Premortem treats an audit as a governed pipeline, not a single LLM call, with persisted artifacts at every stage.
Who this is forAnyone who needs to understand run states, checkpoints, and review gates.
Audit runs
- runStatus: queued → running → paused | completed | failed | cancelled.
- Checkpoints stored in summary.checkpoint: phase, completedSpecialists, counts.
- Pause/resume is cooperative: orchestrator saves state between pipeline stages.
Agent swarm
- Specialist agents run sequentially with per-agent persistence.
- finding_synthesizer_agent merges clusters into issue candidates.
- issue_validator_agent enforces schema before review queue admission.
Artifacts
- Findings → dedupe clusters → issue candidates → published issues.
- Graph snapshots capture repo/CI topology for cross-agent context.
- Lineage links prompt → finding → cluster → candidate → GitLab issue.
Was this article helpful?
