8cb8b9128e
When retainToolCalls is enabled (new default), the retention transcript is output as JSON with full message structure including tool_use blocks (Edit, Read, Bash, Grep, etc.) and their complete input dicts, plus tool_result blocks (truncated at 2k chars). This preserves the context of what actions the assistant actually took, not just its narration. Hindsight MCP tools (recall/retain/reflect) are excluded to prevent feedback loops. Channel message tools still get their text extracted inline. Setting retainToolCalls=false falls back to the legacy text format. |
||
|---|---|---|
| .. | ||
| agno | ||
| ai-sdk | ||
| chat | ||
| claude-code | ||
| crewai | ||
| hermes | ||
| langgraph | ||
| litellm | ||
| nemoclaw | ||
| openclaw | ||
| pydantic-ai | ||
| strands | ||