When using claude-code, each agent runs inside a fresh git worktree
it has never previously visited. Claude shows an interactive trust
prompt on first entry to any new directory and blocks waiting for a
keypress. Without --dangerously-skip-permissions, every spawned agent
silently hangs at this prompt. The dashboard shows all sessions stuck
at "spawning" indefinitely and no work is ever performed.
Root cause: handleAutoMode() in init.ts built the config object without
the agentConfig.permissions field, so the claude-code plugin defaulted
to "default" (no flag) on every spawn.
Fix:
- Add agentConfig.permissions: "skip" to the defaults block emitted by
ao init --auto, with an inline comment explaining why it is required
- Update README Configuration example to show the field and explain
the trust-dialog failure mode
- Add the scenario to the Troubleshooting quick-reference so users
who hit it after manual config edits can find the fix immediately
The fix is safe: AO creates worktrees from the user's own repository,
so the implicit trust granted by --dangerously-skip-permissions is
equivalent to the user manually accepting the prompt for every session.
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
1cf11b6d9f