@genclass/runtime 0.2.1

• Telemetry: no page text by default. telemetry.include.situation now defaults to false: decision events
  carry the trigger, diagnosis, confidence, action, gate, outcome, latency and counts, but not the situation text
  the model read. Opt in with GenClass.init({ telemetry: { include: { situation: true } } }); TELEMETRY.md now
  says exactly what that text can contain (request names with query strings, route changes, the page title, short
  error messages, short typed values, short store values). The console notice changed accordingly.
• Telemetry stays off on local and private hosts (localhost, *.localhost, *.local, *.test,
  *.internal, loopback, private IPv4 ranges, link-local, file:); runtime.telemetry.reason is local. An
  explicit telemetry: true / { ... } still sends from them.
• Fix: the first salient request after the model loaded was never held in guard or heal mode. The hold
  budget and the expected model latency used the first warm-up pass, which on WebGPU includes pipeline compilation
  and reads as seconds, so the first decision was taken in the background and could not act. They now use the
  warm pass timed after compilation (status.latency), so a duplicate submit right after Model ready can be
  coalesced.
• Fix: the "Model ready" console line repeated (every 5 s, with each latency notification). Printed once per
  transition into ready.
• Docs: Zustand's devtools middleware is off in production builds unless enabled: true, so production Zustand
  stores are not discovered without it; ?genclass=guard|heal only raise the mode within limits.

From the compatibility matrix (compat/RESULTS.md)

Seven framework apps (React + Vite, Next.js 16, Vue + Pinia, SvelteKit 3, Angular 22, Solid, plain HTML) and 15 data
layers, every mode, 10 seeds: no bug introduced; observe identical to no GenClass except one seed of an app race (SWR
optimistic rollback) that GenClass's timing tipped to the correct outcome. No situation text or other model-visible
text changes; no telemetry change.

Install: npm i @genclass/runtime@0.2.1 or <script src="https://cdn.jsdelivr.net/npm/@genclass/runtime@0.2.1"></script>. Model: @genclass/runtime-model@0.2.0. Published from CI with npm provenance.

@genclass/runtime 0.1.0-beta.1 + model 0.2.0

npm install @genclass/runtime

• Model 0.2.0 (10 MB, local WebGPU/WASM): gain gate with fitted cautious / balanced / eager profiles. 0% false interventions on held-out real apps in every profile; balanced catches 7.5% (guard) / 10.4% (heal) of real-app problems, eager 21% / 24%. Full tables: docs/runtime/RESULTS.md.
• aggressiveness option (named or 0–1), plus 19 new options: per-route modes, protected/ignored endpoints, auto-downgrade breaker, shadow mode, onBeforeAction veto, action limits, sinks, session summary, model-loading controls (docs/runtime/OPTIONS-SPEC.md).
• One-command install: npx genclass-runtime init (or npx @genclass/runtime init).
• Verified from the registry: guard fixes an out-of-order typeahead 6/6; clean typing makes 0 model calls.

@genclass/runtime 0.1.0-alpha.0

First alpha of @genclass/runtime, a client-side AI runtime that runs inside your web app. It records what the app is doing (user actions, async operations, network requests, state changes, errors, timing and causality), and a small local GenClass model decides whether to step in.

npm install @genclass/runtime

import { GenClass } from "@genclass/runtime";
GenClass.init();

In this alpha

• Observers: fetch, XHR, WebSocket, DOM user actions, errors, navigation, storage, long tasks and timers, with causal links across await.
• State: stores through atoms, guard, and the React, Redux and Zustand adapters. Writes are versioned per field and can be held.
• Learning: invariants (e.g. total == sum(price·qty)), per-operation transition profiles, and latency, error and rate baselines.
• Triage: generic facts decide which situations go to the model; normal traffic never does.
• Modes and actions: observe, guard (default) and heal, with the summed-probability policy gate and the built-in actions.
• Observability: console reports, explain(id), undo, and the kill switch ?genclass=off.
• Model host: a Web Worker running ONNX Runtime on WebGPU or WASM, with Cache Storage and sha256 verification, idle or lazy loading, and the genclass-runtime fetch-model CLI.
• Devtools overlay: @genclass/runtime/devtools.

Not in this alpha. The runtime-specialist model is still training. Until @genclass/runtime-model is published, the runtime observes and records but does not act yet: every decision takes the passive action. Remaining work is listed in OPEN_TASKS.md.

Docs: package README · API · Architecture