Agent remote-write

Three metrics live only in the agent process — they describe activity that doesn't appear in the Redis event stream:

• komputer_agent_steering_total — count of follow-up messages mid-task
• komputer_agent_mcp_connector_status — health of each MCP connector (Slack, GitHub, etc.)
• komputer_agent_subagent_wait_seconds — wall-clock time spent in wait_for_agents.py

When agents sleep (the common case), their /metrics endpoint disappears before Prometheus's next scrape. To capture these reliably, configure the agent to push them via Prometheus remote-write:

metrics:
  agent:
    remoteWrite:
      enabled: true
      url: http://prometheus.monitoring.svc.cluster.local:9090/api/v1/write
      intervalSeconds: 15
      # Optional bearer token (from a K8s Secret with key 'token'):
      bearerTokenSecret: prometheus-remote-write-token

The chart injects KOMPUTER_METRICS_REMOTE_WRITE_URL, KOMPUTER_METRICS_REMOTE_WRITE_INTERVAL, optionally KOMPUTER_METRICS_REMOTE_WRITE_TOKEN, and (when metrics.perAgentLabels=true) KOMPUTER_METRICS_PER_AGENT=true into the default KomputerAgentClusterTemplate, so every agent pod inherits them automatically.

Failures during flush are logged but never crash the agent.

Your remote-write target must have receiver enabled. For Prometheus: --web.enable-remote-write-receiver. For Mimir/Cortex/Thanos: native support.

For local kind dev, the sample KomputerAgentClusterTemplate already wires the agent to the local Prometheus from monitoring/docker-compose.yml. See local development and the local monitoring stack.