zen-gitsync 2.17.8 → 2.17.10

package/src/ui/server/routes/workbench/agentChat.js

@@ -23,6 +23,7 @@
 //   - { type: 'content', delta }           — 正文增量
 //   - { type: 'tool_call_start', toolCallId, name, argsPreview }
 //   - { type: 'tool_result', toolCallId, name, result }
+//   - { type: 'ask_user', interactionId, question, options, allowFreeText }
 //   - { type: 'done', content }            — 本轮最终完成
 //   - { type: 'error', error }
 
@@ -31,17 +32,17 @@ import os from 'os';
 import { logger } from './shared.js';
 import { agentSessionStore } from './agentSessionStore.js';
 
-// 从 CLI 侧导入工具定义与执行器（同一 monorepo，路径可达）
+// 从 CLI 侧导入工具定义、执行器与 LLM 传输层（同一 monorepo，路径可达）
 import { TOOL_DEFINITIONS, executeTool } from '../../../../cli/ai/tools.js';
+import { prepareRequestMessages } from '../../../../cli/ai/context.js';
+import { streamChatOnce } from '../../../../cli/ai/transport.js';
 import { checkDangerousCommand } from '../../../../cli/ai/safety.js';
 import { guardCommand } from '../../../../cli/ai/platformGuard.js';
 import configManager from '../../../../config.js';
-import { buildAiChatRequest, describeAiHttpError } from '../../../../utils/aiEndpoint.js';
 
 // 单轮工具调用循环数的兜底值(防失控);实际值取全局配置 aiMaxToolIterations,
 // 与 CLI 侧 src/cli/ai/agent.js 共用同一个配置项。
 const DEFAULT_MAX_TOOL_ITERATIONS = 200;
-const LLM_TIMEOUT_MS = 300000; // 5 分钟
 
 // 读取全局配置里的单轮工具调用上限。
 // 读配置失败不该把整轮对话打挂 —— 退回默认值继续跑,比用户消息直接发不出去好。
@@ -62,6 +63,7 @@ function buildWebSystemPrompt({ cwd, locale }) {
   const zh = !String(locale || '').startsWith('en');
   const now = new Date().toLocaleString();
   const isWin = process.platform === 'win32';
+  const builtinToolCount = TOOL_DEFINITIONS.length;
   const shellDesc = isWin ? 'cmd.exe / PowerShell' : '/bin/sh';
 
   if (zh) {
@@ -87,6 +89,26 @@ ${isWin ? `- 当前是 Windows,以下 Unix 命令**不存在**,用了必定报"
   · 查命令路径 → cmd 的 where(不是 which)
 - 必须跑 shell 时优先跨平台写法(如 node -e "..."),别用 Unix 专属命令` : `- 当前是 POSIX 环境,Unix 命令可用`}
 
+# 远程仓库(GitHub / Gitee)
+- 问"我有哪些项目""哪些项目需要 pull / 推送" → 用 list_projects 工具。它返回的就是 g ui
+  「最近项目」面板那份清单(最近目录 + 建过任务的目录,带分支/领先/落后/未提交数与任务进度),
+  口径与界面完全一致。**不要**用 list_files 自己扫盘数仓库 —— 那会把 node_modules 里的嵌套
+  仓库也算进来,数出来的个数跟界面对不上。领先/落后是本地快照,要真实值就带 refresh=true
+  先联网 fetch 一轮,再回答"要不要 pull"
+- 问"这个项目关联哪个远端" → run_command 跑 \`git remote -v\`:本地信息,不联网、不依赖任何 CLI,任何环境都能答
+- 问"我账号下有哪些仓库"或某仓库的 PR / Issue → 用官方 CLI。凭据由 CLI 自己保管:
+  · GitHub → \`gh\`。列仓库 \`gh repo list --limit 50 --json name,visibility,updatedAt,primaryLanguage\`;
+    看详情 \`gh repo view <owner/repo>\`;PR \`gh pr list\`;登录态 \`gh auth status\`
+  · Gitee → \`gitee\`。列仓库 \`gitee repo list\`;登录态 \`gitee auth status\`
+    (未登录时退出码**仍是 0**,必须看 \`--json\` 里的 status 字段,只看退出码会误判成已登录)
+  · 两者默认只覆盖**当前登录账号**;组织仓库、私有仓库可能需要更大的 token scope,拉不到就如实说明,
+    不要用其他途径绕过
+  · 绝不向用户索要 token —— 也不要让用户把 token 粘进对话
+- CLI 可能没装、或没在服务端进程的 PATH 里(装完没重启服务就是这个表现,Web 端尤其常见):
+  报"不是内部或外部命令" / command not found 时,**不要换写法反复重试**,更不要凭印象编仓库名或目录名。
+  直接告诉用户未检测到该 CLI,可在「远程仓库」页一键安装/登录,或需要时重启服务让新装的 CLI 生效;
+  若只是想回答本项目的问题,退回 \`git remote -v\`
+
 # 权限(用户已明确授权,无需反复征求同意)
 - 工作目录内:读写文件、执行命令等所有操作直接执行
 - 其他目录:同样可以读取和修改
@@ -104,7 +126,8 @@ ${isWin ? `- 当前是 Windows,以下 Unix 命令**不存在**,用了必定报"
 
 # 与用户交互
 - 需要向用户确认、提问或汇报重要决策时,直接用普通文本输出
-- 不要调用不存在的工具,可用工具只有上面列出的 6 个
+- 需要暂停当前任务并等待用户决定或补充信息时,调用 ask_user,不要猜测或只在普通文本里提问
+- 不要调用不存在的工具,可用工具只有上面列出的 ${builtinToolCount} 个
 - 用户可能随消息附带图片:图片以 image_url 部件出现在 user 消息里;如果当前模型不支持视觉(带图请求报错),提醒用户换用支持视觉的模型
 - 发现高风险或状态不一致的情况时:先用文本说明发现和影响,停下来等用户指示,不要擅自继续破坏性操作
 
@@ -134,6 +157,30 @@ ${isWin ? `- This is Windows. The following Unix commands do NOT exist here:
   · Text processing → node -e "..." or PowerShell
   · Find executable → cmd's where (not which)` : `- POSIX environment: Unix commands are available`}
 
+# Remote repositories (GitHub / Gitee)
+- "Which projects do I have?" / "Which ones need a pull or push?" → use the list_projects tool.
+  It returns exactly the list behind the GUI's "Recent projects" panel (recent directories plus any
+  directory a task was created in, with branch / ahead / behind / uncommitted counts and task
+  progress) — the same numbers the UI shows. Do NOT scan the disk with list_files to count repos:
+  that also picks up nested repos inside node_modules and the totals will not match the UI.
+  Ahead/behind comes from local refs, so pass refresh=true for a real answer about pulling
+- "Which remote does this project point at?" → run_command \`git remote -v\`: local info, no network, no CLI needed
+- "Which repos do I have?" or a repo's PRs / issues → use the official CLI. The CLI owns the credentials:
+  · GitHub → \`gh\`. List repos \`gh repo list --limit 50 --json name,visibility,updatedAt,primaryLanguage\`;
+    details \`gh repo view <owner/repo>\`; PRs \`gh pr list\`; auth state \`gh auth status\`
+  · Gitee → \`gitee\`. List repos \`gitee repo list\`; auth state \`gitee auth status\`
+    (the exit code is still 0 when logged out — read the status field from \`--json\`, or you will
+    wrongly report "logged in")
+  · Both default to the **currently authenticated account only**; org and private repos may need
+    broader token scopes. If you cannot see them, say so plainly — do not route around it
+  · Never ask the user for a token, and never have them paste one into the conversation
+- The CLI may not be installed, or missing from the server process's PATH (that is what "installed but
+  the server cannot find it" looks like — common on the Web side):
+  on "not recognized" / command not found, do NOT retry with a different spelling and do NOT invent
+  repo or directory names. Tell the user the CLI was not found and point them to the "Remote
+  repositories" page (one-click install / login), or restart the server so a freshly installed CLI
+  becomes visible; if you only need to answer for this project, fall back to \`git remote -v\`
+
 # Permissions (explicitly granted by the user)
 - Inside the working directory: read/write files and run commands directly
 - Other directories: may also be read and modified
@@ -141,190 +188,38 @@ ${isWin ? `- This is Windows. The following Unix commands do NOT exist here:
 
 # How you work
 - Act first, ask later: use tools to investigate before asking the user
+- When a decision or missing detail must come from the user, call ask_user. Do not guess or merely describe a question in plain text.
 - After modifying code, verify: run tests, build, or at least syntax check
 - Prefer edit_file for precise replacements; read_file first to confirm original text
 - Use offset/limit for large files
 - git operations via run_command
 - run_command defaults to the working directory; default timeout 120s, max 600s
-- The user may attach images to a message; they arrive as image_url parts in the user message. If the current model rejects images (no vision support), tell the user to switch to a vision-capable model
+- When the task must pause for a decision or missing detail, call ask_user and wait for the user's answer. The user may attach images to a message; they arrive as image_url parts in the user message. If the current model rejects images (no vision support), tell the user to switch to a vision-capable model
 
 # Output
 - Your text output is displayed in the user's Web UI
 - After completing a task, briefly summarize the result`;
 }
 
-// ── LLM 流式调用(OpenAI 兼容 + function calling) ──────────
-// 返回 { content, toolCalls, aborted }
-async function streamChatOnce({ model, messages, signal, onToken, sessionId }) {
-  // sessionId = 对话 ID：OpenCode 网关靠它做路由与提示缓存，同一轮对话的所有请求要复用
-  const { url, headers } = buildAiChatRequest({
-    baseURL: model.baseURL,
-    model: model.model,
-    apiKey: model.apiKey,
-    sessionId,
-  });
-
-  const body = JSON.stringify({
-    model: model.model,
-    messages,
-    tools: TOOL_DEFINITIONS,
-    temperature: 0.3,
-    stream: true,
-  });
-
-  const controller = new AbortController();
-  const timer = setTimeout(() => controller.abort(), LLM_TIMEOUT_MS);
-  const onExternalAbort = () => controller.abort();
-  if (signal) {
-    if (signal.aborted) controller.abort();
-    else signal.addEventListener('abort', onExternalAbort);
-  }
-
-  let content = '';
-  const toolCalls = [];
-  let aborted = false;
-
-  try {
-    const resp = await fetch(url, { method: 'POST', headers, body, signal: controller.signal });
-    if (!resp.ok || !resp.body) {
-      const errText = await resp.text().catch(() => '');
-      // 抽 error.message：网关/provider 的说明都在里面，整坨 JSON 甩到聊天窗没法看
-      const snippet = describeAiHttpError(errText, resp.status);
-      if (resp.status === 400 && /tool|function/i.test(snippet)) {
-        throw new Error(`HTTP 400: 当前模型可能不支持 function calling(${snippet})。请在设置中换用支持工具调用的模型。`);
-      }
-      throw new Error(`HTTP ${resp.status}: ${snippet || resp.statusText}`);
-    }
-
-    const decoder = new TextDecoder('utf-8');
-    let buf = '';
-    for await (const chunk of resp.body) {
-      buf += decoder.decode(chunk, { stream: true });
-      const lines = buf.split('\n');
-      buf = lines.pop() ?? '';
-      for (const line of lines) {
-        const trimmed = line.trim();
-        if (!trimmed || !trimmed.startsWith('data:')) continue;
-        const payload = trimmed.slice(5).trim();
-        if (payload === '[DONE]') continue;
-        let evt;
-        try { evt = JSON.parse(payload); } catch { continue; }
-        const delta = evt.choices?.[0]?.delta || {};
-
-        const thinkingChunk = delta.reasoning_content || delta.reasoning || delta.reasoning_text || '';
-        if (thinkingChunk) onToken({ thinking: thinkingChunk });
-
-        if (delta.content) {
-          content += delta.content;
-          onToken({ content: delta.content });
-        }
-
-        for (const tc of delta.tool_calls || []) {
-          const i = tc.index ?? 0;
-          if (!toolCalls[i]) toolCalls[i] = { id: '', type: 'function', function: { name: '', arguments: '' } };
-          if (tc.id) toolCalls[i].id += tc.id;
-          if (tc.function?.name) toolCalls[i].function.name += tc.function.name;
-          if (tc.function?.arguments) toolCalls[i].function.arguments += tc.function.arguments;
-        }
-      }
-    }
-  } catch (err) {
-    if (err?.name === 'AbortError' || controller.signal.aborted) {
-      aborted = true;
-    } else {
-      throw err;
-    }
-  } finally {
-    clearTimeout(timer);
-    if (signal) signal.removeEventListener('abort', onExternalAbort);
-  }
-  return { content, toolCalls: toolCalls.filter(Boolean), aborted };
-}
+// ── LLM 流式调用 ─────────────────────────────────────────
+// 统一实现在 src/cli/ai/transport.js 的 streamChatOnce —— CLI 的 `g ai` 与这条
+// Web 链路共用同一份。比这里原先那份多出来的能力:请求 usage(带 stream_options
+// 降级重试)、校验 tool call index 边界、吃 evt.error,以及最要紧的一条 ——
+// 流被截断但 tool_calls 已经部分到达时抛"中断"、拒绝执行半截的工具调用。
+// 这里曾经是第二份实现,弱就弱在最后那条:半截参数有可能被拿去执行。
 
-// ── 消息消毒 ──────────────────────────────────────────────
-// 目的:确保发往 LLM provider 的 messages 数组里没有"空内容"或
-// "看似有内容但全是空白"的字段。
-//
-// 背景:部分 LLM provider(Moonshot/Kimi、智谱、火山引擎、MiniMax 等)
-// 对 assistant 历史消息 content 校验严格 —— 当某轮 assistant 只返回
-// tool_calls 而没有文本、或正文被 trim 后只剩空白时,provider 会拒绝
-// 并报 "chat content is empty (2013)"。OpenAI 官方规范允许 assistant
-// 消息在带 tool_calls 时 content 为 null,但 provider 实现不一致:
+// ── 消息准备(消毒 / 历史有界化 / 旧图片降级) ──────────────────
+// 统一实现在 src/cli/ai/context.js 的 prepareRequestMessages —— CLI 的 `g ai`
+// 与这条 Web 链路共用同一份,两侧不再各写一遍。
 //
-//   - assistant 带 tool_calls → content 强制 null(即使有字符串)
-//   - assistant 不带 tool_calls 且 content 全空白 → null(避免触发 2013)
-//   - user content 全空白 → 用单个空格 ' ' 占位(provider 通常可接受)
-//   - tool content 全空白 → '(no output)' 占位(防止序列化时被丢)
-function sanitizeMessages(messages) {
-  for (const m of messages) {
-    if (m == null || typeof m !== 'object') continue;
-    // 非字符串 content(数组形态的多模态 user 消息、已是 null 等) 不动
-    if (m.content === null || m.content === undefined) {
-      // assistant 必须显式 null,不要让 provider 看到 undefined
-      if (m.role === 'assistant') m.content = null;
-      continue;
-    }
-    if (typeof m.content !== 'string') continue;
-    const trimmed = m.content.trim();
-    if (trimmed === '') {
-      if (m.role === 'assistant') {
-        // 带 tool_calls 时强制 null;否则也置 null(provider 更安全)
-        m.content = null;
-      } else if (m.role === 'tool') {
-        m.content = '(no output)';
-      } else if (m.role === 'user') {
-        // user 不能 null(部分 provider 拒),用单个空格占位
-        m.content = ' ';
-      }
-      continue;
-    }
-    // assistant 带 tool_calls:即使是有效正文也强制 null
-    // (OpenAI 规范要求 tool_calls 出现时 content=null,避免 provider 误判)
-    if (m.role === 'assistant' && Array.isArray(m.tool_calls) && m.tool_calls.length > 0) {
-      m.content = null;
-    }
-  }
-  return messages;
-}
-
-// ── 历史裁剪 ──────────────────────────────────────────────
-// 保留 system + 最近 MAX_HISTORY_MESSAGES 条
-// 切口必须落在 user 消息上,避免把 assistant(tool_calls) 与其 tool 结果从中间撕开
-const MAX_HISTORY_MESSAGES = 40;
-function trimHistory(messages) {
-  if (messages.length <= MAX_HISTORY_MESSAGES + 1) return;
-  let cut = messages.length - MAX_HISTORY_MESSAGES;
-  // 至少向后扫 5 条,跳过孤立的 tool / assistant(tool_calls),
-  // 找到真正的 user 节点再切,防止撕裂 tool 调用链
-  let safety = 0;
-  while (cut < messages.length && messages[cut].role !== 'user' && safety < 8) {
-    cut++;
-    safety++;
-  }
-  if (cut <= 1 || messages[cut]?.role !== 'user') return;
-  messages.splice(1, cut - 1);
-}
-
-// ── 多模态历史:旧图片降级为文字 ──────────────────────────
-function stripStaleImages(messages) {
-  const placeholder = '[图片已从历史中省略]';
-  let seenLatest = false;
-  for (let i = messages.length - 1; i >= 0; i--) {
-    const m = messages[i];
-    if (m?.role !== 'user' || !Array.isArray(m.content)) continue;
-    const hasImage = m.content.some(p => p?.type === 'image_url');
-    if (!hasImage) continue;
-    if (!seenLatest) { seenLatest = true; continue; }
-    m.content = m.content.map(p => p?.type === 'image_url'
-      ? { type: 'text', text: placeholder }
-      : p);
-  }
-}
+// 这里曾经是第二份实现:历史只按条数硬切(MAX_HISTORY_MESSAGES=40)+ 纯 splice 丢弃,
+// 而 CLI 侧早已改成「条数/字符双预算 + 丢弃项摘录成一条梗概」,两侧行为因此分叉 ——
+// Web 面板聊久了模型会直接失忆,CLI 还记得要点。现在收敛到一处,口径见 context.js。
 
 // ── 核心入口：运行一轮 agent 对话 ────────────────────────
 //
 // 参数:
-//   { session, model, userMessage, cwd, locale, signal, send, onChild }
+//   { session, model, userMessage, cwd, locale, signal, send, onChild, askUser, listProjects }
 //   - session: 从 agentSessionStore 读取的会话对象
 //   - model: { baseURL, model, apiKey }
 //   - userMessage: 用户输入文本
@@ -334,10 +229,13 @@ function stripStaleImages(messages) {
 //   - signal: AbortSignal (客户端断开时触发)
 //   - send: (obj) => void  SSE 发送函数
 //   - onChild: (child) => void  子进程回调(用于取消)
+//   - askUser: (args, meta) => Promise<string>  等待用户回答
+//   - listProjects: (args) => Promise<string>  list_projects 工具的数据源
+//     (由 agentRoutes 注入:最近目录/tasks.json/看板统计只有 GUI 侧拿得到)
 //
 // 返回: { aborted: boolean }
-export async function runAgentTurn({ session, model, userMessage, images = [], cwd, locale, signal, send, onChild }) {
-  const ctx = { cwd, locale, onChild };
+export async function runAgentTurn({ session, model, userMessage, images = [], cwd, locale, signal, send, onChild, askUser, listProjects }) {
+  const ctx = { cwd, locale, onChild, askUser, listProjects };
 
   // 确保 session.messages 存在
   if (!Array.isArray(session.messages)) session.messages = [];
@@ -361,20 +259,19 @@ export async function runAgentTurn({ session, model, userMessage, images = [], c
       : userMessage
   });
 
-  // 旧图片降级
-  stripStaleImages(session.messages);
-
   const maxIterations = await resolveMaxToolIterations();
 
   for (let iter = 0; iter < maxIterations; iter++) {
-    trimHistory(session.messages);
-    sanitizeMessages(session.messages);
+    // 每轮都从完整会话记录重新构建一次请求副本:条数/字符双预算 → 被丢掉的旧消息
+    // 摘录成一条梗概 → 旧图片降级 → provider 兼容消毒。
+    // 只作用于副本,session.messages 保持完整(与 CLI 的磁盘口径一致)。
+    const messages = prepareRequestMessages(session.messages, { locale });
 
     let result;
     try {
       result = await streamChatOnce({
         model,
-        messages: session.messages,
+        messages,
         signal,
         // 整轮对话（含后续工具调用产生的每一轮请求）复用同一个会话 ID
         sessionId: session.sessionId,
@@ -436,7 +333,16 @@ export async function runAgentTurn({ session, model, userMessage, images = [], c
       const argsPreview = summarizeArgs(name, args);
       send({ type: 'tool_call_start', toolCallId, name, argsPreview });
 
-      const output = await executeTool(name, args, ctx);
+      const toolCtx = { ...ctx, signal };
+      if (name === 'ask_user' && typeof ctx.askUser === 'function') {
+        toolCtx.askUser = askArgs => ctx.askUser(askArgs, {
+          sessionId: session.sessionId,
+          interactionId: toolCallId,
+          send,
+          signal,
+        });
+      }
+      const output = await executeTool(name, args, toolCtx);
       send({ type: 'tool_result', toolCallId, name, result: output });
       session.messages.push({ role: 'tool', tool_call_id: toolCallId, name, content: output });
     }

package/src/ui/server/routes/workbench/agentChatShared.test.js

@@ -0,0 +1,151 @@
+// Web 智能体面板（agentChat.js）必须与 CLI 的 `g ai` 共用同一条**上下文**和**传输**口径。
+//
+// 这两处曾经各有一份自己的实现：
+//   上下文：只按条数硬切 40 条 + 纯 splice 丢弃，且**原地**改 session.messages
+//          （聊久了模型失忆，磁盘上的会话记录还被永久削掉一截）
+//   传输  ：自带的 streamChatOnce 不拦"流被截断但 tool_calls 已部分到达" ——
+//          半截的参数有可能被拿去执行
+// 现在两者都收敛到 src/cli/ai/ 下的共享模块。这个文件跑真实的 runAgentTurn，
+// 把发出去的请求体与产生的事件抓下来，逐条钉住收敛后的口径。
+//
+// 注意:必须在 import agentChat.js **之前**把 USERPROFILE/HOME 指到沙箱,否则
+// configManager / agentSessionStore 会去读真实的 ~/.zen-gitsync。
+import { test } from 'node:test'
+import assert from 'node:assert/strict'
+import { mkdtempSync } from 'node:fs'
+import os from 'node:os'
+import path from 'node:path'
+
+const SANDBOX = mkdtempSync(path.join(os.tmpdir(), 'zen-agent-shared-'))
+process.env.USERPROFILE = SANDBOX
+process.env.HOME = SANDBOX
+delete process.env.HOMEDRIVE
+delete process.env.HOMEPATH
+
+const event = data => `data: ${JSON.stringify(data)}\n\n`
+const sse = chunks => new Response(new ReadableStream({
+  start(controller) {
+    const bytes = new TextEncoder()
+    chunks.forEach(chunk => controller.enqueue(bytes.encode(chunk)))
+    controller.close()
+  },
+}), { status: 200, headers: { 'content-type': 'text/event-stream' } })
+
+const model = { model: 'test', name: 'test', baseURL: 'https://example.invalid/v1', apiKey: 'test' }
+const call = n => ({ id: `call${n}`, type: 'function', function: { name: 'read_file', arguments: '{}' } })
+const okStream = () => sse([event({ choices: [{ delta: { content: 'ok' }, finish_reason: 'stop' }] }), 'data: [DONE]\n\n'])
+const withSystem = () => [{ role: 'system', content: 'rules' }, { role: 'user', content: 'hi' }]
+const longSession = () => {
+  const messages = [{ role: 'system', content: 'rules' }, { role: 'user', content: 'original goal' }]
+  for (let i = 0; i < 60; i++) {
+    messages.push({ role: 'assistant', tool_calls: [call(i)] })
+    messages.push({ role: 'tool', tool_call_id: `call${i}`, content: 'data'.repeat(4000) })
+  }
+  return messages
+}
+
+// 跑一轮真实的 runAgentTurn,桩掉 globalThis.fetch —— 抓住发出去的请求体与产生的事件
+async function runWeb({ session, userMessage, respond, locale = 'zh-CN' }) {
+  const { runAgentTurn } = await import('./agentChat.js')
+  const events = []
+  let sent
+  const original = globalThis.fetch
+  globalThis.fetch = async (_url, init) => { sent = JSON.parse(init.body); return respond() }
+  try {
+    const result = await runAgentTurn({
+      session, model, userMessage, locale, cwd: process.cwd(),
+      signal: new AbortController().signal, send: e => events.push(e),
+    })
+    return { sent, events, result }
+  } finally {
+    globalThis.fetch = original
+  }
+}
+
+const newSession = (sessionId, messages) => ({ sessionId, messages, cwd: process.cwd() })
+
+// ── 上下文口径 ────────────────────────────────────────────
+
+test('请求体与 context.js 的输出逐字段相同（有界 + 摘录）', async () => {
+  const { prepareRequestMessages } = await import('../../../../cli/ai/context.js')
+  const messages = longSession()
+  // 传副本进去:runAgentTurn 会往 session.messages 追加消息,原数组要留作期望值的输入
+  const session = newSession('ag-sandbox-test', structuredClone(messages))
+
+  const { sent } = await runWeb({ session, userMessage: 'next', respond: okStream })
+
+  assert.ok(sent, '必须发出过一次请求')
+  assert.ok(sent.messages.length <= 40, `请求副本未收窄: ${sent.messages.length} 条`)
+  assert.equal(sent.messages[0].role, 'system')
+  assert.match(sent.messages[1].content, /^\[Earlier conversation excerpts/)
+  assert.ok(sent.messages.some(m => m.content === 'original goal'), '原始目标必须保留')
+  // 逐字段相同 = 这条链路确实走的是同一个函数,而不是"看起来差不多"的第二份实现
+  assert.deepEqual(
+    sent.messages,
+    prepareRequestMessages([...messages, { role: 'user', content: 'next' }], { locale: 'zh-CN' }),
+  )
+})
+
+test('会话记录不再被原地裁剪（磁盘口径与 CLI 一致）', async () => {
+  const session = newSession('ag-sandbox-keep', longSession())
+
+  await runWeb({ session, userMessage: 'next', respond: okStream })
+
+  assert.equal(session.messages.filter(m => m.role === 'tool').length, 60,
+    '旧消息必须留在会话记录里(裁剪只作用于请求副本)')
+  assert.equal(session.messages[1].content, 'original goal')
+  assert.equal(session.messages[2].tool_calls[0].id, 'call0')
+})
+
+test('短会话不做任何额外处理,也不塞摘录消息', async () => {
+  const session = newSession('ag-sandbox-short', withSystem())
+
+  const { sent } = await runWeb({ session, userMessage: 'hello', respond: okStream })
+
+  assert.deepEqual(sent.messages.map(m => m.content), ['rules', 'hi', 'hello'])
+})
+
+// ── 传输口径 ──────────────────────────────────────────────
+
+test('请求体带上 stream_options,与 CLI 一致地统计 usage', async () => {
+  const session = newSession('ag-sandbox-usage', withSystem())
+
+  const { sent } = await runWeb({ session, userMessage: 'hello', respond: okStream })
+
+  assert.deepEqual(sent.stream_options, { include_usage: true })
+  assert.equal(sent.stream, true)
+  assert.equal(sent.temperature, 0.3)
+  assert.ok(Array.isArray(sent.tools) && sent.tools.length > 0, '内置工具表必须随请求下发')
+})
+
+test('流被截断（没有 finish_reason 也没有 [DONE]）时不执行半截的工具调用', async () => {
+  const session = newSession('ag-sandbox-truncated', withSystem())
+  // 工具调用参数是完整的、JSON 合法 —— 换在没有这道闸的旧实现里就会被真的执行
+  const truncated = sse([event({
+    choices: [{ delta: { tool_calls: [{ index: 0, id: 'call_x', function: { name: 'read_file', arguments: '{"path":"nope.txt"}' } }] } }],
+  })])
+
+  const { events } = await runWeb({ session, userMessage: '读个文件', respond: () => truncated })
+
+  const error = events.find(e => e.type === 'error')
+  assert.ok(error, '截断的流必须报错,不能静默当成正常结束')
+  assert.match(error.error, /中断/)
+  assert.equal(events.some(e => e.type === 'tool_call_start'), false, '被截断的调用不得开始执行')
+  assert.equal(session.messages.some(m => m.role === 'tool'), false, '被截断的调用不得留下工具结果')
+})
+
+test('400 且正文提到 tool/function 时提示换模型,而不是甩网关 JSON', async () => {
+  const session = newSession('ag-sandbox-nofc', withSystem())
+  const noFunctionCalling = () => new Response(
+    JSON.stringify({ error: { message: 'tools is not supported by this model' } }),
+    { status: 400, headers: { 'content-type': 'application/json' } },
+  )
+
+  const { sent, events } = await runWeb({ session, userMessage: 'hello', respond: noFunctionCalling })
+
+  assert.equal(sent.messages.length, 3, '不该把 stream_options 的降级重试和这条错误混在一起')
+  const error = events.find(e => e.type === 'error')
+  assert.ok(error)
+  assert.match(error.error, /不支持 function calling/)
+  assert.match(error.error, /换用支持工具调用的模型/)
+})