Document AgentErrorEvent vs ConversationErrorEvent in Events architecture

[enyst]

Document AgentErrorEvent vs ConversationErrorEvent in Events architecture

This PR adds a concise section to the Events architecture page documenting the two essential error events and their differences.

Changes

• sdk/arch/events.mdx: New section "Error Events: Agent vs Conversation" explaining:
  • AgentErrorEvent: LLM-convertible, tool-scoped (tool_name/tool_call_id), source="agent", conversation continues, sent to LLM as a tool message
  • ConversationErrorEvent: not LLM-convertible, conversation-level failure (no tool fields), typically source="environment", transitions run loop to ERROR and raises ConversationRunError; surface as top-level UI error

Notes

• The per-package API docs (sdk/api-reference/openhands.sdk.event.mdx) are auto-generated via scripts/generate-api-docs.py. Since ConversationErrorEvent is not re-exported in openhands.sdk.event.init, it doesn’t currently appear there. If we want it listed in that page, we can re-export it in the SDK and regenerate.

Co-authored-by: openhands openhands@all-hands.dev