@mushi-mushi/wasm-classifier

Optional on-device junk filter for bug reports. It stops empty or nonsense reports (“hi”, “asdf”) in the browser, before they are sent. The default backend is a heuristic with no dependencies; an optional ONNX backend runs a model you host, through onnxruntime-web.

pnpm add @mushi-mushi/wasm-classifier
# Optional, only for the ONNX backend:
pnpm add onnxruntime-web

Use it with the web SDK

The web SDK does not load the classifier on its own. Pass it in preFilter.wasmClassifier:

import { Mushi } from 'mushi-mushi'
import { createHeuristicClassifier } from '@mushi-mushi/wasm-classifier'
 
Mushi.init({
  projectId: 'YOUR_PROJECT_ID',
  apiKey: 'mushi_...',
  preFilter: { wasmClassifier: createHeuristicClassifier() },
})

It runs after the SDK’s built-in pattern filter. A block verdict stops the report on the device; pass and unsure send it, and the server does the full classification.

Call it directly

import { createHeuristicClassifier } from '@mushi-mushi/wasm-classifier'
 
const classifier = createHeuristicClassifier({ blockThreshold: 0.2, passThreshold: 0.55 })
const result = await classifier.classify({ description: 'asdf qwer' })
// → { verdict: 'pass' | 'block' | 'unsure', confidence, reason, modelId: 'heuristic-v1', durationMs }

classify() also takes the signals the web SDK passes: category, url, hasScreenshot, hasSelectedElement, hasNetworkErrors, hasConsoleErrors and proactiveTrigger.

ONNX backend (experimental)

import { createOnnxClassifier } from '@mushi-mushi/wasm-classifier'
 
const classifier = await createOnnxClassifier({
  modelUrl: 'https://cdn.your-app.com/mushi/model.onnx',
  cacheKey: 'model-v1',
  preload: true,
  classifyTimeoutMs: 750,
})

If onnxruntime-web is missing, the model fetch fails, or a call exceeds classifyTimeoutMs, it falls back to the heuristic backend. The package does not ship a model.