Card guide

Independent agent monitor

agent.monitor watches recorded conditions and schedules reviews without using a model to poll the laser.

Configure review triggers

The monitor can route a review to a registered task card or an available resident. It records its schedule, conditions, outstanding work and generation independently of the agent conversation. The deterministic laser lock remains a separate control process.

id: health-monitor
operation: agent.monitor
uses: []
parameters:
  agent_card_id: gateway-agent
  task_card_id: diagnosis-task
  interval_s: 3600
  lookback_s: 3600
  sample_interval_s: 2
  conditions: []
  enabled: true
  background:
    enabled: false

gateway-agent is the example gateway card ID. Use the ID configured in your laser deck; the UI shows its role as Gateway agent. Display names do not change routing IDs.

Parameter	Meaning
`interval_s`	Periodic review interval. Zero disables the periodic trigger; configured conditions may still trigger reviews.
`sample_interval_s`	Recorded-state polling interval. This is not the frequency of model calls.
`conditions`	Metric rules with an ID, threshold, clearing threshold, persistence, cooldown and maximum observation age. Repeated timestamps do not count as new evidence.
`task_card_id`	Registered task to run for reviews. If empty, `agent_card_id` selects resident review.
`history_task_card_id`, `history_interval_s`, `history_min_new_samples`	Optional historical-review card, interval and minimum accumulated new valid recorded metric samples.
`background`	Disabled by default. Enabling requires explicit approved span, point, lifetime and handoff bounds plus an adjustment task.

Each condition supplies id, metric, above, clear_below, persistence_s, cooldown_s and max_age_s. An optional nonzero baseline selects absolute fractional change from that baseline. Use the laser’s declared metric meanings and units; historical examples are not current thresholds.

Review, notification and intervention are separate

A condition may justify diagnosis without a physical change or a Teams post. Notification policy decides which future events are reported. Unchanged healthy reviews remain local; actionable failures and human decisions are reported promptly. The monitor does not replace an owner watchdog.

One pending task or resident request is tracked at a time. Unresolved or unknown work is not replayed, and missed review intervals do not accumulate into a burst of model calls. Stale or missing recorded data is not proof of a stopped lock.

Turning the monitor off or stopping its card prevents new background trials and revokes its adjustment grant. An admitted adjustment uses the owner’s bounded restoration path; foreground work waits for verified restoration when required. A model response alone cannot declare the handoff safe.

Read agent-monitor-status.json in the monitor job for current state, outstanding request/task, recorded conditions and restoration status. Inspect owner receipts independently when a physical action occurred.