Files
boshu2__agentops/docs/INCIDENT-RUNBOOK.md
2026-07-15 00:21:02 -04:00

38 lines
1.2 KiB
Markdown

# AgentOps Incident Runbook
## Preserve evidence
Copy the relevant PlanPacket, CandidatePacket, manifest, command output, and
verdict before changing the subject. Do not treat an incomplete or corrupt
artifact as PASS.
## Classify the boundary
- `FAIL`: a concrete acceptance defect was proven.
- `NOT_PROVEN`: identity, freshness, scope coverage, evidence, or persistence
was insufficient.
- deterministic command failure: the repository check failed independently of
semantic validation.
- installation drift: generated projections or installed skill links differ
from their source metadata.
AgentOps reports and stops. The operator or calling system chooses recovery,
revision, rollback, Git, CI, and release actions.
## Repository diagnostics
```bash
python3 scripts/check-cathedral-cut-conformance.py
scripts/regen-all.sh --check
cd cli && go test ./...
```
Run only the command needed to reproduce the observed failure before expanding
to the full deterministic suite.
## Escalation
Escalate to the operator when authority, credentials, safety judgment, or an
external system change is required. A failed check or reviewer disagreement is
evidence to report, not an AgentOps-controlled retry state.