operator runbooks

Intended Documentation

Control Center Operations

Monitor the authority runtime from the console control center and the supporting health, metrics, and live-decision surfaces — runtime pressure, decision throughput, connector health, and the flat watch CLI. No fictional health-check command.

intermediate4 min readimplemented

Control Center Operations#

Implemented capability

The control center is live in the console at /account/authority/control. It aggregates runtime pressure, graph health, anomalies, decision throughput, and connector health, and links directly to the intervention and emergency-control actions.

The control center is the primary surface for watching the authority runtime. This runbook covers what it shows, and the supporting health, metrics, and live-decision surfaces you can script against from outside the console.

No `intended health check` command

The Intended CLI is flat and hyphenated. There is no intended health check. Health and readiness are plain HTTP endpoints (/health, /ready); live decisions stream through the real intended watch command. Use those, or the console.

What the control center shows#

Policy status#

The active policy version per pack, the last deployment, and any drafts pending review. Deep-links into the policy lifecycle.

Decision throughput#

Live evaluation volume broken down by outcome. Note the two vocabularies: POST /intent returns APPROVED / ESCALATED / DENIED, while POST /authority/evaluate returns ALLOW / DENY / REQUIRE_APPROVAL / ESCALATE. The dashboard labels each surface with its own enum — don't conflate them.

Runtime health#

Component-level health: the policy engine, the token signer, audit-pipeline lag, and per-connector health for each enabled integration.

Anomalies and pressure#

Surfaced runtime pressure (load, latency) and flagged anomalies, with links straight to the intervention and emergency-control actions when you need to contain something.

Supporting surfaces outside the console#

The console is the primary surface. These endpoints exist for automation and external observability pipelines.

Health and readiness#

Both are public (no auth) and used by load balancers and uptime checks.

bash

# Liveness — the process is up
curl https://api.intended.so/health

# Readiness — dependencies (DB, etc.) are reachable
curl https://api.intended.so/ready

Metrics#

GET /metrics exposes runtime metrics for your monitoring pipeline.

bash

curl -H "Authorization: Bearer $INTENDED_API_KEY" \
  -H "x-tenant-id: tenant_acme_prod" \
  https://api.intended.so/metrics

Note

The exact /metrics payload (format and field set) is environment-dependent and evolving — treat the shape as subject to change and scrape it rather than hard-coding fields. The control center is the supported, stable view of the same signals.

Live decision stream#

The real-time equivalent of watching the throughput panel is the flat watch CLI, which connects to the decision SSE stream and prints decisions as they happen — color-coded by outcome.

$intended watch

Stream live authority decisions over SSE, color-coded by decision.

--tenantstring *

Tenant id (or configure tenantId / use --session).

--decisionstring

Filter: ALLOW | DENY | ESCALATE.

--domainstring

Filter by domain substring.

--actorstring

Filter by actor substring.

--min-riskstring

Only show decisions at or above this risk score.

bash

# Watch every high-risk denial as it happens
intended watch --tenant tenant_acme_prod --decision DENY --min-risk 60

text

INTENDED watch — streaming decisions for tenant tenant_acme_prod
Connecting to https://api.intended.so/api/stream/decisions ...
Filter: decision=DENY
Filter: min-risk=60

[10:32:01] DENY     dispatch release workflow      risk:82  agent:svc-ci-bot
[10:32:15] DENY     delete production table        risk:95  agent:svc-etl-bot

What to watch, and where it leads#

Signal	Where	If it spikes
`DENIED` / `DENY` rate	Throughput panel / `watch --decision DENY`	Suspect a recent policy deploy — cross-check Deploy and Rollback history, roll back if correlated.
`ESCALATED` rate	Throughput panel	Approval load rose; confirm it matches the impact summary you approved.
Connector unhealthy	Runtime health	The integration may be down or paused; check the connector and emergency-control state.
Audit-pipeline lag	Runtime health	Evidence is falling behind; treat as a containment-relevant signal during an incident.
Token-signer degraded	Runtime health	Token minting may stall; APPROVED intents could fail to produce a usable token.

Threshold alerting is API/console-driven

Configurable threshold alerts (latency, error-rate, deploy-failure, signer-degradation, audit-lag) are managed through the console and admin API rather than a CLI. Per-channel alert routing (Slack/PagerDuty) beyond what the console surfaces is Roadmap.

Next steps#

Incident Response — what to do when a signal here turns into an incident.
Operational Readiness — validate runtime health before go-live.
Deploy and Rollback — correlate anomalies with recent deploys.