Runtime Metrics

GET /runtime/metrics returns operational metrics computed fresh on every request from existing session/decision/policy state — nothing is cached or separately persisted. Every field is traceable back to a real session, decision, or policy record.

Scope

Scoped to Runtime Sessions created via the v0.2+ API and to the resolved tenant (see Security for what tenant scoping does and doesn't guarantee today). A decision created via the raw legacy /runtime/requests path (no session) is not included.

Filters

Query param	Effect
agent_id	Narrow to one agent
task_type	Narrow to one workflow
from / to	ISO 8601 timestamp range on started_at

X-Tenant-ID always scopes the result — it is not a query parameter.

Response shape (summary)

• summary — aggregate counts: sessions by state, human reviews required, approved/rejected/retried, risk distribution, evidence failures, policy hits, Agent Run counts (total_agent_runs, internal_agent_runs, external_agent_runs, average_runs_per_session, retry_rate)
• agents — per-agent breakdown, including zero-session registered agents
• workflows — per-task-type breakdown
• policies — per-policy hit/block/review/pause/flag counts
• recent_interventions — the most recent 10–20 meaningful interventions (human review, block, unsupported claim, medication intervention, high risk, pause, retry, rejection), newest first
• top_agents_by_volume, top_model_providers, top_model_versions

SDK usage

const metrics = await runtime.getMetrics({ agent_id: "acme-radiology-copilot" });
console.log(metrics.summary.human_reviews_required);
console.log(metrics.recent_interventions);

Known limitation

Because metrics are computed on read rather than pre-aggregated, very large session volumes will make this endpoint slower over time. That tradeoff is intentional for v1 — see the root README's production roadmap for the pre-aggregation plan.