@chatpanel/gateway 0.6.23 → 0.6.25

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chatpanel/gateway",
-  "version": "0.6.23",
+  "version": "0.6.25",
   "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/diarize-engine.js

@@ -43,6 +43,7 @@ export function modelOnDisk(modelId = DIARIZE_MODEL, dtype = runtimeDtype()) {
   catch { return false; }
 }
 
+/** @param {{ log?: (m: string) => void, allowDownload?: boolean }} [opts] */
 async function load({ log = () => {}, allowDownload = true } = {}) {
   let lib;
   try { lib = await ensureLib(); }
@@ -79,12 +80,25 @@ async function load({ log = () => {}, allowDownload = true } = {}) {
 }
 
 export function init(cfg = {}) {
-  if (_initPromise) return _initPromise;
+  // Retry after a FAILED attempt (single-flight only while loading/ready) — else a
+  // transient failure (e.g. downloads were disabled, or the model wasn't on disk
+  // yet) would be cached forever and never self-heal once the model is present.
+  if (_initPromise && _state !== 'error') return _initPromise;
   _state = 'loading';
   _initPromise = load({ log: cfg.onLog, allowDownload: cfg.allowDownload !== false });
   return _initPromise;
 }
 
+// Explicit download (the Gateway-tab "Download" button): force the fetch even if
+// the config disabled auto-downloads, and clear any cached failure so it retries.
+/** @param {{ onLog?: (m: string) => void }} [opts] */
+export function download({ onLog } = {}) {
+  _initPromise = null;
+  _state = 'loading';
+  _initPromise = load({ log: onLog, allowDownload: true });
+  return _initPromise;
+}
+
 // Embed a mono 16 kHz Float32 segment → Float32Array x-vector, or null if not ready.
 export async function embed(audio) {
   if (!isReady() || !(audio instanceof Float32Array) || !audio.length) return null;
@@ -108,10 +122,13 @@ function cosine(a, b) {
   return denom ? d / denom : 0;
 }
 
-// A per-session speaker tracker. `threshold` = min cosine to be the SAME speaker;
-// lower → fewer speakers (merges more), higher → more speakers (splits more).
+// A per-session speaker tracker. NOT limited to 2 speakers — this is embedding +
+// online clustering, so it scales to N (a 10-person meeting is fine; accuracy just
+// softens as more voices sound alike). `threshold` = min cosine to be the SAME
+// speaker (lower → fewer speakers/more merging, higher → more speakers/more
+// splitting); `maxSpeakers` caps the roster (extra turns fold into the nearest).
 export class Diarizer {
-  constructor({ threshold = 0.75, maxSpeakers = 8 } = {}) {
+  constructor({ threshold = 0.75, maxSpeakers = 12 } = {}) {
     this.threshold = threshold;
     this.maxSpeakers = maxSpeakers;
     this.centroids = []; // { id, label, vec: Float32Array, n }

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 = '0.6.23';
+export const VERSION = '0.6.25';
 
 // 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 },
       });
     }
 
@@ -698,6 +700,24 @@ export function createGateway(cfg = loadConfig()) {
         return sendJson(res, 202, { accepted: true, active: id, state: sttEngine.state(), progress: sttEngine.progress() });
       }
     }
+    // Speaker (diarization) model manager — the "who said what" x-vector model.
+    // GET → status + install state; POST → force-download it (ignores the
+    // allowDownload config flag; explicit user action from the Gateway tab).
+    if (pathname === '/diarize/model') {
+      if (req.method === 'GET') {
+        const h = diarizeEngine.health();
+        return sendJson(res, 200, {
+          active: diarizeEngine.DIARIZE_MODEL,
+          state: h.state,
+          progress: diarizeEngine.progress(),
+          available: [{ id: diarizeEngine.DIARIZE_MODEL, label: 'Speaker embeddings (wavlm)', lang: 'any', approxMB: 100, note: 'Tells voices apart for meeting transcription (who said what).', installed: diarizeEngine.modelOnDisk() }],
+        });
+      }
+      if (req.method === 'POST') {
+        diarizeEngine.download({ onLog: (m) => console.log(m) });
+        return sendJson(res, 202, { accepted: true, active: diarizeEngine.DIARIZE_MODEL, state: diarizeEngine.state(), progress: diarizeEngine.progress() });
+      }
+    }
     if (pathname === '/stt/sessions' && req.method === 'POST') {
       if (cfg.stt?.enabled === false) return sendJson(res, 403, { error: { message: 'STT disabled in gateway config', type: 'stt_disabled' } });
       let body = null;

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,