Observe

v2.0.0 adds the OBSERVE primitive: a local-first observability layer. It evaluates alert rules, renders pipelines, and compares skill versions — all in-process, no network. Alerts route through the v1.2 EventBus; a host delivers them.

Alerting rules

checkAlerts(metrics, rules) fires the rules whose metric comparison holds, carrying the observed value and a severity.

import { checkAlerts } from "skillweave";

checkAlerts({ pass_rate: 0.4 }, [
  { id: "low-pass", metric: "pass_rate", op: "<", threshold: 0.5, severity: "alert" },
]);
// → [{ id: "low-pass", metric: "pass_rate", value: 0.4, op: "<", threshold: 0.5, severity: "alert" }]

Operators: > · >= · < · <= · == · !=. Severity defaults to warning. A rule whose metric is absent never fires. Route the fired alerts through an EventBus for delivery.

Pipeline visualiser

visualise(pipeline, { format }) renders a loaded pipeline as an ASCII flow (default) or a Mermaid flowchart. The CLI wraps it as skillweave visualise.

import { visualise } from "skillweave";
visualise(pipeline);                      // ASCII: trigger → steps + events
visualise(pipeline, { format: "mermaid" }); // Mermaid flowchart

A/B skill testing

abTest(scoreA, scoreB) compares two skill versions by their boundary-judge score — the higher wins (equal is a tie), with the delta.

import { abTest } from "skillweave";
abTest(0.91, 0.88);   // { winner: "a", a: 0.91, b: 0.88, delta: 0.03 }

Deferred

The hosted observability dashboard and live alert/webhook delivery are a host responsibility (network), and orchestrator-level DAG auto-execution and published performance benchmarks are planned follow-ups — kept out of the runtime to preserve the offline-by-default, zero-shell posture.