Composer 2 outputs foreign language with completely unrelated code

Where does the bug appear (feature/product)?

Cursor IDE

Describe the Bug

I chatted with Composer 2 (and “fast” variant as well) in the same chat sessions then switched to plan mode and asked Composer 2 to write a plan, and it gave an output in a strange non-English language (our project is all in English and always was and will be). the it mixed up very weird code inside the plan! looks like a leak from someone’s else training data:

# Takumi チャット / Zoom 連携 — 実施プラン

## 目的
- **8766 の本番 UI**で Takumi の応答・音声を固める。
- 問題なければ **Zoom(RTMS・BlackHole・ビジュアライザ)**に載せる。
- 今回入れた修正が **意図どおり動く**ことを確認する。

---

## フェーズ A — ローカルチャット(必須)

1. **サーバー**
   - `TAKUMI_API_PORT=8766` で `npm run chat`(または `.env` に固定)。
   - `curl http://127.0.0.1:8766/health` → 200。

2. **ブラウザ**
   - **Chrome** で `http://localhost:8766/` を開く(Web Speech の切り分け用)。
   - ハードリロード(Cmd+Shift+R)。

3. **マイク**
   - マイク ON → OS のマイク許可。
   - ブロック時はアラート文言が出ることを確認(`getUserMedia` / `onerror`)。

4. **Tandem(任意・別検証)**
   - `npm run open:tandem-chat`(`--fresh` でキャッシュバスト)。
   - Tandem で音声が効かない場合は **Chrome 正常 = Tandem 側の Web Speech 制限**として記録(仕様として許容するか判断)。

5. **パフォーマンス**
   - ドラッグ中、DevTools Performance で **毎フレームの React 再レンダーが過剰でない**こと(`scheduleFlexUpdate` の効果)。
   - マウスアップ後、`flexValues` が最終値に追いついていること。

6. **`onResizeEnd`**
   - 親から **毎レンダー新しいコールバック**を渡すテスト用コンポーネントで、マウスアップ時に **最新のコールバック**が呼ばれることを確認(ref 修正の回帰防止)。

---

## フェーズ B — klive-zoom / ビジュアライザ(Zoom 前)

1. **`npm run klive:zoom`**(8770)。
2. **`npm run klive:zoom:open-visualizer`** または Tandem で `http://localhost:8770/zoom/visualizer`。
3. **Test (play sample)** で波形・TTS・表示の確認。
4. **ピクセル min が両パネルで足りない**設定で、コンソールに **相互不満足警告**が出るか(開発時のみ想定)。

---

## フェーズ C — Zoom 本番相当(RTMS)

1. **ngrok**: `ngrok http 8770`、Zoom アプリの webhook URL を登録。
2. **会議**: ホストで RTMS 開始。
3. **マイク**: Zoom → BlackHole(または Aggregate)をドキュメントどおりに設定。
4. **ビデオ**: ビジュアライザタブ共有 or OBS 経由(既存手順)。
5. **会話**: 発話 → 文字起こし → Takumi 応答 → TTS が Zoom に乗るか。
6. **catch-up / 制約**: 境界で止まったあと、マウスが「戻る」挙動と **flex-grow 制約**が矛盾しないか(catch-up を `calculateFlexValues` に寄せた部分)。

---

## フェーズ D — 仕上げ

1. **ドキュメント**
   - `HOW-TO-TALK-WITH-TAKUMI` か `KLIVE-ZOOM-SETUP` に以下を 1 行ずつ追記してよい:
     - `npm run open:tandem-chat`
     - `TAKUMI_API_PORT=8766` の推奨
     - Tandem で音声が無い場合は Chrome で試す、など。

2. **Git**
   - 変更ファイルをコミット(ルール: amend しない)。

3. **(任意)テスト**
   - `usePanelResize` / チャット UI の E2E が無ければ、最低限 **手動チェックリスト**を `analysis_temp/` か issue に残す。

---

## リスク・判断ポイント

| 項目 | 対応 |
|------|------|
| Tandem で Speech 非対応 | チャットは Chrome 推奨を明記 |
| ピクセル min 同時充足不可 | 警告のみ・UI は「後勝ち」前提でよいかプロダクト判断 |
| `flexValues` と DOM のずれ | マウスアップ直後に `flushPendingFlex` が走るか再確認 |

---

## 完了の定義

- フェーズ A の 1–5 が通る。
- フェーズ B でビジュアライザ + サンプル TTS が問題ない。
- フェーズ C を実施する場合は、少なくとも **1 回**エンドツーエンドで聞こえ・表示が揃う。

Steps to Reproduce

I don’t know but this is a major major bug

Expected Behavior

English with RELEAVNT code to the task

Screenshots / Screen Recordings

Operating System

Windows 10/11

Version Information

Version: 2.6.20 (user setup)
VSCode Version: 1.105.1
Commit: b29eb4ee5f9f6d1cb2afbc09070198d3ea6ad760
Date: 2026-03-17T01:50:02.404Z
Build Type: Stable
Release Track: Early Access
Electron: 39.8.1
Chromium: 142.0.7444.265
Node.js: 22.22.1
V8: 14.2.231.22-electron.0
OS: Windows_NT x64 10.0.26200

For AI issues: which model did you use?

Composer 2 & Fast mixed, but for the final prompt just Composer 2

For AI issues: add Request ID with privacy disabled

ebd3cc36-0672-41a3-a661-992c8c23e4c8

Does this stop you from using Cursor

No - Cursor works, but with this issue

Hey, thanks for the detailed report and the request ID, we can confirm this on our end.

This is a known issue with the Composer models where they sometimes output CJK text with unrelated content. Quick note though, this is not a leak of another user’s data. The model is hallucinating and generating plausible looking but completely fictional content like project names, commands, and configs. None of it comes from another user’s actual codebase.

The team is aware and tracking this. If it happens again, starting a new chat session should help. Switching to a non-Composer model like Claude or GPT also works as a workaround.

Let me know if you hit this again, fresh request IDs help us track the scope.

1 Like

This topic was automatically closed 22 days after the last reply. New replies are no longer allowed.