@aituber-onair/chat 0.47.0 → 0.49.0

package/README.ja.md

@@ -571,30 +571,57 @@ CI/ローカルで再現性を高める場合は
 ```typescript
 const claudeService = ChatServiceFactory.createChatService('claude', {
   apiKey: process.env.ANTHROPIC_API_KEY,
-  model: 'claude-sonnet-5'
+  model: 'claude-opus-5',
+  reasoning_effort: 'low'
 });
 ```
 
+`claude-opus-5` は高性能用途向けの明示的な選択肢として利用でき、パッケージの
+既定値にはしていません。Claude Opus 5 は adaptive thinking が既定で有効です。
+対応するClaudeモデルでは `reasoning_effort` を指定でき、Anthropic APIの
+`output_config.effort` に変換されます。省略時のClaude API既定値は `high` です。
+
+- Claude Opus 5、Sonnet 5、Opus 4.8、Opus 4.7:
+  `low`、`medium`、`high`、`xhigh`、`max`
+- Claude Opus 4.6、Sonnet 4.6:
+  `low`、`medium`、`high`、`max`
+- Claude Opus 4.5: `low`、`medium`、`high`
+
+effortを下げるとレイテンシーとトークン効率を優先できますが、表示される回答の
+短さは保証されません。短い回答が必要な場合は、回答長の指示も併用してください。
+複数ターンのツール呼び出しでは provider 固有の thinking / tool-use ブロックを
+保持するため、返された `completion.assistant_message` を会話履歴へ追加してください。
+
 #### Google Gemini
 
 ```typescript
 const geminiService = ChatServiceFactory.createChatService('gemini', {
   apiKey: process.env.GOOGLE_API_KEY,
-  model: 'gemini-3.1-flash-lite'
+  model: 'gemini-3.1-flash-lite',
+  reasoning_effort: 'minimal'
 });
 ```
 
-`gemini-3.1-flash-lite` が推奨の安定版 Flash-Lite モデルです。
-現在の安定版 Flash モデルとして `gemini-3.5-flash` も利用できます。
+`gemini-3.1-flash-lite` は既定の Flash-Lite モデルとして維持します。
+最新の安定版・低遅延 Flash-Lite として `gemini-3.5-flash-lite`、
+より高度なエージェント・マルチモーダル用途向けに `gemini-3.6-flash`、
+安定版 Flash として `gemini-3.5-flash` も利用できます。
 `gemini-3.1-flash-lite-preview`、`gemini-3-pro-preview`、
 `gemini-2.5-flash-lite-preview-06-17` などの preview / deprecated
 モデルは後方互換のため明示的な model string では利用できますが、本番用途では
 新しいモデルへ移行してください。
 
-チャット用途では、`gemini-3.5-flash` に対して Gemini の
-`thinkingConfig`（`thinkingLevel: 'MINIMAL'`、`includeThoughts: false`）を
-自動で送信します。利用者側で追加オプションを指定する必要はなく、既定で高速な
-チャット応答と hidden thinking token の抑制を優先します。
+Gemini 3 モデルでは `reasoning_effort` を指定できます。この値は Gemini の
+`thinkingConfig.thinkingLevel` に変換され、`includeThoughts: false` を維持します。
+
+- Gemini 3 Flash / Flash-Lite: `minimal`, `low`, `medium`, `high`。
+  低遅延なチャット向けの既定値は `minimal` です。
+- Gemini 3 Pro: `low`, `medium`, `high`。Pro は `minimal` 非対応のため、
+  既定値は `low` です。
+
+これにより Gemini 3.6/3.5 Flash の medium thinking 既定値を上書きし、短い
+出力上限を hidden thinking が使い切るリスクを抑えます。Gemini 2.5 は
+`thinkingBudget` を使うため、`reasoning_effort` は送信しません。
 
 #### OpenRouter
 
@@ -609,26 +636,36 @@ const openRouterService = ChatServiceFactory.createChatService('openrouter', {
 ```
 
 **OpenRouterの重要な注意事項:**
-- `gpt-oss-20b:free`モデルでは、技術的制限によりトークン制限が自動的に無効化されます
-- 応答の長さを制御するには、プロンプト内で指示してください（例：「40文字以内で返答してください」）
+- `openrouter/auto` と `openrouter/auto-beta` では、ルーティング先のreasoningモデルが可視テキストの生成前に出力予算を使い切ることを避けるため、`responseLength` 由来のトークン制限を自動的に無効化します。明示的に指定した `maxTokens` はそのまま使用します。
+- `gpt-oss-20b:free` と `z-ai/glm-5.2` では、すべてのトークン制限を引き続き無効化します。これらのモデルと動的Routerでは、応答の長さをプロンプトで指示してください（例:「40文字以内で回答してください」）。
 - 無料階層にはレート制限があります（20リクエスト/分）
 - 無料モデル判定はモデルID末尾の `:free` で行います（動的取得した `:free` も同様にレート制限対象）
+- `openrouter/auto-beta` はタスクを判定するBetaルーターです。リクエストごとにモデルを選択し、そのモデルの料金が適用されます。選択結果はレスポンスの `model` またはOpenRouter Activityで確認できます。
 - `openrouter/fusion` は複数モデルのパネルとジャッジモデルを実行します。単一モデルの固定単価ではなく、内部で使われた各モデル呼び出しと web search/fetch 利用分の合算で課金されます。
-- `z-ai/glm-5.2` は OpenRouter reasoning を `none` にし、reasoning token が出力上限を使い切って可視テキストが空になることを避けるため、トークン制限も自動的に無効化します。
+- `z-ai/glm-5.2` は OpenRouter reasoning も `none` にします。
+- コーディング特化モデルはデフォルトではなく明示的な選択肢です。`kwaipilot/kat-coder-air-v2.5` と `kwaipilot/kat-coder-pro-v2.5` はテキスト専用です。
+- `moonshotai/kimi-k3` の可用性は上流キャパシティに依存し、混雑時はOpenRouterから429レスポンスが返る場合があります。
+- `x-ai/grok-4.5` には現在のEU制限を含む地域別の可用性制限があります。`~x-ai/grok-latest` がGrok 4.5へ解決される場合も同じ制限を受けます。
 - サポート対象モデル（キュレーション済み）:
-  - `openrouter/auto`
+  - `openrouter/auto`, `openrouter/auto-beta`
   - `openrouter/fusion`
   - `openai/gpt-oss-20b:free`
-  - `~openai/gpt-latest`, `~openai/gpt-mini-latest`, `openai/gpt-5.5-pro`, `openai/gpt-5.5`
+  - `~openai/gpt-latest`, `~openai/gpt-mini-latest`
+  - `openai/gpt-5.6-sol`, `openai/gpt-5.6-terra`, `openai/gpt-5.6-luna`
+  - `openai/gpt-5.5-pro`, `openai/gpt-5.5`
   - `openai/gpt-5.1-chat`, `openai/gpt-5.1-codex`, `openai/gpt-5-mini`, `openai/gpt-5-nano`
   - `openai/gpt-4o`, `openai/gpt-4.1-mini`, `openai/gpt-4.1-nano`
   - `~anthropic/claude-sonnet-latest`, `~anthropic/claude-haiku-latest`
+  - `anthropic/claude-opus-5`
   - `anthropic/claude-opus-4`, `anthropic/claude-sonnet-4`
   - `anthropic/claude-3.7-sonnet`, `anthropic/claude-3.5-sonnet`, `anthropic/claude-haiku-4.5`
   - `~google/gemini-pro-latest`, `~google/gemini-flash-latest`
+  - `google/gemini-3.6-flash`, `google/gemini-3.5-flash-lite`
   - `google/gemini-2.5-pro`, `google/gemini-2.5-flash`, `google/gemini-2.5-flash-lite-preview-09-2025`
   - `z-ai/glm-5.2`, `z-ai/glm-4.7-flash`, `z-ai/glm-4.5-air`, `z-ai/glm-4.5-air:free`
-  - `~moonshotai/kimi-latest`, `moonshotai/kimi-k2.7-code`, `moonshotai/kimi-k2.5`
+  - `~x-ai/grok-latest`, `x-ai/grok-4.5`
+  - `~moonshotai/kimi-latest`, `moonshotai/kimi-k3`, `moonshotai/kimi-k2.7-code`, `moonshotai/kimi-k2.5`
+  - `kwaipilot/kat-coder-air-v2.5`, `kwaipilot/kat-coder-pro-v2.5`
 
 **OpenRouter freeモデルの動的リフレッシュ**
 
@@ -697,18 +734,21 @@ const xaiService = ChatServiceFactory.createChatService('xai', {
 ```typescript
 const kimiService = ChatServiceFactory.createChatService('kimi', {
   apiKey: process.env.MOONSHOT_API_KEY,
-  model: 'kimi-k2.6',
+  model: 'kimi-k3',
   // Optional: エンドポイント/ベースURLの上書き
   // endpoint: 'https://api.moonshot.ai/v1/chat/completions',
   // baseUrl: 'https://api.moonshot.ai/v1',
-  thinking: { type: 'enabled' }
+  reasoning_effort: 'low'
 });
 ```
 
 注意:
 - KimiはOpenAI互換のChat Completionsを利用します。
-- 対応モデル: `kimi-k2.7-code`, `kimi-k2.7-code-highspeed`, `kimi-k2.6`, `kimi-k2.5`
+- 対応モデル: `kimi-k3`, `kimi-k2.7-code`, `kimi-k2.7-code-highspeed`, `kimi-k2.6`, `kimi-k2.5`
 - チャット用途向けのデフォルトモデルは引き続き `kimi-k2.6` です。
+- Kimi K3は明示的に選択するReasoningモデルです。`reasoning_effort: 'low' | 'high' | 'max'` に対応し、デフォルトは `max` です。Reasoning自体は無効化できず、K2系の `thinking` オプションも利用できません。`none`、`minimal`、`medium` は非対応です。
+- Kimi K3の出力上限には `max_completion_tokens` を使用します。
+- Kimi K3の複数ターン・ツール呼び出しでは、`reasoning_content` と `tool_calls` を保持するため、返された `completion.assistant_message` を会話履歴へ追加してください。
 - Kimi K2.7 Code 系モデルはコーディング寄りで、thinking mode が必須のため、ツール使用時も `thinking` を有効のままにします。
 - Kimi K2.7 Code 系モデルで `thinking: { type: 'disabled' }` を明示指定すると、リクエスト送信前にエラーになります。
 - 旧 Kimi モデルでは、ツール使用時に `thinking` を `{ type: 'disabled' }` に強制します。
@@ -791,7 +831,7 @@ const sakanaService = ChatServiceFactory.createChatService('sakana', {
 - 対応モデル: `fugu`（デフォルト）, `fugu-ultra`, `fugu-ultra-20260615`
 - Sakanaは新しいChat Completions連携では`max_completion_tokens`を推奨していますが、legacyな`max_tokens`も受け付けます。このproviderでは既存OpenAI互換providerの挙動に合わせて`max_tokens`を使います。
 - Sakanaは最高性能にはResponses APIを推奨していますが、このproviderはpackageのOpenAI互換chat経路に合わせてChat Completionsを利用します。
-- ブラウザから直接呼ぶ場合、Sakana側がoriginに対してCORSを許可していないと失敗します。ブラウザ単体ではなく、Node.js、backend/serverless proxy、または`examples/sakana-basic`を利用してください。
+- ブラウザから直接呼ぶ場合、Sakana側がoriginに対してCORSを許可していないと失敗します。ブラウザ単体ではなく、Node.js、backend/serverless proxy、または`examples/node-basic/sakana-example.js`を利用してください。
 
 #### PLaMo
 
@@ -814,12 +854,33 @@ const plamoService = ChatServiceFactory.createChatService('plamo', {
 
 ```typescript
 const geminiNanoService = ChatServiceFactory.createChatService('gemini-nano', {
-  responseLength: 'medium'
+  responseLength: 'short',
+  initialPrompts: [
+    {
+      role: 'system',
+      content: 'あなたは明るく自然な話し方をするキャラクターです。'
+    },
+    { role: 'user', content: '今日はどんな気分？' },
+    { role: 'assistant', content: '元気いっぱいで、お話しするのが楽しみだよ！' },
+    { role: 'user', content: '準備できた？' },
+    { role: 'assistant', content: 'うん、いつでも始められるよ！' }
+  ]
 });
 ```
 
 注意:
 - APIキー不要 — ChromeのLanguageModel API（Prompt API）を使用します。
+- システム指示、設定した会話例、直近の会話履歴は、roleを維持した
+  `initialPrompts`として渡されます。
+- 設定した`initialPrompts`にsystemメッセージがある場合は、必ず先頭へ
+  正規化されます。会話例はオンデバイスモデルのコンテキストを消費するため、
+  短い例を使用してください。
+- Gemini Nanoでは、トークン数の目安に加えて、`veryShort`は最大1文、
+  `short`は最大2文、`medium`は最大3文、`long`は最大5文、
+  `veryLong`は最大10文となるよう具体的に指示します。`deep`は約5000
+  トークンの目安だけを残し、文数は制限しません。ただし、出力長は厳密保証
+  ではありません。
+- 構造化された履歴として、直近20件までのuser/assistantメッセージを渡します。
 - **Chrome 138以降**で、以下のフラグを有効にする必要があります:
   - `chrome://flags/#optimization-guide-on-device-model` → Enabled
   - `chrome://flags/#prompt-api-for-gemini-nano` → Enabled
@@ -828,6 +889,50 @@ const geminiNanoService = ChatServiceFactory.createChatService('gemini-nano', {
 - ビジョン（画像入力）は非対応です。
 - 初回のモデルダウンロードにはユーザー操作が必要で、数分かかる場合があります。
 
+##### Tips: 会話例で応答長の安定性を高める
+
+`responseLength`を設定すると、Gemini Nano向けの文数、書式、トークン数の
+目安が自動的に指示へ追加されます。さらに、短いuser/assistantの会話例を
+2〜3組ほど`initialPrompts`で渡すと、指示をより強く補強できます。
+
+```typescript
+import {
+  ChatServiceFactory,
+  type GeminiNanoInitialPrompt
+} from '@aituber-onair/chat';
+
+const veryShortExamples: GeminiNanoInitialPrompt[] = [
+  { role: 'user', content: '今日はどんな気分？' },
+  { role: 'assistant', content: 'みんなと話せて元気いっぱいだよ！' },
+  { role: 'user', content: '好きな食べ物は？' },
+  { role: 'assistant', content: '特にサーモンのお寿司が好き！' }
+];
+
+const service = ChatServiceFactory.createChatService('gemini-nano', {
+  responseLength: 'veryShort',
+  initialPrompts: veryShortExamples
+});
+```
+
+会話例は任意です。`initialPrompts`を省略しても、パッケージ側の応答長指示は
+適用されます。会話例を加えることで、期待する文章量、口調、リアクション、
+会話のテンポをオンデバイスモデルへ具体的に伝えられます。
+
+- assistantの例文は選択したプリセットに合わせてください。`veryShort`なら
+  1文、`short`なら最大2文にします。
+- 例文は対象キャラクターの口調で記述してください。汎用的な固定例文は
+  キャラクター性を変える可能性があるため、パッケージからは自動追加しません。
+- 明確に長い回答が必要な`medium`以上へ1文の例を流用すると、短い回答へ
+  偏る可能性があります。
+- 例文はシステムプロンプトや会話履歴と同じコンテキストを消費するため、
+  少数かつ短く保ってください。
+- `responseLength`または`initialPrompts`を変更した場合は、チャットサービスを
+  作り直してください。React Basic Exampleではこれを自動的に行い、短い例文は
+  `veryShort`と`short`だけに適用しています。
+- これは応答長の安定性を高める制御であり、厳密な出力上限ではありません。
+  ChromeのLanguageModel APIにはmax output tokens相当の設定がないため、
+  指定した文数を超える場合があります。
+
 ### ビジョンチャット
 
 組み込み provider のうち、対応モデル一覧を持つものは vision 対応を
@@ -1136,12 +1241,12 @@ vision、JSON mode、reasoning 設定を使うべきかを provider 固有ロジ
 
 - **OpenAI**: GPT-5.6（Sol/Terra/Luna）、GPT-5.5、GPT-5.4 Pro、GPT-5.4、GPT-5.4 Mini、GPT-5.4 Nano、GPT-5.1、GPT-5（Nano/Mini/Standard）、GPT-4.1(miniとnanoを含む), GPT-4, GPT-4o-mini, O3-mini, o1, o1-miniのモデルをサポート
 - **OpenAI-Compatible**: OpenAI互換 endpoint 経由で任意のローカル/セルフホスト model ID を利用できます。vision 対応可否は endpoint ごとに差があるため、原則 `unknown` 扱いです
-- **Gemini**: Gemini 3.5 Flash、Gemini 3.1 Flash-Lite、Gemini 3.1 Pro Preview、Gemini 3 Flash Preview、Gemini 2.5 Pro、Gemini 2.5 Flash、Gemini 2.5 Flash Lite、Gemma 4 31B IT、Gemma 4 26B A4B IT などの推奨モデルをサポート。Gemini 3.5 Flash はチャット用途向けに minimal thinking を自動適用します。Gemini 3.1 Flash-Lite Preview、Gemini 3 Pro Preview、Gemini 2.5 Flash Lite Preview などの lifecycle 上 deprecated なモデルは明示指定用に export を残しています
-- **Claude**: Claude Sonnet 5, Claude Opus 4.8, Claude Opus 4.7, Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.6, Claude Sonnet 4.5, Claude Haiku 4.5 に加え、まだ利用可能だが非推奨の Claude 4 Opus, Claude 4 Sonnet, Claude 3 Haiku をサポート
-- **OpenRouter**: OpenRouterのキュレーション済みモデル一覧（OpenAI/Claude/Gemini/Z.ai/Kimi）をサポート。モデルIDはOpenRouter節を参照してください
+- **Gemini**: Gemini 3.6 Flash、Gemini 3.5 Flash、Gemini 3.5 Flash-Lite、Gemini 3.1 Flash-Lite、Gemini 3.1 Pro Preview、Gemini 3 Flash Preview、Gemini 2.5 Pro、Gemini 2.5 Flash、Gemini 2.5 Flash Lite、Gemma 4 31B IT、Gemma 4 26B A4B IT などの推奨モデルをサポート。チャット用途向けに Gemini 3 Flash は minimal thinking、Gemini 3 Pro は low thinking を既定値にします。Gemini 3.1 Flash-Lite Preview、Gemini 3 Pro Preview、Gemini 2.5 Flash Lite Preview などの lifecycle 上 deprecated なモデルは明示指定用に export を残しています
+- **Claude**: Claude Opus 5, Claude Sonnet 5, Claude Opus 4.8, Claude Opus 4.7, Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.6, Claude Sonnet 4.5, Claude Haiku 4.5 に加え、まだ利用可能だが非推奨の Claude 4 Opus, Claude 4 Sonnet, Claude 3 Haiku をサポート。調整可能な `reasoning_effort` は対応モデルに限り `output_config.effort` として送信します
+- **OpenRouter**: OpenRouterのキュレーション済みモデル一覧（OpenAI/Claude/Gemini/Z.ai/xAI/Kimi/Kwaipilot）をサポート。モデルIDはOpenRouter節を参照してください
 - **Z.ai**: GLM-5.2/GLM-5.1/GLM-5/GLM-5-Turbo（テキスト）、GLM-4.7/4.6（テキスト）、GLM-5V-Turbo/GLM-4.6V系（ビジョン）をサポート
 - **xAI**: Grok 4.5 をチャット用途向けに `reasoning_effort: 'low'` デフォルトでサポート。加えて Grok 4.3、Grok 4.20 の Reasoning/Non-Reasoning、Grok 4-1 Fast の Reasoning/Non-Reasoning をサポートし、全モデルでビジョン対応
-- **Kimi**: Kimi K2.7 Code（`kimi-k2.7-code`）、Kimi K2.7 Code HighSpeed（`kimi-k2.7-code-highspeed`）、Kimi K2.6（`kimi-k2.6`、デフォルト）、Kimi K2.5（`kimi-k2.5`、いずれもビジョン対応）をサポート
+- **Kimi**: Kimi K3（`kimi-k3`、`low` / `high` / `max` reasoning、デフォルトは `max`）、Kimi K2.7 Code（`kimi-k2.7-code`）、Kimi K2.7 Code HighSpeed（`kimi-k2.7-code-highspeed`）、Kimi K2.6（`kimi-k2.6`、デフォルト）、Kimi K2.5（`kimi-k2.5`、いずれもビジョン対応）をサポート
 - **DeepSeek**: DeepSeek V4 Flash（`deepseek-v4-flash`）と DeepSeek V4 Pro（`deepseek-v4-pro`）をOpenAI互換Chat Completions経由でサポート。legacy alias の`deepseek-chat`と`deepseek-reasoner`はDeepSeek側で非推奨です
 - **Mistral**: Ministral 3系（`ministral-3b-2512`, `ministral-8b-2512`, `ministral-14b-2512`）と現行generalist modelをサポートし、streamingとvisionにも対応。adjustable `reasoning_effort`は対応モデルにだけ送信します
 - **Sakana AI**: Fugu（`fugu`）と Fugu Ultra（`fugu-ultra`, `fugu-ultra-20260615`）をOpenAI互換Chat Completions経由でサポート

package/README.md

@@ -585,30 +585,61 @@ For CI/local deterministic runs, pair it with `examples/mock-openai-server`.
 ```typescript
 const claudeService = ChatServiceFactory.createChatService('claude', {
   apiKey: process.env.ANTHROPIC_API_KEY,
-  model: 'claude-sonnet-5'
+  model: 'claude-opus-5',
+  reasoning_effort: 'low'
 });
 ```
 
+`claude-opus-5` is available as an explicit high-capability option and is not
+the package default. Claude Opus 5 has adaptive thinking enabled by default.
+Supported Claude models accept `reasoning_effort`, which maps to Anthropic's
+`output_config.effort`. The Claude API defaults to `high` when it is omitted:
+
+- Claude Opus 5, Sonnet 5, Opus 4.8, and Opus 4.7:
+  `low`, `medium`, `high`, `xhigh`, or `max`.
+- Claude Opus 4.6 and Sonnet 4.6:
+  `low`, `medium`, `high`, or `max`.
+- Claude Opus 4.5: `low`, `medium`, or `high`.
+
+Lower effort prioritizes latency and token efficiency, but does not guarantee
+a shorter visible response. Use response-length prompting as well when concise
+output is required.
+For multi-turn tool calls, append the returned
+`completion.assistant_message` to history so its provider-native thinking and
+tool-use blocks are preserved.
+
 #### Google Gemini
 
 ```typescript
 const geminiService = ChatServiceFactory.createChatService('gemini', {
   apiKey: process.env.GOOGLE_API_KEY,
-  model: 'gemini-3.1-flash-lite'
+  model: 'gemini-3.1-flash-lite',
+  reasoning_effort: 'minimal'
 });
 ```
 
-`gemini-3.1-flash-lite` is the recommended stable Flash-Lite model. Deprecated
+`gemini-3.1-flash-lite` remains the default Flash-Lite model.
+`gemini-3.5-flash-lite` is available as the latest stable, low-latency
+Flash-Lite option, while `gemini-3.6-flash` is available for stronger agentic
+and multimodal tasks. Deprecated
 preview and shutdown-scheduled models such as `gemini-3.1-flash-lite-preview`,
 `gemini-3-pro-preview`, and `gemini-2.5-flash-lite-preview-06-17` remain usable
 by explicit model string for backward compatibility, but are no longer
 advertised in the standard supported-model list for production use.
-`gemini-3.5-flash` is also available as the current stable Flash model.
+`gemini-3.5-flash` is also available as a stable Flash model.
+
+Gemini 3 models accept `reasoning_effort`, which maps to Gemini
+`thinkingConfig.thinkingLevel` while keeping `includeThoughts: false`:
+
+- Gemini 3 Flash / Flash-Lite: `minimal`, `low`, `medium`, or `high`;
+  defaults to `minimal` for low-latency chat.
+- Gemini 3 Pro: `low`, `medium`, or `high`; defaults to `low` because Pro does
+  not support `minimal`.
 
-For chat-style usage, `gemini-3.5-flash` automatically sends Gemini
-`thinkingConfig` with `thinkingLevel: 'MINIMAL'` and `includeThoughts: false`.
-No user-facing option is required; this keeps default chat responses fast and
-reduces hidden thinking token usage.
+This overrides the medium thinking default of Gemini 3.6/3.5 Flash and reduces
+the risk of hidden thinking exhausting short output limits. Gemini 2.5 uses
+`thinkingBudget` instead, so `reasoning_effort` is intentionally not sent for
+those models.
 
 #### OpenRouter
 
@@ -623,26 +654,36 @@ const openRouterService = ChatServiceFactory.createChatService('openrouter', {
 ```
 
 **Important Notes for OpenRouter:**
-- Token limits are automatically disabled for the `gpt-oss-20b:free` model due to technical limitations
-- To control response length, include instructions in your prompt (e.g., "Please respond in 40 characters or less")
+- Automatic token limits from `responseLength` are disabled for `openrouter/auto` and `openrouter/auto-beta` because a routed reasoning model can spend the entire output budget before emitting visible content. An explicitly supplied `maxTokens` is still honored.
+- All token limits remain disabled for `gpt-oss-20b:free` and `z-ai/glm-5.2`. For these models and the dynamic routers, control response length with prompt instructions (e.g., "Please respond in 40 characters or less").
 - Free tier has rate limits (20 requests/minute)
 - Free tier detection is based on the model ID suffix `:free` (dynamic `:free` IDs are also rate-limited)
+- `openrouter/auto-beta` is a Beta task-aware router. It chooses a model for each request and charges that routed model's rate; inspect the response `model` field or OpenRouter Activity to see the selection.
 - `openrouter/fusion` runs a multi-model panel plus a judge model; OpenRouter bills the sum of the underlying model calls and any enabled web search/fetch usage, not a single fixed model rate.
-- For `z-ai/glm-5.2`, OpenRouter reasoning defaults to `none` and token limits are automatically disabled because reasoning tokens can otherwise exhaust the output budget before visible content is emitted.
+- For `z-ai/glm-5.2`, OpenRouter reasoning also defaults to `none`.
+- Specialized coding models are explicit options rather than defaults. This includes `kwaipilot/kat-coder-air-v2.5` and `kwaipilot/kat-coder-pro-v2.5`, which are text-only.
+- `moonshotai/kimi-k3` availability depends on upstream capacity; OpenRouter may return 429 responses when capacity is constrained.
+- `x-ai/grok-4.5` has region-specific availability, including a current EU limitation. `~x-ai/grok-latest` inherits those limits when it resolves to Grok 4.5.
 - Supported models (curated list):
-  - `openrouter/auto`
+  - `openrouter/auto`, `openrouter/auto-beta`
   - `openrouter/fusion`
   - `openai/gpt-oss-20b:free`
-  - `~openai/gpt-latest`, `~openai/gpt-mini-latest`, `openai/gpt-5.5-pro`, `openai/gpt-5.5`
+  - `~openai/gpt-latest`, `~openai/gpt-mini-latest`
+  - `openai/gpt-5.6-sol`, `openai/gpt-5.6-terra`, `openai/gpt-5.6-luna`
+  - `openai/gpt-5.5-pro`, `openai/gpt-5.5`
   - `openai/gpt-5.1-chat`, `openai/gpt-5.1-codex`, `openai/gpt-5-mini`, `openai/gpt-5-nano`
   - `openai/gpt-4o`, `openai/gpt-4.1-mini`, `openai/gpt-4.1-nano`
   - `~anthropic/claude-sonnet-latest`, `~anthropic/claude-haiku-latest`
+  - `anthropic/claude-opus-5`
   - `anthropic/claude-opus-4`, `anthropic/claude-sonnet-4`
   - `anthropic/claude-3.7-sonnet`, `anthropic/claude-3.5-sonnet`, `anthropic/claude-haiku-4.5`
   - `~google/gemini-pro-latest`, `~google/gemini-flash-latest`
+  - `google/gemini-3.6-flash`, `google/gemini-3.5-flash-lite`
   - `google/gemini-2.5-pro`, `google/gemini-2.5-flash`, `google/gemini-2.5-flash-lite-preview-09-2025`
   - `z-ai/glm-5.2`, `z-ai/glm-4.7-flash`, `z-ai/glm-4.5-air`, `z-ai/glm-4.5-air:free`
-  - `~moonshotai/kimi-latest`, `moonshotai/kimi-k2.7-code`, `moonshotai/kimi-k2.5`
+  - `~x-ai/grok-latest`, `x-ai/grok-4.5`
+  - `~moonshotai/kimi-latest`, `moonshotai/kimi-k3`, `moonshotai/kimi-k2.7-code`, `moonshotai/kimi-k2.5`
+  - `kwaipilot/kat-coder-air-v2.5`, `kwaipilot/kat-coder-pro-v2.5`
 
 **Dynamic OpenRouter free model refresh**
 
@@ -710,18 +751,21 @@ Notes:
 ```typescript
 const kimiService = ChatServiceFactory.createChatService('kimi', {
   apiKey: process.env.MOONSHOT_API_KEY,
-  model: 'kimi-k2.6',
+  model: 'kimi-k3',
   // Optional: override endpoint or baseUrl
   // endpoint: 'https://api.moonshot.ai/v1/chat/completions',
   // baseUrl: 'https://api.moonshot.ai/v1',
-  thinking: { type: 'enabled' }
+  reasoning_effort: 'low'
 });
 ```
 
 Notes:
 - Kimi uses OpenAI-compatible Chat Completions.
-- Supported models: `kimi-k2.7-code`, `kimi-k2.7-code-highspeed`, `kimi-k2.6`, `kimi-k2.5`
+- Supported models: `kimi-k3`, `kimi-k2.7-code`, `kimi-k2.7-code-highspeed`, `kimi-k2.6`, `kimi-k2.5`
 - `kimi-k2.6` remains the default model for chat-oriented usage.
+- Kimi K3 is an explicit reasoning model. It accepts `reasoning_effort: 'low' | 'high' | 'max'`, defaults to `max`, always reasons, and does not accept the K2.x `thinking` option. `none`, `minimal`, and `medium` are not supported.
+- Kimi K3 uses `max_completion_tokens` for configured response limits.
+- For Kimi K3 multi-turn and tool-call flows, append the returned `completion.assistant_message` to history so `reasoning_content` and `tool_calls` are preserved.
 - Kimi K2.7 Code models are coding-oriented and require thinking mode, so they keep `thinking` enabled even when tools are used.
 - Explicitly setting `thinking: { type: 'disabled' }` with Kimi K2.7 Code models throws before sending the request.
 - For older Kimi models, when tools are enabled, `thinking` is forced to `{ type: 'disabled' }`.
@@ -804,7 +848,7 @@ Notes:
 - Supported models: `fugu` (default), `fugu-ultra`, and `fugu-ultra-20260615`.
 - Sakana recommends `max_completion_tokens` for new Chat Completions integrations, but also accepts legacy `max_tokens`. This provider keeps `max_tokens` to match existing OpenAI-compatible provider behavior.
 - Sakana recommends the Responses API for best performance, but this provider uses Chat Completions because it matches the package's OpenAI-compatible chat path.
-- Direct browser usage may fail with CORS unless Sakana enables CORS for your origin. Use Node.js, a backend/serverless proxy, or `examples/sakana-basic` instead of calling Sakana directly from browser-only apps.
+- Direct browser usage may fail with CORS unless Sakana enables CORS for your origin. Use Node.js, a backend/serverless proxy, or `examples/node-basic/sakana-example.js` instead of calling Sakana directly from browser-only apps.
 
 #### PLaMo
 
@@ -829,12 +873,34 @@ Notes:
 
 ```typescript
 const geminiNanoService = ChatServiceFactory.createChatService('gemini-nano', {
-  responseLength: 'medium'
+  responseLength: 'short',
+  initialPrompts: [
+    {
+      role: 'system',
+      content: 'You are a cheerful character who speaks naturally.'
+    },
+    { role: 'user', content: 'How are you feeling today?' },
+    { role: 'assistant', content: 'I feel great and ready to chat!' },
+    { role: 'user', content: 'Are you ready?' },
+    { role: 'assistant', content: 'Yes, we can start anytime!' }
+  ]
 });
 ```
 
 Notes:
 - No API key required — uses Chrome's built-in LanguageModel API (Prompt API).
+- System instructions, configured examples, and the most recent conversation
+  history are passed through `initialPrompts` with their roles preserved.
+- If configured `initialPrompts` contain a system message, it is normalized to
+  the first entry. Keep few-shot examples short because they consume the
+  on-device model's context window.
+- Gemini Nano uses concrete sentence-count guidance in addition to the soft
+  token budget: `veryShort` up to 1 sentence, `short` up to 2, `medium` up to
+  3, `long` up to 5, and `veryLong` up to 10. `deep` keeps its approximately
+  5000-token guidance without a sentence-count limit. Output length is still
+  best effort.
+- Up to the most recent 20 user/assistant messages are included as structured
+  history.
 - Requires **Chrome 138+** with the following flags enabled:
   - `chrome://flags/#optimization-guide-on-device-model` → Enabled
   - `chrome://flags/#prompt-api-for-gemini-nano` → Enabled
@@ -843,6 +909,54 @@ Notes:
 - Vision is not supported.
 - The initial model download requires a user action and may take a few minutes.
 
+##### Tip: improve response-length consistency with examples
+
+Setting `responseLength` automatically adds Gemini Nano-specific sentence-count,
+formatting, and soft token-budget instructions. You can reinforce those
+instructions by also passing two or three short user/assistant examples through
+`initialPrompts`:
+
+```typescript
+import {
+  ChatServiceFactory,
+  type GeminiNanoInitialPrompt
+} from '@aituber-onair/chat';
+
+const veryShortExamples: GeminiNanoInitialPrompt[] = [
+  { role: 'user', content: 'How are you feeling today?' },
+  { role: 'assistant', content: 'I feel great and ready to chat!' },
+  { role: 'user', content: 'What food do you like?' },
+  { role: 'assistant', content: 'Salmon sushi is my favorite!' }
+];
+
+const service = ChatServiceFactory.createChatService('gemini-nano', {
+  responseLength: 'veryShort',
+  initialPrompts: veryShortExamples
+});
+```
+
+The examples are optional; the package applies the length instruction even
+when `initialPrompts` is omitted. Examples help the on-device model learn the
+desired response size, speaking style, reaction level, and tempo more
+consistently.
+
+- Match every assistant example to the selected preset. For `veryShort`, use
+  one sentence; for `short`, use no more than two.
+- Write examples in the target character's voice. Hardcoded generic examples
+  are not added by the package because they could change the character's
+  personality.
+- Do not reuse one-sentence examples for `medium` or longer responses when you
+  want visibly longer answers; those examples can bias the model toward short
+  output.
+- Keep the set small because examples share the on-device context window with
+  the system prompt and conversation history.
+- Recreate the chat service when `responseLength` or `initialPrompts` changes.
+  The React basic example does this automatically and applies its short
+  examples only to `veryShort` and `short`.
+- These controls improve consistency but do not provide a strict output limit.
+  Chrome's LanguageModel API currently has no max-output-token option, so
+  Gemini Nano may occasionally exceed the requested sentence count.
+
 ### Vision Chat
 
 For built-in providers with curated model lists, the library pre-validates
@@ -1152,12 +1266,12 @@ Currently, the following AI providers are built-in:
 
 - **OpenAI**: Supports models like GPT-5.6 (Sol/Terra/Luna), GPT-5.5, GPT-5.4 Pro, GPT-5.4, GPT-5.4 Mini, GPT-5.4 Nano, GPT-5.1, GPT-5 (Nano/Mini/Standard), GPT-4.1 (including mini and nano), GPT-4, GPT-4o-mini, O3-mini, o1, o1-mini
 - **OpenAI-Compatible**: Supports arbitrary local/self-hosted model IDs via OpenAI-compatible endpoints. Vision capability is treated as `unknown` unless your app knows the endpoint-specific model catalog.
-- **Gemini**: Supports recommended models like Gemini 3.5 Flash, Gemini 3.1 Flash-Lite, Gemini 3.1 Pro Preview, Gemini 3 Flash Preview, Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.5 Flash Lite, Gemma 4 31B IT, and Gemma 4 26B A4B IT. Gemini 3.5 Flash automatically uses minimal thinking for chat-style responses. Deprecated lifecycle models such as Gemini 3.1 Flash-Lite Preview, Gemini 3 Pro Preview, and Gemini 2.5 Flash Lite Preview remain exported for explicit use.
-- **Claude**: Supports current Claude API model IDs including Claude Sonnet 5, Claude Opus 4.8, Claude Opus 4.7, Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.6, Claude Sonnet 4.5, Claude Haiku 4.5, plus deprecated-but-still-available Claude 4 Opus, Claude 4 Sonnet, and Claude 3 Haiku
-- **OpenRouter**: Supports a curated OpenRouter model list (OpenAI/Claude/Gemini/Z.ai/Kimi). See the OpenRouter section for model IDs.
+- **Gemini**: Supports recommended models like Gemini 3.6 Flash, Gemini 3.5 Flash, Gemini 3.5 Flash-Lite, Gemini 3.1 Flash-Lite, Gemini 3.1 Pro Preview, Gemini 3 Flash Preview, Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.5 Flash Lite, Gemma 4 31B IT, and Gemma 4 26B A4B IT. Gemini 3 Flash models default to minimal thinking for chat-style responses, while Gemini 3 Pro models default to low. Deprecated lifecycle models such as Gemini 3.1 Flash-Lite Preview, Gemini 3 Pro Preview, and Gemini 2.5 Flash Lite Preview remain exported for explicit use.
+- **Claude**: Supports current Claude API model IDs including Claude Opus 5, Claude Sonnet 5, Claude Opus 4.8, Claude Opus 4.7, Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 4.6, Claude Sonnet 4.5, Claude Haiku 4.5, plus deprecated-but-still-available Claude 4 Opus, Claude 4 Sonnet, and Claude 3 Haiku. Adjustable `reasoning_effort` is sent as `output_config.effort` only for models that support it
+- **OpenRouter**: Supports a curated OpenRouter model list (OpenAI/Claude/Gemini/Z.ai/xAI/Kimi/Kwaipilot). See the OpenRouter section for model IDs.
 - **Z.ai**: Supports GLM-5.2/GLM-5.1/GLM-5/GLM-5-Turbo (text), GLM-4.7/4.6 (text), and GLM-5V-Turbo/GLM-4.6V family (vision)
 - **xAI**: Supports Grok 4.5 with `reasoning_effort: 'low'` by default for chat-style responses, plus Grok 4.3, Grok 4.20 Reasoning/Non-Reasoning, and Grok 4-1 Fast Reasoning/Non-Reasoning, all with vision support.
-- **Kimi**: Supports Kimi K2.7 Code (`kimi-k2.7-code`), Kimi K2.7 Code HighSpeed (`kimi-k2.7-code-highspeed`), Kimi K2.6 (`kimi-k2.6`, default), and Kimi K2.5 (`kimi-k2.5`) with vision support
+- **Kimi**: Supports Kimi K3 (`kimi-k3`, `low` / `high` / `max` reasoning with `max` as the default), Kimi K2.7 Code (`kimi-k2.7-code`), Kimi K2.7 Code HighSpeed (`kimi-k2.7-code-highspeed`), Kimi K2.6 (`kimi-k2.6`, default), and Kimi K2.5 (`kimi-k2.5`) with vision support
 - **DeepSeek**: Supports DeepSeek V4 Flash (`deepseek-v4-flash`) and DeepSeek V4 Pro (`deepseek-v4-pro`) via OpenAI-compatible Chat Completions. Legacy aliases `deepseek-chat` and `deepseek-reasoner` are deprecated by DeepSeek.
 - **Mistral**: Supports the Ministral 3 family (`ministral-3b-2512`, `ministral-8b-2512`, `ministral-14b-2512`) and current Mistral generalist models, with streaming and vision support. Adjustable `reasoning_effort` is only sent for supported models.
 - **Sakana AI**: Supports Fugu (`fugu`) and Fugu Ultra (`fugu-ultra`, `fugu-ultra-20260615`) via OpenAI-compatible Chat Completions.

package/dist/cjs/constants/claude.d.ts

@@ -13,5 +13,20 @@ export declare const MODEL_CLAUDE_4_6_OPUS = "claude-opus-4-6";
 export declare const MODEL_CLAUDE_4_7_OPUS = "claude-opus-4-7";
 export declare const MODEL_CLAUDE_4_8_OPUS = "claude-opus-4-8";
 export declare const MODEL_CLAUDE_5_SONNET = "claude-sonnet-5";
+export declare const MODEL_CLAUDE_5_OPUS = "claude-opus-5";
+export type ClaudeReasoningEffort = 'low' | 'medium' | 'high' | 'xhigh' | 'max';
 export declare const CLAUDE_VISION_SUPPORTED_MODELS: string[];
+/**
+ * Return the effort values supported by a Claude model.
+ */
+export declare function getClaudeSupportedReasoningEfforts(model: string): readonly ClaudeReasoningEffort[];
+export declare function isClaudeReasoningEffortModel(model: string): boolean;
+/**
+ * Claude API defaults to high effort when output_config.effort is omitted.
+ */
+export declare function getDefaultClaudeReasoningEffort(model: string): ClaudeReasoningEffort | undefined;
+/**
+ * Normalize UI state when switching between Claude models.
+ */
+export declare function normalizeClaudeReasoningEffort(model: string, effort?: ClaudeReasoningEffort): ClaudeReasoningEffort | undefined;
 //# sourceMappingURL=claude.d.ts.map

package/dist/cjs/constants/claude.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"claude.d.ts","sourceRoot":"","sources":["../../../src/constants/claude.ts"],"names":[],"mappings":"AAAA,eAAO,MAAM,mBAAmB,0CAA0C,CAAC;AAG3E,eAAO,MAAM,oBAAoB,4BAA4B,CAAC;AAC9D,eAAO,MAAM,sBAAsB,8BAA8B,CAAC;AAClE,eAAO,MAAM,uBAAuB,+BAA+B,CAAC;AACpE,eAAO,MAAM,uBAAuB,+BAA+B,CAAC;AACpE,eAAO,MAAM,qBAAqB,6BAA6B,CAAC;AAChE,eAAO,MAAM,mBAAmB,2BAA2B,CAAC;AAC5D,eAAO,MAAM,uBAAuB,+BAA+B,CAAC;AACpE,eAAO,MAAM,sBAAsB,8BAA8B,CAAC;AAClE,eAAO,MAAM,qBAAqB,6BAA6B,CAAC;AAChE,eAAO,MAAM,uBAAuB,sBAAsB,CAAC;AAC3D,eAAO,MAAM,qBAAqB,oBAAoB,CAAC;AACvD,eAAO,MAAM,qBAAqB,oBAAoB,CAAC;AACvD,eAAO,MAAM,qBAAqB,oBAAoB,CAAC;AACvD,eAAO,MAAM,qBAAqB,oBAAoB,CAAC;AAEvD,eAAO,MAAM,8BAA8B,UAY1C,CAAC"}
+{"version":3,"file":"claude.d.ts","sourceRoot":"","sources":["../../../src/constants/claude.ts"],"names":[],"mappings":"AAAA,eAAO,MAAM,mBAAmB,0CAA0C,CAAC;AAG3E,eAAO,MAAM,oBAAoB,4BAA4B,CAAC;AAC9D,eAAO,MAAM,sBAAsB,8BAA8B,CAAC;AAClE,eAAO,MAAM,uBAAuB,+BAA+B,CAAC;AACpE,eAAO,MAAM,uBAAuB,+BAA+B,CAAC;AACpE,eAAO,MAAM,qBAAqB,6BAA6B,CAAC;AAChE,eAAO,MAAM,mBAAmB,2BAA2B,CAAC;AAC5D,eAAO,MAAM,uBAAuB,+BAA+B,CAAC;AACpE,eAAO,MAAM,sBAAsB,8BAA8B,CAAC;AAClE,eAAO,MAAM,qBAAqB,6BAA6B,CAAC;AAChE,eAAO,MAAM,uBAAuB,sBAAsB,CAAC;AAC3D,eAAO,MAAM,qBAAqB,oBAAoB,CAAC;AACvD,eAAO,MAAM,qBAAqB,oBAAoB,CAAC;AACvD,eAAO,MAAM,qBAAqB,oBAAoB,CAAC;AACvD,eAAO,MAAM,qBAAqB,oBAAoB,CAAC;AACvD,eAAO,MAAM,mBAAmB,kBAAkB,CAAC;AAEnD,MAAM,MAAM,qBAAqB,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;AAgChF,eAAO,MAAM,8BAA8B,UAa1C,CAAC;AAEF;;GAEG;AACH,wBAAgB,kCAAkC,CAChD,KAAK,EAAE,MAAM,GACZ,SAAS,qBAAqB,EAAE,CAElC;AAED,wBAAgB,4BAA4B,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAEnE;AAED;;GAEG;AACH,wBAAgB,+BAA+B,CAC7C,KAAK,EAAE,MAAM,GACZ,qBAAqB,GAAG,SAAS,CAEnC;AAED;;GAEG;AACH,wBAAgB,8BAA8B,CAC5C,KAAK,EAAE,MAAM,EACb,MAAM,CAAC,EAAE,qBAAqB,GAC7B,qBAAqB,GAAG,SAAS,CAenC"}

package/dist/cjs/constants/claude.js

@@ -1,6 +1,10 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.CLAUDE_VISION_SUPPORTED_MODELS = exports.MODEL_CLAUDE_5_SONNET = exports.MODEL_CLAUDE_4_8_OPUS = exports.MODEL_CLAUDE_4_7_OPUS = exports.MODEL_CLAUDE_4_6_OPUS = exports.MODEL_CLAUDE_4_6_SONNET = exports.MODEL_CLAUDE_4_5_OPUS = exports.MODEL_CLAUDE_4_5_HAIKU = exports.MODEL_CLAUDE_4_5_SONNET = exports.MODEL_CLAUDE_4_OPUS = exports.MODEL_CLAUDE_4_SONNET = exports.MODEL_CLAUDE_3_7_SONNET = exports.MODEL_CLAUDE_3_5_SONNET = exports.MODEL_CLAUDE_3_5_HAIKU = exports.MODEL_CLAUDE_3_HAIKU = exports.ENDPOINT_CLAUDE_API = void 0;
+exports.CLAUDE_VISION_SUPPORTED_MODELS = exports.MODEL_CLAUDE_5_OPUS = exports.MODEL_CLAUDE_5_SONNET = exports.MODEL_CLAUDE_4_8_OPUS = exports.MODEL_CLAUDE_4_7_OPUS = exports.MODEL_CLAUDE_4_6_OPUS = exports.MODEL_CLAUDE_4_6_SONNET = exports.MODEL_CLAUDE_4_5_OPUS = exports.MODEL_CLAUDE_4_5_HAIKU = exports.MODEL_CLAUDE_4_5_SONNET = exports.MODEL_CLAUDE_4_OPUS = exports.MODEL_CLAUDE_4_SONNET = exports.MODEL_CLAUDE_3_7_SONNET = exports.MODEL_CLAUDE_3_5_SONNET = exports.MODEL_CLAUDE_3_5_HAIKU = exports.MODEL_CLAUDE_3_HAIKU = exports.ENDPOINT_CLAUDE_API = void 0;
+exports.getClaudeSupportedReasoningEfforts = getClaudeSupportedReasoningEfforts;
+exports.isClaudeReasoningEffortModel = isClaudeReasoningEffortModel;
+exports.getDefaultClaudeReasoningEffort = getDefaultClaudeReasoningEffort;
+exports.normalizeClaudeReasoningEffort = normalizeClaudeReasoningEffort;
 exports.ENDPOINT_CLAUDE_API = 'https://api.anthropic.com/v1/messages';
 // claude model
 exports.MODEL_CLAUDE_3_HAIKU = 'claude-3-haiku-20240307';
@@ -17,6 +21,30 @@ exports.MODEL_CLAUDE_4_6_OPUS = 'claude-opus-4-6';
 exports.MODEL_CLAUDE_4_7_OPUS = 'claude-opus-4-7';
 exports.MODEL_CLAUDE_4_8_OPUS = 'claude-opus-4-8';
 exports.MODEL_CLAUDE_5_SONNET = 'claude-sonnet-5';
+exports.MODEL_CLAUDE_5_OPUS = 'claude-opus-5';
+const CLAUDE_STANDARD_REASONING_EFFORTS = [
+    'low',
+    'medium',
+    'high',
+];
+const CLAUDE_MAX_REASONING_EFFORTS = [
+    ...CLAUDE_STANDARD_REASONING_EFFORTS,
+    'max',
+];
+const CLAUDE_XHIGH_REASONING_EFFORTS = [
+    ...CLAUDE_STANDARD_REASONING_EFFORTS,
+    'xhigh',
+    'max',
+];
+const CLAUDE_REASONING_EFFORTS_BY_MODEL = {
+    [exports.MODEL_CLAUDE_4_5_OPUS]: CLAUDE_STANDARD_REASONING_EFFORTS,
+    [exports.MODEL_CLAUDE_4_6_SONNET]: CLAUDE_MAX_REASONING_EFFORTS,
+    [exports.MODEL_CLAUDE_4_6_OPUS]: CLAUDE_MAX_REASONING_EFFORTS,
+    [exports.MODEL_CLAUDE_4_7_OPUS]: CLAUDE_XHIGH_REASONING_EFFORTS,
+    [exports.MODEL_CLAUDE_4_8_OPUS]: CLAUDE_XHIGH_REASONING_EFFORTS,
+    [exports.MODEL_CLAUDE_5_SONNET]: CLAUDE_XHIGH_REASONING_EFFORTS,
+    [exports.MODEL_CLAUDE_5_OPUS]: CLAUDE_XHIGH_REASONING_EFFORTS,
+};
 exports.CLAUDE_VISION_SUPPORTED_MODELS = [
     exports.MODEL_CLAUDE_3_HAIKU,
     exports.MODEL_CLAUDE_4_SONNET,
@@ -29,5 +57,37 @@ exports.CLAUDE_VISION_SUPPORTED_MODELS = [
     exports.MODEL_CLAUDE_4_7_OPUS,
     exports.MODEL_CLAUDE_4_8_OPUS,
     exports.MODEL_CLAUDE_5_SONNET,
+    exports.MODEL_CLAUDE_5_OPUS,
 ];
+/**
+ * Return the effort values supported by a Claude model.
+ */
+function getClaudeSupportedReasoningEfforts(model) {
+    return CLAUDE_REASONING_EFFORTS_BY_MODEL[model] ?? [];
+}
+function isClaudeReasoningEffortModel(model) {
+    return getClaudeSupportedReasoningEfforts(model).length > 0;
+}
+/**
+ * Claude API defaults to high effort when output_config.effort is omitted.
+ */
+function getDefaultClaudeReasoningEffort(model) {
+    return isClaudeReasoningEffortModel(model) ? 'high' : undefined;
+}
+/**
+ * Normalize UI state when switching between Claude models.
+ */
+function normalizeClaudeReasoningEffort(model, effort) {
+    const supported = getClaudeSupportedReasoningEfforts(model);
+    if (supported.length === 0) {
+        return undefined;
+    }
+    if (!effort) {
+        return getDefaultClaudeReasoningEffort(model);
+    }
+    if (supported.includes(effort)) {
+        return effort;
+    }
+    return 'high';
+}
 //# sourceMappingURL=claude.js.map

package/dist/cjs/constants/claude.js.map

@@ -1 +1 @@
-{"version":3,"file":"claude.js","sourceRoot":"","sources":["../../../src/constants/claude.ts"],"names":[],"mappings":";;;AAAa,QAAA,mBAAmB,GAAG,uCAAuC,CAAC;AAE3E,eAAe;AACF,QAAA,oBAAoB,GAAG,yBAAyB,CAAC;AACjD,QAAA,sBAAsB,GAAG,2BAA2B,CAAC;AACrD,QAAA,uBAAuB,GAAG,4BAA4B,CAAC;AACvD,QAAA,uBAAuB,GAAG,4BAA4B,CAAC;AACvD,QAAA,qBAAqB,GAAG,0BAA0B,CAAC;AACnD,QAAA,mBAAmB,GAAG,wBAAwB,CAAC;AAC/C,QAAA,uBAAuB,GAAG,4BAA4B,CAAC;AACvD,QAAA,sBAAsB,GAAG,2BAA2B,CAAC;AACrD,QAAA,qBAAqB,GAAG,0BAA0B,CAAC;AACnD,QAAA,uBAAuB,GAAG,mBAAmB,CAAC;AAC9C,QAAA,qBAAqB,GAAG,iBAAiB,CAAC;AAC1C,QAAA,qBAAqB,GAAG,iBAAiB,CAAC;AAC1C,QAAA,qBAAqB,GAAG,iBAAiB,CAAC;AAC1C,QAAA,qBAAqB,GAAG,iBAAiB,CAAC;AAE1C,QAAA,8BAA8B,GAAG;IAC5C,4BAAoB;IACpB,6BAAqB;IACrB,2BAAmB;IACnB,+BAAuB;IACvB,8BAAsB;IACtB,6BAAqB;IACrB,+BAAuB;IACvB,6BAAqB;IACrB,6BAAqB;IACrB,6BAAqB;IACrB,6BAAqB;CACtB,CAAC"}
+{"version":3,"file":"claude.js","sourceRoot":"","sources":["../../../src/constants/claude.ts"],"names":[],"mappings":";;;AAqEA,gFAIC;AAED,oEAEC;AAKD,0EAIC;AAKD,wEAkBC;AA7GY,QAAA,mBAAmB,GAAG,uCAAuC,CAAC;AAE3E,eAAe;AACF,QAAA,oBAAoB,GAAG,yBAAyB,CAAC;AACjD,QAAA,sBAAsB,GAAG,2BAA2B,CAAC;AACrD,QAAA,uBAAuB,GAAG,4BAA4B,CAAC;AACvD,QAAA,uBAAuB,GAAG,4BAA4B,CAAC;AACvD,QAAA,qBAAqB,GAAG,0BAA0B,CAAC;AACnD,QAAA,mBAAmB,GAAG,wBAAwB,CAAC;AAC/C,QAAA,uBAAuB,GAAG,4BAA4B,CAAC;AACvD,QAAA,sBAAsB,GAAG,2BAA2B,CAAC;AACrD,QAAA,qBAAqB,GAAG,0BAA0B,CAAC;AACnD,QAAA,uBAAuB,GAAG,mBAAmB,CAAC;AAC9C,QAAA,qBAAqB,GAAG,iBAAiB,CAAC;AAC1C,QAAA,qBAAqB,GAAG,iBAAiB,CAAC;AAC1C,QAAA,qBAAqB,GAAG,iBAAiB,CAAC;AAC1C,QAAA,qBAAqB,GAAG,iBAAiB,CAAC;AAC1C,QAAA,mBAAmB,GAAG,eAAe,CAAC;AAInD,MAAM,iCAAiC,GAAqC;IAC1E,KAAK;IACL,QAAQ;IACR,MAAM;CACP,CAAC;AAEF,MAAM,4BAA4B,GAAqC;IACrE,GAAG,iCAAiC;IACpC,KAAK;CACN,CAAC;AAEF,MAAM,8BAA8B,GAAqC;IACvE,GAAG,iCAAiC;IACpC,OAAO;IACP,KAAK;CACN,CAAC;AAEF,MAAM,iCAAiC,GAGnC;IACF,CAAC,6BAAqB,CAAC,EAAE,iCAAiC;IAC1D,CAAC,+BAAuB,CAAC,EAAE,4BAA4B;IACvD,CAAC,6BAAqB,CAAC,EAAE,4BAA4B;IACrD,CAAC,6BAAqB,CAAC,EAAE,8BAA8B;IACvD,CAAC,6BAAqB,CAAC,EAAE,8BAA8B;IACvD,CAAC,6BAAqB,CAAC,EAAE,8BAA8B;IACvD,CAAC,2BAAmB,CAAC,EAAE,8BAA8B;CACtD,CAAC;AAEW,QAAA,8BAA8B,GAAG;IAC5C,4BAAoB;IACpB,6BAAqB;IACrB,2BAAmB;IACnB,+BAAuB;IACvB,8BAAsB;IACtB,6BAAqB;IACrB,+BAAuB;IACvB,6BAAqB;IACrB,6BAAqB;IACrB,6BAAqB;IACrB,6BAAqB;IACrB,2BAAmB;CACpB,CAAC;AAEF;;GAEG;AACH,SAAgB,kCAAkC,CAChD,KAAa;IAEb,OAAO,iCAAiC,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;AACxD,CAAC;AAED,SAAgB,4BAA4B,CAAC,KAAa;IACxD,OAAO,kCAAkC,CAAC,KAAK,CAAC,CAAC,MAAM,GAAG,CAAC,CAAC;AAC9D,CAAC;AAED;;GAEG;AACH,SAAgB,+BAA+B,CAC7C,KAAa;IAEb,OAAO,4BAA4B,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,SAAS,CAAC;AAClE,CAAC;AAED;;GAEG;AACH,SAAgB,8BAA8B,CAC5C,KAAa,EACb,MAA8B;IAE9B,MAAM,SAAS,GAAG,kCAAkC,CAAC,KAAK,CAAC,CAAC;IAC5D,IAAI,SAAS,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QAC3B,OAAO,SAAS,CAAC;IACnB,CAAC;IAED,IAAI,CAAC,MAAM,EAAE,CAAC;QACZ,OAAO,+BAA+B,CAAC,KAAK,CAAC,CAAC;IAChD,CAAC;IAED,IAAI,SAAS,CAAC,QAAQ,CAAC,MAAM,CAAC,EAAE,CAAC;QAC/B,OAAO,MAAM,CAAC;IAChB,CAAC;IAED,OAAO,MAAM,CAAC;AAChB,CAAC"}