@chatpanel/gateway 0.6.24 → 0.6.26

package/README.md

@@ -54,6 +54,15 @@ chatpanel-gateway
 #     backend  : bridge (agent: codex, via http://127.0.0.1:4319)
 ```
 
+> **Binary vs. npm — same features, very different local-AI speed.** Both run
+> identical redaction/routing. But the standalone binary runs the local models
+> (speech-to-text, diarization, NER) on the **WASM** runtime — **fp32-only,
+> single-threaded, slow**. The **npm** install uses the **native** runtime with
+> **quantized (q8)** models — in our tests **~10× faster** speech-to-text
+> (real-time even on larger, more accurate models). If you'll use voice/meeting
+> features, install via **npm**. Don't run both — they can shadow each other on
+> `PATH`; check `GET /health` → `stt.runtime` (`native` vs `wasm`).
+
 Then point your front-end agent at it. **opencode** (`opencode.json`):
 
 ```jsonc

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chatpanel/gateway",
-  "version": "0.6.24",
+  "version": "0.6.26",
   "description": "Local privacy gateway — redacts PII out of OpenAI/Anthropic API traffic before it reaches a model, then restores it in the reply. Point opencode, codex, aider, Claude Code, etc. at it.",
   "type": "module",
   "bin": {

package/src/ner-engine.js

@@ -179,6 +179,7 @@ export async function ensureLib() {
 // (Re)load a specific model into _pipe. Downloads it first if missing (and allowed).
 // Reusable by init() and setModel(). Fail-open: on error, state='error', _pipe stays
 // whatever it was (so a failed SWITCH doesn't kill a working detector).
+/** @param {string} modelId @param {{ log?: (m: string) => void, allowDownload?: boolean }} [opts] */
 async function loadModel(modelId, { log = () => {}, allowDownload = true } = {}) {
   const prevPipe = _pipe;
   const prevModel = _model;
@@ -211,18 +212,26 @@ async function loadModel(modelId, { log = () => {}, allowDownload = true } = {})
   if (!haveLocal) { _progress = { model: modelId, file: null, pct: 0 }; log(`[ner] downloading model ${modelId} (one-time${isCustom ? ', from Hugging Face' : ''})…`); }
 
   try {
-    const pipe = await lib.pipeline('token-classification', modelId, {
-      dtype: 'q8',
-      progress_callback: (p) => {
-        if (!p) return;
-        const pct = typeof p.progress === 'number' ? Math.round(p.progress) : (_progress?.pct ?? 0);
-        if (p.status === 'progress' || p.status === 'download' || p.status === 'initiate') {
-          _progress = { model: modelId, file: p.file || _progress?.file || null, pct };
-        } else if (p.status === 'done' && p.file) {
-          log(`[ner] fetched ${p.file}`);
-        }
-      },
-    });
+    const progress_callback = (p) => {
+      if (!p) return;
+      const pct = typeof p.progress === 'number' ? Math.round(p.progress) : (_progress?.pct ?? 0);
+      if (p.status === 'progress' || p.status === 'download' || p.status === 'initiate') {
+        _progress = { model: modelId, file: p.file || _progress?.file || null, pct };
+      } else if (p.status === 'done' && p.file) {
+        log(`[ner] fetched ${p.file}`);
+      }
+    };
+    let pipe;
+    try {
+      pipe = await lib.pipeline('token-classification', modelId, { dtype: 'q8', progress_callback });
+    } catch (e) {
+      // Mirror gap → fall back to Hugging Face (same as stt-engine).
+      if (haveLocal || isCustom || !allowDownload) throw e;
+      log(`[ner] mirror fetch failed (${String(e.message).slice(0, 80)}) — retrying from Hugging Face…`);
+      lib.env.remoteHost = 'https://huggingface.co/';
+      lib.env.allowRemoteModels = true;
+      pipe = await lib.pipeline('token-classification', modelId, { dtype: 'q8', progress_callback });
+    }
     // Swap in the new pipeline, dispose the old one (free its WASM/native session).
     _pipe = pipe; _model = modelId; _state = 'ready'; _err = null; _progress = null;
     if (prevPipe && prevPipe !== pipe) { try { await prevPipe.dispose?.(); } catch { /* ignore */ } }

package/src/server.js

@@ -43,7 +43,7 @@ import * as openai from './openai.js';
 import * as responses from './responses.js';
 import * as anthropic from './anthropic.js';
 
-export const VERSION = '0.6.24';
+export const VERSION = '0.6.26';
 
 // WARM search tier — SQLite + FTS5 record store (falls back to an encrypted-JSON
 // store if SQLite can't load), fed by the extension's ingest sync + backup-ingest.
@@ -499,7 +499,9 @@ export function createGateway(cfg = loadConfig()) {
       const stt = sttEngine.health();
       return sendJson(res, 200, {
         ok: true, version: VERSION, backend: cfg.backend, tier: cfg.redaction.tier,
-        stt: { enabled: cfg.stt?.enabled !== false, state: stt.state, ready: stt.ok, model: stt.model || cfg.stt?.model || DEFAULT_STT_MODEL },
+        // `runtime` = 'native' (npm, fast quantized) | 'wasm' (binary, slow fp32) —
+        // the extension uses it to advise the far-faster native gateway.
+        stt: { enabled: cfg.stt?.enabled !== false, state: stt.state, ready: stt.ok, model: stt.model || cfg.stt?.model || DEFAULT_STT_MODEL, runtime: stt.runtime, dtype: stt.dtype },
       });
     }
 

package/src/stt-engine.js

@@ -39,6 +39,14 @@ export function runtimeDtype() {
   return globalThis.__CHATPANEL_WASM_PATHS__ ? 'fp32' : 'q8';
 }
 
+// Which ONNX runtime is active: the standalone binary uses onnxruntime-web WASM
+// (fp32-only, single-thread — ~10× slower); the npm package uses onnxruntime-node
+// (native, quantized q8). The extension surfaces this so users on the slow WASM
+// build know the native gateway is far faster.
+export function runtimeName() {
+  return globalThis.__CHATPANEL_WASM_PATHS__ ? 'wasm' : 'native';
+}
+
 let _state = 'off';        // 'off' | 'loading' | 'downloading' | 'ready' | 'error'
 let _model = null;         // active model id
 let _pipe = null;          // the loaded automatic-speech-recognition pipeline
@@ -72,7 +80,7 @@ export function isReady() { return _state === 'ready' && !!_pipe; }
 export function progress() { return _progress; }
 
 export function health() {
-  return { configured: _state !== 'off', ok: isReady(), state: _state, model: _model, dtype: _dtype, error: _err };
+  return { configured: _state !== 'off', ok: isReady(), state: _state, model: _model, dtype: _dtype, runtime: runtimeName(), error: _err };
 }
 
 // (Re)load a model into _pipe. Same contract as ner-engine.loadModel: fail-open,
@@ -111,18 +119,28 @@ async function loadModel(modelId, { log = () => {}, allowDownload = true } = {})
   try {
     // Per-model override wins (catalog `dtype`), else the runtime-appropriate default.
     const dtype = sttModelDtype(modelId) || runtimeDtype();
-    const pipe = await lib.pipeline('automatic-speech-recognition', modelId, {
-      dtype,
-      progress_callback: (p) => {
-        if (!p) return;
-        const pct = typeof p.progress === 'number' ? Math.round(p.progress) : (_progress?.pct ?? 0);
-        if (p.status === 'progress' || p.status === 'download' || p.status === 'initiate') {
-          _progress = { model: modelId, file: p.file || _progress?.file || null, pct };
-        } else if (p.status === 'done' && p.file) {
-          log(`[stt] fetched ${p.file}`);
-        }
-      },
-    });
+    const progress_callback = (p) => {
+      if (!p) return;
+      const pct = typeof p.progress === 'number' ? Math.round(p.progress) : (_progress?.pct ?? 0);
+      if (p.status === 'progress' || p.status === 'download' || p.status === 'initiate') {
+        _progress = { model: modelId, file: p.file || _progress?.file || null, pct };
+      } else if (p.status === 'done' && p.file) {
+        log(`[stt] fetched ${p.file}`);
+      }
+    };
+    let pipe;
+    try {
+      pipe = await lib.pipeline('automatic-speech-recognition', modelId, { dtype, progress_callback });
+    } catch (e) {
+      // The dl.chatpanel.net mirror only proxies an allowlist; a catalog model
+      // missing from it (or any mirror hiccup) 403s. Fall back to Hugging Face so
+      // a download is never blocked by a mirror gap. (Custom ids already use HF.)
+      if (haveLocal || isCustom || !allowDownload) throw e;
+      log(`[stt] mirror fetch failed (${String(e.message).slice(0, 80)}) — retrying from Hugging Face…`);
+      lib.env.remoteHost = 'https://huggingface.co/';
+      lib.env.allowRemoteModels = true;
+      pipe = await lib.pipeline('automatic-speech-recognition', modelId, { dtype, progress_callback });
+    }
     _pipe = pipe; _model = modelId; _state = 'ready'; _err = null; _progress = null; _dtype = dtype;
     if (prevPipe && prevPipe !== pipe) { try { await prevPipe.dispose?.(); } catch { /* ignore */ } }
     log(`[stt] ready — model ${modelId} @ ${dtype} (in-process, offline) — local dictation active`);