@chatpanel/gateway 0.6.24 → 0.6.25

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chatpanel/gateway",
-  "version": "0.6.24",
+  "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/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 = '0.6.25';
 
 // 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 },
       });
     }
 

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,