Production Guide

Goal

Move from prototype to stable service behavior.

P0 safeguards

1. Budget limits configured (steps, runtime, tokens).
2. Stop reasons monitored.
3. Trace persisted for every production run.
4. Failure report retained for debugging.

Suggested deployment checklist

1. Pin model + parser versions.
2. Add regression taskset (small but representative).
3. Run daily/PR regression with threshold gates.
4. Add runtime dashboards from trace summaries.

Incident debugging playbook

1. Locate failing run_id.
2. Read manifest.summary.stop_reason.
3. Find first ok=false event in events.jsonl.
4. Inspect corresponding step in steps.jsonl.
5. Classify root cause:
6. parser robustness
7. tool input schema
8. env capability/config
9. model output drift

Guardrails worth adding early

1. Strict parser mode for critical actions.
2. Tool allowlist per agent.
3. Max retries per action.
4. Safety filters for shell/web tools.

Source Index

• qitos/engine/stop_criteria.py
• qitos/engine/recovery.py
• qitos/engine/hooks.py
• qitos/trace/schema.py
• qitos/qita/cli.py