diff --git a/.plans/speech-to-text-client.md b/.plans/speech-to-text-client.md
deleted file mode 100644
index 609ad69c086..00000000000
--- a/.plans/speech-to-text-client.md
+++ /dev/null
@@ -1,32 +0,0 @@
-# Speech input client plan
-
-## Assumptions
-
-- Backend exposes `POST /api/gateway/v1/audio/transcriptions` through the local CLI's Kilo routes.
-- Request body is JSON: `{ model, input_audio: { data, format }, language? }`.
-- Response body is JSON: `{ text, usage? }`.
-- Client MVP is request/response transcription after the user stops recording, not live partial transcription.
-- VS Code webviews cannot access `navigator.mediaDevices.getUserMedia`, so microphone capture runs in the extension host.
-
-## Client scope
-
-1. Add an experimental Speech to Text setting, default off.
-2. Add an extension-side STT handler that uses the existing Kilo auth stored in the CLI backend.
-3. Record microphone audio in the extension host with a bundled FFmpeg helper, falling back to system FFmpeg if needed.
-4. Add a shared webview hook for `idle -> recording -> transcribing -> error` state.
-5. Add a shared prompt action button that starts/stops recording and shows clean state.
-6. Wire the same hook/button into `PromptInput` and Agent Manager `NewWorktreeDialog`.
-7. Insert successful transcripts at the current caret, preserve existing prompt content, and never auto-submit.
-
-## Error states
-
-- Experiment disabled: hide the button.
-- Missing Kilo login or disabled Kilo provider: disable settings and hide the button.
-- Missing bundled helper or OS microphone denial: show a local recording error.
-- Empty recording or empty transcript: show a local error.
-- Gateway/network failure: show the backend error and keep input unchanged.
-
-## Follow-up
-
-- Replace the static model list with `/api/gateway/v1/transcription-models` discovery if we want automatic model updates.
-- Add microphone device selection if users need to choose a non-default input.