mocode-ai 1.2.6 → 1.2.7

package/dist/llm/providers/anthropic.js

@@ -0,0 +1,370 @@
+import { config, getActiveModel } from '../../config/index.js';
+let fetchImplOverride = null;
+/** 仅供单测注入；生产路径使用 Node 18+ 全局 fetch。 */
+export function __setAnthropicFetchImpl(impl) {
+    fetchImplOverride = impl;
+}
+function parseJsonObject(value) {
+    if (!value.trim())
+        return {};
+    try {
+        const parsed = JSON.parse(value);
+        return parsed && typeof parsed === 'object' && !Array.isArray(parsed)
+            ? parsed
+            : {};
+    }
+    catch {
+        return {};
+    }
+}
+function textBlocks(content) {
+    if (content == null)
+        return [];
+    if (typeof content === 'string')
+        return content ? [{ type: 'text', text: content }] : [];
+    if (!Array.isArray(content))
+        return [{ type: 'text', text: String(content) }];
+    const blocks = [];
+    for (const raw of content) {
+        if (!raw || typeof raw !== 'object')
+            continue;
+        const part = raw;
+        if (part.type === 'text' && typeof part.text === 'string') {
+            blocks.push({ type: 'text', text: part.text });
+            continue;
+        }
+        if (part.type !== 'image_url')
+            continue;
+        const image = part.image_url;
+        if (typeof image?.url !== 'string')
+            continue;
+        const match = /^data:(image\/(?:jpeg|png|gif|webp));base64,(.+)$/s.exec(image.url);
+        if (!match)
+            continue;
+        blocks.push({
+            type: 'image',
+            source: { type: 'base64', media_type: match[1], data: match[2] },
+        });
+    }
+    return blocks;
+}
+function appendMessage(messages, role, blocks, cacheCandidates, cacheable = true) {
+    if (blocks.length === 0)
+        return;
+    const previous = messages[messages.length - 1];
+    if (previous?.role === role)
+        previous.content.push(...blocks);
+    else
+        messages.push({ role, content: [...blocks] });
+    if (cacheable)
+        cacheCandidates.push(...blocks);
+}
+/**
+ * 把会话落盘使用的 OpenAI 兼容消息转换为 Anthropic Messages 内容块。
+ * history 保持旧格式以兼容已有 session；provider 边界负责 tool_use/tool_result 编码。
+ */
+export function encodeAnthropicMessages(input, promptCache = config.anthropicPromptCache) {
+    const system = [];
+    const messages = [];
+    const cacheCandidates = [];
+    let dialogStarted = false;
+    for (const raw of input) {
+        const message = raw;
+        if (message.role === 'system' && !dialogStarted) {
+            system.push(...textBlocks(message.content));
+            continue;
+        }
+        dialogStarted = true;
+        if (message.role === 'system') {
+            const blocks = textBlocks(message.content).map((block) => ({
+                ...block,
+                text: `[System reminder]\n${String(block.text ?? '')}`,
+            }));
+            // 动态尾部 reminder 每步可能变化，不把 cache breakpoint 放在它之后。
+            appendMessage(messages, 'user', blocks, cacheCandidates, false);
+            continue;
+        }
+        if (message.role === 'user') {
+            appendMessage(messages, 'user', textBlocks(message.content), cacheCandidates);
+            continue;
+        }
+        if (message.role === 'assistant') {
+            const blocks = textBlocks(message.content);
+            for (const call of message.tool_calls ?? []) {
+                blocks.push({
+                    type: 'tool_use',
+                    id: call.id ?? '',
+                    name: call.function?.name ?? '',
+                    input: parseJsonObject(call.function?.arguments ?? ''),
+                });
+            }
+            appendMessage(messages, 'assistant', blocks, cacheCandidates);
+            continue;
+        }
+        if (message.role === 'tool') {
+            appendMessage(messages, 'user', [{
+                    type: 'tool_result',
+                    tool_use_id: message.tool_call_id ?? '',
+                    content: typeof message.content === 'string'
+                        ? message.content
+                        : JSON.stringify(message.content ?? ''),
+                }], cacheCandidates);
+        }
+    }
+    if (promptCache) {
+        // 最后一个稳定消息块形成会话前缀断点；动态 session reminder 明确不参与。
+        const last = cacheCandidates[cacheCandidates.length - 1];
+        if (last)
+            last.cache_control = { type: 'ephemeral' };
+        // 没有对话时仍缓存稳定 system prompt。
+        else if (system.length > 0)
+            system[system.length - 1].cache_control = { type: 'ephemeral' };
+    }
+    return { system, messages };
+}
+export function encodeAnthropicTools(tools, promptCache = config.anthropicPromptCache) {
+    const encoded = tools.map((tool) => ({
+        name: tool.function.name,
+        description: tool.function.description,
+        input_schema: tool.function.parameters ?? { type: 'object', properties: {} },
+    }));
+    if (promptCache && encoded.length > 0) {
+        encoded[encoded.length - 1] = {
+            ...encoded[encoded.length - 1],
+            cache_control: { type: 'ephemeral' },
+        };
+    }
+    return encoded;
+}
+function endpoint(baseURL) {
+    const base = baseURL.replace(/\/+$/, '');
+    if (/\/v1$/i.test(base))
+        return `${base}/messages`;
+    if (/\/v1\/messages$/i.test(base))
+        return base;
+    return `${base}/v1/messages`;
+}
+export function buildAnthropicRequest(messages, tools) {
+    const encoded = encodeAnthropicMessages(messages);
+    const anthropicTools = encodeAnthropicTools(tools);
+    return {
+        model: getActiveModel(),
+        max_tokens: config.maxTokens ?? 8192,
+        stream: true,
+        system: encoded.system,
+        messages: encoded.messages,
+        ...(anthropicTools.length > 0 ? { tools: anthropicTools } : {}),
+    };
+}
+class AnthropicHttpError extends Error {
+    status;
+    code;
+    headers;
+    constructor(status, message, headers, code) {
+        super(message);
+        this.name = 'AnthropicAPIError';
+        this.status = status;
+        this.code = code;
+        this.headers = headers;
+    }
+}
+async function throwResponseError(response) {
+    let message = `Anthropic API HTTP ${response.status}`;
+    let code;
+    try {
+        const body = JSON.parse(await response.text());
+        if (typeof body.error?.message === 'string')
+            message = body.error.message;
+        if (typeof body.error?.type === 'string')
+            code = body.error.type;
+    }
+    catch {
+        // 非 JSON 错误页只保留状态码，避免把代理 HTML 大段写进 TUI。
+    }
+    throw new AnthropicHttpError(response.status, message, response.headers, code);
+}
+function decodeSseRecord(record) {
+    let event = '';
+    const data = [];
+    for (const line of record.split(/\r?\n/)) {
+        if (line.startsWith('event:'))
+            event = line.slice(6).trim();
+        else if (line.startsWith('data:'))
+            data.push(line.slice(5).trimStart());
+    }
+    if (data.length === 0)
+        return null;
+    return { event, data: data.join('\n') };
+}
+async function* readSse(response) {
+    if (!response.body)
+        throw new Error('Anthropic 流式响应缺少 body');
+    const reader = response.body.getReader();
+    const decoder = new TextDecoder();
+    let buffer = '';
+    while (true) {
+        const { value, done } = await reader.read();
+        buffer += decoder.decode(value, { stream: !done });
+        let match;
+        while ((match = /\r?\n\r?\n/.exec(buffer)) !== null) {
+            const record = buffer.slice(0, match.index);
+            buffer = buffer.slice(match.index + match[0].length);
+            const parsed = decodeSseRecord(record);
+            if (parsed)
+                yield parsed;
+        }
+        if (done)
+            break;
+    }
+    const parsed = decodeSseRecord(buffer);
+    if (parsed)
+        yield parsed;
+}
+function finiteToken(value) {
+    return typeof value === 'number' && Number.isFinite(value) && value >= 0 ? value : 0;
+}
+function usageFromAnthropic(raw) {
+    const direct = finiteToken(raw.input_tokens);
+    const cacheRead = finiteToken(raw.cache_read_input_tokens);
+    const cacheCreation = finiteToken(raw.cache_creation_input_tokens);
+    const output = finiteToken(raw.output_tokens);
+    const prompt = direct + cacheRead + cacheCreation;
+    return {
+        promptTokens: prompt,
+        completionTokens: output,
+        totalTokens: prompt + output,
+        cachedTokens: cacheRead,
+        cacheCreationTokens: cacheCreation,
+        reasoningTokens: 0,
+    };
+}
+function addLiveCount(text, state) {
+    for (const ch of text) {
+        const cp = ch.codePointAt(0) ?? 0;
+        if ((cp >= 0x4e00 && cp <= 0x9fff)
+            || (cp >= 0x3400 && cp <= 0x4dbf)
+            || (cp >= 0x3040 && cp <= 0x30ff)
+            || (cp >= 0xac00 && cp <= 0xd7a3))
+            state.cjk++;
+        else
+            state.other++;
+    }
+}
+/** Anthropic Messages API 单次请求；外层 chat() 继续统一负责重试。 */
+export async function anthropicChatOnce(messages, handlers, signal, tools) {
+    const fetchImpl = fetchImplOverride ?? fetch;
+    const response = await fetchImpl(endpoint(config.baseURL), {
+        method: 'POST',
+        headers: {
+            'content-type': 'application/json',
+            accept: 'text/event-stream',
+            'x-api-key': config.apiKey,
+            'anthropic-version': process.env.ANTHROPIC_VERSION || '2023-06-01',
+        },
+        body: JSON.stringify(buildAnthropicRequest(messages, tools)),
+        signal,
+    });
+    if (!response.ok)
+        await throwResponseError(response);
+    let content = '';
+    let usage;
+    const toolAcc = new Map();
+    const live = { cjk: 0, other: 0 };
+    const reportLive = () => {
+        handlers.onProgress?.({ completionTokens: Math.ceil(live.cjk + live.other / 4) });
+    };
+    for await (const record of readSse(response)) {
+        if (record.data === '[DONE]')
+            continue;
+        let event;
+        try {
+            event = JSON.parse(record.data);
+        }
+        catch {
+            continue;
+        }
+        const type = typeof event.type === 'string' ? event.type : record.event;
+        if (type === 'error') {
+            const error = event.error;
+            const thrown = new Error(typeof error?.message === 'string' ? error.message : 'Anthropic stream error');
+            thrown.name = typeof error?.type === 'string' ? error.type : 'AnthropicStreamError';
+            throw thrown;
+        }
+        if (type === 'message_start') {
+            const message = event.message;
+            if (message?.usage)
+                usage = usageFromAnthropic(message.usage);
+            continue;
+        }
+        if (type === 'content_block_start') {
+            const index = finiteToken(event.index);
+            const block = event.content_block;
+            if (block?.type === 'text' && typeof block.text === 'string' && block.text) {
+                content += block.text;
+                addLiveCount(block.text, live);
+                handlers.onText?.(block.text);
+                reportLive();
+            }
+            else if (block?.type === 'tool_use') {
+                const initialInput = block.input && typeof block.input === 'object'
+                    ? block.input
+                    : undefined;
+                const call = {
+                    id: typeof block.id === 'string' ? block.id : '',
+                    name: typeof block.name === 'string' ? block.name : '',
+                    // 流式 tool_use 通常以 input:{} 开始，真正 JSON 紧随 input_json_delta；
+                    // 只有非空 input 才直接采用，避免累加成 `{}{...}`。
+                    arguments: initialInput && Object.keys(initialInput).length > 0
+                        ? JSON.stringify(initialInput)
+                        : '',
+                };
+                toolAcc.set(index, call);
+                if (call.name)
+                    handlers.onToolCall?.(call.name);
+            }
+            continue;
+        }
+        if (type === 'content_block_delta') {
+            const index = finiteToken(event.index);
+            const delta = event.delta;
+            if (delta?.type === 'text_delta' && typeof delta.text === 'string') {
+                content += delta.text;
+                addLiveCount(delta.text, live);
+                handlers.onText?.(delta.text);
+                reportLive();
+            }
+            else if (delta?.type === 'input_json_delta' && typeof delta.partial_json === 'string') {
+                const call = toolAcc.get(index) ?? { id: '', name: '', arguments: '' };
+                call.arguments += delta.partial_json;
+                toolAcc.set(index, call);
+                addLiveCount(delta.partial_json, live);
+                reportLive();
+            }
+            continue;
+        }
+        if (type === 'message_delta') {
+            const raw = event.usage;
+            if (raw) {
+                const previous = usage ?? usageFromAnthropic({});
+                const output = finiteToken(raw.output_tokens);
+                usage = {
+                    ...previous,
+                    completionTokens: output,
+                    totalTokens: previous.promptTokens + output,
+                };
+            }
+        }
+    }
+    if (usage) {
+        handlers.onProgress?.({
+            completionTokens: usage.completionTokens,
+            promptTokens: usage.promptTokens,
+            cachedTokens: usage.cachedTokens,
+            cacheCreationTokens: usage.cacheCreationTokens,
+        });
+    }
+    const toolCalls = [...toolAcc.entries()]
+        .sort((a, b) => a[0] - b[0])
+        .map(([, call]) => ({ ...call, arguments: call.arguments || '{}' }));
+    return { content: content || null, toolCalls, usage };
+}

package/dist/repl/index.js

@@ -1,7 +1,7 @@
 import readline from 'node:readline/promises';
 import { emitKeypressEvents } from 'node:readline';
 import { stdin, stdout } from 'node:process';
-import { config, updateModelConfig, isModelConfigured, updateMemoryConfig, isMemoryEnabled, isSubAgentEnabled, updateSubAgentConfig, isFrontendToolsEnabled, updateFrontendToolsConfig, updateLanguageConfig, languageFromShell, buildBasePrompt, getPlanModeSuffix, hasCodegraphIndex, reinjectSessionStateIntoSystem, DEFAULT_CONTEXT_WINDOW_TOKENS, } from '../config/index.js';
+import { config, updateModelConfig, isModelConfigured, updateMemoryConfig, isMemoryEnabled, isSubAgentEnabled, updateSubAgentConfig, isFrontendToolsEnabled, updateFrontendToolsConfig, updateLanguageConfig, languageFromShell, buildBasePrompt, getPlanModeSuffix, hasCodegraphIndex, DEFAULT_CONTEXT_WINDOW_TOKENS, pinSessionModel, } from '../config/index.js';
 import { getLanguage, normalizeLanguage, t, } from '../i18n/index.js';
 import { DEFAULT_BUDGET_POLICY } from '../context/budget.js';
 import { updateConfigKey, writeConfigKeys, CONFIG_PATH } from '../config/file.js';
@@ -152,17 +152,18 @@ function themeDescription(name) {
         rose: 1, emerald: 1, amber: 1, lavender: 1, sunset: 1,
     } ? t(key) : '';
 }
-/** /model 预设后端:选一个预填 baseURL,仍可逐项改。base_url 取自 README 常见表。 */
+/** /model 预设后端：同时声明原生协议；旧服务继续走 OpenAI-compatible。 */
 const MODEL_PRESETS = [
-    { label: 'GLM(智谱)', baseURL: 'https://open.bigmodel.cn/api/v3', model: 'glm-4.6', window: DEFAULT_CONTEXT_WINDOW_TOKENS },
-    { label: 'DeepSeek', baseURL: 'https://api.deepseek.com', model: 'deepseek-chat', window: DEFAULT_CONTEXT_WINDOW_TOKENS },
-    { label: 'Qwen(阿里)', baseURL: 'https://dashscope.aliyuncs.com/compatible-mode/v1', model: 'qwen-plus', window: DEFAULT_CONTEXT_WINDOW_TOKENS },
+    { label: 'Anthropic Claude', provider: 'anthropic', baseURL: 'https://api.anthropic.com', model: 'claude-sonnet-4-5', window: 200000, anthropicPromptCache: true },
+    { label: 'GLM(智谱)', provider: 'openai', baseURL: 'https://open.bigmodel.cn/api/v3', model: 'glm-4.6', window: DEFAULT_CONTEXT_WINDOW_TOKENS, anthropicPromptCache: false },
+    { label: 'DeepSeek', provider: 'openai', baseURL: 'https://api.deepseek.com', model: 'deepseek-chat', window: DEFAULT_CONTEXT_WINDOW_TOKENS, anthropicPromptCache: false },
+    { label: 'Qwen(阿里)', provider: 'openai', baseURL: 'https://dashscope.aliyuncs.com/compatible-mode/v1', model: 'qwen-plus', window: DEFAULT_CONTEXT_WINDOW_TOKENS, anthropicPromptCache: false },
     // MiniMax OpenAI 兼容端点(https://platform.minimax.io/docs/api-reference/text-openai-api)。
     // MiniMax-M3 为唯一支持图片/视频输入的模型;M2 系列纯文本(见 llm/capabilities.ts KNOWN_TEXT_ONLY_PREFIXES)。
-    { label: 'MiniMax', baseURL: 'https://api.minimax.io/v1', model: 'MiniMax-M3', window: DEFAULT_CONTEXT_WINDOW_TOKENS },
-    { label: '本地 Ollama', baseURL: 'http://localhost:11434/v1', model: 'qwen2.5:7b', window: DEFAULT_CONTEXT_WINDOW_TOKENS },
-    { label: '本地 vLLM', baseURL: 'http://localhost:8000/v1', model: 'default', window: DEFAULT_CONTEXT_WINDOW_TOKENS },
-    { label: '自定义 base_url', baseURL: '', model: '', window: DEFAULT_CONTEXT_WINDOW_TOKENS },
+    { label: 'MiniMax', provider: 'openai', baseURL: 'https://api.minimax.io/v1', model: 'MiniMax-M3', window: DEFAULT_CONTEXT_WINDOW_TOKENS, anthropicPromptCache: false },
+    { label: '本地 Ollama', provider: 'openai', baseURL: 'http://localhost:11434/v1', model: 'qwen2.5:7b', window: DEFAULT_CONTEXT_WINDOW_TOKENS, anthropicPromptCache: false },
+    { label: '本地 vLLM', provider: 'openai', baseURL: 'http://localhost:8000/v1', model: 'default', window: DEFAULT_CONTEXT_WINDOW_TOKENS, anthropicPromptCache: false },
+    { label: '自定义 base_url', provider: 'openai', baseURL: '', model: '', window: DEFAULT_CONTEXT_WINDOW_TOKENS, anthropicPromptCache: false },
 ];
 /** apiKey 脱敏:只露末 4 位,前面打星号(显示用,绝不把明文 key 写进内容区)。 */
 function maskKey(k) {
@@ -750,6 +751,8 @@ export function renderHistory(history) {
 export async function startRepl(initialHistory, sessionId, sandboxRootOverride, initialQueryHistory) {
     // 模式重置:agentMode 不落盘,每个 REPL 会话从 auto 开始(/resume / --resume 亦重置)。
     setAgentMode('auto');
+    // 钉死本会话模型：之后运行中 agent 一律用此值，其它窗口的 /model switch 不会影响本窗口。
+    pinSessionModel();
     // 沙箱根:文件操作边界。优先级 --sandbox-root > SANDBOX_ROOT env > process.cwd()。
     // 纯边界记录(不 chdir),jail.ts 内部 resolve。子 agent 同进程继承全局 root。
     setSandboxRoot(sandboxRootOverride ?? config.sandboxRoot ?? process.cwd());
@@ -834,10 +837,12 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
         if (listPresets().length === 0) {
             try {
                 const migrated = migrateCurrentToPreset({
+                    provider: config.provider,
                     baseURL: config.baseURL,
                     apiKey: config.apiKey,
                     model: config.model,
                     contextWindow: config.contextWindowTokens,
+                    anthropicPromptCache: config.anthropicPromptCache,
                 });
                 if (migrated) {
                     layout.contentWrite(`${ui.dim}  ↳ 检测到老配置,已自动迁为预设 “${migrated}”(${ui.cyan}/model list${ui.dim} 查看 · /model switch 切换)${ui.reset}\n`);
@@ -1545,9 +1550,9 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
             // focus 透传到 compact_history action 的 LLM 摘要 prompt。
             // 返回 SchedulerRunLog 给 UI 显示决策;退化路径(开关关时)在 manualCompact 内部走 compactHistory。
             const log = await manualCompact(history, focus, { force });
-            // ② compact 后把会话状态(plan + 笔记段)重注入系统提示（history[0]），避免 agent 因上下文压缩丢失计划与笔记。
-            if (log.compactHistoryCalled)
-                reinjectSessionStateIntoSystem(history);
+            // 会话状态(plan + 笔记段)不在此处回写 history[0]:agent/core 每步都在 requestHistory
+            // 末尾注入最新副本(buildSessionStateReminder),压缩后下一步自然恢复,且系统提示保持
+            // 逐字节稳定以命中 prompt 缓存。
             const d = log.compactDetail;
             appendCurrentSessionRuntimeEvent('compact', {
                 source: 'manual',
@@ -1716,16 +1721,20 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
             // 共用:apply 一个预设到 config + 持久化 + 重建 client + 重显横幅。无参 /model 选菜单和 /model use 都走这里。
             const applyPresetAndPersist = (target) => {
                 updateModelConfig({
+                    provider: target.provider,
                     model: target.model,
                     baseURL: target.baseURL,
                     apiKey: target.apiKey,
                     contextWindowTokens: target.contextWindow,
+                    anthropicPromptCache: target.anthropicPromptCache,
                 });
                 writeConfigKeys({
+                    LLM_PROVIDER: target.provider,
                     LLM_BASE_URL: target.baseURL,
                     LLM_API_KEY: target.apiKey,
                     LLM_MODEL: target.model,
                     CONTEXT_WINDOW_TOKENS: String(target.contextWindow),
+                    ANTHROPIC_PROMPT_CACHE: target.anthropicPromptCache ? 'true' : 'false',
                 });
                 reconfigureClient();
                 refreshStatusBase(history);
@@ -1736,20 +1745,22 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
                 else {
                     layout.writeBanner(bannerLines(banner()));
                 }
-                layout.contentWrite(`${ui.dim}(已切换到预设 “${target.name}” → ${target.model} @ ${target.baseURL})${ui.reset}\n`);
+                const cacheLabel = target.provider === 'anthropic'
+                    ? ` · Prompt Cache ${target.anthropicPromptCache ? 'on' : 'off'}`
+                    : '';
+                layout.contentWrite(`${ui.dim}(已切换到预设 “${target.name}” → ${target.model} · ${target.provider}${cacheLabel} @ ${target.baseURL})${ui.reset}\n`);
                 if (config.llmKeysFromShell.length > 0) {
                     layout.contentWrite(`${ui.dim}(shell 环境变量已设 ${config.llmKeysFromShell.join(' / ')},文件写入下次启动被其覆盖)${ui.reset}\n`);
                 }
             };
-            // 决定自动存的预设名:用 desired(model 字段),若与已有预设四元组完全相同则不重复存(返 null);
-            // 否则若 desired 已存在则追加 -2/-3/...。desired 含非法字符(如 glm-4.6 的 '.')时先 sanitize(. → -),
-            // sanitize 后仍空才退化到 'preset'。
-            const uniquePresetName = (desired, baseURL, apiKey, model, contextWindow) => {
+            // 决定自动存的预设名:协议、缓存配置和连接四元组都一致时不重复存。
+            const uniquePresetName = (desired, provider, baseURL, apiKey, model, contextWindow, anthropicPromptCache) => {
                 const existing = listPresets();
-                const sameEntry = existing.find((p) => p.baseURL === baseURL && p.apiKey === apiKey && p.model === model && p.contextWindow === contextWindow);
+                const sameEntry = existing.find((p) => p.provider === provider && p.baseURL === baseURL && p.apiKey === apiKey
+                    && p.model === model && p.contextWindow === contextWindow
+                    && p.anthropicPromptCache === anthropicPromptCache);
                 if (sameEntry)
-                    return null; // 完全相同,不重复存
-                // sanitize:把非 [a-zA-Z0-9_-] 字符(如 glm-4.6 的 '.')替换为 -,压缩两端 -,裁 1-32。
+                    return null;
                 const sanitized = desired
                     .replace(/[^a-zA-Z0-9_-]+/g, '-')
                     .replace(/^-+|-+$/g, '')
@@ -1773,14 +1784,17 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
                     layout.contentWrite(`${ui.dim}(还没有预设;先跑 /model 添加一个)${ui.reset}\n`);
                     continue;
                 }
-                const isCurrent = (p) => p.baseURL === config.baseURL &&
+                const isCurrent = (p) => p.provider === config.provider &&
+                    p.baseURL === config.baseURL &&
                     p.apiKey === config.apiKey &&
                     p.model === config.model &&
-                    p.contextWindow === config.contextWindowTokens;
+                    p.contextWindow === config.contextWindowTokens &&
+                    p.anthropicPromptCache === (config.provider === 'anthropic' && config.anthropicPromptCache);
                 const cols = layout.getGeo().cols;
                 const labelFor = (p) => {
                     const tag = isCurrent(p) ? ' ★current' : '';
-                    const right = `${p.model} @ ${p.baseURL}`;
+                    const cache = p.provider === 'anthropic' ? ` · cache ${p.anthropicPromptCache ? 'on' : 'off'}` : '';
+                    const right = `${p.provider}${cache} · ${p.model} @ ${p.baseURL}`;
                     const left = `${p.name}${tag}`;
                     const sep = left.length + 1 + right.length;
                     if (sep <= cols - 2)
@@ -1790,7 +1804,7 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
                 const choice = await promptIntervention({
                     type: 'choice',
                     title: '切换模型预设',
-                    detail: `当前: ${config.model} @ ${config.baseURL}(★ = 已匹配)`,
+                    detail: `当前: ${config.provider} · ${config.model} @ ${config.baseURL}(★ = 已匹配)`,
                     options: presets.map(labelFor),
                     allowCustom: false, // 纯切换,不需要「其他」干扰
                 });
@@ -1812,19 +1826,30 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
                 }
                 layout.contentWrite(`${ui.dim}已配置 ${ps.length} 个预设:${ui.reset}\n`);
                 for (const p of ps) {
-                    const star = p.baseURL === config.baseURL && p.apiKey === config.apiKey && p.model === config.model ? ' ★' : '';
-                    layout.contentWrite(`  ${ui.accent}${p.name}${ui.reset}${star}  ${ui.dim}${p.model} @ ${p.baseURL}${ui.reset}\n`);
+                    const current = p.provider === config.provider
+                        && p.baseURL === config.baseURL
+                        && p.apiKey === config.apiKey
+                        && p.model === config.model
+                        && p.contextWindow === config.contextWindowTokens
+                        && p.anthropicPromptCache === (config.provider === 'anthropic' && config.anthropicPromptCache);
+                    const star = current ? ' ★' : '';
+                    const cache = p.provider === 'anthropic' ? ` · cache ${p.anthropicPromptCache ? 'on' : 'off'}` : '';
+                    layout.contentWrite(`  ${ui.accent}${p.name}${ui.reset}${star}  ${ui.dim}${p.provider}${cache} · ${p.model} @ ${p.baseURL}${ui.reset}\n`);
                 }
-                layout.contentWrite(`${ui.dim}(★ = 与当前一致;切换用 /model switch)${ui.reset}\n`);
+                layout.contentWrite(`${ui.dim}(★ = 与当前协议及缓存配置一致;切换用 /model switch)${ui.reset}\n`);
                 continue;
             }
-            // /model show:显示当前四项配置(apiKey 脱敏)。
+            // /model show:显示当前协议、连接与缓存配置(apiKey 脱敏)。
             if (arg === 'show') {
                 layout.contentWrite(`${ui.dim}当前模型配置:${ui.reset}\n`);
-                layout.contentWrite(`  ${ui.accent}baseURL${ui.reset}  ${config.baseURL}\n`);
-                layout.contentWrite(`  ${ui.accent}apiKey ${ui.reset}  ${maskKey(config.apiKey)}\n`);
-                layout.contentWrite(`  ${ui.accent}model  ${ui.reset}  ${config.model}\n`);
-                layout.contentWrite(`  ${ui.accent}窗口   ${ui.reset}  ${config.contextWindowTokens} tokens\n`);
+                layout.contentWrite(`  ${ui.accent}provider${ui.reset}  ${config.provider}\n`);
+                layout.contentWrite(`  ${ui.accent}baseURL ${ui.reset}  ${config.baseURL}\n`);
+                layout.contentWrite(`  ${ui.accent}apiKey  ${ui.reset}  ${maskKey(config.apiKey)}\n`);
+                layout.contentWrite(`  ${ui.accent}model   ${ui.reset}  ${config.model}\n`);
+                layout.contentWrite(`  ${ui.accent}窗口    ${ui.reset}  ${config.contextWindowTokens} tokens\n`);
+                if (config.provider === 'anthropic') {
+                    layout.contentWrite(`  ${ui.accent}缓存    ${ui.reset}  Prompt Cache ${config.anthropicPromptCache ? 'on' : 'off'}\n`);
+                }
                 layout.contentWrite(`${ui.dim}(配置文件: ${CONFIG_PATH})${ui.reset}\n`);
                 continue;
             }
@@ -2017,25 +2042,42 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
                     }
                 }
             }
-            // 3) 应用:内存 config + env(updateModelConfig)→ 持久化(writeConfigKeys)→ 重建 client(reconfigureClient)。
-            updateModelConfig({ model, baseURL, apiKey, contextWindowTokens: window });
+            // 3) 应用协议、连接与缓存配置；Anthropic 原生协议无需重建 OpenAI client，但统一刷新无害。
+            const provider = preset.provider;
+            const anthropicPromptCache = provider === 'anthropic' && preset.anthropicPromptCache;
+            updateModelConfig({
+                provider,
+                model,
+                baseURL,
+                apiKey,
+                contextWindowTokens: window,
+                anthropicPromptCache,
+            });
             writeConfigKeys({
+                LLM_PROVIDER: provider,
                 LLM_BASE_URL: baseURL,
                 LLM_API_KEY: apiKey,
                 LLM_MODEL: model,
                 CONTEXT_WINDOW_TOKENS: String(window),
+                ANTHROPIC_PROMPT_CACHE: anthropicPromptCache ? 'true' : 'false',
             });
             reconfigureClient();
-            // 3.5) 自动存为命名预设:用 model 字段,重名追加 -2/-3,完全相同的四元组不重复存。
-            //     让 /model 跑一次就多一份可切换的预设,/model switch 切回去。
+            // 3.5) 自动存为命名预设；协议与缓存策略也是去重键的一部分。
             let savedName = null;
             try {
-                const finalName = uniquePresetName(model, baseURL, apiKey, model, window);
+                const finalName = uniquePresetName(model, provider, baseURL, apiKey, model, window, anthropicPromptCache);
                 if (finalName) {
-                    savePreset({ name: finalName, baseURL, apiKey, model, contextWindow: window });
+                    savePreset({
+                        name: finalName,
+                        provider,
+                        baseURL,
+                        apiKey,
+                        model,
+                        contextWindow: window,
+                        anthropicPromptCache,
+                    });
                     savedName = finalName;
                 }
-                // finalName === null 表示与某个已存在预设完全一致,不再重复保存。
             }
             catch (e) {
                 layout.contentWrite(`${ui.red}保存预设失败: ${e.message}${ui.reset}\n`);
@@ -2049,7 +2091,10 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
             else {
                 layout.writeBanner(bannerLines(banner()));
             }
-            layout.contentWrite(`${ui.dim}(已切换模型 → ${model} @ ${baseURL})${ui.reset}\n`);
+            const cacheLabel = provider === 'anthropic'
+                ? ` · Prompt Cache ${anthropicPromptCache ? 'on' : 'off'}`
+                : '';
+            layout.contentWrite(`${ui.dim}(已切换模型 → ${model} · ${provider}${cacheLabel} @ ${baseURL})${ui.reset}\n`);
             if (savedName) {
                 layout.contentWrite(`${ui.dim}(已保存为预设 “${savedName}”,下次 /model use ${savedName} 一键切回)${ui.reset}\n`);
             }