OpenHands Tool Call: Reasoning to Runtime Diagram

The boundary

Reasoning ends when an ActionEvent is persisted; everything after runs in the ToolExecutor
Executors bind to the conversation's workspace at create(conv_state) time
Locally the shell is tmux/subprocess in the workspace; remotely the same executors run inside the agent-server

Failure handling

Unknown tool, bad JSON or schema errors: ActionEvent(action=None) + AgentErrorEvent
ValueError from an executor becomes AgentErrorEvent; the loop continues
Cancelled before start: a synthetic error result keeps tool_call/result pairing

Output bounding

tmux history-limit 10,000 lines; session-level maybe_truncate to 30,000 chars
to_llm_content keeps head + tail within 30,000 chars and saves the text under observations/