fix: deduplicate observations with same tool_call_id

[xingyaoww]

Summary

Fixes Anthropic API error: "each tool_use must have a single result. Found multiple tool_result blocks with id: <tool_call_id>"

Root Cause Analysis

Investigation of trajectory fcbbaec269bc4903ad4027c3bef0bc57 revealed a complex interaction:

What Happened

1. The agent ran a terminal command that started a nested agent-server (running examples/02_remote_agent_server/01_convo_with_local_agent_server.py)
2. The nested server shared the same persistence directory as the outer conversation
3. When the nested server started, it resumed the same conversation and found:
   • execution_status = RUNNING
   • An unmatched action (the terminal command that WAS STILL running in the outer context)
4. The nested server created an AgentErrorEvent for the "interrupted" action
5. The terminal command completed normally, creating an ObservationEvent
6. Both events had the same tool_call_id

Timeline from Events

09:48:05.911 - Event 61: ActionEvent (terminal command starts)
09:48:16.554 - Event 62: Nested server resumes, sets status=ERROR
09:48:16.588 - Event 63: AgentErrorEvent (nested server thinks tool crashed)
09:48:18.847 - Event 65: ObservationEvent (terminal command completes)

The log output in Event 65 shows:

"Resumed conversation fcbbaec2-69bc-4903-ad40-27c3bef0bc57 from persistent storage"

This confirms the nested server resumed the same conversation that was actively running.

Why This Causes the API Error

When the events are converted to LLM messages via events_to_messages(), both AgentErrorEvent and ObservationEvent get converted to tool_result blocks with the same ID. The Anthropic API rejects this:

"each tool_use must have a single result. Found multiple tool_result blocks with id: toolu_01EvREYhc5WD2xswPAvEc8ir"

Solution

Added a _deduplicate_observations() method to the View class that:

1. Identifies the best observation for each tool_call_id based on priority:
   • ObservationEvent (highest - contains actual tool result)
   • UserRejectObservation (user action)
   • AgentErrorEvent (lowest - error notification)
2. Emits only one observation per tool_call_id, preserving event order

This is a defensive fix at the View level. The event_service cannot prevent this because:

• At the time of creating the AgentErrorEvent, the observation hasn't arrived yet (it comes ~2 seconds later)
• The nested server has no way to know the original tool is still running

Alternative Considerations

Other potential fixes (not implemented, for future consideration):

• Documentation: Warn that nested servers should use different persistence directories
• Architecture: Use process-level locking to prevent multiple servers from writing to the same conversation
• State: Include AgentErrorEvent in get_unmatched_actions() to prevent duplicate error events

Testing

• Added 6 new tests in test_view_duplicate_observations.py
• All 87 existing view tests continue to pass
• Pre-commit hooks pass

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
openhands-sdk/openhands/sdk/context
view.py	222	5	97%	227, 263, 416, 419, 514
TOTAL	18256	5560	69%