Monitoring and runbook boundaries

LMP records evidence for a decision. It is not currently a hosted monitoring service and does not provide a production alerting backend.

What is recorded locally

An evaluation artifact can retain:

• package identity, version, digest, and signature status;
• evaluator version and mode;
• workspace identity and Git state;
• selected and excluded scope;
• findings, severities, rationale, remediation, and limitations;
• skipped checks and parser boundaries;
• remediation loop transitions;
• privacy metadata and command-execution results.

The artifact is the primary investigation record. Store it with the review or CI run and keep the exact profile and source revision alongside it.

Daemon health

For a supervised local daemon, use its PID-file lifecycle commands:

lmpd --pid-file /tmp/lmpd.pid --status
lmpd --pid-file /tmp/lmpd.pid --reload
lmpd --pid-file /tmp/lmpd.pid --stop

--reload requests a fresh evaluation. --stop requests clean shutdown and PID-file cleanup. A process that cannot report status or remove its PID file is an operational failure, not a healthy idle state.

The resource gate measures bounded idle RSS and CPU for the daemon. It is a qualification check, not a long-running production service-level objective.

Failure triage

1. Preserve the artifact and command that produced the failure.
2. Classify the result as pass, needs_revision, blocked, or evaluation_error.
3. For blocked, repair package integrity, scope, runtime, or execution prerequisites before interpreting policy findings.
4. For needs_revision, inspect the rule ID, selected file, evidence, and remediation before changing the code or policy.
5. For evaluation_error, preserve the runtime error and input artifact, repair the evaluator boundary, and rerun before interpreting any findings.
6. Re-run with the same package digest and deliberate scope change, then compare the new artifact with the previous one.

Transport errors, malformed MCP requests, and unavailable processes are boundary errors. They are not evidence that the source violated a policy rule.

Production monitoring gap

The repository does not currently ship hosted metrics, alert routing, retention policies, or automatic rollback. A production operator must provide those systems and record their relationship to LMP artifacts before claiming continuous monitoring or operational observability.

The existing qualification and benchmark reports are strong local evidence, but they do not replace uptime monitoring, alert tests, incident records, or long-running workload measurements.