ai-notify 0.2.1 → 0.4.0

package/README.ja.md

@@ -0,0 +1,140 @@
+# ai-notify
+
+[English](README.md) · **日本語**
+
+**ターミナルのAIエージェントが「あなたを必要とした瞬間」を逃さない** — Claude Code・Codex などのエージェントがターンを終えた／入力を求めた瞬間に、音・読み上げ・デスクトップ通知で知らせます。**全エージェント・全ターミナルを1つのスイッチで一括ミュート**。デーモンも常駐プロセスも無し。
+
+![ai-notify demo](https://raw.githubusercontent.com/unoryota/ai-notify/main/assets/hero-ja.gif)
+
+```sh
+brew install unoryota/tap/ai-notify   # macOS（Homebrew）
+# または:  npm i -g ai-notify
+
+ai-notify init        # インストール済みのエージェントを自動検出して配線
+```
+
+## 何が違うか
+
+エージェントは数分間も沈黙しがち。ai-notify は適切な瞬間に呼び戻します。とくに **AIを並列でたくさん動かす運用**のために作られています：
+
+- 🎙️ **ターミナルごとに声を変えられる。** ペインごとに別の声を割り当てれば、**どの窓が終わったか**を聞くだけで分かります — `export AI_NOTIFY_VOICE=Eddy`（または [VOICEVOX](#-voicevox-キャラクターボイス) のキャラ）。
+- 🌐 **あなたの言語で読み上げ。** エージェントの英語の返信・プロンプトを翻訳してから読み上げ／表示（キーレス・無料）。日本人にうれしい。
+- 📝 **「何をしたか」を教える。** 完了通知は「終わりました」だけでなく、エージェントの最後の返信（transcript）から作業内容を要約します。
+- 🔕 **1スイッチで全部ミュート。** 全エージェント・全ターミナルが同じフラグを読むので、会議中はワンタップで全部静かに。
+- 🔔 **ネイティブのメニューバーも内蔵。** `ai-notify menubar install` — Hammerspoon/SwiftBar 不要。
+
+## 対応エージェント
+
+| エージェント | 状態 | 配線方法 |
+| ----- | ------ | -------------- |
+| Claude Code | ✅ | `~/.claude/settings.json` の `Notification` + `Stop` フック |
+| Codex CLI | ✅ | `~/.codex/config.toml` の `notify`（`agent-turn-complete`） |
+| Gemini CLI | 🧪 検出のみ・フック作業中 | PR歓迎 |
+
+別のエージェント（aider, opencode, amp …）の追加は小さなPRで可能：`src/providers/` にファイルを1つ置くだけ。[CONTRIBUTING](CONTRIBUTING.md) 参照。
+
+## コマンド
+
+```sh
+ai-notify init [--dry-run] [--only claude,codex]   # 検出したエージェントを配線
+ai-notify toggle | on | off | status               # ミュートスイッチ
+ai-notify volume [0.0-2.0]                          # 音量の取得／設定
+ai-notify voice [number|name|preview|default]      # 読み上げ音声を選ぶ
+ai-notify voicevox [on <id>|off|speakers|test]     # VOICEVOXの声で読み上げ
+ai-notify tsundere [on|off|level <0-1>|test]       # ツンデレ口調（緊急度でツン⇄デレ）
+ai-notify translate [on <lang>|off|test]           # エージェントの文章を自分の言語で
+ai-notify menubar [install|uninstall|status]       # ネイティブのメニューバー（macOS）
+ai-notify doctor                                   # 依存・配線の確認
+ai-notify uninstall                                # 配線をきれいに削除
+```
+
+ペイン別の上書き — エージェントを起動する**前**に、その端末で `export` する：
+
+```sh
+AI_NOTIFY_LABEL=api               # この窓の読み上げ／通知での名前
+AI_NOTIFY_VOICE=Eddy              # この窓の `say` 音声
+AI_NOTIFY_VOICEVOX_SPEAKER=3      # この窓の VOICEVOX 話者ID
+AI_NOTIFY_VOLUME=0.5              # この窓の音量（0.0〜2.0）
+AI_NOTIFY_TSUNDERE_LEVEL=0.8     # この窓のツンデレ既定値（0=デレ〜1=ツン）
+```
+
+## 🎛️ ネイティブのメニューバー — ミュート・音量・声
+
+エージェントが走っているターミナルにはコマンドを打てないので、**メニューバー**から全部操作します：
+
+```sh
+ai-notify menubar install   # ネイティブのメニューバーアプリ・ログイン時に自動起動
+```
+
+モノクロの波形アイコンが**状態を色で**表します（Adobe風）：通常はシルエットのみ、入力待ちがあると**黄ドット**、ミュート中は**赤＋斜線**。
+
+- **左クリック** → メニュー：**音量スライダー**、**ツンデレ**トグル＋デレ⇄ツンスライダー、**声の一覧**（システム＋VOICEVOX）、**ペイン別**設定（開いている各ターミナルに個別の声と音量）。
+- **右クリック** → 即ミュート切替。
+
+第三者アプリ不要。別の方法が好みなら、**Hammerspoon**・**SwiftBar/xbar**・**Raycast**・標準の**ショートカット**用レシピが [`recipes/`](recipes/) にあります。`ai-notify status --icon` は `🔔`/`🔕` だけを出力するので、tmux・プロンプト・Claude Code のステータスラインに埋め込めます。
+
+> 切替は実行中でも効きます：次にエージェントが発火した時にフラグを読むので、トグルした瞬間に全稼働エージェントへ反映されます。
+
+## 🎙️ VOICEVOX キャラクターボイス
+
+通知を [VOICEVOX](https://voicevox.hiroshiba.jp/) のキャラ声（例：ずんだもん）で読み上げられます（無料・ローカル・オフライン）。
+
+> **VOICEVOXアプリのインストールと起動が必要です。** ai-notify はローカルのエンジンを叩くだけで、音声データは同梱していません。未起動なら ai-notify は**OS標準の音声**（Samantha, Kyoko …）を使います（こちらは設定不要ですぐ動きます）。
+
+`ai-notify voicevox setup` が案内します — ダウンロードページを開き、インストール済みならアプリを起動してエンジンの起動を待ちます。その後：
+
+```sh
+ai-notify voicevox setup        # VOICEVOXの導入／起動
+ai-notify voicevox speakers     # 利用可能なキャラとIDの一覧
+ai-notify voicevox on 3         # 話者3（例：ずんだもん）を使う
+```
+
+`AI_NOTIFY_VOICEVOX_SPEAKER` で各ペインに別キャラを割り当て可能。エンジンが起動していなければ自動でOS音声にフォールバックします。
+*VOICEVOXのキャラには利用規約があります。録画などを共有する場合は [VOICEVOXのガイドライン](https://voicevox.hiroshiba.jp/term/) に従ってクレジットしてください。*
+
+## 🌐 あなたの言語で読み上げ
+
+```sh
+ai-notify translate on ja       # エージェントのメッセージを翻訳してから読み上げ
+ai-notify translate test "I fixed the auth bug and added 3 tests."
+```
+
+キーレス・無料（HTTP 1リクエスト。オフライン時はローカルの定型文にフォールバック）。デスクトップ通知には原文も表示されます。
+
+## 💢 ツンデレモード（任意・遊び心）
+
+読み上げに「ツンデレ」人格を載せ、**事象の緊急度で口調が変わります**：
+
+- **失敗・危険な許可待ち** → 声大きめの鋭い**ツン**で「ちょっと！ビルドが失敗じゃない。早く直しなさいよね！」
+- **問題なしのパス** → やさしい**デレ**で「…ふふ、よくやったじゃない。べ、別に褒めてないんだからね…えらいえらい。」
+
+```sh
+ai-notify tsundere on            # 既定はOFF
+ai-notify tsundere level 0.6     # 既定の強さ 0（デレ）〜1（ツン）。メニューバーにもスライダー
+ai-notify tsundere test          # T3/T2/T1/T0 のサンプルを試聴
+```
+
+**無API・決定論・オフライン**（テンプレートで生成。課金ゼロ）。緊急度はエージェントの文面からのキーワード推定（厳密な重大度ではなくベストエフォート）で、デスクトップ通知は素の文面のまま。**VOICEVOX**利用時は、強さに応じて同じキャラの**ツンツン／あまあま**スタイルを選ぶので、声色そのものがツン・デレに変わります。`lang` は `ja` / `en` 対応。
+
+## ⏳ どの窓が・何を求めているか
+
+各通知のタイトルに窓ラベルが付きます — 入力待ちは `⏳ <label>`、完了は `✓ <label>`。本文には**何を**（翻訳されたプロンプト、または作業内容の要約）が出ます。各ペインに短い `AI_NOTIFY_LABEL` を設定すれば、10個のターミナルもひと目で見分けられます。
+
+## 仕組み
+
+XDGパス配下の単一のミュートフラグと設定だけ — デーモンも調整も無し：
+
+```
+${XDG_STATE_HOME:-~/.local/state}/ai-notify/muted     # 存在＝ミュート
+${XDG_CONFIG_HOME:-~/.config}/ai-notify/config.json   # 音・声・各種オプション
+```
+
+各エージェントのフックが `ai-notify hook --source <agent>` を呼び、発火時にこの1つのフラグを読みます。`ai-notify config init` で編集可能な設定（エージェント別の音・声・TTSバックエンド・翻訳・テンプレート）を書き出せます。
+
+## 対応プラットフォーム
+
+macOS は完全対応（`afplay` / `say` / VOICEVOX / `terminal-notifier` / ネイティブメニューバー）。Linux はベストエフォート（`paplay`/`canberra`, `notify-send`, `spd-say`/`espeak`, VOICEVOX）。Windows はビープ＋PowerShell読み上げ。利用できないバックエンドは静かに縮退し、エラーにはなりません。
+
+## ライセンス
+
+[MIT](LICENSE)。ランタイム依存ゼロ。

package/README.md

@@ -1,11 +1,15 @@
 # ai-notify
 
+**English** · [日本語](README.ja.md)
+
 **Know the moment your terminal AI agent needs you** — a sound, a spoken read-out, and a desktop banner the instant Claude Code, Codex, or another agent finishes a turn or asks for input. One mute switch covers **all of them, across every terminal**. No daemon, no background process.
 
-![ai-notify demo](https://raw.githubusercontent.com/unoryota/ai-notify/main/assets/demo.gif)
+![ai-notify demo](https://raw.githubusercontent.com/unoryota/ai-notify/main/assets/hero-en.gif)
 
 ```sh
-npm i -g ai-notify
+brew install unoryota/tap/ai-notify   # macOS (Homebrew)
+# or:  npm i -g ai-notify
+
 ai-notify init        # auto-detects your agents and wires them
 ```
 
@@ -19,10 +23,6 @@ Plenty of agents go quiet for minutes. ai-notify pulls you back at the right mom
 - 🔕 **One switch mutes everything.** Every agent in every terminal reads the same flag — one tap silences them all for a meeting.
 - 🔔 **A real menu bar bell, built in.** `ai-notify menubar install` — no Hammerspoon/SwiftBar required.
 
-> ### 日本語
-> 複数のAIエージェント（Claude Code / Codex …）を**並列で動かすと、どのターミナルの通知か分からない**——を解決する通知ツール。
-> **ペインごとに声を変えられる**（VOICEVOXのキャラ声も）／**英語の出力を日本語に翻訳して読み上げ**／**完了通知に作業内容の要約**／**1タップで全部ミュート**（MTG用）／**メニューバーのベルも内蔵**。
-
 ## Supported agents
 
 | Agent | Status | How it's wired |
@@ -41,6 +41,7 @@ ai-notify toggle | on | off | status               # the mute switch
 ai-notify volume [0.0-2.0]                          # get/set output volume
 ai-notify voice [number|name|preview|default]      # pick the spoken voice
 ai-notify voicevox [on <id>|off|speakers|test]     # speak in VOICEVOX voices
+ai-notify tsundere [on|off|level <0-1>|test]       # tsundere persona (ツン⇄デレ by urgency)
 ai-notify translate [on <lang>|off|test]           # speak agent text in your language
 ai-notify menubar [install|uninstall|status]       # native menu bar app (macOS)
 ai-notify doctor                                   # check deps & wiring
@@ -53,6 +54,7 @@ Per-window overrides — `export` these in a terminal *before* launching the age
 AI_NOTIFY_LABEL=api               # name this window in the read-out / notification
 AI_NOTIFY_VOICE=Eddy              # this window's `say` voice
 AI_NOTIFY_VOICEVOX_SPEAKER=3      # this window's VOICEVOX speaker id
+AI_NOTIFY_TSUNDERE_LEVEL=0.8      # this window's tsundere baseline (0=デレ … 1=ツン)
 AI_NOTIFY_VOLUME=0.5              # this window's volume (0.0–2.0)
 ```
 
@@ -66,7 +68,7 @@ ai-notify menubar install   # native menu bar app, starts at login
 
 A monochrome waveform icon shows status by color (Adobe-style): plain when idle, a **yellow** dot when an agent is waiting for you, **red + slash** when muted.
 
-- **Left-click** → menu: a **volume slider**, the **voice list** (system + VOICEVOX), and **per-pane** controls — each open terminal gets its own voice *and* volume.
+- **Left-click** → menu: a **volume slider**, a **tsundere** toggle + デレ⇄ツン slider, the **voice list** (system + VOICEVOX), and **per-pane** controls — each open terminal gets its own voice *and* volume.
 - **Right-click** → instant mute toggle.
 
 No third-party app needed. Prefer something else? There are drop-in recipes for **Hammerspoon**, **SwiftBar/xbar**, **Raycast**, and the built-in **macOS Shortcuts** in [`recipes/`](recipes/). `ai-notify status --icon` prints just `🔔`/`🔕` to embed in tmux / your prompt / Claude Code's status line.
@@ -75,9 +77,14 @@ No third-party app needed. Prefer something else? There are drop-in recipes for
 
 ## 🎙️ VOICEVOX character voices
 
-Speak your notifications in [VOICEVOX](https://voicevox.hiroshiba.jp/) character voices (free, local, offline). Run the VOICEVOX app, then:
+Optionally speak your notifications in [VOICEVOX](https://voicevox.hiroshiba.jp/) character voices (e.g. ずんだもん) — free, local, offline.
+
+> **Needs the VOICEVOX app installed and running.** ai-notify calls its local engine; it does not bundle the voices. Without it, ai-notify just uses your OS voice (Samantha, Kyoko, …) — no setup required.
+
+`ai-notify voicevox setup` walks you through it — it opens the download page, or launches the app and waits for the engine if it's already installed. Then:
 
 ```sh
+ai-notify voicevox setup        # install / launch VOICEVOX
 ai-notify voicevox speakers     # list available characters + ids
 ai-notify voicevox on 3         # use speaker 3 (e.g. ずんだもん)
 ```
@@ -94,6 +101,21 @@ ai-notify translate test "I fixed the auth bug and added 3 tests."
 
 Key-less and no cost (one HTTP request; falls back to a localized template offline). The desktop banner still shows the original text.
 
+## 💢 Tsundere mode (optional, fun)
+
+Give the spoken read-out a tsundere persona whose tone tracks **how urgent the event is**:
+
+- a **failure / dangerous approval** → a louder, sharp **ツン** scolding ("Hey! The build failed — don't just sit there, fix it!")
+- a **clean pass / no issues** → a warm **デレ** "good job" ("...heh, not bad. N-not that I'm impressed or anything.")
+
+```sh
+ai-notify tsundere on            # off by default
+ai-notify tsundere level 0.6     # baseline 0 (デレ) … 1 (ツン); the menu bar has a slider
+ai-notify tsundere test          # hear T3/T2/T1/T0 samples
+```
+
+It's **deterministic and offline** — phrase banks, no API, no cost. The urgency is a keyword heuristic over the agent's text (so it's best-effort, not a real severity signal), and the desktop banner stays factual. With **VOICEVOX** the level also picks the character's own **ツンツン / あまあま** style, so the same character actually *sounds* harsher or sweeter. `lang` supports `ja` and `en`.
+
 ## ⏳ Which window, and what it's asking
 
 Each notification is titled with the window label — `⏳ <label>` when an agent is waiting, `✓ <label>` when it's done — and the body says **what** (the translated prompt, or a summary of what was done). Set a short `AI_NOTIFY_LABEL` per pane and you can tell ten terminals apart at a glance.

package/menubar/AiNotifyMenuBar.swift

@@ -46,6 +46,12 @@ enum State {
         try? String(format: "%.2f", v).write(toFile: file("volume"), atomically: true, encoding: .utf8)
     }
 
+    // Tsundere baseline level 0.0 (デレ) – 1.0 (ツン). Same file the CLI reads.
+    static func setTsundereLevel(_ v: Double) {
+        try? FileManager.default.createDirectory(atPath: dir(), withIntermediateDirectories: true)
+        try? String(format: "%.2f", v).write(toFile: file("tsundere-level"), atomically: true, encoding: .utf8)
+    }
+
     @discardableResult
     static func cli(_ args: [String], capture: Bool = false) -> String? {
         let launcher = file("cli")
@@ -130,9 +136,36 @@ final class AppDelegate: NSObject, NSApplicationDelegate {
     @objc private func quit() { NSApp.terminate(nil) }
 
     @objc private func volumeChanged(_ s: NSSlider) { State.setVolume(s.doubleValue) }
+    // Slider is shown reversed (left = ツン, right = デレ) but the file keeps the
+    // canonical scale (0 = デレ, 1 = ツン), so write back 1 - position.
+    @objc private func tsundereLevelChanged(_ s: NSSlider) { State.setTsundereLevel(1 - s.doubleValue) }
+    @objc private func tsundereToggled(_ b: NSButton) { State.cli(["tsundere", "toggle"]) }
+    @objc private func paneTsundereChanged(_ s: NSSlider) {
+        if let tty = s.identifier?.rawValue { State.cli(["tsundere-pane", tty, String(format: "%.2f", 1 - s.doubleValue)]) }
+    }
     @objc private func paneVolumeChanged(_ s: NSSlider) {
         if let tty = s.identifier?.rawValue { State.cli(["volume-pane", tty, String(format: "%.2f", s.doubleValue)]) }
     }
+    // identifier carries the prosody key (speed | pitch | intonation).
+    @objc private func prosodyChanged(_ s: NSSlider) {
+        if let key = s.identifier?.rawValue { State.cli(["voice-prosody", key, String(format: "%.3f", s.doubleValue)]) }
+    }
+
+    // A labeled VOICEVOX base-prosody slider (speed / pitch / intonation). Applied
+    // on release (one subprocess per drag avoided). The key rides in the identifier.
+    private func prosodyRow(label: String, value: Double, lo: Double, hi: Double, key: String) -> NSMenuItem {
+        let row = NSView(frame: NSRect(x: 0, y: 0, width: 240, height: 24))
+        let cap = NSTextField(labelWithString: label)
+        cap.frame = NSRect(x: 12, y: 4, width: 48, height: 16)
+        cap.font = .systemFont(ofSize: 11); cap.textColor = .secondaryLabelColor
+        let slider = NSSlider(value: value, minValue: lo, maxValue: hi, target: self, action: #selector(prosodyChanged(_:)))
+        slider.frame = NSRect(x: 62, y: 3, width: 162, height: 20)
+        slider.isContinuous = false
+        slider.identifier = NSUserInterfaceItemIdentifier(key)
+        row.addSubview(cap); row.addSubview(slider)
+        let item = NSMenuItem(); item.view = row
+        return item
+    }
 
     // A 🔊 + slider row. identifier == nil => global (live); otherwise a pane tty
     // (applied on release to avoid a subprocess per drag tick).
@@ -149,6 +182,44 @@ final class AppDelegate: NSObject, NSApplicationDelegate {
         return item
     }
 
+    // A ツン ⇄ デレ slider for the tsundere baseline level. Shown reversed (left =
+    // ツン, right = デレ) for intuition, while the file keeps 0 = デレ, 1 = ツン — so
+    // the knob sits at 1 - value and writes back 1 - position. Continuous.
+    private func tsundereRow(value: Double, identifier: String? = nil) -> NSMenuItem {
+        let row = NSView(frame: NSRect(x: 0, y: 0, width: 220, height: 26))
+        let left = NSTextField(labelWithString: "ツン")
+        left.frame = NSRect(x: 12, y: 5, width: 30, height: 16)
+        left.font = .systemFont(ofSize: 10); left.textColor = .secondaryLabelColor
+        // identifier == nil => global (live, writes the level file); a pane tty =>
+        // per-pane override applied on release (one subprocess per drag avoided).
+        let action: Selector = identifier == nil ? #selector(tsundereLevelChanged(_:)) : #selector(paneTsundereChanged(_:))
+        let slider = NSSlider(value: 1 - value, minValue: 0, maxValue: 1, target: self, action: action)
+        slider.frame = NSRect(x: 46, y: 3, width: 128, height: 20)
+        slider.isContinuous = (identifier == nil)
+        slider.trackFillColor = .systemPink
+        if let id = identifier { slider.identifier = NSUserInterfaceItemIdentifier(id) }
+        let right = NSTextField(labelWithString: "デレ")
+        right.frame = NSRect(x: 178, y: 5, width: 30, height: 16)
+        right.font = .systemFont(ofSize: 10); right.textColor = .secondaryLabelColor
+        row.addSubview(left); row.addSubview(slider); row.addSubview(right)
+        let item = NSMenuItem(); item.view = row
+        return item
+    }
+
+    // ツンデレモード on/off as a checkbox living inside a view row, so a click
+    // toggles in place instead of dismissing the menu (a normal menu item closes
+    // on click). The level slider below stays mounted regardless of this state, so
+    // the menu height never jumps.
+    private func tsundereToggleRow(on: Bool) -> NSMenuItem {
+        let row = NSView(frame: NSRect(x: 0, y: 0, width: 220, height: 24))
+        let btn = NSButton(checkboxWithTitle: "ツンデレモード", target: self, action: #selector(tsundereToggled(_:)))
+        btn.frame = NSRect(x: 12, y: 2, width: 196, height: 20)
+        btn.state = on ? .on : .off
+        row.addSubview(btn)
+        let item = NSMenuItem(); item.view = row
+        return item
+    }
+
     // representedObject is the full CLI arg array to run.
     @objc private func runItem(_ item: NSMenuItem) {
         if let cmd = item.representedObject as? [String] { State.cli(cmd) }
@@ -163,16 +234,49 @@ final class AppDelegate: NSObject, NSApplicationDelegate {
     private func showMenu() {
         let menu = NSMenu()
 
-        // Global volume slider.
-        menu.addItem(sliderRow(value: State.volume, action: #selector(volumeChanged(_:)), identifier: nil))
-        menu.addItem(.separator())
-
         // Parse menu-json once.
         let json = (State.cli(["menu-json"], capture: true)?.data(using: .utf8))
             .flatMap { try? JSONSerialization.jsonObject(with: $0) as? [String: Any] }
         let voices = (json?["voices"] as? [[String: Any]]) ?? []
         let panes = (json?["panes"] as? [[String: Any]]) ?? []
 
+        // Global volume slider.
+        menu.addItem(sliderRow(value: State.volume, action: #selector(volumeChanged(_:)), identifier: nil))
+
+        // Tsundere mode: checkbox toggle + ツン⇄デレ baseline slider. Both live in
+        // view rows and are always mounted, so toggling never closes the menu nor
+        // shifts its height.
+        let tsun = json?["tsundere"] as? [String: Any]
+        let tsunOn = (tsun?["enabled"] as? Bool) ?? false
+        let tsunLevel = (tsun?["level"] as? Double) ?? 0.5
+        menu.addItem(tsundereToggleRow(on: tsunOn))
+        menu.addItem(tsundereRow(value: tsunLevel))
+        menu.addItem(.separator())
+
+        // VOICEVOX base prosody (speed / pitch / intonation) — only when VOICEVOX
+        // is the active TTS, since these are VOICEVOX audio_query scales.
+        if (json?["tts"] as? String) == "voicevox" {
+            let pr = json?["prosody"] as? [String: Any] ?? [:]
+            let range = json?["prosodyRange"] as? [String: Any] ?? [:]
+            let bounds: (String, Double, Double) -> (Double, Double) = { key, dlo, dhi in
+                let r = range[key] as? [Any]
+                let lo = (r?.first as? Double) ?? dlo
+                let hi = (r?.last as? Double) ?? dhi
+                return (lo, hi)
+            }
+            menu.addItem(disabledHeader("読み上げ（VOICEVOX）"))
+            for (key, label, dlo, dhi, dflt) in [
+                ("speed", "速さ", 0.5, 1.5, 1.0),
+                ("pitch", "高さ", -0.15, 0.15, 0.0),
+                ("intonation", "抑揚", 0.0, 1.5, 1.0),
+            ] {
+                let (lo, hi) = bounds(key, dlo, dhi)
+                let v = (pr[key] as? Double) ?? dflt
+                menu.addItem(prosodyRow(label: label, value: v, lo: lo, hi: hi, key: key))
+            }
+            menu.addItem(.separator())
+        }
+
         if voices.isEmpty {
             menu.addItem(disabledHeader("(声の一覧を取得できません)"))
         } else {
@@ -200,6 +304,15 @@ final class AppDelegate: NSObject, NSApplicationDelegate {
                 volDef.state = (p["volumeSet"] as? Bool ?? false) ? .off : .on
                 sub.addItem(volDef)
                 sub.addItem(.separator())
+                // Per-pane tsundere baseline (same ツン⇄デレ slider as global).
+                sub.addItem(disabledHeader("ツンデレ"))
+                let pts = (p["tsundere"] as? Double) ?? tsunLevel
+                sub.addItem(tsundereRow(value: pts, identifier: tty))
+                let tsDef = NSMenuItem(title: "強さを全体に従う", action: #selector(runItem(_:)), keyEquivalent: "")
+                tsDef.target = self; tsDef.representedObject = ["tsundere-pane", tty, "clear"]
+                tsDef.state = (p["tsundereSet"] as? Bool ?? false) ? .off : .on
+                sub.addItem(tsDef)
+                sub.addItem(.separator())
                 // Per-pane voice.
                 sub.addItem(disabledHeader("声"))
                 let def = NSMenuItem(title: "デフォルト（全体に従う）", action: #selector(runItem(_:)), keyEquivalent: "")

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai-notify",
-  "version": "0.2.1",
+  "version": "0.4.0",
   "description": "Desktop, sound, and spoken notifications for terminal AI coding agents (Claude Code, Codex, Gemini, ...) — with one mute switch that covers all of them, across every terminal.",
   "type": "module",
   "bin": {
@@ -15,6 +15,7 @@
     "menubar/README.md",
     "menubar/dist",
     "README.md",
+    "README.ja.md",
     "LICENSE"
   ],
   "scripts": {

package/src/cli.mjs

@@ -3,7 +3,7 @@
 // One mute switch for all of them, across every terminal. No daemon.
 
 import { readFileSync } from 'node:fs';
-import { execSync } from 'node:child_process';
+import { execSync, execFileSync } from 'node:child_process';
 import { providers, byId } from './providers/index.mjs';
 import { emit } from './notify.mjs';
 import { deriveLabel, cliInvocation, isEphemeralInstall } from './util.mjs';
@@ -12,6 +12,7 @@ import * as menubar from './menubar.mjs';
 import { translate } from './translate.mjs';
 import { diagnose as highlightDiagnose, clearHighlight } from './highlight.mjs';
 import * as voicevox from './voicevox.mjs';
+import * as tsundere from './tsundere.mjs';
 import {
   isMuted,
   setMuted,
@@ -22,12 +23,18 @@ import {
   DEFAULT_CONFIG,
   readVolume,
   setVolume,
+  readTsundereLevel,
+  setTsundereLevel,
+  readVoiceProsody,
+  setVoiceProsody,
+  resetVoiceProsody,
+  VOICE_PROSODY_RANGE,
   readPanes,
   readPaneSetting,
   updatePaneSetting,
 } from './state.mjs';
 
-const VERSION = '0.2.1';
+const VERSION = '0.3.0';
 
 const args = process.argv.slice(2);
 const cmd = args[0];
@@ -253,6 +260,33 @@ const cmds = {
     const config = readConfig();
     const url = config.voicevox?.url || voicevox.DEFAULT_URL;
 
+    if (sub === 'setup') {
+      if (voicevox.isAvailable(url)) {
+        log('✓ VOICEVOX engine is already running.');
+        return log('Enable it:  ai-notify voicevox on    (list voices: ai-notify voicevox speakers)');
+      }
+      if (process.platform !== 'darwin') {
+        log(`VOICEVOX is not running. Install it from ${voicevox.DOWNLOAD_URL} and launch the app, then:`);
+        return log('  ai-notify voicevox on');
+      }
+      if (voicevox.appInstalled()) {
+        log('VOICEVOX is installed but not running. Launching it…');
+        voicevox.launchApp();
+        log('  Waiting for the engine to start (first launch can take ~30s)…');
+        if (voicevox.waitForEngine(url, 45000)) {
+          log('✓ engine ready.');
+          return log('Enable it:  ai-notify voicevox on');
+        }
+        return log('  Still starting. Once VOICEVOX is open, run:  ai-notify voicevox on');
+      }
+      log('VOICEVOX is not installed. Opening the download page…');
+      voicevox.openDownloadPage();
+      log('  1. Download the macOS app and move it to Applications.');
+      log('  2. First launch is Gatekeeper-blocked (it is open-source / unsigned):');
+      log('       xattr -dr com.apple.quarantine /Applications/VOICEVOX.app && open -a VOICEVOX');
+      log('  3. Then run:  ai-notify voicevox setup   (or  ai-notify voicevox on)');
+      return;
+    }
     if (sub === 'speakers') {
       const list = voicevox.listSpeakers(url);
       if (!list.length) return log(`No speakers (is VOICEVOX running at ${url}?).`);
@@ -302,6 +336,92 @@ const cmds = {
     log(`🔊 volume → ${n}`);
   },
 
+  // Tsundere mode: skin the spoken read-out with a tsundere persona that turns
+  // ツン (harsh + louder) on failures and デレ (warm) on clean passes. Offline,
+  // deterministic, no cost.
+  //   tsundere on|off|toggle | level <0-1> | test [t3|t2|t1|t0] | status
+  tsundere() {
+    const sub = positionals[0] || 'status';
+    const config = readConfig();
+    const ts = config.tsundere;
+    const url = config.voicevox?.url || voicevox.DEFAULT_URL;
+
+    const sayText = (text, voice, tone = 'normal') => {
+      try {
+        const t = tsundere.decorateForSay(text, tone); // human contour, not 棒読み
+        execFileSync('say', voice ? ['-v', voice, t] : [t], { stdio: 'ignore' });
+      } catch {
+        /* non-mac / no say — ignore */
+      }
+    };
+
+    if (sub === 'on' || sub === 'off' || sub === 'toggle') {
+      const enabled = sub === 'toggle' ? !ts.enabled : sub === 'on';
+      config.tsundere = { ...ts, enabled };
+      // With VOICEVOX, resolve & cache the character's ツンツン/あまあま style ids
+      // now, so fire-time skips the lookup.
+      if (enabled && config.tts === 'voicevox') {
+        const sm = voicevox.resolveStyles(config.voicevox?.speaker, url);
+        if (sm) config.tsundere.styleMap = sm;
+      }
+      writeConfig(config);
+      log(enabled ? '💢 ツンデレ ON（デレ⇄ツン・緊急度で口調が変化）' : 'ツンデレ OFF');
+      if (enabled) {
+        log('  既定の強さ:  ai-notify tsundere level <0=デレ 〜 1=ツン>');
+        log('  試聴:        ai-notify tsundere test');
+      }
+      return;
+    }
+    if (sub === 'level') {
+      const arg = positionals[1];
+      if (arg === undefined) {
+        const v = readTsundereLevel();
+        return log(`tsundere level: ${v != null ? v : ts.level}  (0=デレ 〜 1=ツン)`);
+      }
+      const n = setTsundereLevel(arg);
+      return log(`💢 tsundere level → ${n}  (0=デレ 〜 1=ツン)`);
+    }
+    if (sub === 'test') {
+      const which = (positionals[1] || '').toLowerCase();
+      const lang = ts.lang || 'ja';
+      const level = readTsundereLevel() != null ? readTsundereLevel() : ts.level;
+      const ja = lang === 'ja';
+      const samples = {
+        t3: { event: 'done', raw: 'Build failed: TypeError in auth.ts', body: ja ? 'ビルドが失敗' : 'the build failed' },
+        t2: { event: 'waiting', raw: 'Claude needs your permission to run a command', body: ja ? '許可待ち' : 'waiting for your input' },
+        t1: { event: 'done', raw: 'Updated three files', body: ja ? '3ファイルを更新' : 'updated three files' },
+        t0: { event: 'done', raw: 'All tests passed, no issues', body: ja ? 'テスト全部パス' : 'all tests passed' },
+      };
+      const keys = samples[which] ? [which] : ['t3', 't2', 't1', 't0'];
+      const sm = config.tts === 'voicevox' ? ts.styleMap || voicevox.resolveStyles(config.voicevox?.speaker, url) : null;
+      log(`tsundere test (level ${level}, lang ${lang}):\n`);
+      for (const k of keys) {
+        const s = samples[k];
+        const tier = tsundere.classifyUrgency(s.event, s.raw, s.body);
+        const eff = tsundere.effectiveLevel(level, tier, ts.urgencyShift !== false);
+        const text = tsundere.wrap(s.body, eff, tier, lang, 0);
+        const mul = tsundere.volumeMul(tier, ts.volumeBoost !== false);
+        const tone = tsundere.axisFor(eff);
+        log(`  [${tier} ×${mul} ${tone}] ${text}`);
+        if (sm) {
+          const speaker = sm[tone] ?? config.voicevox?.speaker;
+          voicevox.speak(text, speaker, url, mul, undefined, tsundere.effectiveProsody(tone, readVoiceProsody()));
+        } else {
+          sayText(text, config.voice || '', tone);
+        }
+      }
+      return;
+    }
+    // status
+    const lvl = readTsundereLevel() != null ? readTsundereLevel() : ts.level;
+    log(`tsundere: ${ts.enabled ? '💢 ON' : 'OFF'}`);
+    log(`  level:        ${lvl}  (0=デレ 〜 1=ツン)`);
+    log(`  urgencyShift: ${ts.urgencyShift !== false ? 'on' : 'off'}  (緊急度で口調を増減)`);
+    log(`  volumeBoost:  ${ts.volumeBoost !== false ? 'on' : 'off'}  (重大時は音量↑)`);
+    log(`  lang:         ${ts.lang || 'ja'}`);
+    if (!ts.enabled) log('\nEnable:  ai-notify tsundere on    試聴:  ai-notify tsundere test');
+  },
+
   // Assign a voice to a specific pane (by tty), from the menu bar.
   //   voice-pane <tty> voicevox <id> | say <name> | clear
   'voice-pane'() {
@@ -340,6 +460,38 @@ const cmds = {
     log(`pane ${tty}: volume ${v}`);
   },
 
+  // Set a specific pane's tsundere baseline level (0=デレ – 1=ツン), or `clear` to
+  // follow the global level.  tsundere-pane <tty> <0-1|clear>
+  'tsundere-pane'() {
+    const [tty, arg] = positionals;
+    if (!tty || arg === undefined) {
+      console.error('usage: tsundere-pane <tty> <0-1|clear>');
+      process.exit(1);
+    }
+    if (arg === 'clear') {
+      updatePaneSetting(tty, { tsundere: null });
+      return log(`pane ${tty}: tsundere level reset to global`);
+    }
+    const v = Math.min(1, Math.max(0, Number(arg)));
+    updatePaneSetting(tty, { tsundere: v });
+    log(`pane ${tty}: tsundere level ${v}`);
+  },
+
+  // Get/set the VOICEVOX base prosody (the normal-tone scales the menu bar
+  // sliders drive). With no args, prints the current values as JSON.
+  //   voice-prosody [speed|pitch|intonation <value> | reset]
+  'voice-prosody'() {
+    const [key, val] = positionals;
+    if (key === 'reset') return log(JSON.stringify(resetVoiceProsody()));
+    if (!key || val === undefined) return log(JSON.stringify(readVoiceProsody()));
+    const next = setVoiceProsody(key, val);
+    if (!next) {
+      console.error('usage: voice-prosody <speed|pitch|intonation> <value> | reset');
+      process.exit(1);
+    }
+    log(`voice prosody ${key} → ${next[key]}`);
+  },
+
   // Machine-readable state for the menu bar agent: mute, volume, the selectable
   // voices, and the recently-active panes (for per-pane assignment). Not human.
   'menu-json'() {
@@ -371,6 +523,7 @@ const cmds = {
     // Panes = live terminals currently running an agent (so they show up before
     // they ever fire a notification) merged with previously-recorded ones.
     const globalVol = readVolume() != null ? readVolume() : typeof config.volume === 'number' ? config.volume : 1;
+    const tsLevel = readTsundereLevel() != null ? readTsundereLevel() : config.tsundere?.level ?? 0.5;
     const recorded = new Map(readPanes().map((p) => [p.tty, p.label]));
     const ttys = new Set([...livePanes(), ...recorded.keys()]);
     const panes = [...ttys].map((tty) => {
@@ -381,6 +534,8 @@ const cmds = {
         current: labelFor(s.tts ? s : null),
         volume: typeof s.volume === 'number' ? s.volume : globalVol,
         volumeSet: typeof s.volume === 'number',
+        tsundere: typeof s.tsundere === 'number' ? s.tsundere : tsLevel,
+        tsundereSet: typeof s.tsundere === 'number',
       };
     });
     log(
@@ -389,6 +544,10 @@ const cmds = {
         volume: readVolume() != null ? readVolume() : typeof config.volume === 'number' ? config.volume : 1,
         voices,
         panes,
+        tsundere: { enabled: !!config.tsundere?.enabled, level: tsLevel },
+        tts: config.tts || 'say',
+        prosody: readVoiceProsody(),
+        prosodyRange: VOICE_PROSODY_RANGE,
       })
     );
   },
@@ -524,7 +683,9 @@ Usage:
   ai-notify toggle | on | off | status               control the mute switch
   ai-notify volume [0.0-2.0]                          get/set output volume
   ai-notify voice [number|name|preview|default]      pick the spoken voice
-  ai-notify voicevox [on <id>|off|speakers|test]     speak in VOICEVOX character voices
+  ai-notify voicevox [setup|on <id>|off|speakers|test]  speak in VOICEVOX character voices
+  ai-notify tsundere [on|off|level <0-1>|test|status]   tsundere persona (ツン⇄デレ by urgency)
+  ai-notify voice-prosody [speed|pitch|intonation <v>|reset]  VOICEVOX read-out tuning
   ai-notify menubar [install|uninstall|status]       native menu bar bell (macOS)
   ai-notify translate [on <lang>|off|test]           speak agent text in your language
   ai-notify doctor                                    check deps & wiring

package/src/notify.mjs

@@ -8,11 +8,22 @@ import { spawn, execFileSync } from 'node:child_process';
 import { existsSync, rmSync } from 'node:fs';
 import { tmpdir } from 'node:os';
 import { join } from 'node:path';
-import { isMuted, readConfig, readVolume, recordPane, readPaneSetting, setPaneWaiting } from './state.mjs';
+import {
+  isMuted,
+  readConfig,
+  readVolume,
+  recordPane,
+  readPaneSetting,
+  setPaneWaiting,
+  readTsundereLevel,
+  readVoiceProsody,
+  nextCounter,
+} from './state.mjs';
 import { controllingTty } from './util.mjs';
 import { translate } from './translate.mjs';
 import { highlightWaiting, clearHighlight } from './highlight.mjs';
 import * as voicevox from './voicevox.mjs';
+import * as tsundere from './tsundere.mjs';
 
 const platform = process.platform; // 'darwin' | 'linux' | 'win32'
 
@@ -73,14 +84,23 @@ const sayWithVolume = (text, voice, vol) => {
   }
 };
 
-const speak = (text, voice, vol = 1) => {
+const speak = (text, voice, vol = 1, tone = 'normal') => {
   if (!text) return;
   if (platform === 'darwin') {
-    if (vol !== 1) return sayWithVolume(text, voice, vol);
-    run('say', voice ? ['-v', voice, text] : [text]);
+    // Give the OS voice human contour (pace/pitch/intonation + real pauses)
+    // instead of a flat 棒読み monotone.
+    const t = tsundere.decorateForSay(text, tone);
+    if (vol !== 1) return sayWithVolume(t, voice, vol);
+    run('say', voice ? ['-v', voice, t] : [t]);
   } else if (platform === 'linux') {
-    if (which('spd-say')) run('spd-say', [text]);
-    else if (which('espeak')) run('espeak', [text]);
+    const e = tsundere.prosodyFor(tone).espeak;
+    if (which('spd-say')) {
+      const r = Math.max(-100, Math.min(100, Math.round((e.speed - 175) / 1.5)));
+      const pch = Math.max(-100, Math.min(100, Math.round((e.pitch - 50) * 2)));
+      run('spd-say', ['-r', String(r), '-p', String(pch), text]);
+    } else if (which('espeak')) {
+      run('espeak', ['-p', String(e.pitch), '-s', String(e.speed), text]);
+    }
   } else if (platform === 'win32') {
     run('powershell', [
       '-NoProfile',
@@ -187,14 +207,47 @@ export const emit = ({ provider = 'default', event = 'done', label = '', message
           ? config.volume
           : 1;
 
+  // Tsundere mode: skin the spoken text, scale volume, and (with VOICEVOX) pick
+  // the character's ツンツン/あまあま style — all driven by the event's urgency.
+  // The banner is left untouched (it stays factual). Off => identical to before.
+  let outText = speakText;
+  let outVol = vol;
+  let outSpeaker = speaker;
+  let speakTone = 'normal'; // delivery contour; tsundere sets it to tsun/dere
+  const ts = config.tsundere;
+  if (ts && ts.enabled) {
+    const tier = tsundere.classifyUrgency(event, message, fullBody);
+    const envLevel = parseFloat(process.env.AI_NOTIFY_TSUNDERE_LEVEL);
+    const baseLevel = Number.isFinite(envLevel)
+      ? Math.min(1, Math.max(0, envLevel))
+      : typeof pane.tsundere === 'number'
+        ? pane.tsundere
+        : readTsundereLevel() != null
+          ? readTsundereLevel()
+          : typeof ts.level === 'number'
+            ? ts.level
+            : 0.5;
+    const eff = tsundere.effectiveLevel(baseLevel, tier, ts.urgencyShift !== false);
+    speakTone = tsundere.axisFor(eff);
+    outVol = Math.min(2, Math.max(0, vol * tsundere.volumeMul(tier, ts.volumeBoost !== false)));
+    outText = tsundere.wrap(spokenBody, eff, tier, ts.lang || 'ja', nextCounter('tsundere'));
+    if (config.speakLabel === true && label) outText = `${label}、${outText}`;
+    if (tts === 'voicevox') {
+      const sm = ts.styleMap || voicevox.resolveStyles(outSpeaker, config.voicevox?.url);
+      const axis = tsundere.axisFor(eff);
+      if (sm && sm[axis] != null) outSpeaker = sm[axis];
+    }
+  }
+
   if (!muted) {
-    playSound(soundName, vol);
-    if (config.speak && vol > 0) {
+    playSound(soundName, outVol);
+    if (config.speak && outVol > 0) {
       let spoken = false;
       if (tts === 'voicevox') {
-        spoken = voicevox.speak(speakText, speaker, config.voicevox?.url, vol);
+        const prosody = tsundere.effectiveProsody(speakTone, readVoiceProsody());
+        spoken = voicevox.speak(outText, outSpeaker, config.voicevox?.url, outVol, undefined, prosody);
       }
-      if (!spoken) speak(speakText, voice, vol); // OS `say` (also the VOICEVOX fallback)
+      if (!spoken) speak(outText, voice, outVol, speakTone); // OS `say` (also the VOICEVOX fallback)
     }
   }
 

package/src/state.mjs

@@ -67,6 +67,98 @@ export const setVolume = (v) => {
   return n;
 };
 
+// --- Tsundere level --------------------------------------------------------
+// A single number 0.0 (full デレ) – 1.0 (full ツン) in a state file, written by
+// the menu bar slider or `ai-notify tsundere level`, read at fire time. Overrides
+// config.tsundere.level; $AI_NOTIFY_TSUNDERE_LEVEL overrides per window.
+
+const tsundereLevelPath = () => join(stateDir(), 'tsundere-level');
+
+export const readTsundereLevel = () => {
+  try {
+    const v = parseFloat(readFileSync(tsundereLevelPath(), 'utf8'));
+    return Number.isFinite(v) ? Math.min(1, Math.max(0, v)) : null;
+  } catch {
+    return null;
+  }
+};
+
+export const setTsundereLevel = (v) => {
+  const n = Math.min(1, Math.max(0, Number(v)));
+  ensureDir(stateDir());
+  writeFileSync(tsundereLevelPath(), String(n));
+  return n;
+};
+
+// --- VOICEVOX base prosody -------------------------------------------------
+// User-tunable BASE scales for the VOICEVOX read-out — the values used at the
+// NORMAL tone; tsundere tones nudge from here. Written by the menu bar sliders /
+// `ai-notify voice-prosody`, read at fire time. One small JSON file so all three
+// stay in sync. Defaults = neutral (identical to no tuning).
+export const VOICE_PROSODY_DEFAULTS = { speed: 1.0, pitch: 0.0, intonation: 1.0 };
+export const VOICE_PROSODY_RANGE = { speed: [0.5, 1.5], pitch: [-0.15, 0.15], intonation: [0.0, 1.5] };
+
+const voiceProsodyPath = () => join(stateDir(), 'voice-prosody.json');
+
+const clampProsody = (key, v) => {
+  const [lo, hi] = VOICE_PROSODY_RANGE[key] || [0, 2];
+  return Math.min(hi, Math.max(lo, Number(v)));
+};
+
+export const readVoiceProsody = () => {
+  let raw = {};
+  try {
+    raw = JSON.parse(readFileSync(voiceProsodyPath(), 'utf8')) || {};
+  } catch {
+    /* missing/corrupt -> defaults */
+  }
+  const out = {};
+  for (const k of Object.keys(VOICE_PROSODY_DEFAULTS)) {
+    out[k] = typeof raw[k] === 'number' ? clampProsody(k, raw[k]) : VOICE_PROSODY_DEFAULTS[k];
+  }
+  return out;
+};
+
+// Set one key (speed | pitch | intonation); returns the full updated object, or
+// null for an unknown key.
+export const setVoiceProsody = (key, value) => {
+  if (!(key in VOICE_PROSODY_DEFAULTS)) return null;
+  const cur = readVoiceProsody();
+  cur[key] = clampProsody(key, value);
+  ensureDir(stateDir());
+  writeFileSync(voiceProsodyPath(), JSON.stringify(cur));
+  return cur;
+};
+
+export const resetVoiceProsody = () => {
+  try {
+    rmSync(voiceProsodyPath(), { force: true });
+  } catch {
+    /* ignore */
+  }
+  return { ...VOICE_PROSODY_DEFAULTS };
+};
+
+// A small persisted counter (per name), so phrase rotation varies across fires
+// even for identical input. Wraps to stay small; best-effort.
+export const nextCounter = (name) => {
+  const p = join(stateDir(), `ctr-${name}`);
+  let n = 0;
+  try {
+    n = parseInt(readFileSync(p, 'utf8'), 10) || 0;
+  } catch {
+    /* first use */
+  }
+  n = (n + 1) % 1000000;
+  try {
+    ensureDir(stateDir());
+    writeFileSync(p, String(n));
+  } catch {
+    /* ignore */
+  }
+  return n;
+};
+
 // --- Per-pane state --------------------------------------------------------
 // Recently-active terminal panes (so the menu bar can offer per-pane voices),
 // and a per-tty voice override. Both are small JSON files in the state dir.
@@ -114,7 +206,8 @@ export const readPanes = () =>
     .map(([tty, v]) => ({ tty, label: v.label || '', ts: v.ts || 0 }))
     .sort((a, b) => b.ts - a.ts);
 
-// Per-pane settings: { tts, speaker, voice, volume }. Any subset may be set.
+// Per-pane settings: { tts, speaker, voice, volume, tsundere }. Any subset may
+// be set (tsundere = a 0–1 baseline level override; null/absent = follow global).
 export const readPaneSetting = (tty) => (tty ? readJson(paneVoicesPath(), {})[tty] || {} : {});
 
 // Merge `patch` into the pane's settings; keys set to null are removed; an empty
@@ -176,6 +269,21 @@ export const DEFAULT_CONFIG = {
   // Per window: $AI_NOTIFY_VOICEVOX_SPEAKER overrides the speaker id.
   tts: 'say',
   voicevox: { url: 'http://127.0.0.1:50021', speaker: 3 },
+  // Tsundere mode: skin the SPOKEN read-out with a tsundere persona whose
+  // harshness (ツン) ⇄ sweetness (デレ) tracks the event's urgency — high-urgency
+  // failures get a louder ツン scolding, clean passes get a デレ "good job".
+  // Off by default. `level` is the baseline 0 (デレ) – 1 (ツン); the menu bar
+  // slider / `ai-notify tsundere level` write a state file that overrides it.
+  // With VOICEVOX, the level also picks the character's ツンツン/あまあま style
+  // (cached in `styleMap`). No API, no cost — deterministic phrase banks.
+  tsundere: {
+    enabled: false,
+    level: 0.5,
+    urgencyShift: true, // modulate the level by the event's urgency
+    volumeBoost: true, // louder on high-urgency events
+    lang: 'ja', // phrase bank language (ja | en)
+    styleMap: null, // { normal, tsun, dere } VOICEVOX style ids; auto-resolved
+  },
   // Spoken read-out templates for agent events. The window label is added
   // separately (speakLabel), so leave {label} out here to avoid doubling it.
   // Override per language (e.g. Japanese) in config.json. An agent that supplies
@@ -193,7 +301,12 @@ export const DEFAULT_CONFIG = {
 export const readConfig = () => {
   try {
     const raw = JSON.parse(readFileSync(configPath(), 'utf8'));
-    return { ...DEFAULT_CONFIG, ...raw, providers: { ...DEFAULT_CONFIG.providers, ...(raw.providers || {}) } };
+    return {
+      ...DEFAULT_CONFIG,
+      ...raw,
+      providers: { ...DEFAULT_CONFIG.providers, ...(raw.providers || {}) },
+      tsundere: { ...DEFAULT_CONFIG.tsundere, ...(raw.tsundere || {}) },
+    };
   } catch {
     return DEFAULT_CONFIG;
   }

package/src/tsundere.mjs

@@ -0,0 +1,218 @@
+// Tsundere mode: skin the spoken read-out with a tsundere persona whose harshness
+// (ツン) ⇄ sweetness (デレ) tracks the event's urgency.
+//
+//   high urgency (error / failure / dangerous approval)  -> ツン  + louder
+//   low  urgency (tests passed / no issues / approved)    -> デレ  (warm)
+//
+// Everything here is deterministic and offline — phrase banks, no API, no cost.
+// The banner (visual) is never skinned; only the spoken text is wrapped.
+
+// --- Urgency classifier ----------------------------------------------------
+// We only see the agent's notification text, so urgency is a heuristic. Order
+// matters: a "no errors" / "tests passed" message must read as POSITIVE even
+// though it contains the word "error".
+
+const POSITIVE =
+  /\b(passed|all tests? pass(ed)?|no (issues?|errors?|problems?|failures?)|looks good|lgtm|approved?|success(ful|fully)?|succeeded|completed successfully)\b|✅|問題(は)?(あ?り?ま?せん|な(い|し))|エラー(は)?(あり|出て)?(ま|い)?せん|テスト.*(成功|通過|パス|通り)|レビュー.*(通|問題|OK)|承認|無事(完了|成功)/i;
+
+// Critical = a failure, OR an approval for a DESTRUCTIVE action. A generic
+// "permission to run a command" is just a wait (T2) — only destructive verbs /
+// dangerous commands escalate to T3.
+const CRITICAL =
+  /\b(failed|failing|failure|crash(ed|ing)?|exception|panic|fatal|unrecoverable|aborted|broke(n)?|blocked)\b|❌|🛑|\b(permission|approval)\b[^.!?\n]*\b(delete|remove|overwrite|reset|drop|truncate|force|rm)\b|rm\s+-rf|force[- ]?push|git\s+push\s+-f|drop\s+table|truncate\b|エラーが|失敗|クラッシュ|例外が|落ちて|中断され|危険なコマンド/i;
+
+// Returns one of 'T3' (critical) | 'T2' (waiting) | 'T1' (neutral done) | 'T0' (positive).
+// `raw` is the agent's original text (pre-translation); `core` is the formatted
+// body. We test the raw text first for accuracy.
+export const classifyUrgency = (event = 'done', raw = '', core = '') => {
+  const text = `${raw || ''} ${core || ''}`;
+  if (POSITIVE.test(text) && !CRITICAL.test(text)) return 'T0';
+  if (CRITICAL.test(text)) return 'T3';
+  if (event === 'waiting') return 'T2';
+  return 'T1';
+};
+
+// Per-tier modulation: nudge the baseline tsun level toward ツン (positive bias)
+// or デレ (negative bias), and scale the volume. Kept small (±0.25) so the
+// SLIDER stays in charge — at either extreme the slider wins (even a success
+// reads ツン at max, even a failure reads デレ at min); near the middle the
+// urgency nudge is what tips the tone. T0 never lowers the volume.
+const BIAS = { T3: 0.25, T2: 0.1, T1: 0, T0: -0.25 };
+const VOLMUL = { T3: 1.3, T2: 1.1, T1: 1, T0: 1 };
+
+export const effectiveLevel = (level, tier, urgencyShift = true) => {
+  const base = Number.isFinite(level) ? level : 0.5;
+  return Math.min(1, Math.max(0, base + (urgencyShift ? BIAS[tier] || 0 : 0)));
+};
+
+export const volumeMul = (tier, volumeBoost = true) => (volumeBoost ? VOLMUL[tier] || 1 : 1);
+
+// eff >= 0.6 => ツン, <= 0.4 => デレ, else ノーマル. A narrow ノーマル band (only
+// the genuinely-neutral middle) so the contrast between ツン and デレ is obvious
+// instead of everything collapsing into a bland middle. Used for both the phrase
+// tone and the VOICEVOX style pick.
+export const axisFor = (eff) => (eff >= 0.6 ? 'tsun' : eff <= 0.4 ? 'dere' : 'normal');
+
+// --- Phrase banks ----------------------------------------------------------
+// BANK[lang][tone] = { <tier>: [...], default: [...] }. `{body}` is the task
+// gist (kept, so the read-out is still informative). Tasteful, short, SFW.
+
+const BANK = {
+  ja: {
+    // ツン: 冷たい・とげとげ・素直じゃない。失敗には容赦なく、成功も渋々。
+    tsun: {
+      T3: [
+        'ちょっと！また{body}じゃない。…ぼーっとしてないで早く直しなさいよ！',
+        'はぁ？{body}って…どこ見てたのよ。さっさと直す！',
+        'べ、別に心配なんてしてないけど…{body}よ。早くなんとかしなさいよね！',
+      ],
+      T2: [
+        '…{body}。あんたの判断待ちなの。さっさと決めなさいよ。',
+        'ねえ、{body}でしょ。…わたしに聞いてないで自分で決めなさい。',
+      ],
+      T1: [
+        'ふん、{body}。…言われなくてもやっといたわよ。',
+        '{body}。…別にあんたのためじゃないんだからね。',
+      ],
+      T0: [
+        '{body}…ま、まあ及第点ね。べ、別に褒めてないんだからね！',
+        'ふん、{body}じゃない。…ちょっとは見直したけど、調子に乗らないでよね。',
+      ],
+      default: ['{body}。…さっさと次いきなさいよ。'],
+    },
+    // ノーマル: 中央のニュートラル帯だけ。素っ気なく事実だけ。
+    normal: {
+      default: ['{body}。', '{body}。…以上よ。'],
+    },
+    // デレ: あまあま・素直・openly 心配＆応援。失敗にも寄り添う。
+    dere: {
+      T3: [
+        'あっ、{body}…！大丈夫？あわてなくていいから、一緒に直そ？',
+        '{body}みたい…。落ち込まないで、ね？あなたならきっと直せるよ。',
+      ],
+      T2: [
+        'ねぇ、{body}だって。…あなたの答え、ここで待ってるね。',
+        '{body}…どうするか、ゆっくり決めていいからね。',
+      ],
+      T1: [
+        '{body}、おしまい。…おつかれさま、えらいよ。',
+        '{body}。…ちゃんとできてる、すごいね。',
+      ],
+      T0: [
+        '{body}…！やったね、すごいすごい！わたし、ほんとに嬉しい！',
+        'わぁ、{body}だって！さすがだなぁ、大好き…！',
+        'お疲れさま。{body}…できるって信じてたよ、ほんとえらい！',
+      ],
+      default: ['{body}。…よくがんばったね。'],
+    },
+  },
+  en: {
+    tsun: {
+      T3: [
+        "Hey! {body} again?! ...Don't just sit there — fix it!",
+        'Seriously? {body}. Clean it up, now.',
+        "I-it's not like I was worried, but... {body}. Deal with it.",
+      ],
+      T2: ['...{body}. It needs your call. Hurry up and decide already.'],
+      T1: [
+        'Hmph. {body}. ...I did it without being asked, obviously.',
+        "{body}. ...Not that I did it for you or anything.",
+      ],
+      T0: [
+        "{body}... fine, that's passable. N-not that I'm impressed!",
+        'Hmph, {body}. ...A little better, I guess. Don’t let it go to your head.',
+      ],
+      default: ['{body}. ...Get on with the next one.'],
+    },
+    normal: { default: ['{body}.', "{body}. ...That's that."] },
+    dere: {
+      T3: [
+        "Oh no, {body}...! Are you okay? Don't panic — let's fix it together, okay?",
+        "{body}, huh... Don't be down. You've got this, I know it.",
+      ],
+      T2: [
+        "Hey, {body}. ...I'll be right here waiting for your call.",
+        '{body}... take your time deciding, okay?',
+      ],
+      T1: [
+        '{body}, all done. ...Nice work, you did great.',
+        "{body}. ...You really pulled it off. I'm proud of you.",
+      ],
+      T0: [
+        "{body}...! You did it! Amazing, amazing! I'm so happy for you!",
+        "Wow, {body}! That's incredible — good job!",
+        'Nice work. {body}... I always knew you could do it.',
+      ],
+      default: ['{body}. ...You did your best, well done.'],
+    },
+  },
+};
+
+export const isLangSupported = (lang) => !!BANK[lang];
+
+// Wrap `body` with a tsundere line. `rot` rotates phrase choice so repeats vary.
+// Unsupported language => body is returned unchanged (volume/voice still apply).
+export const wrap = (body, eff, tier, lang = 'ja', rot = 0) => {
+  const bank = BANK[lang];
+  if (!bank || !body) return body;
+  const tone = axisFor(eff);
+  const group = bank[tone] || bank.normal;
+  const arr = (group && (group[tier] || group.default)) || ['{body}'];
+  const phrase = arr[((rot % arr.length) + arr.length) % arr.length];
+  return phrase.replace('{body}', body);
+};
+
+// --- Delivery / prosody ----------------------------------------------------
+// The persona's VOICE, not just its words: how it's spoken, so the read-out has
+// human contour instead of a flat 棒読み monotone. Each tone gets its own pace,
+// pitch, and intonation range.
+//   say.*  : macOS `say` embedded-command deltas, RELATIVE to the voice's own
+//            natural setting (rate wpm, pbas pitch base, pmod pitch range) — so
+//            it works on any voice without knowing its defaults.
+//   vv.*   : VOICEVOX audio_query scales (speed/pitch/intonation; 1.0 = default).
+//   espeak : { pitch 0-99, speed wpm } for the Linux fallback.
+//   tsun  = quick, higher, sharp swings (agitated scolding).
+//   dere  = slower, gentle, wide warm intonation + longer pauses (affectionate).
+//   normal= mild, just enough lilt to not sound robotic.
+// Kept deliberately SUBTLE: VOICEVOX is already expressive, so over-driving the
+// scales (intonation ≫1.2, any pitch shift) is what makes it sound warbly and
+// unnatural. The real ツン/デレ contrast comes from the character STYLE
+// (ツンツン/あまあま, see voicevox.resolveStyles) — these scales only add a light
+// pace/lilt on top, staying inside natural ranges. Same idea for `say`: a small
+// pmod, not a big one (heavy pitch-modulation = robotic warble).
+const PROSODY = {
+  tsun: { say: { rate: 16, pbas: 3, pmod: 3 }, vv: { speed: 1.06, pitch: 0.0, intonation: 1.2 }, espeak: { pitch: 56, speed: 190 } },
+  normal: { say: { rate: 0, pbas: 0, pmod: 2 }, vv: { speed: 1.0, pitch: 0.0, intonation: 1.0 }, espeak: { pitch: 50, speed: 175 } },
+  dere: { say: { rate: -12, pbas: 1, pmod: 4 }, vv: { speed: 0.96, pitch: 0.0, intonation: 1.1 }, espeak: { pitch: 46, speed: 160 } },
+};
+
+export const prosodyFor = (tone) => PROSODY[tone] || PROSODY.normal;
+
+// Combine the user's GUI-tunable BASE scales (the normal-tone values) with this
+// tone's nudge, for the VOICEVOX read-out. speed & intonation are scales (they
+// multiply), pitch is an offset (it adds). base = {} => pure tone prosody, which
+// equals the old behaviour. Returns { speed, pitch, intonation }.
+export const effectiveProsody = (tone, base = {}) => {
+  const t = prosodyFor(tone).vv;
+  const b = { speed: 1, pitch: 0, intonation: 1, ...base };
+  return {
+    speed: b.speed * t.speed,
+    pitch: b.pitch + t.pitch,
+    intonation: b.intonation * t.intonation,
+  };
+};
+
+const sgn = (n) => (n >= 0 ? `+${n}` : `${n}`);
+
+// Wrap spoken text with macOS `say` embedded commands for the given tone, and
+// turn ellipses / commas into real beats so the line breathes. Stray `[[`/`]]`
+// in the dynamic text is neutralized first so it can't inject its own commands.
+export const decorateForSay = (text, tone = 'normal') => {
+  if (!text) return text;
+  const p = prosodyFor(tone).say;
+  const body = String(text)
+    .replace(/\[\[|\]\]/g, '') // can't let task text smuggle in commands
+    .replace(/[…⋯]+|・・・+|\.{3,}/g, ' [[slnc 220]] ') // a short beat where it trails off
+    .replace(/([、,])\s*/g, '$1 [[slnc 70]] '); // commas breathe a touch
+  return `[[rate ${sgn(p.rate)}]] [[pbas ${sgn(p.pbas)}]] [[pmod ${sgn(p.pmod)}]] ${body}`;
+};

package/src/voicevox.mjs

@@ -9,15 +9,57 @@
 // return false and the caller falls back to the OS `say` voice.
 
 import { execSync, execFileSync } from 'node:child_process';
-import { existsSync, statSync, mkdtempSync, rmSync, appendFileSync } from 'node:fs';
+import { existsSync, statSync, mkdtempSync, rmSync, appendFileSync, readFileSync, writeFileSync } from 'node:fs';
 import { join } from 'node:path';
-import { tmpdir } from 'node:os';
+import { tmpdir, homedir } from 'node:os';
 import { stateDir } from './state.mjs';
 
 export const DEFAULT_URL = 'http://127.0.0.1:50021';
+export const DOWNLOAD_URL = 'https://voicevox.hiroshiba.jp/';
 
 const platform = process.platform;
 
+const sleep = (ms) => {
+  try {
+    Atomics.wait(new Int32Array(new SharedArrayBuffer(4)), 0, 0, ms);
+  } catch {
+    /* SharedArrayBuffer unavailable — skip the wait */
+  }
+};
+
+// Is the VOICEVOX app installed (macOS)?
+export const appInstalled = () => {
+  if (platform !== 'darwin') return false;
+  return ['/Applications/VOICEVOX.app', join(homedir(), 'Applications/VOICEVOX.app')].some((p) => existsSync(p));
+};
+
+export const launchApp = () => {
+  try {
+    if (platform === 'darwin') execFileSync('open', ['-a', 'VOICEVOX']);
+  } catch {
+    /* ignore */
+  }
+};
+
+export const openDownloadPage = () => {
+  try {
+    if (platform === 'darwin') execFileSync('open', [DOWNLOAD_URL]);
+    else if (platform === 'linux') execFileSync('xdg-open', [DOWNLOAD_URL]);
+  } catch {
+    /* ignore */
+  }
+};
+
+// Poll until the engine answers, or timeout.
+export const waitForEngine = (url = DEFAULT_URL, timeoutMs = 40000) => {
+  const start = Date.now();
+  while (Date.now() - start < timeoutMs) {
+    if (isAvailable(url, 1500)) return true;
+    sleep(2000);
+  }
+  return false;
+};
+
 // Record why a synthesis fell back to the OS voice, so intermittent fallbacks
 // are diagnosable instead of silent. Best-effort.
 const logFail = (reason) => {
@@ -73,6 +115,34 @@ export const listCharacters = (url = DEFAULT_URL) => {
   }
 };
 
+// For tsundere mode: given a speaker id, find the character that owns it and map
+// its styles to { normal, tsun, dere } speaker ids (so the SAME character can
+// speak in a ツンツン or あまあま voice). Missing styles fall back to normal.
+export const resolveStyles = (speakerId, url = DEFAULT_URL) => {
+  try {
+    const out = execFileSync('curl', ['-s', '-m', '4', `${url}/speakers`], { encoding: 'utf8', timeout: 5000 });
+    const data = JSON.parse(out);
+    const sid = Number(speakerId);
+    for (const sp of data) {
+      const styles = sp.styles || [];
+      if (!styles.some((s) => Number(s.id) === sid)) continue;
+      const find = (re) => {
+        const m = styles.find((s) => re.test(s.name || ''));
+        return m ? Number(m.id) : null;
+      };
+      const normal = find(/ノーマル|普通/) ?? sid;
+      return {
+        normal,
+        tsun: find(/ツンツン|ツン/) ?? normal,
+        dere: find(/あまあま|甘え|デレ|ささやき/) ?? normal,
+      };
+    }
+  } catch {
+    /* engine down / parse error — caller falls back to the base speaker */
+  }
+  return null;
+};
+
 const playWav = (wav, vol = 1) => {
   if (platform === 'darwin') execFileSync('afplay', ['-v', String(vol), wav], { timeout: 30000 });
   else if (platform === 'linux') {
@@ -84,8 +154,25 @@ const playWav = (wav, vol = 1) => {
   }
 };
 
+// Apply a prosody profile to a VOICEVOX audio_query JSON in place, so the
+// read-out has human contour (pace/pitch/intonation) instead of a flat 棒読み.
+// Only the small query JSON passes through Node; the WAV never does.
+const applyProsody = (queryPath, prosody) => {
+  if (!prosody) return;
+  try {
+    const q = JSON.parse(readFileSync(queryPath, 'utf8'));
+    if (typeof prosody.speed === 'number') q.speedScale = prosody.speed;
+    if (typeof prosody.pitch === 'number') q.pitchScale = prosody.pitch;
+    if (typeof prosody.intonation === 'number') q.intonationScale = prosody.intonation;
+    writeFileSync(queryPath, JSON.stringify(q));
+  } catch {
+    /* leave the query untouched on any parse/IO error */
+  }
+};
+
 // Synthesize and play. Returns true if it spoke, false to fall back to `say`.
-export const speak = (text, speaker = 3, url = DEFAULT_URL, vol = 1, timeoutMs = 15000) => {
+// `prosody` (optional) = { speed, pitch, intonation } audio_query scale overrides.
+export const speak = (text, speaker = 3, url = DEFAULT_URL, vol = 1, timeoutMs = 15000, prosody = null) => {
   if (!text) return false;
   let dir;
   try {
@@ -93,12 +180,26 @@ export const speak = (text, speaker = 3, url = DEFAULT_URL, vol = 1, timeoutMs =
     const wav = join(dir, 'v.wav');
     const sec = String(Math.max(2, Math.ceil(timeoutMs / 1000)));
     const enc = encodeURIComponent(text); // URL-encoded -> no shell metacharacters
-    // Pipe audio_query straight into synthesis. execSync uses /bin/sh for the pipe.
-    const cmd =
-      `curl -s -m ${sec} -X POST "${url}/audio_query?speaker=${speaker}&text=${enc}" | ` +
-      `curl -s -m ${sec} -X POST -H "Content-Type: application/json" -d @- ` +
-      `"${url}/synthesis?speaker=${speaker}" -o "${wav}"`;
-    execSync(cmd, { timeout: timeoutMs + 1000, stdio: 'ignore' });
+    if (prosody) {
+      // Two steps so we can tune the query JSON between them (still no WAV in Node).
+      const q = join(dir, 'q.json');
+      execSync(`curl -s -m ${sec} -X POST "${url}/audio_query?speaker=${speaker}&text=${enc}" -o "${q}"`, {
+        timeout: timeoutMs + 1000,
+        stdio: 'ignore',
+      });
+      applyProsody(q, prosody);
+      execSync(
+        `curl -s -m ${sec} -X POST -H "Content-Type: application/json" -d @"${q}" "${url}/synthesis?speaker=${speaker}" -o "${wav}"`,
+        { timeout: timeoutMs + 1000, stdio: 'ignore' }
+      );
+    } else {
+      // Pipe audio_query straight into synthesis. execSync uses /bin/sh for the pipe.
+      const cmd =
+        `curl -s -m ${sec} -X POST "${url}/audio_query?speaker=${speaker}&text=${enc}" | ` +
+        `curl -s -m ${sec} -X POST -H "Content-Type: application/json" -d @- ` +
+        `"${url}/synthesis?speaker=${speaker}" -o "${wav}"`;
+      execSync(cmd, { timeout: timeoutMs + 1000, stdio: 'ignore' });
+    }
     if (!existsSync(wav) || statSync(wav).size < 1000) {
       logFail(`empty/short wav (speaker ${speaker}, ${text.length} chars)`);
       return false;