feat: transcribe dictation into the channel composer

.lane/memory/apps/app/src/composer.tsx/01M4YQG7BVWEFZPD56A2BA0VE9-mode-ace-mode-channels-id-an.md

@@ -4,9 +4,9 @@ anchor: function Composer
 created: 2026-10-15T02:50:44Z
 norm: '1'
 sig: 6e0e6c5c8afbd01a
-body_hash: 192a2a69ff8a183f
-raw_hash: 4c35effb9b4cdea0
-vouched: 2026-10-16T22:11:30Z
+body_hash: 3ecd8d58953b3ffd
+raw_hash: 8d76c6254cce3b7d
+vouched: 2026-10-16T22:22:52Z
 lines: 99-282
 ---
 

.lane/memory/apps/app/src/composer.tsx/01M4ZSEZBJJ3AMEBTR7SH36FW4-models-are-read-only-while-t.md

@@ -4,9 +4,9 @@ anchor: function Composer
 created: 2026-10-15T12:44:15Z
 norm: '1'
 sig: 6e0e6c5c8afbd01a
-body_hash: 192a2a69ff8a183f
-raw_hash: 4c35effb9b4cdea0
-vouched: 2026-10-16T22:11:30Z
+body_hash: 3ecd8d58953b3ffd
+raw_hash: 8d76c6254cce3b7d
+vouched: 2026-10-16T22:22:52Z
 lines: 101-290
 ---
 

.lane/memory/apps/app/src/composer.tsx/01M4ZYQVGB6C87WBEV12WPQ8PF-a-withdrawn-message-returns.md

@@ -4,9 +4,9 @@ anchor: function Composer
 created: 2026-10-15T14:16:28Z
 norm: '1'
 sig: 6e0e6c5c8afbd01a
-body_hash: 192a2a69ff8a183f
-raw_hash: 4c35effb9b4cdea0
-vouched: 2026-10-16T22:11:30Z
+body_hash: 3ecd8d58953b3ffd
+raw_hash: 8d76c6254cce3b7d
+vouched: 2026-10-16T22:22:52Z
 lines: 104-326
 ---
 

.lane/memory/apps/desktop/electrobun.config.ts/01M53CZ2S58V1DTJKB64V1SVRD-wkwebview-dictation-needs-ns.md

@@ -0,0 +1,12 @@
+---
+id: 01M53CZ2S58V1DTJKB64V1SVRD
+anchor: '@file'
+created: 2026-10-16T22:22:49Z
+norm: '1'
+sig: eb0482a0ddee93a6
+body_hash: de216f5680cfb1c1
+raw_hash: 6c0b5799010be95c
+lines: 1-42
+---
+
+WKWebView dictation needs NSMicrophoneUsageDescription and NSSpeechRecognitionUsageDescription in Info.plist, or macOS ends the app on first use

.lane/memory/packages/ui/src/hooks/use-dictation.ts/01M53CZ2RRXDYDPZWZ8AKNMFH0-deactivating-aborts-and-deta.md

@@ -0,0 +1,12 @@
+---
+id: 01M53CZ2RRXDYDPZWZ8AKNMFH0
+anchor: function useDictation
+created: 2026-10-16T22:22:49Z
+norm: '1'
+sig: de94e6fe0d9626d6
+body_hash: 78e3734a05619b3c
+raw_hash: 85cbb253122dfd32
+lines: 28-51
+---
+
+deactivating aborts and detaches handlers, so no result lands after Stop or Send moves the editor on

apps/app/src/composer.tsx

@@ -293,6 +293,7 @@ export function Composer(
 				accessory={accessory}
 				scope={scope}
 				tools={!phone}
+				mic
 				modes={blocked ? BLOCKED_MODES : MODES}
 				mode={mode}
 				onModeChange={changeMode}

apps/desktop/electrobun.config.ts

@@ -26,6 +26,14 @@ export default {
 			codesign: false,
 			notarize: false,
 			createDmg: false,
+			// Electrobun writes these usage strings into Info.plist; macOS ends the app when the composer's
+			// dictation asks for the microphone or speech recognition without them. Signing uses
+			// native/entitlements.plist, which carries the hardened-runtime audio-input entitlement.
+			entitlements: {
+				"com.apple.security.device.audio-input": "Ace listens while you dictate a message.",
+				"com.apple.security.personal-information.speech-recognition":
+					"Ace turns your dictation into message text.",
+			},
 		},
 	},
 	release: { generatePatch: false },

apps/desktop/native/entitlements.plist

@@ -4,4 +4,5 @@
 	<key>com.apple.security.cs.allow-jit</key><true/>
 	<key>com.apple.security.cs.allow-unsigned-executable-memory</key><true/>
 	<key>com.apple.security.cs.disable-library-validation</key><true/>
+	<key>com.apple.security.device.audio-input</key><true/>
 </dict></plist>

packages/ui/src/components/composer/chat-composer.tsx

@@ -18,6 +18,7 @@ import type { MarkType, Node as ProseMirrorNode, NodeType } from "prosemirror-mo
 import { TextSelection } from "prosemirror-state";
 import type { EditorView } from "prosemirror-view";
 
+import { canDictate, useDictation } from "../../hooks/use-dictation";
 import { useDraft } from "../../hooks/use-draft";
 import { useHistory } from "../../hooks/use-history";
 import { useMic } from "../../hooks/use-mic";
@@ -177,7 +178,7 @@ type Props = {
 	agents?: Agent[];
 	/** Overrides the active mode's placeholder. */
 	placeholder?: string;
-	/** Enables the microphone + recording UI. Defaults to false. */
+	/** Enables dictation where the runtime can transcribe speech. Defaults to false. */
 	mic?: boolean;
 	mode?: string;
 	onModeChange?: (mode: string) => void;
@@ -598,7 +599,7 @@ function useComposer({
 	models,
 	agents,
 	placeholder,
-	mic = false,
+	mic: dictation = false,
 	mode,
 	onModeChange,
 	model,
@@ -673,7 +674,15 @@ function useComposer({
 	let draft = useDraft(scope);
 	let restoredMode = useRef(draft.defaultValue ? curMode : undefined);
 	let history = useHistory(scope);
+	let mic = dictation && canDictate;
 	let analyser = useMic(mic && recording);
+	let dictated = useRef({ from: 0, size: 0 });
+	useDictation(mic && recording, (text) => {
+		let v = view.current!;
+		let { from, size } = dictated.current;
+		v.dispatch(v.state.tr.insertText(text, from, from + size));
+		dictated.current.size = text.length;
+	}, () => setRecording(false));
 
 	let view = useRef<EditorView | null>(null);
 	let file = useRef<HTMLInputElement | null>(null);
@@ -957,9 +966,15 @@ function useComposer({
 		v.focus();
 	}
 
+	function record() {
+		dictated.current = { from: view.current!.state.selection.from, size: 0 };
+		setRecording(true);
+	}
+
 	function send() {
 		let v = view.current;
 		if (!v) return;
+		setRecording(false);
 		let text = v.state.doc.textContent;
 		let sentMode = modeRef.current;
 		let mode = latest.current.modes.find(item => item.id === sentMode);
@@ -1183,6 +1198,7 @@ function useComposer({
 		empty,
 		filled,
 		planning,
+		record,
 		setRecording,
 		cancel,
 		link,
@@ -1635,6 +1651,7 @@ function ComposerSend({ state }: { state: ComposerState }) {
 		submitBusy,
 		submitLabel,
 		planning,
+		record,
 		setRecording,
 		send,
 		sendable,
@@ -1650,9 +1667,7 @@ function ComposerSend({ state }: { state: ComposerState }) {
 			busyLabel={submitLabel}
 			plan={planning}
 			className="relative z-10"
-			onRecord={mic
-				? () => setRecording(true)
-				: undefined}
+			onRecord={mic ? record : undefined}
 			onStop={mic ? () => setRecording(false) : undefined}
 			onSend={send}
 			disabled={!sendable && (!mic || busy || !canSend)}

packages/ui/src/hooks/use-dictation.ts

@@ -0,0 +1,53 @@
+import { useEffect, useRef } from "react";
+
+type Recognition = {
+	continuous: boolean;
+	interimResults: boolean;
+	lang: string;
+	onresult: ((event: { results: ArrayLike<ArrayLike<{ transcript: string }>> }) => void) | null;
+	onend: (() => void) | null;
+	start(): void;
+	abort(): void;
+};
+
+// WebKit (Safari and the desktop app's WKWebView) and Chromium ship only the prefixed constructor.
+const Speech = typeof window === "undefined"
+	? undefined
+	: (window as unknown as { webkitSpeechRecognition?: new() => Recognition })
+		.webkitSpeechRecognition;
+
+/** Whether this runtime can transcribe speech. */
+const canDictate = Boolean(Speech);
+
+/**
+ * Transcribe the microphone while `active`. `onText` receives the whole session's transcript,
+ * interim words included, each time it changes; `onEnd` runs when recognition stops on its own
+ * (silence, error, or a denied permission). Deactivating aborts, so the last interim words stand
+ * and no late result lands after the caller has moved on.
+ */
+function useDictation(active: boolean, onText: (text: string) => void, onEnd: () => void) {
+	let handlers = useRef({ onText, onEnd });
+	handlers.current = { onText, onEnd };
+
+	useEffect(() => {
+		if (!active || !Speech) return;
+		let recognition = new Speech();
+		recognition.continuous = true;
+		recognition.interimResults = true;
+		recognition.lang = navigator.language;
+		recognition.onresult = ({ results }) => {
+			let text = "";
+			for (let i = 0; i < results.length; i++) text += results[i]![0]!.transcript;
+			handlers.current.onText(text);
+		};
+		recognition.onend = () => handlers.current.onEnd();
+		recognition.start();
+		return () => {
+			recognition.onresult = null;
+			recognition.onend = null;
+			recognition.abort();
+		};
+	}, [active]);
+}
+
+export { canDictate, useDictation };