Skip to content

VM Health

VM health is a live typed summary, not a raw SQL view. capsem-process maintains in-memory counters from accepted resolved security events. Persistent VMs seed/recompute from session.db once at load time; hot status reads do not scan SQLite.

CategoryExamples
Profileprofile_id, profile_revision, profile_status, package/asset pin state.
HTTP/DNS/MCPrequest counts, denied counts, MCP calls, DNS queries.
Modelprovider/model, model call count, input/output tokens, estimated cost.
File/processfile event count, process event count, exec count.
Securitytotal security events, enforcement decisions, blocks, detection findings, latest block, latest detection.

Host-owned AI calls can correlate with a VM/session/profile for explanation, but they charge host/service counters, not VM counters.

  • capsem status --json
  • capsem list and capsem info
  • gateway /status
  • service /info/{id}
  • Settings -> Policy and Sessions UI panels
  • future /metrics and OpenTelemetry exporters

Metrics use bounded labels: profile id, profile revision, event family, decision, provider, model, rule id where cardinality is controlled. Full local evidence stays in timeline/backtest/hunt/session APIs, not in metric labels.

Rate-limit and budget enforcement is reserved for S22. The bedrock release exposes the quota dimensions and counters needed for that later sprint; it does not claim budget enforcement.