Workflow Execution Troubleshooting

Classify failures from evidence:

Class	Signal
Environment	Backend CLI crash, missing auth, or non-zero exit before useful agent output.
Baseline repair	Initial test, lint, or verify gate is red and the generated artifact is repairing it before main work.
Gate failure	A test, lint, or build command failed and the flow is repairing.
Non-convergence	A loop or fix loop hit stuck, timeout, or ceiling.
Crash	Partial state or commits exist but the process died mid-flow.
Served child	orcats serve stays alive while one child firing fails.

Recovery steps:

1. Re-run the backend doctor for environment failures.
2. Let baseline repair and gate failures continue until they converge or hit a guard.
3. Re-run after crashes when persistent plans or state stores are present.
4. Reproduce served-child failures with ORCA_LOOP_EVENT='...' orcats run <loop>.

Do not retry a dirty-worktree rejection with --baseline=accept-dirty unless the operator explicitly asks. That mode writes a dirty baseline snapshot before backend repair work.

Stop and ask before destructive git operations.