fix: remove hardcoded voice:"default" that hangs Kokoro models

generateAudioViaProxy() hardcoded voice:"default" in the request body.
Kokoro models do not ship a voices/default.safetensors file, causing the
server to throw RuntimeError during streaming. Since the HTTP status was
already sent, the client never receives an error and hangs until timeout.

Use buildInjectedParams(cfg) instead, which omits voice and lets the
server use its own default. This aligns the direct-generate path with
the proxy path behavior.

Fixes #3

Author: cosformula

index.ts

@@ -5,7 +5,7 @@
  * Zero API key, zero cloud dependency.
  */
 
-import { resolveConfig, resolvePortBinding, type MlxAudioConfig } from "./src/config.js";
+import { resolveConfig, resolvePortBinding, buildInjectedParams, type MlxAudioConfig } from "./src/config.js";
 import { ProcessManager } from "./src/process-manager.js";
 import { TtsProxy } from "./src/proxy.js";
 import { HealthChecker } from "./src/health.js";
@@ -523,7 +523,7 @@ export default function register(api: PluginApi) {
         settled = true;
         resolve(result);
       };
-      const body = JSON.stringify({ model: cfg.model, input: text, voice: "default" });
+      const body = JSON.stringify({ ...buildInjectedParams(cfg), input: text });
       const req = http.request(
         {
           hostname: "127.0.0.1",

package-lock.json

@@ -1,6 +1,6 @@
 {
   "name": "@cosformula/openclaw-mlx-audio",
-  "version": "0.1.6",
+  "version": "0.1.7",
   "description": "OpenClaw local TTS plugin powered by mlx-audio — zero API key, zero cloud dependency",
   "type": "module",
   "main": "./dist/index.js",