Summary
When a working agent finishes its task, the Claude Code stop hook fires but fails with:
Stop hook error: Failed with non-blocking status code: No stderr output
As a result, no mng/agent_states completion event is emitted to the parent mind, so the mind never learns the agent finished.
Steps to reproduce
- Create a working agent via
mng create <name> worker --message-file ...
- Agent completes its task successfully (writes output files, prints summary)
- Observe stop hook error in
mng capture <agent> output
- Parent mind never receives a completion event for the agent
Expected behavior
Stop hook runs successfully → completion event emitted → parent mind receives mng/agent_states event with the agent's final state.
Actual behavior
Stop hook fails silently. Parent mind never notified. User had to manually inform the mind that the agent was done.
Environment
- Agent type: worker (claude-mind harness)
- Observed on:
explore-slack-summarize agent (agent-3dc672424d774aa0b751dfb3b67514a9)
- The agent itself completed successfully — output files were written, final summary was printed
- Error appears at the very end of the agent's tmux output after "✻ Brewed for 4m 52s"