When conversation history exceeds max_history, the blind slice in trim_messages
could cut between an assistant(tool_calls) message and its tool result messages.
DeepSeek/OpenAI requires every role:"tool" message to follow a preceding
assistant message with matching tool_calls.
The fix detects when the slice boundary lands on a tool message, then scans
backwards to include the parent assistant(tool_calls) message(s). If the parent
cannot be found (e.g. cut at a user boundary), the orphaned tool messages are
removed. This fixes the "Messages with role 'tool' must be a response to a
preceding message with 'tool_calls'" error from DeepSeek.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- 4.1 Fallback chain: LLM fallback_llm config in AgentLLMConfig, retry with alternate model on API failure; Agent fallback_agent in DAG nodes
- 4.2 Knowledge sharing: GlobalKnowledge model with embedding-based semantic search, auto-extraction of tool names as tags after execution
- 4.3 Async execution: execute_agent_task fully implemented with AgentRuntime, scheduler dual-path for workflow/non-workflow agents
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
