

No config. Thinking blocks are known to be non-essential after the turn completes; stripping them is the single highest-ROI shrink for reasoning-heavy agent loops.

Constants

KEEP_LAST_MESSAGES

Keep thinking blocks in this many trailing messages.

Functions

prune_thinking

Strip Thinking parts from older messages and drop any messages that become empty as a result.