mirror of
https://github.com/crewAIInc/crewAI.git
synced 2026-07-03 14:09:24 +00:00
1c6a190046
The "extract inner content" approach broke agents badly: after context summarization, the LLM sometimes echoes the <summary> block from the summary user message. Extracting the inner content treated that echo as the "real response", causing the agent to dump its entire system prompt and conversation history to the user. Reverted to the original stripping approach: <summary> blocks are removed from the response. Also reverted the streaming filter from the preflight-buffer design (which broke streaming for non-summary models by accumulating everything until task end) back to the simple character-level suppression filter that streams normally and only suppresses <summary>...</summary> spans. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>