mocode-ai 1.4.2 → 1.4.3

package/dist/llm/index.js

@@ -4,7 +4,7 @@ import { tools } from '../tools/registry.js';
 import { getPlanDisabledTools, getProfileDisabledTools } from '../tools/constants.js';
 import { ThinkTagFilter } from './think-filter.js';
 import { sanitizeToolSchemas } from './tool-schema.js';
-import { anthropicChatOnce } from './providers/anthropic.js';
+import { defaultAnthropicFetch, anthropicChatOnce } from './providers/anthropic.js';
 import { registerModelProvider, getModelProvider, listModelProviders } from './provider.js';
 // 强制关闭第三方调试日志泄漏:openai SDK 在 process.env.DEBUG === 'true' 时用裸
 // console.log 把请求/响应直写 stdout,会污染 TUI 输入框(并泄露 headers/URL)。
@@ -33,11 +33,31 @@ const RETRY_JITTER = 0.2;
  * 与 OpenAI 兼容协议的独立 `reasoning_content` 字段不同,这些模型把 thinking 直接嵌进 content
  * 字符串,期间不调 onText(spinner 持续转 ⠹ 思考中…),也不写入可见 content(history 不被思考段污染)。
  */
-let client = new OpenAI({
-    baseURL: config.baseURL,
-    apiKey: config.apiKey,
-    maxRetries: 0,
-});
+function newOpenAIClient(runtimeConfig) {
+    return new OpenAI({
+        baseURL: runtimeConfig.baseURL,
+        apiKey: runtimeConfig.apiKey,
+        maxRetries: 0,
+    });
+}
+/** Build mutable client state owned by one runtime; no module-global overrides are consulted. */
+export function createChatClientState(runtimeConfig, overrides = {}) {
+    return {
+        openAI: newOpenAIClient(runtimeConfig),
+        openAICreateImpl: overrides.openAICreateImpl ?? null,
+        anthropicFetchImpl: overrides.anthropicFetchImpl ?? fetch,
+    };
+}
+/** Rebuild only the OpenAI SDK client while preserving runtime-local test/transport overrides. */
+export function reconfigureChatClient(state, runtimeConfig) {
+    state.openAI = newOpenAIClient(runtimeConfig);
+}
+const defaultClientState = createChatClientState(config, { anthropicFetchImpl: defaultAnthropicFetch });
+const defaultProviderRuntime = {
+    config,
+    getModel: getActiveModel,
+    clientState: defaultClientState,
+};
 /**
  * 运行时重建 OpenAI 客户端(/model 切换 baseURL/apiKey 后调)。
  * config.model 已在 chat() 每次读取(热切),但 client 的 baseURL/apiKey 是构造时固化的实例字段,
@@ -45,11 +65,7 @@ let client = new OpenAI({
  * 子 agent 复用本模块 chat(),故只此一处重建即全链路生效。
  */
 export function reconfigureClient() {
-    client = new OpenAI({
-        baseURL: config.baseURL,
-        apiKey: config.apiKey,
-        maxRetries: 0,
-    });
+    reconfigureChatClient(defaultClientState, config);
 }
 // ── 重试 helper(纯函数 + sleep,被 chat() 调用;亦可单测导入验证)──────
 /** 可被 AbortSignal 取消的 sleep;signal 已 abort 时立即抛 AbortError。 */
@@ -209,10 +225,9 @@ function logRetry(attempt, err, waitMs) {
     // 不写前导 \n —— contentWrite 由续写位管位置,前置换行会留空行;结尾 \n 由劫持逻辑补。
     console.error(`[llm] 第 ${attempt}/${RETRY_MAX_ATTEMPTS} 次失败(${tag}: ${msg}),${(waitMs / 1000).toFixed(1)}s 后重试…`);
 }
-let createImplOverride = null;
 /** 仅供单测用:覆盖 chat() 内部实际调用的 create 桩。生产代码不要碰。 */
 export function __setChatCreateImpl(impl) {
-    createImplOverride = impl;
+    defaultClientState.openAICreateImpl = impl;
 }
 /**
  * 工具 schema 保持稳定数组引用，再在 MCP 发现或扩展变动后原地刷新；
@@ -249,11 +264,11 @@ export function refreshChatTools() {
 // 新增 provider 只需 registerModelProvider,不必改 chat()。
 registerModelProvider({
     name: 'openai',
-    chatOnce: (messages, handlers, signal, tools) => chatOnce(messages, handlers, signal, tools),
+    chatOnce: (messages, handlers, signal, tools, runtime) => chatOnce(messages, handlers, signal, tools, runtime),
 });
 registerModelProvider({
     name: 'anthropic',
-    chatOnce: (messages, handlers, signal, tools) => anthropicChatOnce(messages, handlers, signal, tools ?? chatTools),
+    chatOnce: (messages, handlers, signal, tools, runtime) => anthropicChatOnce(messages, handlers, signal, tools ?? chatTools, runtime),
 });
 /**
  * 多 provider 兼容的 cache / reasoning 字段提取。
@@ -325,29 +340,28 @@ function retryErrorCode(error) {
         return `HTTP_${value.status}`;
     return value.code ?? value.name ?? 'RETRYABLE_ERROR';
 }
-/**
- * 流式调一次 LLM:增量回调文本,内部累加 tool_calls 片段。
- * tool_calls 跨 chunk 按 index 累加(id / name / arguments 拼接)。
- * include_usage 时末尾 chunk 携带 usage,先读再 continue(末尾 chunk 无 delta)。
- *
- * 包了一层重试:429/5xx/timeout/网络错按指数退避重试(默认 4 次),400/401/用户中断立即抛。
- * 重试由 chat() 统一管,chatOnce() 只负责单次请求,职责单一便于单测。
- */
-export async function chat(messages, handlers = {}, signal, 
+/** Bind chat dispatch, retries and built-in providers to one explicit runtime. */
+export function createChatTransport(runtime) {
+    return (messages, handlers = {}, signal, toolsOverride) => chatWithRuntime(runtime, messages, handlers, signal, toolsOverride);
+}
+export function chat(messages, handlers = {}, signal, 
 /** 覆盖默认工具 schema;plan 模式传 planChatTools(只读子集),缺省=全量 chatTools。 */
 toolsOverride) {
+    return chatWithRuntime(defaultProviderRuntime, messages, handlers, signal, toolsOverride);
+}
+async function chatWithRuntime(runtime, messages, handlers, signal, toolsOverride) {
     let lastErr;
     for (let attempt = 1; attempt <= RETRY_MAX_ATTEMPTS; attempt++) {
         if (signal?.aborted) {
             throw new DOMException('This operation was aborted', 'AbortError');
         }
         try {
-            const provider = getModelProvider(config.provider);
+            const provider = getModelProvider(runtime.config.provider);
             if (!provider) {
                 // 未注册的 provider:给出可用名单,避免悄悄落到错误实现。
-                throw new Error(`未知的 LLM provider "${config.provider}";已注册:${listModelProviders().join(', ') || '(空)'}`);
+                throw new Error(`未知的 LLM provider "${runtime.config.provider}";已注册:${listModelProviders().join(', ') || '(空)'}`);
             }
-            return await provider.chatOnce(messages, handlers, signal, toolsOverride);
+            return await provider.chatOnce(messages, handlers, signal, toolsOverride, runtime);
         }
         catch (err) {
             lastErr = err;
@@ -411,7 +425,7 @@ export function normalizeImageDetail(messages) {
     return changed ? next : messages;
 }
 /** 单次流式 LLM 请求(无重试);chat() 的内部实现,可被 __setChatCreateImpl 注入桩以做单测。 */
-async function chatOnce(messages, handlers, signal, toolsOverride) {
+async function chatOnce(messages, handlers, signal, toolsOverride, runtime = defaultProviderRuntime) {
     // 防御:messages 必须至少含一条非空 user 消息,否则 OpenAI/Anthropic 都会 400。
     // compact force 分支曾把所有 user 丢进摘要 → 重建 history 无 user → 下一轮 400。
     // 在 transport 边界拦住所有类似回归(compact / 外部注入 / resume 损坏)。
@@ -429,15 +443,17 @@ async function chatOnce(messages, handlers, signal, toolsOverride) {
         throw new Error('messages must contain at least one non-empty user message');
     }
     // signal 透传给 SDK 第二参(RequestOptions);abort 后 for await 抛错,chat 不 catch,透传 runAgent 处理。
-    // createImplOverride 的 body 类型故意宽成 Record(测试桩用),生产走 client 分支时由 OpenAI 自己的类型守门。
-    const create = createImplOverride
-        ? createImplOverride
-        : (body, opts) => client.chat.completions.create(body, opts);
+    // openAICreateImpl 的 body 类型故意宽成 Record(测试桩用),生产走 client 分支时由 OpenAI 自己的类型守门。
+    const runtimeConfig = runtime.config;
+    const runtimeClientState = runtime.clientState;
+    const create = runtimeClientState.openAICreateImpl
+        ? runtimeClientState.openAICreateImpl
+        : (body, opts) => runtimeClientState.openAI.chat.completions.create(body, opts);
     // transport 边界消毒:剔除部分后端(kimi-k3@dashscope 实测)整请求 400 拒绝的
     // uniqueItems 关键字。无需改写时返回原引用,前缀缓存逐字节稳定不受影响。
     const activeTools = sanitizeToolSchemas(toolsOverride ?? chatTools);
     const stream = await create({
-        model: getActiveModel(),
+        model: runtime.getModel(),
         messages: normalizeImageDetail(messages),
         stream: true,
         // 空工具表(如摘要请求)不发 tools 字段——部分网关拒绝空数组。
@@ -449,8 +465,8 @@ async function chatOnce(messages, handlers, signal, toolsOverride) {
                 parallel_tool_calls: true,
             }
             : {}),
-        ...(config.maxTokens ? { max_tokens: config.maxTokens } : {}),
-        ...(config.includeUsage ? { stream_options: { include_usage: true } } : {}),
+        ...(runtimeConfig.maxTokens ? { max_tokens: runtimeConfig.maxTokens } : {}),
+        ...(runtimeConfig.includeUsage ? { stream_options: { include_usage: true } } : {}),
     }, signal ? { signal } : undefined);
     // content 内嵌 think 标签由独立增量状态机过滤。它只暂存“可能组成标签”的后缀，
     // 因而既能覆盖标签任意位置跨 chunk，也不会让普通正文固定延迟数个字符。

package/dist/llm/providers/anthropic.js

@@ -4,6 +4,10 @@ let fetchImplOverride = null;
 export function __setAnthropicFetchImpl(impl) {
     fetchImplOverride = impl;
 }
+/** Default runtime keeps the historical process-level test override behavior. */
+export function defaultAnthropicFetch(input, init) {
+    return (fetchImplOverride ?? fetch)(input, init);
+}
 function parseJsonObject(value) {
     if (!value.trim())
         return {};
@@ -142,12 +146,13 @@ function endpoint(baseURL) {
         return base;
     return `${base}/v1/messages`;
 }
-export function buildAnthropicRequest(messages, tools) {
-    const encoded = encodeAnthropicMessages(messages);
-    const anthropicTools = encodeAnthropicTools(tools);
+export function buildAnthropicRequest(messages, tools, runtime) {
+    const runtimeConfig = runtime?.config ?? config;
+    const encoded = encodeAnthropicMessages(messages, runtimeConfig.anthropicPromptCache);
+    const anthropicTools = encodeAnthropicTools(tools, runtimeConfig.anthropicPromptCache);
     return {
-        model: getActiveModel(),
-        max_tokens: config.maxTokens ?? 8192,
+        model: runtime?.getModel() ?? getActiveModel(),
+        max_tokens: runtimeConfig.maxTokens ?? 8192,
         stream: true,
         system: encoded.system,
         messages: encoded.messages,
@@ -249,17 +254,18 @@ function addLiveCount(text, state) {
     }
 }
 /** 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), {
+export async function anthropicChatOnce(messages, handlers, signal, tools, runtime) {
+    const runtimeConfig = runtime?.config ?? config;
+    const fetchImpl = runtime?.clientState.anthropicFetchImpl ?? defaultAnthropicFetch;
+    const response = await fetchImpl(endpoint(runtimeConfig.baseURL), {
         method: 'POST',
         headers: {
             'content-type': 'application/json',
             accept: 'text/event-stream',
-            'x-api-key': config.apiKey,
+            'x-api-key': runtimeConfig.apiKey,
             'anthropic-version': process.env.ANTHROPIC_VERSION || '2023-06-01',
         },
-        body: JSON.stringify(buildAnthropicRequest(messages, tools)),
+        body: JSON.stringify(buildAnthropicRequest(messages, tools, runtime)),
         signal,
     });
     if (!response.ok)

package/dist/llm/runtime.js

@@ -0,0 +1 @@
+export {};

package/dist/permissions/index.js

@@ -147,8 +147,8 @@ function matches(grant, tool, fingerprint, projectRoot) {
         grant.fingerprint === fingerprint &&
         (grant.scope !== 'project' || grant.projectRoot === projectRoot));
 }
-export async function checkPermission(tool, args, signal, options = {}) {
-    if (!config.permissionEnabled || getToolRisk(tool) === 'safe')
+async function checkPermissionWithSessionGrants(runtimeConfig, runtimeSessionGrants, tool, args, signal, options = {}) {
+    if (!runtimeConfig.permissionEnabled || getToolRisk(tool) === 'safe')
         return 'allow';
     if (signal?.aborted)
         return 'deny';
@@ -165,13 +165,13 @@ export async function checkPermission(tool, args, signal, options = {}) {
     const fingerprint = permissionFingerprint(tool, args);
     const projectRoot = canonicalProjectRoot(options.projectRoot ?? getSandboxRoot() ?? process.cwd());
     if (!forceOnce) {
-        if (sessionGrants.some((grant) => matches(grant, tool.name, fingerprint, projectRoot)))
+        if (runtimeSessionGrants.some((grant) => matches(grant, tool.name, fingerprint, projectRoot)))
             return 'allow';
         if (permanentGrants.some((grant) => matches(grant, tool.name, fingerprint, projectRoot)))
             return 'allow';
     }
     // CI/pipes must fail closed. Operators can deliberately restore unattended behavior.
-    if (!process.stdin.isTTY && !config.permissionNonInteractiveAllow && !options.prompt)
+    if (!process.stdin.isTTY && !runtimeConfig.permissionNonInteractiveAllow && !options.prompt)
         return 'deny';
     if (signal?.aborted)
         return 'deny';
@@ -214,7 +214,7 @@ export async function checkPermission(tool, args, signal, options = {}) {
     if (signal?.aborted || result.action === 'cancelled' || result.value === denyOption)
         return 'deny';
     if (result.value === sessionOption) {
-        sessionGrants.push({ tool: tool.name, fingerprint, scope: 'session' });
+        runtimeSessionGrants.push({ tool: tool.name, fingerprint, scope: 'session' });
     }
     else if (result.value === projectOption) {
         const grant = { tool: tool.name, fingerprint, scope: 'project', projectRoot };
@@ -230,6 +230,22 @@ export async function checkPermission(tool, args, signal, options = {}) {
     }
     return 'allow';
 }
+/** 兼容入口：显式 config 仍共享进程级 session grants。 */
+export function checkPermissionWithConfig(runtimeConfig, tool, args, signal, options = {}) {
+    return checkPermissionWithSessionGrants(runtimeConfig, sessionGrants, tool, args, signal, options);
+}
+/** 为独立 RuntimeContext 创建私有 session grants；项目/永久授权仍按原语义进程共享。 */
+export function createPermissionChecker(runtimeConfig, defaultProjectRoot) {
+    const runtimeSessionGrants = [];
+    return (tool, args, signal, options = {}) => checkPermissionWithSessionGrants(runtimeConfig, runtimeSessionGrants, tool, args, signal, {
+        ...options,
+        projectRoot: options.projectRoot ?? defaultProjectRoot,
+    });
+}
+/** 默认兼容入口：继续读取进程级 config。独立 RuntimeContext 使用 checkPermissionWithConfig。 */
+export function checkPermission(tool, args, signal, options = {}) {
+    return checkPermissionWithConfig(config, tool, args, signal, options);
+}
 export function revokePermanentAllow(toolName, fingerprint) {
     loadPermanent();
     permanentGrants = permanentGrants.filter((grant) => grant.tool !== toolName || (fingerprint !== undefined && grant.fingerprint !== fingerprint));

package/dist/repl/commands/compact.js

@@ -12,7 +12,7 @@
 import * as layout from '../../ui/layout.js';
 import { ui } from '../../ui/theme.js';
 import { t } from '../../i18n/index.js';
-import { manualCompact, appendCurrentSessionRuntimeEvent, hashTraceValue } from '../../session/index.js';
+import { appendCurrentSessionRuntimeEvent, hashTraceValue } from '../../session/index.js';
 import { startRunningListener, stopRunningListener } from '../running-input.js';
 import { unhandled, next } from './types.js';
 export const compactCommands = [
@@ -38,7 +38,7 @@ export const compactCommands = [
         const log = await (async () => {
             const signal = startRunningListener(t('running.compacting'));
             try {
-                return await manualCompact(ctx.history, focus, { force, signal });
+                return await ctx.runtime.compact(ctx.history, { focus, force, signal, contextState: ctx.contextState });
             }
             finally {
                 stopRunningListener();

package/dist/repl/commands/session.js

@@ -8,9 +8,6 @@ import * as layout from '../../ui/layout.js';
 import { ui } from '../../ui/theme.js';
 import { bannerLines } from '../../ui/render.js';
 import { t } from '../../i18n/index.js';
-import { newSessionId, listSessions } from '../../session/index.js';
-import { resetState } from '../../rollback/index.js';
-import { setCurrentSessionId } from '../../session/state.js';
 import { promptSessionPicker } from '../../ui/prompt.js';
 import { unhandled, next } from './types.js';
 export const sessionCommands = [
@@ -18,13 +15,7 @@ export const sessionCommands = [
         if (ctx.line !== '/clear')
             return unhandled();
         ctx.history.length = 1; // 保留 system 提示
-        resetState(); // 同步清空回滚轮次/快照
-        // /clear 立刻换新会话 id:之前延后到 turn 收尾分配,但 runAgent 期间模型
-        // 调 plan_update / note_append 会命中「no active session」——它们写 notes.md 靠
-        // getNotesFilePath(getCurrentSessionId()),而下一个 turn 还没走完。改为与启动对齐
-        // 立即分配:notes.md 由写入时按需建文件,这里预分配 id 不触盘。
-        ctx.state.currentSessionId = newSessionId();
-        setCurrentSessionId(ctx.state.currentSessionId, process.cwd()); // 同步到 session/state
+        ctx.state.currentSessionId = ctx.runtime.session.clear();
         ctx.state.turnCount = 0; // 反思 cadence 重新计数
         ctx.contextState.lastUsage = undefined;
         ctx.contextState.lifecycleStats = undefined;
@@ -44,7 +35,7 @@ export const sessionCommands = [
     async (ctx) => {
         if (ctx.line !== '/sessions')
             return unhandled();
-        const sessions = listSessions(); // 不传 limit = 全量
+        const sessions = ctx.runtime.session.list(); // 不传 limit = 全量
         if (sessions.length === 0) {
             layout.contentWrite(`${ui.dim}(没有已保存的会话)${ui.reset}\n`);
             return next();
@@ -71,7 +62,7 @@ export const sessionCommands = [
     async (ctx) => {
         if (ctx.line !== '/resume')
             return unhandled();
-        const sessions = listSessions(10);
+        const sessions = ctx.runtime.session.list(10);
         if (sessions.length === 0) {
             layout.contentWrite(`${ui.dim}(没有已保存的会话)${ui.reset}\n`);
             return next();

package/dist/repl/message-format.js

@@ -57,10 +57,15 @@ export function renderHistory(history) {
     let pendingBatches = [];
     let normalBatch = null;
     const flushBatch = () => {
+        if (pendingBatches.length === 0)
+            return false;
+        // 气泡/正文 → 摘要行边界:尾部视觉空行幂等收成 1 条(与实时 writeToolHeader 同语义)。
+        layout.normalizeMutationBoundary();
         for (const entries of pendingBatches)
             batch.writeSummaryOnly(entries, layout);
         pendingBatches = [];
         normalBatch = null;
+        return true;
     };
     for (let idx = 0; idx < history.length; idx++) {
         const m = history[idx];