Observe the system
Connect commands, workflows, attempts, runtimes, artifacts, and claims without logging secrets.
Outcome
Operators can trace one user command through durable execution and evidence while sensitive values and hidden reasoning remain absent from diagnostics.
Prerequisites
Choose approved telemetry destinations and retention. Define redaction tests before enabling an external sink.
Steps
- Propagate trace ID plus stable project, run, task, attempt, and execution IDs.
- Record command rejection reasons, workflow retries, stream lag, sandbox lifecycle, kernel output, artifact finalization, and budget variance.
- Export provider telemetry because provider dashboards are not canonical retention.
- Store reasoning summaries only when the provider explicitly supplies them; never hidden chain of thought.
- Test canary secrets against logs, errors, traces, prompts, and artifacts.
Slice 0 exposes stable trace and correlation fields in events but has no production telemetry claim.
Verify
Trace one deterministic run from command to claim and confirm every hop shares stable identity. Search the captured telemetry for the canary and local credential values; any match blocks activation.
Recover
On leakage, stop the sink, revoke exposed credentials, preserve access-controlled incident evidence, fix redaction at the source, and replay only sanitized diagnostics where policy permits.