@convai/embody 0.4.0 → 0.4.1

package/CHANGELOG.md

@@ -1,5 +1,11 @@
 # @convai/embody
 
+## 0.4.1
+
+### Patch Changes
+
+- 62deba9: `useWidgetClient` hides the server's `user-transcription` echo of a typed message, so `<ConvaiWidget/>` shows each sent message once instead of twice.
+
 ## 0.4.0
 
 ### Minor Changes

package/dist/react/useWidgetClient.d.ts

@@ -6,6 +6,31 @@ export type WidgetClient = ConvaiClient & {
     isVideoEnabled: boolean;
     isScreenShareActive: boolean;
 };
+type ChatMessage = ConvaiClient['chatMessages'][number];
+/**
+ * How long after a typed message its server echo can still arrive. Measured
+ * at ~2 s on staging; the window is generous because a slow turn is still an
+ * echo, while the same words said aloud half a minute later are a new message.
+ */
+export declare const TYPED_ECHO_WINDOW_MS = 30000;
+/**
+ * The transcript without the server's echo of typed messages.
+ *
+ * `sendUserTextMessage` appends the text locally as `user-llm-text`, and the
+ * SDK drops the server's `user-llm-text` echo so the message appears once.
+ * But the server ALSO answers typed text with a final `user-transcription`
+ * carrying the same words — with no `logicalTurnId` to correlate on — which
+ * the SDK stores as a new message and `<ConvaiWidget/>` renders as a second
+ * "You" bubble. Every typed message showed twice; every reply once. Verified
+ * on web-sdk 1.8.0 and 1.9.0-beta.3, which filter identically.
+ *
+ * A transcription is an echo when an unmatched typed message with the same
+ * trimmed text precedes it within the window; each typed message cancels at
+ * most one. Spoken turns never produce a `user-llm-text`, so a real
+ * transcription is never dropped. An echo that streams in chunks can show
+ * its partial text until the final chunk matches — a flicker, not a duplicate.
+ */
+export declare function withoutTypedEchoes(messages: readonly ChatMessage[]): ChatMessage[];
 /**
  * ```tsx
  * const { client } = useCharacter({ apiKey, characterId })
@@ -16,3 +41,4 @@ export type WidgetClient = ConvaiClient & {
  * Returns `null` until the character's connection exists.
  */
 export declare function useWidgetClient(client: ConvaiClient | null): WidgetClient | null;
+export {};

package/dist/react/useWidgetClient.js

@@ -12,6 +12,47 @@
  * for activity, and the three control emitters for the toggles.
  */
 import { useEffect, useMemo, useState } from 'react';
+/**
+ * How long after a typed message its server echo can still arrive. Measured
+ * at ~2 s on staging; the window is generous because a slow turn is still an
+ * echo, while the same words said aloud half a minute later are a new message.
+ */
+export const TYPED_ECHO_WINDOW_MS = 30_000;
+/**
+ * The transcript without the server's echo of typed messages.
+ *
+ * `sendUserTextMessage` appends the text locally as `user-llm-text`, and the
+ * SDK drops the server's `user-llm-text` echo so the message appears once.
+ * But the server ALSO answers typed text with a final `user-transcription`
+ * carrying the same words — with no `logicalTurnId` to correlate on — which
+ * the SDK stores as a new message and `<ConvaiWidget/>` renders as a second
+ * "You" bubble. Every typed message showed twice; every reply once. Verified
+ * on web-sdk 1.8.0 and 1.9.0-beta.3, which filter identically.
+ *
+ * A transcription is an echo when an unmatched typed message with the same
+ * trimmed text precedes it within the window; each typed message cancels at
+ * most one. Spoken turns never produce a `user-llm-text`, so a real
+ * transcription is never dropped. An echo that streams in chunks can show
+ * its partial text until the final chunk matches — a flicker, not a duplicate.
+ */
+export function withoutTypedEchoes(messages) {
+    const pending = [];
+    return messages.filter((m) => {
+        const content = m.content.trim();
+        const at = Date.parse(m.timestamp);
+        if (m.type === 'user-llm-text') {
+            pending.push({ content, at });
+            return true;
+        }
+        if (m.type !== 'user-transcription')
+            return true;
+        const i = pending.findIndex((t) => t.content === content && at - t.at >= 0 && at - t.at <= TYPED_ECHO_WINDOW_MS);
+        if (i === -1)
+            return true;
+        pending.splice(i, 1);
+        return false;
+    });
+}
 /** Priority order matches the SDK's: speaking > listening > thinking > connected. */
 function activityOf(state) {
     if (state.isSpeaking)
@@ -89,6 +130,9 @@ export function useWidgetClient(client) {
         };
         return new Proxy(client, {
             get(target, prop, receiver) {
+                // The widget renders straight from this; see withoutTypedEchoes.
+                if (prop === 'chatMessages')
+                    return withoutTypedEchoes(Reflect.get(target, prop, target));
                 if (typeof prop === 'string' && prop in overrides)
                     return overrides[prop];
                 // Receiver is the TARGET, not the proxy, so getters and methods run

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@convai/embody",
-  "version": "0.4.0",
+  "version": "0.4.1",
   "description": "Give a Convai character a body: resolve 3D characters by characterId and drive them with lipsync and animation. Layered on @convai/web-sdk.",
   "license": "Apache-2.0",
   "author": "Convai <support@convai.com>",