Worker sub-agents fail to start: 'Agent did not signal readiness within 10.0s' and empty Claude transcript

[joshalbrecht]

Summary

When using create_working_agent.sh (which wraps mng create --env ROLE=working --type worker) to spawn a worker sub-agent, the command consistently fails with:

imbue.mng.errors.AgentStartError: Failed to start agent <name>: Agent did not signal readiness within 10.0s. This may indicate a trust dialog appeared or Claude Code failed to start.

The agent process is created and reaches WAITING state with waiting_reason: END_OF_TURN, but the Claude Code transcript is completely empty — meaning Claude never actually ran.

Steps to Reproduce

1. From a running claude-mind agent (mind type), run create_working_agent.sh <name> <message-file>
2. mng create is invoked with --env ROLE=working --type worker --message-file ...
3. The command times out after 10s with the AgentStartError
4. Inspecting the created agent: state is WAITING, plugin.claude.waiting_reason = "END_OF_TURN", but logs/claude_transcript/events.jsonl is 0 bytes

Expected Behavior

The worker agent should start Claude Code, process the initial message, and signal readiness. The transcript should contain the agent's work.

Actual Behavior

• Claude Code never executes (empty transcript)
• Agent reaches WAITING/END_OF_TURN without doing any work
• mng create times out waiting for the readiness signal

Environment

• Host: huge-bumblebee-of-fruition (local provider)
• Mind agent: product_manager (agent-771572cfe4a04fc58fe8a52044a93c9f), type claude-mind, state RUNNING
• Failed worker agents:
  • setup-target-repo (agent-48f256e6b051489682d62b64b965b24d), type worker, now archived/STOPPED
  • setup-github-repo (agent-f00e4801b41d4606b52ee31947575a81), type worker, state WAITING
• Date: 2026-03-27 ~00:59–01:00 UTC

Additional Context

This occurred during first-run onboarding when attempting to delegate the "set up target GitHub repo" task. Both attempts with different agent names failed identically. The mind agent itself is running fine; only worker sub-agents fail to start.

The error message suggests "a trust dialog appeared" — it's possible Claude Code is prompting for permission approval when launched as a sub-agent, which would block the readiness signal. The --dangerously-skip-permissions flag is present in the command, but may not be taking effect in this context.