Live telemetry

Search-to-Fill Rate· skil.ski

% of AI agent requests that successfully find, bind, and execute a tool without human intervention. The operator's pricing power metric.

Fill Rate (24h)

0.00%

Invocations (24h)

P50 latency

—

P95 latency

—

No failure data in the last 24 hours. Telemetry starts accumulating as agent requests flow through the MCP endpoint.

How we count

A request is counted as successful when the MCP tools/call completes without error: the skill slug is resolved from the tool registry, the execution path reaches a response body, and no auth or schema fault is emitted. Federated stub results (tools routed to external origins) are counted separately with their upstream outcome.

Latency is measured as wall-clock time from the moment the POST body is fully parsed to the moment the JSON-RPC response is emitted. This includes DB lookup, skill body load, and any upstream proxy time. Rows without a latency value (legacy pre-migration records) are excluded from percentile computation.

Failure modes are only counted from rows that carry the new outcomecolumn (added in migration 036). Pre-migration rows are counted in totals but not broken down by failure mode. Target fill rate is ≥99% — the threshold at which agent-side confidence is high enough to price and commit to SLA tiers.

Machine-readable endpoint

curl skill.ski/api/registry/fill-rate | jq

Supports ?window=24h (default), ?window=7d, ?window=30d. CORS-open. Cached 5 minutes.