Voice dictation mic button silently fails — audio captured but no transcript inserted (desktop app, Windows)

Summary

On the Windows desktop app, clicking the microphone button in the input toolbar appears to do nothing. Investigation via the app's own logs shows the click handler IS firing and audio capture starts (Web Audio ScriptProcessorNode activates), but no transcript is ever returned to the input box. Windows-side voice typing (Win+H) works fine into the same input, confirming mic hardware and OS permissions are healthy.

Environment

• OS: Windows 11 Enterprise 10.0.26200
• Claude Desktop app version: 1.7196.0 (MS Store / packaged)
• Bundled CCD version: 2.1.138
• Node: 24.15.0
• Auth: OAuth (signed in via Claude.ai account, not API key)

Permission state (all confirmed green)

| Scope | Value |
|---|---|
| HKCU mic ConsentStore | Allow |
| HKCU mic NonPackaged | Allow |
| HKLM mic ConsentStore | Allow |
| Per-app: Claude_pzs8sxrjxfjjc | Allow |

Evidence the click handler fires

logs/claude.ai-web.log shows a ScriptProcessorNode is deprecated warning at every mic-button click time:

2026-05-17 09:39:47 [warn] The ScriptProcessorNode is deprecated. Use AudioWorkletNode instead.
2026-05-17 09:40:05 [warn] The ScriptProcessorNode is deprecated. Use AudioWorkletNode instead.
2026-05-17 09:40:31 [warn] The ScriptProcessorNode is deprecated. Use AudioWorkletNode instead.
2026-05-17 09:44:18 [warn] The ScriptProcessorNode is deprecated. Use AudioWorkletNode instead.
2026-05-17 09:45:07 [warn] The ScriptProcessorNode is deprecated. Use AudioWorkletNode instead.
2026-05-17 09:48:45 [warn] The ScriptProcessorNode is deprecated. Use AudioWorkletNode instead.
2026-05-17 09:50:20 [warn] The ScriptProcessorNode is deprecated. Use AudioWorkletNode instead.
2026-05-17 09:51:54 [warn] The ScriptProcessorNode is deprecated. Use AudioWorkletNode instead.

No corresponding transcript ever appears in the input field, and no transcription error surfaces in the UI.

Possibly related warning

Also seen once during a mic interaction:

[warn] Blocked aria-hidden on an element because its descendant retained focus.
Element with focus: <div.tiptap ProseMirror ProseMirror-focused>
Ancestor with aria-hidden: <div#root>

If the input loses focus mid-dictation due to a modal/overlay setting aria-hidden on #root, the insertion target may be unreachable when the transcript returns.

Reproduction

1. Open Claude Desktop on Windows.
2. Click into the input box.
3. Click the microphone icon in the toolbar.
4. Speak.
5. Stop / wait.

Expected: transcript appears in the input box.
Actual: nothing appears. No error UI. The audio pipeline starts (per logs) but no text materializes.

Workaround

Win+H (Windows built-in voice typing) works correctly into the same input.