fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets - #1710

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

…#1710)

## Problem

Desktop dictation fails with "Could not prepare voice transcription." on the release `t3` runtime.
That runtime is a Node 26 SEA: undici 8 defaults `allowH2` on, so the global `WebSocket` negotiates RFC 8441 WebSocket over HTTP/2 with `api.meta.ai`.
undici 8.10.2 then fails Meta's valid h2 response because `Sec-WebSocket-Accept` is missing (close 1006); the undici fix, `nodejs/undici#5774`, closed unmerged.

Dev and packaged-desktop backends run Electron's Node 24 (undici 7, HTTP/1.1), so they never reproduced it.

## Change

The realtime route provides `NodeSocket.layerWebSocketConstructorWS` in place of the global constructor.
The `ws` package is already bundled and stays on HTTP/1.1.

```diff
-        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructor), Effect.ignoreCause);
+          // `ws` stays on HTTP/1.1. Node 26's undici negotiates WebSocket over HTTP/2 with Meta, then
+          // fails the valid RFC 8441 response for lacking Sec-WebSocket-Accept.
+        ).pipe(Effect.provide(NodeSocket.layerWebSocketConstructorWS), Effect.ignoreCause);
```

## Scope and approval

Maintainer reported the regression, tested the fix live in T3 Code, and authorized this PR and merge.
One-line transport swap inside the fork voice-input route.

## Verification

- [x] Maintainer dictated successfully against a Node 26 dev server with this fix.
- [x] Node 26 dev server: `/api/fork/voice-input/realtime` upgraded 101, and Meta acknowledged the session.
- [x] Direct Meta probe: 2s of PCM plus `endStream` returned a final transcript and closed 1000.
- [x] `vp test run apps/server/src/voice-input`: 45/45 passed.
- [x] `vp run typecheck` (apps/server): exit 0.
- [x] `vp run fork:ci`: exit 0.

Follow-up, not in this PR: the client maps every realtime failure, including Meta's own error message, to the generic "Could not prepare" text.

Claude Opus 5.5 via Claude Code in T3 Code.

## Fork trailers

Fork-Domain: voice-input
Fork-Tier: bugfix
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>

Squashes:

- 26171fbd8fa feat(server): scaffold external voice dictation (#1672) (#1672)
- da9e3d289f0 feat(server): transcribe recordings with Meta Muse (#1673) (#1673)
- c7b4f781cf5 feat(desktop): record dictation into editable chat drafts (#1675) (#1675)
- 9d5dcb319c9 fix(desktop): keep dictation transcribing across thread switches (#1684) (#1684)
- 7782b1f05b3 feat(desktop): stream live Meta dictation into the composer (#1686) (#1686)
- 80071d08816 fix(voice-input): keep Meta realtime dictation on HTTP/1.1 WebSockets (#1710) (#1710)
- 7edffba3c2b feat(voice-input): live dictation preview for any OpenAI Realtime endpoint (#1775) (#1775)
- 2283697771d fix(voice-input): use InlineButton for processing error dismissal (#1800) (#1800)
- 54c87d908a2 fix(voice-input): one base URL derives dictation upload and live paths (#1881) (#1881)
- bd64d598282 feat(voice-input): send per-project speech terms to the dictation service (#1882) (#1882)

Fork-Domain: voice-input
Fork-Tier: core
Fork-Upstreamable: no
Co-authored-by: donjor <[email protected]>
Co-authored-by: donjor <[email protected]>