diff --git a/grok-voice/skills/add-dictation/SKILL.md b/grok-voice/skills/add-dictation/SKILL.md
index b651704d1..7ebe85233 100644
--- a/grok-voice/skills/add-dictation/SKILL.md
+++ b/grok-voice/skills/add-dictation/SKILL.md
@@ -71,12 +71,13 @@ export async function transcribe(blob: Blob, filename: string) {
 ```
 
 ```ts
-// client
+// client (stream is a dedicated microphone stream owned by this recording)
 const mime = MediaRecorder.isTypeSupported("audio/webm;codecs=opus") ? "audio/webm;codecs=opus" : "audio/mp4";
 const rec = new MediaRecorder(stream, { mimeType: mime });
 const parts: BlobPart[] = [];
 rec.ondataavailable = (e) => parts.push(e.data);
 rec.onstop = async () => {
+  stream.getTracks().forEach((track) => track.stop()); // MediaRecorder.stop() does not release the microphone
   const fd = new FormData();
   fd.append("file", new Blob(parts, { type: mime }), "dictation");
   const { text } = await (await fetch("/api/dictation", { method: "POST", body: fd })).json();
diff --git a/grok-voice/tests/dictation-capture.test.cjs b/grok-voice/tests/dictation-capture.test.cjs
new file mode 100644
index 000000000..d76df84a5
--- /dev/null
+++ b/grok-voice/tests/dictation-capture.test.cjs
@@ -0,0 +1,49 @@
+const assert = require('node:assert/strict');
+const fs = require('node:fs');
+const path = require('node:path');
+const { stripTypeScriptTypes } = require('node:module');
+const test = require('node:test');
+const vm = require('node:vm');
+
+function setup() {
+  const skill = fs.readFileSync(path.join(__dirname, '../skills/add-dictation/SKILL.md'), 'utf8');
+  const snippet = skill.match(/```ts\n\/\/ client[^\n]*\n([\s\S]*?)\n```/)[1];
+  const tracks = [{ live: true, stop() { this.live = false; } }, { live: true, stop() { this.live = false; } }];
+  let recorder, resolveRequest, rejectRequest, body;
+  const inserted = [];
+  class Recorder {
+    static isTypeSupported() { return true; }
+    constructor(stream) { this.stream = stream; recorder = this; }
+    start() {}
+  }
+  const context = vm.createContext({
+    MediaRecorder: Recorder, stream: { getTracks: () => tracks }, Blob, FormData,
+    fetch(url, options) { body = options.body; return new Promise((resolve, reject) => { resolveRequest = resolve; rejectRequest = reject; }); },
+    insertAtCursor(text) { inserted.push(text); },
+  });
+  vm.runInContext(stripTypeScriptTypes(snippet), context);
+  recorder.ondataavailable({ data: new Blob(['recorded audio']) });
+  return { tracks, recorder, inserted, body: () => body,
+    resolve() { resolveRequest(new Response(JSON.stringify({ text: 'hello' }), { status: 200 })); },
+    reject() { rejectRequest(new Error('offline')); },
+  };
+}
+
+test('recording stop releases every microphone track before upload completes', async () => {
+  const s = setup(); const pending = s.recorder.onstop();
+  assert.equal(s.tracks.filter((track) => track.live).length, 0);
+  s.resolve(); await pending;
+});
+
+test('a failed upload leaves no microphone capture running', async () => {
+  const s = setup(); const pending = s.recorder.onstop();
+  s.reject(); await assert.rejects(pending, /offline/);
+  assert.ok(s.tracks.every((track) => !track.live));
+});
+
+test('releasing capture preserves the recorded blob and successful transcription', async () => {
+  const s = setup(); const pending = s.recorder.onstop();
+  assert.equal(await s.body().get('file').text(), 'recorded audio');
+  s.resolve(); await pending;
+  assert.deepEqual(s.inserted, ['hello']);
+});