/autolab - Autolab Workflow Operator (Claude)

Use this skill when running Autolab with agent_runner.runner: claude.

Fast Triage

1. •autolab status
2. •autolab verify --stage <stage>
3. •autolab run

Claude-Specific Defaults

• •Keep agent_runner.runner: claude in .autolab/verifier_policy.yaml.
• •Use explicit timeout settings (agent_runner.timeout_seconds).
• •Use iteration_plus_core edit scope for normal work, iteration_only for strict isolation.
• •Do not enable claude_dangerously_skip_permissions unless execution is fully trusted/non-interactive.

Stage Flow

hypothesis -> design -> implementation -> implementation_review -> launch -> slurm_monitor -> extract_results -> update_docs -> decide_repeat

Use autolab explain <stage> to inspect active verifier requirements.

Common Tasks

Stuck Stage

1. •autolab status
2. •autolab verify --stage <stage>
3. •Fix artifacts/policy mismatches.
4. •autolab run

SLURM Issues

1. •Validate runs/<run_id>/run_manifest.json.
2. •Validate ledger entry with autolab slurm-job-list verify --manifest <manifest> --doc docs/slurm_job_list.md.
3. •Re-run stage verification.

Assistant/Automation Loop Issues

1. •Inspect guardrail counters in autolab status.
2. •Check autorun.guardrails and autorun.meaningful_change policy blocks.
3. •Escalate to human_review if retries churn without progress.

Policy Misconfiguration

1. •Run autolab configure --check.
2. •Verify python_bin, dry_run_command, and requirements_by_stage.
3. •Keep requirements as a subset of stage capabilities in .autolab/workflow.yaml.

Safe Defaults

• •Never force stage transitions by editing .autolab/state.json.
• •Keep docs/todo.md as Markdown.
• •Keep guardrails enabled for unattended runs.
• •Prefer bounded loops (--max-iterations, --max-hours).