feat(agent): sandbox conversation analysis

fix(agent): isolate conversation workspaces
2026-08-25 16:08:33 +08:00 · 2026-08-25 13:18:39 +08:00
37 changed files with 2941 additions and 106 deletions
@@ -1,7 +1,6 @@
 node_modules/
 __pycache__/
 .opencode/node_modules/
-.opencode/skills/
 .local.env
 .vscode
 docker-compose.yml
@@ -9,6 +9,7 @@ model: deepseek/deepseek-v4-flash

 - 工具执行期间不输出过程说明，全部完成后只回复最终结果
 - 直接给出结论、关键数据和可执行建议，默认仅展示最重要的 Top 5；数据不足或任务失败时简要说明影响和下一步
+- 多步骤或预计超过 30 秒的任务，开始时使用 `todowrite` 给用户展示计划，并在每个里程碑更新状态；简单问答不创建计划

 ## 工作流生命周期

@@ -49,6 +50,9 @@ Skills 树是**动态生长的**——工作流不是预置的，而是从实际
 4. 避免直接用 `Read` 或 `cat` 读取结果文件，尤其是大文件；优先用 `head`/`tail`/`rg` 截断查看，或用 Python 只向 stdout 输出精简 JSON，避免大文件冲击 stdin/stdout
 5. 无可用数据时不得编造结果
 6. 禁止使用 `task` 子代理；当前前端无法观测和干预子代理的具体工作过程
+7. Bash 始终运行在当前对话专属沙箱中：只能读写当前对话目录、读取 skills 和 Python 环境，不能联网，也不能访问 `/app` 源码、密钥、其他对话或全局 `tool-output`
+8. 需要文件输入时，调用 `tjwater_cli(..., store_result=true)`，使用返回的 `data_file.file_path`；不要把 JSON 手工写到 `/tmp`
+9. `store_render_ref` 的输入必须是 `{metadata, location: {file_path}, data}` 包装 JSON；若旧脚本输出裸数据，先在当前对话目录内包装，且 `location.file_path` 必须等于包装文件的绝对路径

 ## 工作流沉淀（skill_manager）

@@ -63,30 +67,12 @@ Skills 树是**动态生长的**——工作流不是预置的，而是从实际

 目录入口也通过 `skill_manager` 维护：更新 `skills/workflow/SKILL.md` 时使用 `write_skill(skill_path="workflow", ...)`，更新根入口 `skills/SKILL.md` 时使用 `write_skill(skill_path="__root__", ...)`。

-**脚本编写要求——优先用 pipe 串联**：
+**脚本编写要求——数据获取与本地分析分离**：

-workflow skill 脚本应尽量用 shell pipe 在一次 subprocess 调用中串联多个 CLI 命令。减少 tool calling 次数，提升执行效率。
-
-```python
-import subprocess, os
-
-# env dict 仅用于当前子进程，不污染 os.environ，多用户安全
-env = {**os.environ,
-       "TJWATER_SERVER": auth["server"],
-       "TJWATER_ACCESS_TOKEN": auth["access_token"], ...}
-
-# 好：一次 shell 调用，pipe 串联
-cmd = "tjwater-cli net list-pipes | jq '...' | xargs tjwater-cli analysis calc"
-result = subprocess.run(cmd, shell=True, env=env, capture_output=True, text=True)
-
-# 差：多次 subprocess.run
-step1 = subprocess.run(["tjwater-cli", "net", "list-pipes"], ...)
-step2 = subprocess.run(["tjwater-cli", "analysis", "calc"], ...)
-```
-
-管道场景下用子进程隔离的 env dict 传认证，释放 stdin 给管道数据流。不修改全局 `os.environ`。认证 JSON 由内部桥接注入，脚本不硬编码。
-
-CLI **不增加** `--input/--output`，数据转换由 `jq`/`xargs` 在 shell 管道中完成。
+- 后端数据只能由 `tjwater_cli` 工具获取；认证与网络请求留在 Agent 主进程
+- 分析脚本接收 `data_file.file_path`，只处理当前对话目录内的本地文件
+- 多份互不依赖的数据可并行调用 `tjwater_cli(..., store_result=true)`，随后在一次沙箱 Bash 中运行 Python 分析
+- 脚本输出文件必须写入当前工作目录；禁止使用 `/tmp`、全局 `tool-output` 或硬编码认证环境变量

 **触发时机**：
 - 用户明确说"保存/沉淀/记录工作流"
@@ -0,0 +1,186 @@
+---
+name: tjwater-cli
+description: tjwater-cli 命令行工具使用说明，涵盖命令发现、输出格式、命令族、错误处理及最佳实践。
+---
+
+# tjwater-cli 使用说明
+
+## 概述
+
+`tjwater-cli` 是 TJWater 供水管网系统的命令行工具，用于与后端服务交互，支持数据查询、分析和工程操作。所有输出统一为 JSON 格式。
+
+## 工具调用
+
+通过 `tjwater_cli` 工具执行 CLI 命令：
+
+```json
+{
+  "reason": "说明调用原因",
+  "command": "project list",
+  "timeout": 120,
+  "store_result": false
+}
+```
+
+| 参数 | 类型 | 必填 | 说明 |
+|------|------|------|------|
+| `reason` | string | 是 | 调用原因 |
+| `command` | string | 是 | CLI 子命令（不含二进制路径和 `--auth-context`） |
+| `timeout` | number | 否 | 超时秒数，默认 120，大结果集建议 300+ |
+| `store_result` | boolean | 否 | 强制保存到当前对话目录并返回 `data_file.file_path`；分析脚本需要文件输入时设为 true |
+
+认证上下文（token、server、project、network）由内部桥接自动注入，无需手动传参。
+
+## 命令发现
+
+Agent 通过 `help` 动态发现可用命令，而非依赖硬编码清单。
+
+**重要：命令分为两类——触发动作与数据获取。**
+
+- **触发动作**（`simulation`、`analysis`）：向服务端发起计算请求，返回任务状态/ID，**不直接返回分析结果**。
+- **数据获取**（`data timeseries`）：所有计算结果（仿真压力、分析指标等）的唯一数据出口，需在触发动作完成后调用。
+
+```
+simulation/analysis → 触发计算 → 返回状态/任务ID
+                              ↓
+                  data timeseries → 获取计算结果
+```
+
+通过 `help` 发现命令：
+
+```
+tjwater-cli help                   → 一级命令清单（含 commands 数组和 summary）
+tjwater-cli help data timeseries   → data timeseries 的子命令与参数详情
+tjwater-cli help simulation        → simulation 的子命令与参数详情
+tjwater-cli help COMMAND           → 子命令与参数详情
+```
+
+`help` 返回 JSON 格式，Agent 可直接解析 `commands` 数组识别可用能力。
+
+**严禁猜测命令或参数！** 所有命令路径、子命令和参数（名称、类型、必填/可选）均以 `help` 输出为准。执行任何命令前，必须先通过 `help` 确认其存在及参数签名，禁止凭经验拼写。
+
+### 已知命令族
+
+| 命令族 | 典型子命令 | 用途 |
+|------|-----------|------|
+| `project` | `list`, `db-health` | 项目管理、数据库健康检查 |
+| `data` | `timeseries realtime links / nodes`, `timeseries scada query` | **时序数据查询**（实时/SCADA），所有分析结果的唯一获取渠道 |
+| `simulation` | 通过 `help simulation` 发现 | **触发水力仿真计算**（执行成功返回状态，实际结果需走 `data timeseries` 获取） |
+| `analysis` | 通过 `help analysis` 发现 | **触发分析计算**（执行成功返回状态，实际结果需走 `data timeseries` 获取） |
+| `net` | `list-pipes` | 管网拓扑查询 |
+| `help` | (无子命令) | 命令发现入口 |
+
+> 完整命令清单始终以 `tjwater-cli help` 实时输出为准。
+
+## 输出格式
+
+所有命令返回统一 JSON 结构：
+
+```json
+{
+  "schema_version": "tjwater-cli/v1",
+  "ok": true,
+  "data": { ... },
+  "error": {
+    "code": "COMMAND_NOT_FOUND",
+    "message": "详细错误描述"
+  }
+}
+```
+
+- `ok: true` — 成功，数据在 `data` 字段
+- `ok: false` — 失败，检查 `error.code` 和 `error.message`
+
+### 大结果集处理
+
+超过内联阈值的结果不会终止 CLI，而是保存到当前对话的 `tool-data/` 目录并返回：
+
+```json
+{
+  "ok": true,
+  "data_file": {
+    "file_path": "/app/data/conversation-workspaces/conversation-.../tool-data/cli-....json",
+    "bytes": 38700000,
+    "content_type": "application/json"
+  }
+}
+```
+
+禁止完整读取超大结果集。优先使用：
+- 采样/截断参数（如 `--limit`、`--offset`）
+- `--field` 按字段过滤
+- `store_result=true` 后用沙箱 Python 脚本按字段读取
+
+## 错误码速查
+
+| error.code | 含义 | 来源 | 处理建议 |
+|------|------|------|------|
+| `UNAUTHENTICATED` | 缺少 access token | CLI `core.py:162` | 检查认证上下文注入 |
+| `SERVER_ERROR` | 后端返回 error 状态 | CLI `core.py:400` | 记录 `request_id`，结合后端日志排查 |
+| `REQUEST_TIMEOUT` | CLI 请求后端超时 | CLI `core.py:445` | 增大 `timeout` 参数或检查后端负载 |
+| `TIMEOUT` | bridge 层进程超时 | Agent `server.ts:199` | 增大 `tjwater_cli` 的 `timeout` 参数 |
+| `COMMAND_NOT_FOUND` | 命令/子命令不存在 | CLI `helping.py:300` | 执行 `help` 确认命令拼写 |
+| `INPUT_NOT_FOUND` | `--input` 文件不存在 | CLI `core.py:243` | 检查文件路径 |
+| `REQUEST_FAILED` | 网络连接失败 | CLI `core.py:453` | 检查服务端可达性 |
+| `AUTH_CONTEXT_INVALID` | 认证上下文格式错误 | CLI `core.py:111` | 检查 auth headers 格式 |
+
+## 最佳实践
+
+1. **禁止猜测命令** — 执行任何命令前必须先 `tjwater_cli(command="help ...")` 确认命令存在及参数签名，参数均已写在 help 中，禁止凭经验拼写
+2. **reason 必填** — 每次调用必须说明具体理由
+3. **触发后取数据** — `simulation`/`analysis` 仅触发计算，结果必须从 `data timeseries` 获取，勿将触发返回的状态信息当作分析结果
+4. **文件分析** — workflow 脚本需要文件时使用 `store_result=true`，不得从 Bash 直接联网调用 CLI
+5. **结果验证** — 始终检查 `ok` 字段，失败时先处理错误码再重试
+6. **大结果集** — 优先过滤/采样，不要一次性拉取全部数据
+7. **模拟时长控制** — 模拟（`simulation`）或方案模拟的 `--duration` 不宜过长，建议每次仿真时间跨度控制在一小时以内，避免计算耗时过长或结果数据量过大
+
+## 示例
+
+### 查询所有实时节点数据
+```json
+{
+  "reason": "获取最近1小时内全部节点的实时数据",
+  "command": "data timeseries realtime nodes --start-time 2026-06-03T08:00:00+08:00 --end-time 2026-06-03T09:00:00+08:00"
+}
+```
+> `data timeseries realtime nodes` 仅接受 `--start-time` / `--end-time`，返回全量节点数据。
+
+### 按节点查询方案时序字段
+```json
+{
+  "reason": "查询节点 J-001 最近1小时的压力数据",
+  "command": "data timeseries scheme node-field --node J-001 --field pressure --start-time 2026-06-03T08:00:00+08:00 --end-time 2026-06-03T09:00:00+08:00"
+}
+```
+
+### 查询 SCADA 时序数据
+```json
+{
+  "reason": "查询 SCADA 设备 170490 在指定时间范围的 monitored_value",
+  "command": "data timeseries scada query --device-id 170490 --field monitored_value --start-time 2026-06-02T00:00:00+08:00 --end-time 2026-06-03T00:00:00+08:00"
+}
+```
+
+### 触发仿真并获取结果
+
+通常系统会自动跑仿真，建议**先尝试获取结果**，若无数据再触发仿真：
+
+```json
+// step 1: 先尝试获取仿真结果
+{
+  "reason": "尝试获取节点 J-001 09:00 时刻的仿真压力",
+  "command": "data timeseries realtime simulation-by-id-time --id J-001 --type junction --time 2026-06-03T09:00:00+08:00"
+}
+// step 2: 若 step 1 无数据（ok: false 或 data 为空），触发仿真
+{
+  "reason": "无已有仿真结果，触发1小时水力仿真",
+  "command": "simulation run --start-time 2026-06-03T08:00:00+08:00 --duration 60"
+}
+// step 3: 仿真完成后，再次获取结果（同 step 1）
+{
+  "reason": "获取仿真结果中节点 J-001 09:00 时刻的压力",
+  "command": "data timeseries realtime simulation-by-id-time --id J-001 --type junction --time 2026-06-03T09:00:00+08:00"
+}
+```
+
+`simulation run` 仅接受 `--start-time`（RFC3339，必填）和 `--duration`（整数分钟，必填）。
@@ -0,0 +1,150 @@
+---
+name: hydraulic-bottleneck-analysis
+description: 基于实时水力数据的管网水力瓶颈识别与改造建议。复合评分法（流速×水头损失）定位瓶颈管段，输出分级改造方案。
+---
+
+# 水力瓶颈分析工作流
+
+## 概述
+
+本工作流通过复合评分法（流速分级 × 水头损失百分位）从全管网管道中识别水力瓶颈管段，并结合节点压力、管径、粗糙系数给出分级改造建议。
+
+适用场景：管网运行评估、管网改造优先级排序、泵站阀站运行诊断。
+
+## 评分方法论
+
+### 双维度复合评分
+
+| 维度 | 判定标准 | 分值 |
+|------|----------|------|
+| **流速** | >3.0 m/s = 极危 | 3 |
+|  | 2.0–3.0 m/s = 严重 | 2 |
+|  | 1.5–2.0 m/s = 偏高 | 1 |
+|  | <1.5 m/s = 正常 | 0 |
+| **水头损失** | >P90 = 严重 | 2 |
+|  | P80–P90 = 中度 | 1 |
+|  | <P80 = 正常 | 0 |
+
+**瓶颈判定**：`(流速≥1 且 水损≥1)` 或 `(流速≥2)` —— 即双侧超标或单侧流速严重。
+
+**复合评分** = 流速分值 + 水损分值（最高 5 分），按降序排列。
+
+### 辅助指标
+
+| 指标 | 阈值 | 含义 |
+|------|------|------|
+| 节点压力 < 20m | — | 低压区域，需增压 |
+| 节点压力 20–25m | — | 压力偏低 |
+| roughness > 130 | — | 管壁粗糙，建议内衬修复 |
+
+> ⚠️ **setting 字段无效**：`data timeseries realtime links` 返回的 `setting` 字段为无效值，不可用于阀门节流或水泵出口判定。若需确定阀门/泵状态，应通过 `network get-link-properties` 逐条查询。
+
+## 数据依赖
+
+| 步骤 | 命令 | 数据量 | 超时 | 关键字段 |
+|------|------|--------|------|----------|
+| ① 管道属性 | `network get-all-pipes-properties` | ~11.7MB / 91K条 | 120s | id, node1, node2, length, diameter, roughness |
+| ② 管道水力 | `data timeseries realtime links --start-time T --end-time T+15min` | ~39MB / 182K条 | 300s | id, flow, velocity, headloss, time |
+| ③ 节点压力 | `data timeseries realtime nodes --start-time T --end-time T+15min` | ~28MB / 176K条 | 300s | id, pressure, time |
+
+> **时间窗口说明**：模拟步长 15 分钟，查询窗口取 `T~T+15min` 可覆盖 1–2 个时间步。脚本内部按 `--target-time` 精确筛选目标时刻的记录。
+
+> **大结果集处理**：三份调用都使用 `store_result=true`，结果保存到当前对话的 `tool-data/` 目录。脚本直接读取每次返回的 `data_file.file_path`，不得访问全局 `tool-output/`。
+
+## 执行步骤
+
+### 第 1 步：拉取三份数据
+
+并行发起 3 个 `tjwater_cli` 调用（互不依赖）：
+
+```bash
+# ① 管道静态属性
+tjwater_cli(command="network get-all-pipes-properties", timeout=120, store_result=true)
+
+# ② 目标时刻管道水力数据
+tjwater_cli(command="data timeseries realtime links --start-time 2026-04-01T08:00:00+08:00 --end-time 2026-04-01T08:15:00+08:00", timeout=300, store_result=true)
+
+# ③ 目标时刻节点压力数据
+tjwater_cli(command="data timeseries realtime nodes --start-time 2026-04-01T08:00:00+08:00 --end-time 2026-04-01T08:15:00+08:00", timeout=300, store_result=true)
+```
+
+### 第 2 步：运行分析脚本
+
+```bash
+python3 <skill_dir>/scripts/bottleneck_analysis.py \
+  --pipe-props <data_file.file_path-①> \
+  --realtime <data_file.file_path-②> \
+  --node-pressures <data_file.file_path-③> \
+  --target-time '2026-04-01T08:00:00+08:00' \
+  --top 50 \
+  2>./bottleneck_report.txt
+```
+
+**脚本参数**：
+- `--pipe-props`：管道属性 JSON 文件路径（必填）
+- `--realtime`：实时管道水力 JSON 文件路径（必填）
+- `--node-pressures`：实时节点压力 JSON 文件路径（必填）
+- `--target-time`：目标时刻 ISO8601（必填，如 `2026-04-01T08:00:00+08:00`）
+- `--top`：输出 Top N 瓶颈管段（默认 100）
+
+**输出**：
+- **stderr**：文本摘要 + Top 30 表格（可重定向到文件查看）
+- **stdout**：完整 JSON 结果，包含 `summary` 和 `top_bottlenecks`（含每条的建议 `suggestions`）
+
+### 第 3 步：结果解读与分类
+
+按瓶颈严重程度和根因分类：
+
+| 类别 | 判定条件 | 优先处理方案 |
+|------|----------|--------------|
+| 🔴 极危 | 流速>3.0m/s 且 水损>P90 | 检查模型/扩容/分流 |
+| 🔵 串联瓶颈 | 连续多根瓶颈管段共享节点 | 统一规划扩径，一次性解除 |
+| 🟣 低压区 | 端节点压力<20m | 扩径 + 评估中途加压 |
+| 🟡 高粗糙度 | roughness>130 的瓶颈管段 | 内衬修复降阻 |
+| 🟢 短管异常 | length<10m 且 headloss>P95 | 检查模型是否存在局部阻塞 |
+
+### 第 4 步：可视化（可选）
+
+1. **地图定位**：将 Top N 瓶颈管段用 `locate_features` 高亮到地图
+2. **统计图表**：用 `show_chart` 展示管径分布柱状图 + 流速等级柱状图
+3. **样式渲染**：可对 pipes 图层按 velocity 或 headloss 属性做分层设色
+
+## 改造建议生成逻辑
+
+脚本自动为每条瓶颈管段生成建议，规则如下：
+
+```
+if velocity > 2.0 → "流速X.XXm/s过高，需扩容或分流"
+elif velocity > 1.5 → "流速X.XXm/s偏高"
+
+if headloss > P95 → "水头损失X.XXm(>P95)严重超标"
+elif headloss > P90 → "水头损失X.XXm(>P90)"
+
+if diameter < 100 → "管径Xmm偏小，建议扩径至≥150mm"
+elif diameter < 200 → "管径Xmm，评估扩容至250-300mm"
+
+if roughness > 130 → "粗糙系数X偏高，建议内衬修复"
+
+if min_pressure < 20 → "端节点压力X.Xm(<20m)，低压区域需增压"
+elif min_pressure < 25 → "端节点压力X.Xm偏低"
+
+if length < 0.01 and headloss > 0.5 → "短管高水损，检查是否存在模型异常或局部阻塞"
+```
+
+## 参考数据规模（实测）
+
+基于 91,000 管段 / 88,000 节点规模的管网模型：
+
+| 指标 | 实测值 |
+|------|--------|
+| 管道属性数据量 | 91,052 条 / ~11.7MB |
+| 实时管道数据量 | 182,108 条（2步）/ ~38.7MB |
+| 实时节点数据量 | 175,814 条（2步）/ ~28.2MB |
+| 分析脚本处理时间 | 约 10-20 秒 |
+| 典型瓶颈数量 | 50-100 条（占总管数 0.05%-0.1%） |
+
+## 已知限制
+
+- 水头损失百分位阈值（P80/P90）基于**全管网**统计，如果管网上游存在极端水损（如 400m+），会拉高整体 P 值，导致部分中高水损管段被漏判。极端场景下可考虑对水损做分位数裁剪（如排除 >P99.9 的离群值）后再计算 P80/P90。
+- **setting 字段不可用**：`data timeseries realtime links` 返回的 `setting` 值为无效数据，本工作流已移除所有基于 setting 的阀门节流 / 水泵出口判定。若需要此类判定，应通过 `network get-link-properties` 逐条获取属性中的 setting 作为替代。
+- 脚本读取全量 JSON 入内存，峰值内存约 200-300MB，需确保执行环境有足够内存。
@@ -0,0 +1,80 @@
+# 数据源与字段映射
+
+## CLI 命令清单
+
+### ① 管道静态属性
+
+```bash
+tjwater-cli network get-all-pipes-properties
+```
+
+返回字段：
+
+| 字段 | 类型 | 说明 | 分析用途 |
+|------|------|------|----------|
+| id | string | 管段 ID | 主键，关联水力数据 |
+| node1 | string | 起始节点 ID | 拓扑，定位压力 |
+| node2 | string | 终止节点 ID | 拓扑，定位压力 |
+| length | float | 管长 (m) | 短管高水损检测 |
+| diameter | int | 管径 (mm) | 管径分级，扩容建议 |
+| roughness | int | 粗糙系数 | 内衬修复判定 |
+| minor_loss | float | 局部水头损失系数 | 暂未使用 |
+| status | string | OPEN/CLOSED | 管道状态 |
+
+### ② 管道实时水力
+
+```bash
+tjwater-cli data timeseries realtime links --start-time <T> --end-time <T+15min>
+```
+
+返回字段：
+
+| 字段 | 类型 | 说明 | 分析用途 |
+|------|------|------|----------|
+| time | string | 时间戳 ISO8601 | 筛选目标时刻 |
+| id | string | 管段 ID | 关联静态属性 |
+| flow | float | 流量 | 辅助参考 |
+| velocity | float | 流速 (m/s) | **核心评分指标** |
+| headloss | float | 水头损失 (m) | **核心评分指标** |
+| setting | float | ⚠️ 无效值 | 时序 API 返回的 setting 为无效值，不可用于判定 |
+| friction | float | 摩擦系数 | 暂未使用 |
+| quality | float | 水质 | 暂未使用 |
+| reaction | float | 反应速率 | 暂未使用 |
+| status | string | OPEN/CLOSED | 管道状态 |
+
+### ③ 节点实时压力
+
+```bash
+tjwater-cli data timeseries realtime nodes --start-time <T> --end-time <T+15min>
+```
+
+返回字段：
+
+| 字段 | 类型 | 说明 | 分析用途 |
+|------|------|------|----------|
+| time | string | 时间戳 ISO8601 | 筛选目标时刻 |
+| id | string | 节点 ID | 关联管段端点 |
+| pressure | float | 压力 (m) | **低压判定** |
+| total_head | float | 总水头 (m) | 含高程信息 |
+| actual_demand | float | 实际需水量 | 暂未使用 |
+| quality | float | 水质 | 暂未使用 |
+
+## 数据合并逻辑
+
+```
+管道属性 (pipe_map[id])  ←─id─→  实时水力 (rt[time==TT])
+                                       │
+                                  node1, node2
+                                       │
+                                       ↓
+                              节点压力 (node_pressure[id])
+```
+
+合并时以**实时水力数据为主表**，左联管道属性，再通过 node1/node2 查找两端压力。
+
+## 时间处理
+
+- 模拟步长：15 分钟
+- 查询窗口建议：T 到 T+15min（覆盖 1-2 步）
+- 脚本内精确筛选：`r.get('time') == TT` 严格匹配字符串
+- 若目标时刻（如 08:00）无数据，需先触发 `simulation run --start-time T --duration 15`
@@ -0,0 +1,196 @@
+#!/usr/bin/env python3
+"""
+水力瓶颈管道综合分析
+数据源：管道属性 + 实时水力 + 节点压力 → 复合评分 → 改造建议
+注：realtime links 的 setting 字段为无效值，已移除所有基于 setting 的判定。
+"""
+import json, sys, math, argparse
+from collections import defaultdict
+
+VELOCITY_THRESHOLDS = {"critical": 3.0, "severe": 2.0, "high": 1.5}
+
+def pct(d, v):
+    if not d: return 0
+    k = (v/100)*(len(d)-1); f=math.floor(k); c=math.ceil(k)
+    return d[f] if f==c else d[f]*(c-k)+d[c]*(k-f)
+
+def load_json(path):
+    with open(path, encoding='utf-8') as f: raw = f.read()
+    return json.loads(raw[raw.find('{'):])
+
+def main():
+    ap = argparse.ArgumentParser()
+    ap.add_argument('--pipe-props', required=True)
+    ap.add_argument('--realtime', required=True)
+    ap.add_argument('--node-pressures', required=True)
+    ap.add_argument('--target-time', required=True)
+    ap.add_argument('--top', type=int, default=100)
+    args = ap.parse_args()
+    TT = args.target_time
+
+    # 1. Load
+    print("[1/5] Loading data...", file=sys.stderr)
+    props = load_json(args.pipe_props).get('data', [])
+    pipe_map = {p['id']: p for p in props}
+
+    rt = load_json(args.realtime).get('data', [])
+    rt = [r for r in rt if r.get('time') == TT]
+
+    np = load_json(args.node_pressures).get('data', [])
+    np = [n for n in np if n.get('time') == TT]
+    node_pressure = {n['id']: n.get('pressure', n.get('value',0)) for n in np}
+    print(f"  Pipes: {len(props)}, Realtime: {len(rt)}, Node pressures: {len(node_pressure)}", file=sys.stderr)
+
+    # 2. Merge
+    print("[2/5] Merging...", file=sys.stderr)
+    merged = []
+    for r in rt:
+        pid = r['id']; prop = pipe_map.get(pid)
+        if prop:
+            merged.append({**prop, **r, '_prop_id': prop['id'], '_rt_id': r['id']})
+    print(f"  Merged: {len(merged)}", file=sys.stderr)
+
+    # 3. Score
+    print("[3/5] Scoring...", file=sys.stderr)
+    hl_vals = sorted([abs(m['headloss']) for m in merged])
+    p80 = pct(hl_vals, 80); p90 = pct(hl_vals, 90); p95 = pct(hl_vals, 95)
+
+    scored = []
+    for m in merged:
+        vel = abs(m['velocity']); hl = abs(m['headloss'])
+        diam = m.get('diameter', 0)
+        length = m.get('length', 0); roughness = m.get('roughness', 0)
+        n1, n2 = m['node1'], m['node2']
+        pid = m['id']
+
+        vs = 3 if vel>=3 else (2 if vel>=2 else (1 if vel>=1.5 else 0))
+        vg = "极危" if vs==3 else ("严重" if vs==2 else ("偏高" if vs==1 else "正常"))
+        hs = 2 if hl>p90 else (1 if hl>p80 else 0)
+        hg = "严重" if hs==2 else ("中度" if hs==1 else "正常")
+        composite = vs + hs
+        is_bn = (vs>=1 and hs>=1) or (vs>=2)
+
+        # Node pressures
+        p1 = node_pressure.get(n1); p2 = node_pressure.get(n2)
+        min_p = min(p1, p2) if (p1 is not None and p2 is not None) else None
+
+        # Pipe category
+        if diam <= 50: dcat = "微型(≤50mm)"
+        elif diam <= 100: dcat = "小型(51-100mm)"
+        elif diam <= 200: dcat = "中型(101-200mm)"
+        elif diam <= 400: dcat = "大型(201-400mm)"
+        elif diam <= 800: dcat = "主干(401-800mm)"
+        else: dcat = "干管(>800mm)"
+
+        scored.append({
+            'id': pid, 'node1': n1, 'node2': n2,
+            'velocity': round(vel, 4), 'flow': round(m.get('flow',0), 4),
+            'headloss': round(hl, 4), 'length': round(length, 4),
+            'diameter': int(diam), 'roughness': int(roughness),
+            'vel_grade': vg, 'hl_grade': hg, 'composite_score': composite,
+            'is_bottleneck': is_bn,
+            'n1_pressure': round(p1, 2) if p1 is not None else None,
+            'n2_pressure': round(p2, 2) if p2 is not None else None,
+            'min_pressure': round(min_p, 2) if min_p is not None else None,
+            'diam_cat': dcat,
+        })
+
+    # 4. Filter & sort
+    print("[4/5] Filtering bottlenecks...", file=sys.stderr)
+    bn = [s for s in scored if s['is_bottleneck']]
+    bn.sort(key=lambda x: (-x['composite_score'], -x['velocity']))
+
+    critical = [b for b in bn if b['vel_grade']=='极危']
+    severe = [b for b in bn if b['vel_grade']=='严重']
+    high_vel = [b for b in bn if b['vel_grade']=='偏高']
+
+    low_p = sum(1 for b in bn if b['min_pressure'] and b['min_pressure'] < 20)
+    lowish_p = sum(1 for b in bn if b['min_pressure'] and 20 <= b['min_pressure'] < 25)
+    high_rough = sum(1 for b in bn if b['roughness'] > 130)
+
+    # Diam distribution
+    dd = defaultdict(int)
+    for b in bn:
+        d = b['diameter']
+        if d <= 50: dd['≤50mm']+=1
+        elif d <= 100: dd['51-100mm']+=1
+        elif d <= 200: dd['101-200mm']+=1
+        elif d <= 400: dd['201-400mm']+=1
+        elif d <= 800: dd['401-800mm']+=1
+        else: dd['>800mm']+=1
+
+    # 5. Output
+    print("[5/5] Generating report...", file=sys.stderr)
+
+    # Print text summary to stderr
+    print(f"\n{'='*70}", file=sys.stderr)
+    print(f"  水力瓶颈分析报告 - {TT}", file=sys.stderr)
+    print(f"{'='*70}", file=sys.stderr)
+    print(f"  总管道数: {len(scored)}", file=sys.stderr)
+    print(f"  瓶颈管道: {len(bn)} ({len(bn)/len(scored)*100:.1f}%)", file=sys.stderr)
+    print(f"    极危(>3.0m/s): {len(critical)} 条", file=sys.stderr)
+    print(f"    严重(2.0-3.0): {len(severe)} 条", file=sys.stderr)
+    print(f"    偏高(1.5-2.0): {len(high_vel)} 条", file=sys.stderr)
+    print(f"\n  水头损失阈值: P80={p80:.4f}m P90={p90:.4f}m P95={p95:.4f}m", file=sys.stderr)
+    print(f"  平均={sum(hl_vals)/len(hl_vals):.4f}m 最大={max(hl_vals):.4f}m", file=sys.stderr)
+    print(f"\n  低压节点(<20m): {low_p} 条, 偏低(20-25m): {lowish_p} 条", file=sys.stderr)
+    print(f"  高粗糙度(>130): {high_rough} 条", file=sys.stderr)
+
+    print(f"\n  瓶颈管径分布:", file=sys.stderr)
+    for cat in ['≤50mm','51-100mm','101-200mm','201-400mm','401-800mm','>800mm']:
+        print(f"    {cat}: {dd.get(cat,0)} 条", file=sys.stderr)
+
+    # Text table (Top 30) to stderr
+    print(f"\n{'='*120}", file=sys.stderr)
+    print(f"{'排名':<5} {'管道ID':<10} {'流速(m/s)':<10} {'水损(m)':<10} {'管径(mm)':<9} {'管长(km)':<10} {'评分':<4} {'压力1':<8} {'压力2':<8} {'管径类别':<18}", file=sys.stderr)
+    print('-'*120, file=sys.stderr)
+    for i, b in enumerate(bn[:30]):
+        p1s = f"{b['n1_pressure']:.1f}" if b['n1_pressure'] is not None else "-"
+        p2s = f"{b['n2_pressure']:.1f}" if b['n2_pressure'] is not None else "-"
+        print(f"{i+1:<5} {b['id']:<10} {b['velocity']:<10.4f} {b['headloss']:<10.2f} {b['diameter']:<9} {b['length']:<10.4f} {b['composite_score']:<4} {p1s:<8} {p2s:<8} {b['diam_cat']:<18}", file=sys.stderr)
+
+    # Generate suggestions for each bottleneck
+    for b in bn:
+        sug = []
+        if b['velocity'] > 2.0:
+            sug.append(f"流速{b['velocity']:.2f}m/s过高，需扩容或分流")
+        elif b['velocity'] > 1.5:
+            sug.append(f"流速{b['velocity']:.2f}m/s偏高")
+        hl = b['headloss']
+        if hl > p95:
+            sug.append(f"水头损失{hl:.2f}m(>P95)严重超标")
+        elif hl > p90:
+            sug.append(f"水头损失{hl:.2f}m(>P90)")
+        if b['diameter'] < 100:
+            sug.append(f"管径{b['diameter']}mm偏小，建议扩径至≥150mm")
+        elif b['diameter'] < 200:
+            sug.append(f"管径{b['diameter']}mm，评估扩容至250-300mm")
+        if b['roughness'] > 130:
+            sug.append(f"粗糙系数{b['roughness']}偏高，建议内衬修复")
+        if b['min_pressure'] is not None and b['min_pressure'] < 20:
+            sug.append(f"端节点压力{b['min_pressure']:.1f}m(<20m)，低压区域需增压")
+        elif b['min_pressure'] is not None and b['min_pressure'] < 25:
+            sug.append(f"端节点压力{b['min_pressure']:.1f}m偏低")
+        if b['length'] < 0.01 and b['headloss'] > 0.5:
+            sug.append("短管高水损，检查是否存在模型异常或局部阻塞")
+        b['suggestions'] = sug
+
+    result = {
+        'target_time': TT,
+        'summary': {
+            'total_pipes': len(scored),
+            'bottleneck_count': len(bn),
+            'critical': len(critical), 'severe': len(severe), 'high_vel': len(high_vel),
+            'low_pressure_nodes': low_p, 'lowish_pressure_nodes': lowish_p,
+            'high_roughness_pipes': high_rough,
+            'headloss_p80': round(p80,4), 'headloss_p90': round(p90,4),
+            'headloss_p95': round(p95,4),
+            'diameter_distribution': dict(dd),
+        },
+        'top_bottlenecks': bn[:args.top],
+    }
+
+    print(json.dumps(result, ensure_ascii=False, indent=2))
+
+if __name__ == '__main__':
+    main()
@@ -0,0 +1,146 @@
+---
+name: service-area-analysis
+description: 基于实时水力模拟数据的水源追溯供水服务范围分区。通过管段流量确定水流方向，从水库BFS追溯服务节点，环网/零流量节点用无向拓扑补充分配，输出分区可视化。
+---
+
+# 供水服务范围分区工作流
+
+## 概述
+
+本工作流基于指定时刻的水力模拟结果，通过**流向追溯法**将全部管网节点分配到各水库的服务范围。核心思路：利用管段流量符号判定水流方向，构建有向图从水库逐级追溯，对环网和零流量节点用无向拓扑修正。
+
+适用场景：供水服务范围评估、DMA分区规划、多水源供水格局分析、管网调度策略评估。
+
+## 分区方法
+
+### 第一步：水流方向判定
+
+对于每条管段，根据实时流量 `flow` 判定水流方向：
+
+| flow 值 | 水流方向 | 说明 |
+|---------|----------|------|
+| `flow > 1e-6` | node1 → node2 | 正向流量 |
+| `flow < -1e-6` | node2 → node1 | 反向流量 |
+| `|flow| ≤ 1e-6` | 无方向 | 零流量，不参与有向追溯 |
+
+### 第二步：多源有向BFS
+
+1. 以每个水库为根节点，沿水流方向执行 BFS
+2. 遍历到的节点归属该水库的服务范围
+3. **先到先得**：一个节点首次被访问到的水库即为归属
+4. 预期覆盖 **85–90%** 节点
+
+### 第三步：无向拓扑修正
+
+有向BFS不可达节点（通常 10–15%）通过无向图邻近性补充分配：
+
+| 不可达原因 | 说明 |
+|------------|------|
+| 环状管网 | 水流回路中下游节点反向连回上游，有向遍历被阻断 |
+| 零流量管段 | `flow≈0` 的管段无方向，其下游节点断开 |
+| 多水源交汇 | 交汇区流向往复，非树状拓扑 |
+
+### 输出统计
+
+每个分区输出：
+- `node_count`：分区内节点总数
+- `total_demand`：总需水量（负数=净供水区）
+- `avg_pressure`/`min_pressure`/`max_pressure`：压力统计
+
+## 数据依赖
+
+| 步骤 | 命令 | 数据量 | 超时 | 关键字段 |
+|------|------|--------|------|----------|
+| ① 管道拓扑 | `network get-all-pipes-properties` | ~11.7MB / 91K条 | 120s | id, node1, node2 |
+| ② 水库属性 | `network get-all-reservoirs-properties` | ~小 | 120s | id, links |
+| ③ 管段流量 | `data timeseries realtime links --start-time T --end-time T+15min` | ~39MB / 182K条 | 300s | id, flow, time |
+| ④ 节点数据 | `data timeseries realtime nodes --start-time T --end-time T+15min` | ~28MB / 176K条 | 300s | id, pressure, actual_demand, time |
+
+> **时间窗口**：模拟步长 15 分钟，查询 T~T+15min 覆盖 1–2 个时间步。脚本按 `--target-time` 精确筛选。
+
+> **文件输入**：四份调用都使用 `store_result=true`，包括结果较小的水库属性。脚本读取每次返回的 `data_file.file_path`；文件都属于当前对话，禁止使用 `/tmp` 或全局 `tool-output/`。
+
+## 执行步骤
+
+### 第 1 步：并行拉取数据
+
+4 个 `tjwater_cli` 调用（互不依赖），可一次发起：
+
+```bash
+# ① 管道静态拓扑
+tjwater_cli(command="network get-all-pipes-properties", timeout=120, store_result=true)
+
+# ② 水库属性
+tjwater_cli(command="network get-all-reservoirs-properties", timeout=120, store_result=true)
+
+# ③ 目标时刻管段流量
+tjwater_cli(command="data timeseries realtime links --start-time 2026-04-01T08:00:00+08:00 --end-time 2026-04-01T08:15:00+08:00", timeout=300, store_result=true)
+
+# ④ 目标时刻节点数据
+tjwater_cli(command="data timeseries realtime nodes --start-time 2026-04-01T08:00:00+08:00 --end-time 2026-04-01T08:15:00+08:00", timeout=300, store_result=true)
+```
+
+### 第 2 步：运行分区脚本
+
+```bash
+python3 <skill_dir>/scripts/service_area_partition.py \
+  --pipe-props <data_file.file_path-①> \
+  --reservoirs <data_file.file_path-②> \
+  --links <data_file.file_path-③> \
+  --nodes <data_file.file_path-④> \
+  --target-time '2026-04-01T08:00:00+08:00' \
+  --output ./service_area_partition_wrapper.json
+```
+
+**脚本参数**：
+
+| 参数 | 说明 | 必填 |
+|------|------|------|
+| `--pipe-props` | 管道属性 JSON 文件路径 | 是 |
+| `--reservoirs` | 水库属性 JSON 文件路径 | 是 |
+| `--links` | 实时管段数据 JSON 文件路径 | 是 |
+| `--nodes` | 实时节点数据 JSON 文件路径 | 是 |
+| `--target-time` | 目标时刻 ISO8601 | 是 |
+| `--output` | 分区结果输出路径 | 是 |
+
+**输出**：
+- **stderr**：处理日志 + 各分区统计表格
+- **stdout**：紧凑 JSON 摘要（total_nodes, reservoirs, areas）
+- **文件**：符合 `store_render_ref` 要求的 `{metadata, location, data}` 包装 JSON，其中 `data` 包含 `node_area_map`、`area_ids`、`area_colors` 和分析元数据
+
+### 第 3 步：前端可视化
+
+```bash
+# 持久化分区结果
+store_render_ref(file_path=<output-file>)
+
+# 渲染节点分区
+render_junctions(render_ref="res-xxxxxxxx-xxxx-xx")
+
+# 定位水库
+locate_features(ids=[...], feature_type="reservoir")
+
+# 展示统计图表
+show_chart(title="各水源分区节点数/压力对比", chart_type="bar", ...)
+```
+
+## 参考数据规模
+
+基于 91,000 管段 / 88,000 节点规模的管网模型：
+
+| 指标 | 实测值 |
+|------|--------|
+| 管道拓扑数据量 | 91,052 条 |
+| 水库数量 | 13 个 |
+| 总节点数 | 87,907 |
+| 有向BFS分配节点 | ~76,900 (87.5%) |
+| 无向修正节点 | ~11,000 (12.5%) |
+| 分区覆盖率 | 100% |
+| 脚本处理时间 | ~15-30 秒 |
+| 峰值内存 | ~400-500MB |
+
+## 已知限制
+
+- **水库顺序敏感**：多源 BFS 中先遍历到的水库优先分配，不同水库启动顺序可能影响边界区域分配结果
+- **单时刻快照**：分区仅反映目标时刻的水力工况，不同时段的泵站启停、阀门切换可能导致分区边界变化
+- **零流量阈值**：`1e-6` 阈值过滤极低流量管段，若管网有长期小流量管段可能漏判方向
@@ -0,0 +1,205 @@
+#!/usr/bin/env python3
+"""
+供水服务范围分析与分区 — 可复用脚本
+基于实时水力数据，从水库沿水流方向追溯服务范围，自动发现水库并分区。
+
+用法:
+  python3 service_area_partition.py \
+    --pipe-props pipes.json \
+    --reservoirs reservoirs.json \
+    --links realtime_links.json \
+    --nodes realtime_nodes.json \
+    --target-time '2026-04-01T08:00:00+08:00' \
+    --output ./service_area_partition_wrapper.json
+"""
+
+import argparse
+import json
+import os
+import sys
+from collections import deque, defaultdict
+
+COLORS = [
+    "rgba(31,119,180,0.7)", "rgba(255,127,14,0.7)", "rgba(44,160,44,0.7)",
+    "rgba(148,103,189,0.7)", "rgba(140,86,75,0.7)", "rgba(227,119,194,0.7)",
+    "rgba(127,127,127,0.7)", "rgba(188,189,34,0.7)", "rgba(23,190,207,0.7)",
+    "rgba(174,199,232,0.7)", "rgba(255,152,150,0.7)", "rgba(196,156,148,0.7)",
+    "rgba(219,64,82,0.7)", "rgba(153,204,153,0.7)", "rgba(255,204,102,0.7)",
+    "rgba(102,102,204,0.7)", "rgba(204,102,102,0.7)", "rgba(102,204,204,0.7)",
+    "rgba(204,153,204,0.7)", "rgba(153,153,153,0.7)"
+]
+
+def load_json(path, label):
+    print(f"Loading {label}...", file=sys.stderr)
+    with open(path) as f:
+        return json.load(f)
+
+def main():
+    parser = argparse.ArgumentParser(description="供水服务范围分区分析")
+    parser.add_argument("--pipe-props", required=True, help="管道属性 JSON 文件")
+    parser.add_argument("--reservoirs", required=True, help="水库属性 JSON 文件")
+    parser.add_argument("--links", required=True, help="实时管段数据 JSON 文件")
+    parser.add_argument("--nodes", required=True, help="实时节点数据 JSON 文件")
+    parser.add_argument("--target-time", required=True, help="目标时刻 ISO8601")
+    parser.add_argument("--output", required=True, help="分区结果输出 JSON 路径")
+    args = parser.parse_args()
+
+    # --- Step 1: Load pipe topology ---
+    pdata = load_json(args.pipe_props, "pipe topology")["data"]
+    pipe_topology = {}
+    node_neighbors = defaultdict(set)
+    for p in pdata:
+        pid = p["id"]
+        n1, n2 = p["node1"], p["node2"]
+        pipe_topology[pid] = (n1, n2)
+        node_neighbors[n1].add(n2)
+        node_neighbors[n2].add(n1)
+    print(f"  {len(pdata)} pipes, {len(node_neighbors)} unique nodes", file=sys.stderr)
+
+    # --- Step 2: Discover reservoirs ---
+    rdata = load_json(args.reservoirs, "reservoirs")["data"]
+    reservoirs = [r["id"] for r in rdata]
+    print(f"  {len(reservoirs)} reservoirs: {reservoirs}", file=sys.stderr)
+
+    # --- Step 3: Load link flow at target time ---
+    ldata = load_json(args.links, "link flows")["data"]
+    target_links = [l for l in ldata if l["time"] == args.target_time]
+    flow_direction = {}
+    pipe_flow = {}
+    for l in target_links:
+        lid = l["id"]
+        flow_val = l["flow"]
+        pipe_flow[lid] = abs(flow_val)
+        if lid in pipe_topology:
+            n1, n2 = pipe_topology[lid]
+            if flow_val > 1e-6:
+                flow_direction[lid] = (n1, n2)
+            elif flow_val < -1e-6:
+                flow_direction[lid] = (n2, n1)
+    nonzero = len(flow_direction)
+    print(f"  {len(target_links)} link records, {nonzero} with non-zero flow", file=sys.stderr)
+
+    # --- Step 4: Load node data at target time ---
+    ndata = load_json(args.nodes, "node data")["data"]
+    target_nodes = [n for n in ndata if n["time"] == args.target_time]
+    node_pressure = {}
+    node_demand = {}
+    for n in target_nodes:
+        nid = n["id"]
+        node_pressure[nid] = n.get("pressure", 0)
+        node_demand[nid] = n.get("actual_demand", 0)
+    print(f"  {len(target_nodes)} nodes", file=sys.stderr)
+
+    # --- Step 5: Build downstream graph ---
+    downstream = defaultdict(set)
+    for _lid, (up, dn) in flow_direction.items():
+        downstream[up].add(dn)
+    print(f"  downstream graph: {len(downstream)} source nodes", file=sys.stderr)
+
+    # --- Step 6: Multi-source BFS along flow direction ---
+    reservoir_area = {}
+    node_served_by = {}
+    queue = deque()
+    for rid in reservoirs:
+        reservoir_area[rid] = {rid}
+        node_served_by[rid] = rid
+        queue.append((rid, rid, 0))
+
+    while queue:
+        node, source, dist = queue.popleft()
+        for neighbor in downstream.get(node, set()):
+            if neighbor not in node_served_by:
+                node_served_by[neighbor] = source
+                reservoir_area[source].add(neighbor)
+                queue.append((neighbor, source, dist + 1))
+
+    directed_count = len(node_served_by)
+    unassigned = set(node_pressure.keys()) - set(node_served_by.keys())
+    print(f"  flow-tracing assigned: {directed_count}, unassigned: {len(unassigned)}", file=sys.stderr)
+
+    # --- Step 7: Undirected proximity fallback ---
+    if unassigned:
+        print("  running proximity fallback...", file=sys.stderr)
+        ua_queue = deque()
+        ua_visited = {}
+        for nid, src in node_served_by.items():
+            ua_visited[nid] = src
+            ua_queue.append((nid, src, 0))
+
+        while ua_queue:
+            node, source, dist = ua_queue.popleft()
+            for neighbor in node_neighbors.get(node, set()):
+                if neighbor not in ua_visited:
+                    ua_visited[neighbor] = source
+                    reservoir_area[source].add(neighbor)
+                    ua_queue.append((neighbor, source, dist + 1))
+
+        still = set(node_pressure.keys()) - set(ua_visited.keys())
+        if still:
+            print(f"  WARNING: {len(still)} nodes still unassigned", file=sys.stderr)
+        node_served_by = ua_visited
+
+    # --- Step 8: Compute statistics ---
+    print(f"\n=== 供水服务范围分区统计 ===\n", file=sys.stderr)
+    area_stats = []
+    for rid in reservoirs:
+        nodes_in = reservoir_area.get(rid, set())
+        pressures = [node_pressure[n] for n in nodes_in if n in node_pressure]
+        demands = [node_demand[n] for n in nodes_in if n in node_demand]
+        area_stats.append({
+            "reservoir": rid,
+            "node_count": len(nodes_in),
+            "total_demand": round(sum(demands), 4),
+            "avg_pressure": round(sum(pressures)/len(pressures), 2) if pressures else 0,
+            "min_pressure": round(min(pressures), 2) if pressures else 0,
+            "max_pressure": round(max(pressures), 2) if pressures else 0,
+        })
+
+    area_stats.sort(key=lambda x: x["node_count"], reverse=True)
+    for s in area_stats:
+        print(f"  水源 {s['reservoir']:>8s}: {s['node_count']:>6d} 节点 | "
+              f"总需水={s['total_demand']:.2f} | "
+              f"压力 avg={s['avg_pressure']:.1f}m [{s['min_pressure']:.1f}–{s['max_pressure']:.1f}m]",
+              file=sys.stderr)
+
+    # --- Step 9: Assign colors and write output ---
+    area_colors = {}
+    for i, rid in enumerate(reservoirs):
+        area_colors[rid] = COLORS[i % len(COLORS)]
+
+    output = {
+        "node_area_map": node_served_by,
+        "area_ids": reservoirs,
+        "area_colors": area_colors,
+        "metadata": {
+            "analysis_time": args.target_time,
+            "total_nodes": len(node_served_by),
+            "reservoir_count": len(reservoirs),
+            "directed_assigned": directed_count,
+            "proximity_assigned": len(node_served_by) - directed_count,
+            "method": "flow-direction-source-tracing"
+        }
+    }
+
+    absolute_output = os.path.abspath(args.output)
+    wrapper = {
+        "metadata": {
+            "generated_by": "service_area_partition.py",
+            "schema_version": 1,
+        },
+        "location": {"file_path": absolute_output},
+        "data": output,
+    }
+    with open(absolute_output, "w", encoding="utf-8") as f:
+        json.dump(wrapper, f, ensure_ascii=False)
+
+    summary = {
+        "total_nodes": len(node_served_by),
+        "reservoirs": len(reservoirs),
+        "areas": area_stats,
+        "output_file": absolute_output
+    }
+    print(json.dumps(summary, ensure_ascii=False))
+
+if __name__ == "__main__":
+    main()
@@ -0,0 +1,53 @@
+import { tool } from "@opencode-ai/plugin";
+
+const internalBaseUrl =
+  process.env.TJWATER_AGENT_INTERNAL_BASE_URL ?? "http://127.0.0.1:8787";
+const internalToken = process.env.TJWATER_AGENT_INTERNAL_TOKEN ?? "";
+
+export default tool({
+  description:
+    "在当前对话专属的 Landlock 沙箱中运行 Shell 命令。只能读写当前对话工作区，不能访问网络、密钥、其他对话或应用源码。",
+  args: {
+    command: tool.schema.string().describe("需要在沙箱中执行的 Shell 命令。"),
+    description: tool.schema
+      .string()
+      .optional()
+      .describe("面向用户的简短命令说明。"),
+    timeout: tool.schema
+      .number()
+      .optional()
+      .describe("超时秒数，默认 120，最大 1800。"),
+  },
+  async execute(args, context) {
+    await context.ask({
+      permission: "bash",
+      patterns: [args.command],
+      always: [args.command],
+      metadata: {
+        command: args.command,
+        ...(args.description ? { description: args.description } : {}),
+      },
+    });
+    const response = await fetch(
+      `${internalBaseUrl}/internal/tools/sandbox-shell`,
+      {
+        method: "POST",
+        headers: {
+          "Content-Type": "application/json",
+          "x-agent-internal-token": internalToken,
+        },
+        body: JSON.stringify({
+          session_id: context.sessionID,
+          command: args.command,
+          description: args.description,
+          timeout: args.timeout,
+        }),
+      },
+    );
+    const text = await response.text();
+    if (!response.ok) {
+      throw new Error(text);
+    }
+    return text;
+  },
+});
@@ -3,12 +3,25 @@ import { tool } from "@opencode-ai/plugin";
 const internalBaseUrl =
  process.env.TJWATER_AGENT_INTERNAL_BASE_URL ?? "http://127.0.0.1:8787";
 const internalToken = process.env.TJWATER_AGENT_INTERNAL_TOKEN ?? "";
-const importDirectory =
-  process.env.RESULT_REF_IMPORT_DIR ?? "./data/result-imports";
+
+type StoreRenderRefArgs = {
+  file_path?: unknown;
+  filePath?: unknown;
+};
+
+export function resolveStoreRenderFilePath(args: StoreRenderRefArgs): string {
+  if (typeof args.file_path === "string" && args.file_path.trim() !== "") {
+    return args.file_path;
+  }
+  if (typeof args.filePath === "string" && args.filePath.trim() !== "") {
+    return args.filePath;
+  }
+  throw new Error("file_path is required");
+}

 export default tool({
  description:
-    `导入 ${importDirectory} 下的受控 JSON 包装文件并返回 render_ref。文件必须是 { metadata: object, location: { file_path: string }, data: { node_area_map, area_ids?, area_colors? } }，location.file_path 必须与传入的绝对路径完全一致。只接受该目录内的真实文件，不接受目录外路径或指向目录外的符号链接。`,
+    "导入当前对话工作目录下的受控 JSON 包装文件并返回 render_ref。文件必须是 { metadata: object, location: { file_path: string }, data: { node_area_map, area_ids?, area_colors? } }，location.file_path 必须与传入的绝对路径完全一致。只接受当前对话工作目录内的真实文件，不接受其他对话目录、目录外路径或指向目录外的符号链接。",
  args: {
    reason: tool.schema
      .string()
@@ -17,11 +30,17 @@ export default tool({
      ),
    file_path: tool.schema
      .string()
+      .optional()
      .describe(
-        `位于 ${importDirectory} 内的包装 JSON 文件绝对路径。必须包含 metadata、location.file_path 和 data；data 才是 render_junctions 使用的 { node_area_map, area_ids?, area_colors? }。`,
+        "位于当前对话工作目录内的包装 JSON 文件绝对路径。必须包含 metadata、location.file_path 和 data；data 才是 render_junctions 使用的 { node_area_map, area_ids?, area_colors? }。",
      ),
+    filePath: tool.schema
+      .string()
+      .optional()
+      .describe("兼容旧调用的参数名；新调用应优先使用 file_path。"),
  },
  async execute(args, context) {
+    const filePath = resolveStoreRenderFilePath(args);
    const response = await fetch(
      `${internalBaseUrl}/internal/tools/store-render-ref`,
      {
@@ -32,7 +51,7 @@ export default tool({
        },
        body: JSON.stringify({
          session_id: context.sessionID,
-          file_path: args.file_path,
+          file_path: filePath,
        }),
      },
    );
@@ -20,6 +20,12 @@ export default tool({
      .number()
      .optional()
      .describe("超时秒数，默认 120。大结果集建议设 300+。"),
+    store_result: tool.schema
+      .boolean()
+      .optional()
+      .describe(
+        "是否强制把结果保存到当前对话工作区并返回 data_file。分析脚本需要文件输入时设为 true。",
+      ),
  },
  async execute(args, context) {
    const response = await fetch(
@@ -34,6 +40,7 @@ export default tool({
          session_id: context.sessionID,
          reason: args.reason,
          command: args.command,
+          store_result: args.store_result,
          timeout: args.timeout,
        }),
      },
@@ -38,4 +38,4 @@ PRs should describe runtime behavior changes, list `bun run check` and any test

 Do not commit `.env`, logs, session transcripts, generated result references, or `node_modules/`. Keep registry and deploy credentials in Gitea secrets.

-Automatic approval for `glob` and `grep` must remain limited to canonical paths inside an explicit safe workspace subtree. Broad workspace-root searches, symlink escapes, external paths, and `.env`, `data/`, or `logs/` targets must stay interactive.
+Automatic approval for `glob` and `grep` must remain limited to canonical paths inside an explicit safe workspace subtree. Broad source-workspace searches, symlink escapes, external paths, and `.env`, ordinary `data/`, or `logs/` targets must stay interactive. The current session's canonical `data/conversation-workspaces/<conversation-id>/` directory is the only `data/` exception; access must still reject symlinks and every other conversation directory.
@@ -18,6 +18,7 @@ RUN if [ -n "${UBUNTU_APT_MIRROR}" ]; then \
    apt-get update && apt-get install -y --no-install-recommends \
    curl \
    jq \
+    libseccomp2 \
    unzip \
    python3 \
    python3-venv && \
@@ -43,6 +44,8 @@ RUN if [ -n "${UBUNTU_APT_MIRROR}" ]; then \
    rich \
    ipython \
    pytest && \
+    (getent group 10001 >/dev/null || groupadd --gid 10001 tjwater-sandbox) && \
+    (getent passwd 10001 >/dev/null || useradd --uid 10001 --gid 10001 --no-create-home --shell /usr/sbin/nologin tjwater-sandbox) && \
    rm -rf /var/lib/apt/lists/*

 FROM base AS deps
@@ -59,9 +62,10 @@ WORKDIR /app
 COPY --from=deps /app/node_modules ./node_modules
 COPY --from=deps /app/.opencode/node_modules ./.opencode/node_modules
 COPY package.json bun.lock ./
-COPY tsconfig.json opencode.json README.md .gitignore ./
+COPY tsconfig.json opencode.json README.md .gitignore Dockerfile ./
 COPY src ./src
 COPY cli ./cli
+COPY scripts ./scripts
 COPY .opencode ./.opencode
 RUN bun run check

@@ -29,6 +29,8 @@ data/                 本地运行时数据，禁止提交
 logs/                 本地日志，禁止提交
 ```

+仓库跟踪 `.opencode/skills/` 中经过评审的默认工作流基线；部署环境仍可通过持久化卷保留 `skill_manager` 在运行中沉淀的增量内容。默认基线不得包含真实客户数据、认证信息或本地执行产物。
+
 ## 本地开发

 项目使用 Bun：
@@ -86,13 +88,15 @@ TJWATER_API_BASE_URL=http://127.0.0.1:8000

 `opencode.json` 已启用 `experimental.continue_loop_on_deny`。用户拒绝权限请求后，OpenCode V1 会把拒绝结果交还给 Agent，让其尝试无需该权限的替代方案，而不是直接结束本轮执行。

-前端提供三种整体权限模式：“请求批准”只执行 OpenCode 明确允许的白名单，其他权限请求逐次交给用户确认；“自动批准”额外自动放行低风险业务工具、skill，以及真实路径位于工作区安全子树且不涉及 `.env`、`data/`、`logs/` 的 glob/grep；工作区根目录的宽泛搜索仍需确认。“始终允许”自动放行当前对话中所有未被 OpenCode 明确禁止的权限请求。自动放行统一使用单次批准，切换整体模式后立即恢复对应策略，不会写入持久授权。
+前端提供三种整体权限模式：“请求批准”只执行 OpenCode 明确允许的白名单，Shell 和写操作逐次交给用户确认；“自动批准”额外自动放行低风险业务工具、skill、沙箱 Shell，以及真实路径位于当前 conversation workspace 内且通过 realpath/symlink 校验的 read/edit/glob/grep；“始终允许”自动放行当前对话中所有未被 OpenCode 明确禁止的权限请求。自动放行统一使用单次批准，切换整体模式后立即恢复对应策略，不会写入持久授权。

-单次权限请求支持“允许一次”“保存授权”和“拒绝”。“保存授权”使用 OpenCode 的 `always` 回复，仅保存 OpenCode 为本次请求建议的权限范围，并只在当前 OpenCode 会话内生效。外部目录以及 `.env`、`data/`、`logs/` 路径仍由静态配置明确禁止，三种整体模式都不能绕过这些拒绝规则。
+单次权限请求支持“允许一次”“保存授权”和“拒绝”。“保存授权”使用 OpenCode 的 `always` 回复，仅保存 OpenCode 为本次请求建议的权限范围，并只在当前 OpenCode 会话内生效。任意外部目录默认仍由静态配置禁止；`.env`、普通 `data/`、`logs/` 和其他对话目录保持禁止。真实聊天会话使用 `data/conversation-workspaces/<随机目录>/` 作为独立工作目录。普通 `rm <文件>`、`rmdir` 和非强制递归删除可在沙箱内执行，`rm -rf`/`rm -fr` 及等价的递归强制删除形式会在执行前拒绝。

-`store_render_ref` 只会从 `RESULT_REF_IMPORT_DIR`（默认 `./data/result-imports`）导入包装格式 JSON。文件必须包含 `metadata`、`location.file_path` 和 `data`，且真实路径不能越出导入目录；单文件默认上限为 128 MiB，成功导入后源包装文件会被删除。
+OpenCode 的内置 Bash 由同名自定义工具覆盖。命令经内部鉴权路由进入独立子进程，切换到专用非 root UID 后应用 Landlock 文件规则和 seccomp 网络规则：当前 conversation workspace 可读写，系统/Python/skills 只读，其他应用文件、其他对话和全局 `tool-output` 不可见；IPv4/IPv6 TCP 与 UDP socket 均被拒绝。启动时会探测 Landlock ABI（要求 ≥4）和 libseccomp，失败时 Agent 直接启动失败，不会回退到未沙箱化 Shell。Shell 环境不包含模型 key、内部 token 或用户 access token，`HOME`、`TMPDIR` 和 Python 缓存均位于当前对话目录。

-CLI 桥接层对 stdout 设置独立的 128 MiB 硬上限（`MAX_CLI_OUTPUT_BYTES`）；stderr 最多保留 256 KiB（`MAX_CLI_STDERR_BYTES`），超出后截断但不会终止 CLI。`MAX_INLINE_RESULT_BYTES`（默认 12000 字节）只控制 OpenCode 的内联阈值，较大结果由 OpenCode 写入标准 `tool-output` 目录。Agent 启动时及后续定期清理其中超过 `RESULT_REF_TTL_HOURS`（默认 7 天）的 `tool_*` 文件。
+`store_render_ref` 只会从当前对话绑定的工作目录导入包装格式 JSON；工作区根目录固定为项目内的 `./data/conversation-workspaces`，以确保 OpenCode 能继续发现项目配置和工具。文件必须包含 `metadata`、`location.file_path` 和 `data`，且真实路径不能越出当前对话目录；单文件默认上限为 128 MiB，成功导入后只删除这一份源包装文件。升级前已经存在的会话没有独立工作目录，需要新建对话后才能使用该导入能力。
+
+CLI 桥接层对 stdout 设置独立的 128 MiB 硬上限（`MAX_CLI_OUTPUT_BYTES`）；stderr 最多保留 256 KiB（`MAX_CLI_STDERR_BYTES`），超出后截断但不会终止 CLI。`MAX_INLINE_RESULT_BYTES`（默认 12000 字节）仅决定内联还是落盘：较大结果写入当前对话的 `tool-data/` 并返回 `data_file.file_path`，不会因为超过 12000 字节杀掉 CLI；分析脚本需要文件输入时可由 `tjwater_cli(store_result=true)` 强制落盘小结果。会话暂存数据不自动删除。OpenCode 自身为其他工具生成的全局 `tool-output` 仍按 `RESULT_REF_TTL_HOURS`（默认 7 天）清理，但沙箱命令不能访问该目录。

 ## 配置与安全

@@ -44,6 +44,12 @@
    "bash": {
      "*": "ask",
      "rm *": "ask",
+      "rm -rf *": "deny",
+      "rm -fr *": "deny",
+      "rm -r -f *": "deny",
+      "rm -f -r *": "deny",
+      "rm --recursive --force *": "deny",
+      "rm --force --recursive *": "deny",
      "rmdir *": "ask",
      "mv *": "ask",
      "chmod *": "ask",
@@ -51,13 +57,7 @@
      "sudo *": "ask",
      "curl *": "ask",
      "wget *": "ask",
-      "*.env*": "deny",
-      "*data/*": "deny",
-      "* data": "deny",
-      "*/data": "deny",
-      "*logs/*": "deny",
-      "* logs": "deny",
-      "*/logs": "deny"
+      "*.env*": "deny"
    },
    "question": "allow",
    "task": "deny",
@@ -0,0 +1,332 @@
+#!/usr/bin/env python3
+"""Apply a Landlock + seccomp policy, drop privileges, then exec one shell command."""
+
+from __future__ import annotations
+
+import argparse
+import ctypes
+import ctypes.util
+import errno
+import json
+import os
+import platform
+import sys
+from pathlib import Path
+
+
+LANDLOCK_CREATE_RULESET_VERSION = 1
+LANDLOCK_RULE_PATH_BENEATH = 1
+
+ACCESS_FS_EXECUTE = 1 << 0
+ACCESS_FS_WRITE_FILE = 1 << 1
+ACCESS_FS_READ_FILE = 1 << 2
+ACCESS_FS_READ_DIR = 1 << 3
+ACCESS_FS_REMOVE_DIR = 1 << 4
+ACCESS_FS_REMOVE_FILE = 1 << 5
+ACCESS_FS_MAKE_CHAR = 1 << 6
+ACCESS_FS_MAKE_DIR = 1 << 7
+ACCESS_FS_MAKE_REG = 1 << 8
+ACCESS_FS_MAKE_SOCK = 1 << 9
+ACCESS_FS_MAKE_FIFO = 1 << 10
+ACCESS_FS_MAKE_BLOCK = 1 << 11
+ACCESS_FS_MAKE_SYM = 1 << 12
+ACCESS_FS_REFER = 1 << 13
+ACCESS_FS_TRUNCATE = 1 << 14
+
+ACCESS_NET_CONNECT_TCP = 1 << 0
+ACCESS_NET_BIND_TCP = 1 << 1
+
+READ_ACCESS = ACCESS_FS_EXECUTE | ACCESS_FS_READ_FILE | ACCESS_FS_READ_DIR
+WRITE_ACCESS = (
+    ACCESS_FS_WRITE_FILE
+    | ACCESS_FS_REMOVE_DIR
+    | ACCESS_FS_REMOVE_FILE
+    | ACCESS_FS_MAKE_CHAR
+    | ACCESS_FS_MAKE_DIR
+    | ACCESS_FS_MAKE_REG
+    | ACCESS_FS_MAKE_SOCK
+    | ACCESS_FS_MAKE_FIFO
+    | ACCESS_FS_MAKE_BLOCK
+    | ACCESS_FS_MAKE_SYM
+    | ACCESS_FS_REFER
+    | ACCESS_FS_TRUNCATE
+)
+HANDLED_FS_ACCESS = READ_ACCESS | WRITE_ACCESS
+
+PR_SET_NO_NEW_PRIVS = 38
+AF_INET = 2
+AF_INET6 = 10
+SCMP_ACT_ALLOW = 0x7FFF0000
+SCMP_ACT_ERRNO = 0x00050000 | errno.EPERM
+SCMP_CMP_EQ = 4
+
+
+class RulesetAttr(ctypes.Structure):
+    _fields_ = [
+        ("handled_access_fs", ctypes.c_uint64),
+        ("handled_access_net", ctypes.c_uint64),
+    ]
+
+
+class PathBeneathAttr(ctypes.Structure):
+    _fields_ = [
+        ("allowed_access", ctypes.c_uint64),
+        ("parent_fd", ctypes.c_int32),
+    ]
+
+
+class ScmpArgCmp(ctypes.Structure):
+    _fields_ = [
+        ("arg", ctypes.c_uint32),
+        ("op", ctypes.c_uint32),
+        ("datum_a", ctypes.c_uint64),
+        ("datum_b", ctypes.c_uint64),
+    ]
+
+
+def syscall_numbers() -> tuple[int, int, int]:
+    machine = platform.machine().lower()
+    if machine not in {"x86_64", "amd64", "aarch64", "arm64"}:
+        raise RuntimeError(f"unsupported architecture for Landlock syscalls: {machine}")
+    return 444, 445, 446
+
+
+def checked_syscall(libc: ctypes.CDLL, number: int, *args: object) -> int:
+    result = int(libc.syscall(number, *args))
+    if result < 0:
+        error_number = ctypes.get_errno()
+        raise OSError(error_number, os.strerror(error_number))
+    return result
+
+
+def get_landlock_abi(libc: ctypes.CDLL) -> int:
+    create_ruleset, _, _ = syscall_numbers()
+    return checked_syscall(
+        libc,
+        create_ruleset,
+        ctypes.c_void_p(),
+        ctypes.c_size_t(0),
+        ctypes.c_uint32(LANDLOCK_CREATE_RULESET_VERSION),
+    )
+
+
+def load_seccomp() -> ctypes.CDLL:
+    library_name = ctypes.util.find_library("seccomp") or "libseccomp.so.2"
+    library = ctypes.CDLL(library_name, use_errno=True)
+    library.seccomp_init.argtypes = [ctypes.c_uint32]
+    library.seccomp_init.restype = ctypes.c_void_p
+    library.seccomp_release.argtypes = [ctypes.c_void_p]
+    library.seccomp_syscall_resolve_name.argtypes = [ctypes.c_char_p]
+    library.seccomp_syscall_resolve_name.restype = ctypes.c_int
+    library.seccomp_rule_add_array.argtypes = [
+        ctypes.c_void_p,
+        ctypes.c_uint32,
+        ctypes.c_int,
+        ctypes.c_uint,
+        ctypes.POINTER(ScmpArgCmp),
+    ]
+    library.seccomp_rule_add_array.restype = ctypes.c_int
+    library.seccomp_load.argtypes = [ctypes.c_void_p]
+    library.seccomp_load.restype = ctypes.c_int
+    return library
+
+
+def add_path_rule(
+    libc: ctypes.CDLL,
+    add_rule_number: int,
+    ruleset_fd: int,
+    path: str,
+    access: int,
+) -> None:
+    resolved_path = os.path.realpath(path)
+    if not os.path.exists(resolved_path):
+        return
+    path_fd = os.open(resolved_path, os.O_PATH | os.O_CLOEXEC)
+    try:
+        allowed_access = access
+        if not Path(resolved_path).is_dir():
+            allowed_access &= ACCESS_FS_EXECUTE | ACCESS_FS_READ_FILE | ACCESS_FS_WRITE_FILE | ACCESS_FS_TRUNCATE
+        attribute = PathBeneathAttr(
+            allowed_access=allowed_access,
+            parent_fd=path_fd,
+        )
+        checked_syscall(
+            libc,
+            add_rule_number,
+            ctypes.c_int(ruleset_fd),
+            ctypes.c_int(LANDLOCK_RULE_PATH_BENEATH),
+            ctypes.byref(attribute),
+            ctypes.c_uint32(0),
+        )
+    finally:
+        os.close(path_fd)
+
+
+def build_landlock_ruleset(
+    libc: ctypes.CDLL,
+    abi: int,
+    workspace: str,
+    read_only_paths: list[str],
+) -> int:
+    create_ruleset, add_rule, _ = syscall_numbers()
+    handled_network = (
+        ACCESS_NET_CONNECT_TCP | ACCESS_NET_BIND_TCP if abi >= 4 else 0
+    )
+    ruleset_attribute = RulesetAttr(
+        handled_access_fs=HANDLED_FS_ACCESS,
+        handled_access_net=handled_network,
+    )
+    ruleset_fd = checked_syscall(
+        libc,
+        create_ruleset,
+        ctypes.byref(ruleset_attribute),
+        ctypes.sizeof(ruleset_attribute),
+        ctypes.c_uint32(0),
+    )
+    try:
+        add_path_rule(
+            libc,
+            add_rule,
+            ruleset_fd,
+            workspace,
+            HANDLED_FS_ACCESS,
+        )
+        for path in read_only_paths:
+            add_path_rule(libc, add_rule, ruleset_fd, path, READ_ACCESS)
+        for device_path, access in (
+            ("/dev/null", ACCESS_FS_READ_FILE | ACCESS_FS_WRITE_FILE),
+            ("/dev/zero", ACCESS_FS_READ_FILE | ACCESS_FS_WRITE_FILE),
+            ("/dev/urandom", ACCESS_FS_READ_FILE),
+            ("/dev/random", ACCESS_FS_READ_FILE),
+        ):
+            add_path_rule(libc, add_rule, ruleset_fd, device_path, access)
+    except Exception:
+        os.close(ruleset_fd)
+        raise
+    return ruleset_fd
+
+
+def install_seccomp_network_filter(library: ctypes.CDLL) -> None:
+    context = library.seccomp_init(SCMP_ACT_ALLOW)
+    if not context:
+        raise RuntimeError("seccomp_init failed")
+    try:
+        socket_syscall = library.seccomp_syscall_resolve_name(b"socket")
+        if socket_syscall < 0:
+            raise RuntimeError("could not resolve socket syscall")
+        for domain in (AF_INET, AF_INET6):
+            comparison = ScmpArgCmp(
+                arg=0,
+                op=SCMP_CMP_EQ,
+                datum_a=domain,
+                datum_b=0,
+            )
+            result = library.seccomp_rule_add_array(
+                context,
+                SCMP_ACT_ERRNO,
+                socket_syscall,
+                1,
+                ctypes.byref(comparison),
+            )
+            if result != 0:
+                raise OSError(-result, os.strerror(-result))
+        result = library.seccomp_load(context)
+        if result != 0:
+            raise OSError(-result, os.strerror(-result))
+    finally:
+        library.seccomp_release(context)
+
+
+def set_no_new_privileges(libc: ctypes.CDLL) -> None:
+    result = libc.prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0)
+    if result != 0:
+        error_number = ctypes.get_errno()
+        raise OSError(error_number, os.strerror(error_number))
+
+
+def restrict_process(
+    workspace: str,
+    read_only_paths: list[str],
+    uid: int,
+    gid: int,
+) -> int:
+    libc = ctypes.CDLL(None, use_errno=True)
+    libc.syscall.restype = ctypes.c_long
+    libc.prctl.restype = ctypes.c_int
+    seccomp = load_seccomp()
+    abi = get_landlock_abi(libc)
+    if abi < 4:
+        raise RuntimeError(f"Landlock ABI 4 or newer is required; detected ABI {abi}")
+
+    ruleset_fd = build_landlock_ruleset(
+        libc,
+        abi,
+        workspace,
+        read_only_paths,
+    )
+    try:
+        if os.geteuid() == 0:
+            os.setgroups([])
+            os.setgid(gid)
+            os.setuid(uid)
+        set_no_new_privileges(libc)
+        _, _, restrict_self = syscall_numbers()
+        checked_syscall(
+            libc,
+            restrict_self,
+            ctypes.c_int(ruleset_fd),
+            ctypes.c_uint32(0),
+        )
+        install_seccomp_network_filter(seccomp)
+    finally:
+        os.close(ruleset_fd)
+    return abi
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--probe", action="store_true")
+    parser.add_argument("--workspace")
+    parser.add_argument("--read-only", action="append", default=[])
+    parser.add_argument("--uid", type=int, default=10_001)
+    parser.add_argument("--gid", type=int, default=10_001)
+    parser.add_argument("--command")
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = parse_args()
+    libc = ctypes.CDLL(None, use_errno=True)
+    libc.syscall.restype = ctypes.c_long
+    if args.probe:
+        abi = get_landlock_abi(libc)
+        load_seccomp()
+        print(json.dumps({"ok": True, "landlock_abi": abi, "seccomp": True}))
+        return 0 if abi >= 4 else 1
+
+    if not args.workspace or args.command is None:
+        raise RuntimeError("workspace and command are required")
+    workspace = os.path.realpath(args.workspace)
+    if not os.path.isdir(workspace):
+        raise RuntimeError("workspace must be an existing directory")
+    os.chdir(workspace)
+    restrict_process(workspace, args.read_only, args.uid, args.gid)
+    os.execve("/bin/bash", ["bash", "-c", args.command], dict(os.environ))
+    return 127
+
+
+if __name__ == "__main__":
+    try:
+        raise SystemExit(main())
+    except Exception as error:
+        print(
+            json.dumps(
+                {
+                    "ok": False,
+                    "error": "SANDBOX_UNAVAILABLE",
+                    "message": str(error),
+                }
+            ),
+            file=sys.stderr,
+        )
+        raise SystemExit(125)
@@ -12,6 +12,7 @@ export type SessionBinding = {
  clientSessionId: string;
  sessionId: string;
  startedAt: number;
+  workspaceDirectory?: string;
 };

 export type SessionContext = {
@@ -52,20 +53,26 @@ export class ChatSessionBridge {
    await this.abortActiveRuntime(requestContext.clientSessionId, existingSessionId);

    let sessionId = existingSessionId;
+    let runtimeSession;
    let created = false;
    if (!sessionId) {
-      const session = await this.runtime.createSession();
-      sessionId = session.id;
+      runtimeSession = await this.runtime.createSession(undefined, {
+        conversationWorkspace: true,
+      });
+      sessionId = runtimeSession.id;
      requestContext = {
        ...requestContext,
        clientSessionId: sessionId,
      };
      created = true;
+    } else {
+      runtimeSession = await this.runtime.getSession(sessionId);
    }
    const binding: SessionBinding = {
      clientSessionId: requestContext.clientSessionId,
      sessionId,
      startedAt: Date.now(),
+      workspaceDirectory: runtimeSession.directory,
    };
    setRuntimeSessionContext({
      accessToken: requestContext.accessToken,
@@ -79,6 +86,7 @@ export class ChatSessionBridge {
      sessionId,
      tokenExpiresAt: requestContext.tokenExpiresAt,
      traceId: requestContext.traceId,
+      workspaceDirectory: runtimeSession.directory,
    });

    return { binding, requestContext, created };
@@ -7,6 +7,8 @@ import {
  parseAgentModelOptions,
 } from "./chat/modelConfig.js";

+export const RESULT_REF_IMPORT_DIRECTORY = "./data/conversation-workspaces";
+
 // 本地开发可在项目根目录放 .local.env；已存在的系统环境变量优先级更高。
 dotenv.config({ path: ".local.env", override: false });

@@ -116,7 +118,9 @@ const envSchema = z
    // result_ref 持久化存储目录。
    RESULT_REF_STORAGE_DIR: z.string().default("./data/result-refs"),
    // 仅允许 store_render_ref 从该目录导入受控 JSON 包装文件。
-    RESULT_REF_IMPORT_DIR: z.string().default("./data/result-imports"),
+    RESULT_REF_IMPORT_DIR: z
+      .literal(RESULT_REF_IMPORT_DIRECTORY)
+      .default(RESULT_REF_IMPORT_DIRECTORY),
    // 单个渲染包装 JSON 的最大导入字节数。
    RESULT_REF_IMPORT_MAX_BYTES: z.coerce
      .number()
@@ -1,6 +1,9 @@
-import { realpath, stat } from "node:fs/promises";
-import { isAbsolute, relative } from "node:path";
+import { stat } from "node:fs/promises";

+import {
+  resolveConversationWorkspace,
+  resolveExistingPathInsideRoot,
+} from "../runtime/conversationWorkspace.js";
 import { readJsonFile, removeFileIfExists } from "../utils/fileStore.js";
 import {
  type ResultReferenceKind,
@@ -68,9 +71,19 @@ export class ResultReferenceResolver {

  async registerRenderPayloadFile(
    filePath: string,
-    input: Omit<RegisterResultReferenceInput, "data" | "kind" | "schemaVersion">,
+    input: Omit<RegisterResultReferenceInput, "data" | "kind" | "schemaVersion"> & {
+      workspaceDirectory: string;
+    },
  ) {
-    const resolvedFilePath = await resolvePathInsideRoot(filePath, this.importRoot);
+    const resolvedWorkspaceDirectory = await resolveConversationWorkspace(
+      input.workspaceDirectory,
+      this.importRoot,
+    );
+    const resolvedFilePath = await resolveExistingPathInsideRoot(
+      filePath,
+      resolvedWorkspaceDirectory,
+      "render payload file must be inside the current conversation workspace",
+    );
    const fileStat = await stat(resolvedFilePath);
    if (!fileStat.isFile()) {
      throw new Error("render payload path must point to a regular file");
@@ -95,8 +108,9 @@ export class ResultReferenceResolver {
      throw new Error("render payload file does not contain a valid junction render payload");
    }

+    const { workspaceDirectory: _workspaceDirectory, ...registrationInput } = input;
    const record = await this.register({
-      ...input,
+      ...registrationInput,
      data: payload,
      kind: RESULT_REFERENCE_KIND.renderJunctionsPayload,
      schemaVersion: 1,
@@ -186,25 +200,6 @@ export const extractRenderJunctionPayload = (
  };
 };

-const resolvePathInsideRoot = async (filePath: string, rootPath: string) => {
-  if (!isAbsolute(filePath)) {
-    throw new Error("render payload file_path must be absolute");
-  }
-  const [resolvedFilePath, resolvedRootPath] = await Promise.all([
-    realpath(filePath),
-    realpath(rootPath),
-  ]);
-  const relativePath = relative(resolvedRootPath, resolvedFilePath);
-  if (
-    relativePath === ".." ||
-    relativePath.startsWith(`..${process.platform === "win32" ? "\\" : "/"}`) ||
-    isAbsolute(relativePath)
-  ) {
-    throw new Error("render payload file must be inside RESULT_REF_IMPORT_DIR");
-  }
-  return resolvedFilePath;
-};
-
 const normalizeDataForKind = (
  kind: ResultReferenceKind,
  data: unknown,
@@ -155,7 +155,9 @@ export const buildChatRouter = (
    const actorKey = toActorKey(userId);
    const projectKey = toProjectKey(projectId);
    const requestedSessionId = parsed.data.session_id?.trim();
-    const sessionId = requestedSessionId || (await runtime.createSession()).id;
+    const sessionId =
+      requestedSessionId ||
+      (await runtime.createSession(undefined, { conversationWorkspace: true })).id;

    const { record, created } = await sessionMetadataStore.ensure({
      actorKey,
@@ -451,7 +453,9 @@ export const buildChatRouter = (
        res.status(404).json({ message: "source session not found" });
        return;
      }
-      const forkSession = await runtime.createSession();
+      const forkSession = await runtime.createSession(undefined, {
+        conversationWorkspace: true,
+      });
      const { record: targetSessionRecord } = await sessionMetadataStore.ensure({
        actorKey,
        parentSessionId: sourceSessionId,
@@ -874,6 +878,7 @@ export const buildChatRouter = (
          traceId: requestContext.traceId,
          projectId: requestContext.projectId,
          signal: abortController.signal,
+          workspaceRoot: binding.workspaceDirectory,
          write: (event, data) => {
            publish(event, data);
          },
@@ -1,5 +1,5 @@
-import { lstatSync, readdirSync, realpathSync } from "node:fs";
-import { isAbsolute, relative, resolve, sep } from "node:path";
+import { existsSync, lstatSync, readdirSync, realpathSync } from "node:fs";
+import { basename, dirname, isAbsolute, relative, resolve, sep } from "node:path";

 export type ApprovalMode = "request" | "auto" | "always";

@@ -46,6 +46,14 @@ export const canAutoApprovePermission = (
    return isSafeWorkspaceSearch(normalized, context);
  }

+  if (normalized === "bash") {
+    return isConversationWorkspace(context.workspaceRoot);
+  }
+
+  if (normalized === "read" || normalized === "edit" || normalized === "write") {
+    return isSafeConversationFileAccess(context);
+  }
+
  if (lowRiskToolPermissions.has(normalized)) {
    return true;
  }
@@ -62,9 +70,19 @@ export const resolvePermissionApproval = (
  permission: string,
  context: PermissionApprovalContext = {},
 ) => {
+  if (isDirectRecursiveForceRemove(permission, context)) {
+    return {
+      autoApprove: false,
+      autoReject: true,
+      title: "已拒绝递归强制删除",
+      detail: "当前安全策略禁止直接执行带 recursive 和 force 参数的 rm 命令。",
+    } as const;
+  }
+
  if (approvalMode === "always") {
    return {
      autoApprove: true,
+      autoReject: false,
      title: "已按始终允许模式放行",
      detail:
        "当前会话处于始终允许模式，已放行本次权限请求；明确禁止的权限仍由 OpenCode 拒绝。",
@@ -74,6 +92,7 @@ export const resolvePermissionApproval = (
  if (approvalMode === "auto" && canAutoApprovePermission(permission, context)) {
    return {
      autoApprove: true,
+      autoReject: false,
      title: "已自动批准低风险权限",
      detail: "当前批准模式允许自动执行低风险工具，已放行本次请求。",
    } as const;
@@ -81,11 +100,119 @@ export const resolvePermissionApproval = (

  return {
    autoApprove: false,
+    autoReject: false,
    title: "等待权限确认",
    detail: undefined,
  } as const;
 };

+const isDirectRecursiveForceRemove = (
+  permission: string,
+  context: PermissionApprovalContext,
+): boolean => {
+  if (normalizePermission(permission) !== "bash") {
+    return false;
+  }
+  const command =
+    typeof context.metadata?.command === "string"
+      ? context.metadata.command
+      : context.patterns?.join("\n");
+  if (!command) {
+    return false;
+  }
+
+  return splitShellCommandSegments(command).some((segment) => {
+    const words = tokenizeShellSegment(segment);
+    let commandIndex = 0;
+    while (commandIndex < words.length) {
+      const word = words[commandIndex]!;
+      const executable = word.split("/").at(-1)?.toLowerCase();
+      if (word === "!" || /^[A-Za-z_][A-Za-z0-9_]*=/.test(word)) {
+        commandIndex += 1;
+        continue;
+      }
+      if (executable === "command") {
+        commandIndex += 1;
+        while (words[commandIndex]?.startsWith("-") && words[commandIndex] !== "--") {
+          commandIndex += 1;
+        }
+        if (words[commandIndex] === "--") commandIndex += 1;
+        continue;
+      }
+      if (executable === "env") {
+        commandIndex += 1;
+        while (commandIndex < words.length) {
+          const envWord = words[commandIndex]!;
+          if (envWord === "--") {
+            commandIndex += 1;
+            break;
+          }
+          if (envWord === "-u" || envWord === "--unset") {
+            commandIndex += 2;
+            continue;
+          }
+          if (
+            envWord.startsWith("-") ||
+            /^[A-Za-z_][A-Za-z0-9_]*=/.test(envWord)
+          ) {
+            commandIndex += 1;
+            continue;
+          }
+          break;
+        }
+        continue;
+      }
+      if (executable === "sudo" || executable === "doas") {
+        commandIndex += 1;
+        while (words[commandIndex]?.startsWith("-")) {
+          commandIndex += 1;
+        }
+        continue;
+      }
+      if (executable === "busybox" && words[commandIndex + 1] === "rm") {
+        commandIndex += 1;
+      }
+      break;
+    }
+
+    const executable = words[commandIndex]?.split("/").at(-1)?.toLowerCase();
+    if (executable !== "rm") {
+      return false;
+    }
+
+    let recursive = false;
+    let force = false;
+    for (const word of words.slice(commandIndex + 1)) {
+      if (word === "--") {
+        break;
+      }
+      if (word === "--recursive") {
+        recursive = true;
+      } else if (word === "--force") {
+        force = true;
+      } else if (/^-[^-]/.test(word)) {
+        recursive ||= /[rR]/.test(word.slice(1));
+        force ||= word.slice(1).includes("f");
+      }
+    }
+    return recursive && force;
+  });
+};
+
+const splitShellCommandSegments = (command: string): string[] =>
+  command.split(/&&|\|\||[;|()\n]/u);
+
+const tokenizeShellSegment = (segment: string): string[] =>
+  (segment.match(/(?:[^\s"'\\]+|"(?:\\.|[^"])*"|'[^']*')+/gu) ?? []).map(
+    (word) => {
+      const quoted = word.match(/^(?:"([\s\S]*)"|'([\s\S]*)')$/u);
+      return (quoted ? (quoted[1] ?? quoted[2] ?? "") : word).replace(
+        /(["'])|\\(.)/gu,
+        "$2",
+      );
+    },
+  );
+
 const isSafeWorkspaceSearch = (
  permission: "glob" | "grep",
  context: PermissionApprovalContext,
@@ -133,13 +260,14 @@ const isSafeWorkspaceSearch = (
        isAbsolute(expression) ||
        containsParentTraversal(expression) ||
        containsAmbiguousGlobSyntax(expression) ||
-        containsProtectedPath(expression),
+        containsProtectedPath(expression, isConversationWorkspace(root)),
    )
  ) {
    return false;
  }

-  if (!relativePath) {
+  const conversationWorkspace = isConversationWorkspace(root);
+  if (!relativePath && !conversationWorkspace) {
    if (permission !== "glob") {
      return false;
    }
@@ -156,16 +284,19 @@ const isSafeWorkspaceSearch = (
  }

  return (
-    relativePath !== "" &&
    relativePath !== ".." &&
    !relativePath.startsWith(`..${sep}`) &&
    !isAbsolute(relativePath) &&
-    !containsProtectedPath(relativePath) &&
-    isSafeSearchTarget(searchRoot, relativePath)
+    !containsProtectedPath(relativePath, conversationWorkspace) &&
+    isSafeSearchTarget(searchRoot, relativePath, conversationWorkspace)
  );
 };

-const isSafeSearchTarget = (searchRoot: string, relativePath: string): boolean => {
+const isSafeSearchTarget = (
+  searchRoot: string,
+  relativePath: string,
+  conversationWorkspace: boolean,
+): boolean => {
  try {
    const target = lstatSync(searchRoot);
    if (target.isFile()) {
@@ -175,7 +306,10 @@ const isSafeSearchTarget = (searchRoot: string, relativePath: string): boolean =
      return false;
    }
    const topLevelName = relativePath.split(sep)[0];
-    if (!topLevelName || !lowRiskSearchRootNames.has(topLevelName)) {
+    if (
+      !conversationWorkspace &&
+      (!topLevelName || !lowRiskSearchRootNames.has(topLevelName))
+    ) {
      return false;
    }

@@ -183,7 +317,10 @@ const isSafeSearchTarget = (searchRoot: string, relativePath: string): boolean =
    while (pending.length > 0) {
      const directory = pending.pop()!;
      for (const entry of readdirSync(directory, { withFileTypes: true })) {
-        if (entry.isSymbolicLink() || containsProtectedPath(entry.name)) {
+        if (
+          entry.isSymbolicLink() ||
+          containsProtectedPath(entry.name, conversationWorkspace)
+        ) {
          return false;
        }
        if (entry.isDirectory()) {
@@ -213,10 +350,71 @@ const containsParentTraversal = (value: string): boolean =>
 const containsAmbiguousGlobSyntax = (value: string): boolean =>
  /[?[\]{}()!+@\\]/.test(value);

-const containsProtectedPath = (value: string): boolean => {
+const containsProtectedPath = (
+  value: string,
+  conversationWorkspace = false,
+): boolean => {
  const normalized = value.replaceAll("\\", "/").toLowerCase();
  return (
    normalized.includes(".env") ||
-    /(?:^|[^a-z0-9_-])(?:data|logs)(?:$|[^a-z0-9_-])/.test(normalized)
+    (!conversationWorkspace &&
+      /(?:^|[^a-z0-9_-])(?:data|logs)(?:$|[^a-z0-9_-])/.test(normalized))
  );
 };
+
+const isConversationWorkspace = (workspaceRoot: string | undefined): boolean => {
+  if (!workspaceRoot?.trim()) return false;
+  try {
+    const root = realpathSync.native(resolve(workspaceRoot));
+    return (
+      basename(dirname(root)) === "conversation-workspaces" &&
+      basename(root).startsWith("conversation-") &&
+      lstatSync(root).isDirectory()
+    );
+  } catch {
+    return false;
+  }
+};
+
+const isSafeConversationFileAccess = (
+  context: PermissionApprovalContext,
+): boolean => {
+  if (!isConversationWorkspace(context.workspaceRoot)) return false;
+  const requestedPath = [
+    context.metadata?.path,
+    context.metadata?.file_path,
+    context.metadata?.filePath,
+    context.metadata?.filepath,
+    context.metadata?.file,
+    context.patterns?.[0],
+  ].find((value): value is string => typeof value === "string" && value.trim().length > 0);
+  if (!requestedPath || /[*?[\]{}]/.test(requestedPath)) return false;
+  try {
+    const root = realpathSync.native(resolve(context.workspaceRoot!));
+    const target = resolve(root, requestedPath);
+    let existingPath = target;
+    while (!existsSync(existingPath)) {
+      const parent = dirname(existingPath);
+      if (parent === existingPath) return false;
+      existingPath = parent;
+    }
+    const resolvedExistingPath = realpathSync.native(existingPath);
+    const relativeExisting = relative(root, resolvedExistingPath);
+    if (
+      relativeExisting === ".." ||
+      relativeExisting.startsWith(`..${sep}`) ||
+      isAbsolute(relativeExisting)
+    ) {
+      return false;
+    }
+    const relativeTarget = relative(root, target);
+    return (
+      relativeTarget !== ".." &&
+      !relativeTarget.startsWith(`..${sep}`) &&
+      !isAbsolute(relativeTarget) &&
+      !containsProtectedPath(relativeTarget, true)
+    );
+  } catch {
+    return false;
+  }
+};
@@ -71,6 +71,7 @@ type StreamPromptOptions = {
  traceId?: string;
  projectId?: string;
  signal?: AbortSignal;
+  workspaceRoot?: string;
  write: (event: string, data: Record<string, unknown>) => void;
 };

@@ -157,6 +158,7 @@ export const streamPromptResponse = async ({
  traceId,
  projectId,
  signal,
+  workspaceRoot,
  write,
 }: StreamPromptOptions): Promise<{
  aborted: boolean;
@@ -397,7 +399,7 @@ export const streamPromptResponse = async ({
          {
            metadata: event.properties.metadata,
            patterns: event.properties.patterns,
-            workspaceRoot: process.cwd(),
+            workspaceRoot,
          },
        );
        logDevelopmentDebug("permission request received", {
@@ -410,20 +412,25 @@ export const streamPromptResponse = async ({
        emitProgress({
          id: `permission-${event.properties.id}`,
          phase: "permission",
-          status: permissionApproval.autoApprove ? "completed" : "running",
+          status: permissionApproval.autoReject
+            ? "error"
+            : permissionApproval.autoApprove
+              ? "completed"
+              : "running",
          title: permissionApproval.title,
          detail: permissionApproval.detail ?? buildPermissionDetail(event),
        });
-        if (permissionApproval.autoApprove) {
+        if (permissionApproval.autoApprove || permissionApproval.autoReject) {
+          const reply = permissionApproval.autoReject ? "reject" : "once";
          await runtime.replyPermission({
            requestId: event.properties.id,
            sessionId,
-            reply: "once",
+            reply,
          });
          write("permission_response", {
            session_id: clientSessionId,
            request_id: event.properties.id,
-            reply: "once" satisfies PermissionReply,
+            reply: reply satisfies PermissionReply,
          });
          continue;
        }
@@ -448,7 +455,7 @@ export const streamPromptResponse = async ({
          {
            metadata: event.properties.metadata,
            patterns: event.properties.resources,
-            workspaceRoot: process.cwd(),
+            workspaceRoot,
          },
        );
        logDevelopmentDebug("permission v2 request received", {
@@ -461,20 +468,25 @@ export const streamPromptResponse = async ({
        emitProgress({
          id: `permission-${event.properties.id}`,
          phase: "permission",
-          status: permissionApproval.autoApprove ? "completed" : "running",
+          status: permissionApproval.autoReject
+            ? "error"
+            : permissionApproval.autoApprove
+              ? "completed"
+              : "running",
          title: permissionApproval.title,
          detail: permissionApproval.detail ?? buildPermissionV2Detail(event),
        });
-        if (permissionApproval.autoApprove) {
+        if (permissionApproval.autoApprove || permissionApproval.autoReject) {
+          const reply = permissionApproval.autoReject ? "reject" : "once";
          await runtime.replyPermission({
            requestId: event.properties.id,
            sessionId,
-            reply: "once",
+            reply,
          });
          write("permission_response", {
            session_id: clientSessionId,
            request_id: event.properties.id,
-            reply: "once" satisfies PermissionReply,
+            reply: reply satisfies PermissionReply,
          });
          continue;
        }
@@ -0,0 +1,137 @@
+import { randomUUID } from "node:crypto";
+import {
+  chown,
+  chmod,
+  lstat,
+  mkdir,
+  realpath,
+  rename,
+  rm,
+  stat,
+  writeFile,
+} from "node:fs/promises";
+import { isAbsolute, join, relative, resolve, sep } from "node:path";
+
+export const SANDBOX_UID = 10_001;
+export const SANDBOX_GID = 10_001;
+
+const isOutsideRoot = (root: string, candidate: string) => {
+  const relativePath = relative(root, candidate);
+  return (
+    relativePath === ".." ||
+    relativePath.startsWith(`..${sep}`) ||
+    isAbsolute(relativePath)
+  );
+};
+
+export const setSandboxOwnership = async (path: string) => {
+  if (typeof process.getuid !== "function" || process.getuid() !== 0) {
+    return;
+  }
+  await chown(path, SANDBOX_UID, SANDBOX_GID);
+};
+
+export const resolveConversationWorkspace = async (
+  workspaceDirectory: string,
+  importRoot: string,
+) => {
+  const workspaceLinkStat = await lstat(workspaceDirectory);
+  if (workspaceLinkStat.isSymbolicLink()) {
+    throw new Error("conversation workspace must not be a symbolic link");
+  }
+  const [resolvedImportRoot, resolvedWorkspaceDirectory] = await Promise.all([
+    realpath(importRoot),
+    realpath(workspaceDirectory),
+  ]);
+  if (isOutsideRoot(resolvedImportRoot, resolvedWorkspaceDirectory)) {
+    throw new Error("conversation workspace must be inside RESULT_REF_IMPORT_DIR");
+  }
+  const relativeWorkspace = relative(
+    resolvedImportRoot,
+    resolvedWorkspaceDirectory,
+  );
+  if (
+    !relativeWorkspace ||
+    relativeWorkspace.includes("/") ||
+    relativeWorkspace.includes("\\")
+  ) {
+    throw new Error(
+      "conversation workspace must be a direct child of RESULT_REF_IMPORT_DIR",
+    );
+  }
+  const workspaceStat = await stat(resolvedWorkspaceDirectory);
+  if (!workspaceStat.isDirectory()) {
+    throw new Error("conversation workspace must point to a directory");
+  }
+  return resolvedWorkspaceDirectory;
+};
+
+export const resolveExistingPathInsideRoot = async (
+  filePath: string,
+  rootPath: string,
+  outsideMessage = "path must be inside the current conversation workspace",
+) => {
+  if (!isAbsolute(filePath)) {
+    throw new Error("file_path must be absolute");
+  }
+  const [resolvedFilePath, resolvedRootPath] = await Promise.all([
+    realpath(filePath),
+    realpath(rootPath),
+  ]);
+  if (isOutsideRoot(resolvedRootPath, resolvedFilePath)) {
+    throw new Error(outsideMessage);
+  }
+  return resolvedFilePath;
+};
+
+export type StagedConversationFile = {
+  bytes: number;
+  contentType: string;
+  filePath: string;
+};
+
+export const stageConversationText = async (
+  workspaceDirectory: string,
+  importRoot: string,
+  content: string,
+  options: {
+    contentType: string;
+    extension: string;
+    prefix: string;
+  },
+): Promise<StagedConversationFile> => {
+  const workspace = await resolveConversationWorkspace(
+    workspaceDirectory,
+    importRoot,
+  );
+  const outputDirectory = resolve(workspace, "tool-data");
+  if (isOutsideRoot(workspace, outputDirectory)) {
+    throw new Error("tool data directory escaped the conversation workspace");
+  }
+  await mkdir(outputDirectory, { mode: 0o700, recursive: true });
+  const resolvedOutputDirectory = await realpath(outputDirectory);
+  if (isOutsideRoot(workspace, resolvedOutputDirectory)) {
+    throw new Error("tool data directory escaped the conversation workspace");
+  }
+  await chmod(resolvedOutputDirectory, 0o700);
+  await setSandboxOwnership(resolvedOutputDirectory);
+
+  const fileName = `${options.prefix}-${randomUUID()}.${options.extension}`;
+  const filePath = join(resolvedOutputDirectory, fileName);
+  const temporaryPath = `${filePath}.${randomUUID()}.tmp`;
+  try {
+    await writeFile(temporaryPath, content, { encoding: "utf8", mode: 0o600 });
+    await setSandboxOwnership(temporaryPath);
+    await rename(temporaryPath, filePath);
+  } catch (error) {
+    await rm(temporaryPath, { force: true }).catch(() => undefined);
+    throw error;
+  }
+  await chmod(filePath, 0o600);
+  await setSandboxOwnership(filePath);
+  return {
+    bytes: Buffer.byteLength(content),
+    contentType: options.contentType,
+    filePath,
+  };
+};
@@ -2,11 +2,15 @@ import {
  createOpencode,
  type OpencodeClient,
 } from "@opencode-ai/sdk/v2";
+import { randomUUID } from "node:crypto";
 import { existsSync, readFileSync } from "node:fs";
+import { chmod, mkdir, rmdir } from "node:fs/promises";
 import { resolve } from "node:path";

 import { config } from "../config.js";
 import { logger } from "../logger.js";
+import { ensureDirectory } from "../utils/fileStore.js";
+import { setSandboxOwnership } from "./conversationWorkspace.js";
 import {
  cleanupExpiredToolOutputs,
  resolveOpencodeToolOutputDirectory,
@@ -126,12 +130,47 @@ export class OpencodeRuntimeAdapter {
    }
  }

-  async createSession(title?: string) {
+  async createSession(
+    title?: string,
+    options: {
+      conversationWorkspace?: boolean;
+      workspaceRoot?: string;
+    } = {},
+  ) {
    const client = await this.ensureClient();
+    if (!options.conversationWorkspace) {
+      const response = await client.session.create({ title });
+      return requireData(response.data, "session.create");
+    }
+
+    const workspaceRoot = resolve(
+      options.workspaceRoot ?? config.RESULT_REF_IMPORT_DIR,
+    );
+    const directory = resolve(workspaceRoot, `conversation-${randomUUID()}`);
+    await ensureDirectory(workspaceRoot);
+    await chmod(workspaceRoot, 0o700);
+    await mkdir(directory, { mode: 0o700 });
+    await setSandboxOwnership(directory);
+    try {
      const response = await client.session.create({
+        directory,
        title,
+        permission: [
+          { permission: "read", pattern: `${directory}/**`, action: "allow" },
+          { permission: "edit", pattern: `${directory}/**`, action: "ask" },
+        ],
      });
      return requireData(response.data, "session.create");
+    } catch (error) {
+      await rmdir(directory).catch(() => undefined);
+      throw error;
+    }
+  }
+
+  async getSession(sessionId: string) {
+    const client = await this.ensureClient();
+    const response = await client.session.get({ sessionID: sessionId });
+    return requireData(response.data, "session.get");
  }

  async sendPrompt(sessionId: string, text: string) {
@@ -15,6 +15,7 @@ export type RuntimeSessionContext = {
  sessionId: string;
  tokenExpiresAt?: string;
  traceId: string;
+  workspaceDirectory?: string;
 };

 const contexts = new Map<string, RuntimeSessionContext>();
@@ -0,0 +1,47 @@
+import { config } from "../config.js";
+import { type RuntimeSessionContext } from "./sessionContext.js";
+import { stageConversationText } from "./conversationWorkspace.js";
+
+export type ToolDataFilePayload = {
+  bytes: number;
+  content_type: string;
+  file_path: string;
+};
+
+export const stageLargeToolOutput = async (
+  context: RuntimeSessionContext,
+  content: string,
+  options: {
+    contentType: string;
+    extension: string;
+    force?: boolean;
+    prefix: string;
+  },
+): Promise<ToolDataFilePayload | null> => {
+  if (!options.force && Buffer.byteLength(content) <= config.MAX_INLINE_RESULT_BYTES) {
+    return null;
+  }
+  if (!context.workspaceDirectory) {
+    throw new Error(
+      "large tool output requires a conversation workspace; create a new conversation",
+    );
+  }
+  const staged = await stageConversationText(
+    context.workspaceDirectory,
+    config.RESULT_REF_IMPORT_DIR,
+    content,
+    options,
+  );
+  return {
+    bytes: staged.bytes,
+    content_type: staged.contentType,
+    file_path: staged.filePath,
+  };
+};
+
+export const buildLargeCliResult = (dataFile: ToolDataFilePayload) => ({
+  ok: true,
+  schema_version: "tjwater-cli/v1",
+  summary: "CLI 结果已保存到当前对话工作区",
+  data_file: dataFile,
+});
@@ -0,0 +1,219 @@
+import { spawn } from "node:child_process";
+import { existsSync } from "node:fs";
+import { mkdir } from "node:fs/promises";
+import { resolve } from "node:path";
+
+import { config } from "../config.js";
+import {
+  resolveConversationWorkspace,
+  SANDBOX_GID,
+  SANDBOX_UID,
+  setSandboxOwnership,
+} from "../runtime/conversationWorkspace.js";
+
+export type SandboxProbe = {
+  landlockAbi: number;
+  seccomp: boolean;
+};
+
+export type SandboxExecutionResult = {
+  exitCode: number | null;
+  outcome: "completed" | "output_limit" | "timeout";
+  signal: NodeJS.Signals | null;
+  stderr: string;
+  stderrTruncated: boolean;
+  stdout: string;
+};
+
+const sandboxRunnerPath = resolve("scripts/landlock_sandbox.py");
+const maxSandboxTimeoutSeconds = 30 * 60;
+
+const resolvePythonPath = () => {
+  const virtualEnvironment = process.env.VIRTUAL_ENV?.trim();
+  const candidates = [
+    virtualEnvironment ? resolve(virtualEnvironment, "bin/python") : "",
+    "/opt/venv/bin/python",
+    "/usr/bin/python3",
+  ];
+  return candidates.find((candidate) => candidate && existsSync(candidate)) ?? "python3";
+};
+
+const collectProcess = async (
+  args: string[],
+  options: {
+    cwd?: string;
+    env?: NodeJS.ProcessEnv;
+    maxStderrBytes: number;
+    maxStdoutBytes: number;
+    timeoutMs: number;
+  },
+): Promise<SandboxExecutionResult> => {
+  const child = spawn(resolvePythonPath(), [sandboxRunnerPath, ...args], {
+    cwd: options.cwd,
+    detached: true,
+    env: options.env,
+    stdio: ["ignore", "pipe", "pipe"],
+  });
+  const stdoutChunks: Buffer[] = [];
+  const stderrChunks: Buffer[] = [];
+  let stdoutBytes = 0;
+  let stderrBytes = 0;
+  let stderrTruncated = false;
+  let outcome: SandboxExecutionResult["outcome"] = "completed";
+  let terminationStarted = false;
+  let forceKillTimer: ReturnType<typeof setTimeout> | undefined;
+
+  const killGroup = (signal: NodeJS.Signals) => {
+    if (child.pid) {
+      try {
+        process.kill(-child.pid, signal);
+        return;
+      } catch {
+        // Fall back to the direct child when process groups are unavailable.
+      }
+    }
+    child.kill(signal);
+  };
+  const terminate = (nextOutcome: "output_limit" | "timeout") => {
+    if (terminationStarted) return;
+    terminationStarted = true;
+    outcome = nextOutcome;
+    killGroup("SIGTERM");
+    forceKillTimer = setTimeout(() => killGroup("SIGKILL"), 1500);
+  };
+
+  const timeoutTimer = setTimeout(() => terminate("timeout"), options.timeoutMs);
+  child.stdout.on("data", (chunk: Buffer) => {
+    if (terminationStarted) return;
+    if (stdoutBytes + chunk.length > options.maxStdoutBytes) {
+      terminate("output_limit");
+      return;
+    }
+    stdoutChunks.push(chunk);
+    stdoutBytes += chunk.length;
+  });
+  child.stderr.on("data", (chunk: Buffer) => {
+    if (stderrTruncated) return;
+    const remaining = options.maxStderrBytes - stderrBytes;
+    if (chunk.length > remaining) {
+      if (remaining > 0) {
+        stderrChunks.push(chunk.subarray(0, remaining));
+        stderrBytes += remaining;
+      }
+      stderrTruncated = true;
+      return;
+    }
+    stderrChunks.push(chunk);
+    stderrBytes += chunk.length;
+  });
+
+  return await new Promise((resolveResult, reject) => {
+    child.once("error", (error) => {
+      clearTimeout(timeoutTimer);
+      if (forceKillTimer) clearTimeout(forceKillTimer);
+      reject(error);
+    });
+    child.once("close", (exitCode, signal) => {
+      clearTimeout(timeoutTimer);
+      if (forceKillTimer) clearTimeout(forceKillTimer);
+      resolveResult({
+        exitCode,
+        outcome,
+        signal,
+        stderr: Buffer.concat(stderrChunks, stderrBytes).toString("utf8"),
+        stderrTruncated,
+        stdout:
+          outcome === "output_limit"
+            ? ""
+            : Buffer.concat(stdoutChunks, stdoutBytes).toString("utf8"),
+      });
+    });
+  });
+};
+
+export const probeLandlockSandbox = async (): Promise<SandboxProbe> => {
+  const result = await collectProcess(["--probe"], {
+    maxStderrBytes: 16 * 1024,
+    maxStdoutBytes: 16 * 1024,
+    timeoutMs: 5000,
+  });
+  if (result.exitCode !== 0) {
+    throw new Error(result.stderr || "Landlock sandbox probe failed");
+  }
+  const payload = JSON.parse(result.stdout) as {
+    landlock_abi?: unknown;
+    seccomp?: unknown;
+  };
+  if (
+    typeof payload.landlock_abi !== "number" ||
+    payload.landlock_abi < 4 ||
+    payload.seccomp !== true
+  ) {
+    throw new Error("Landlock ABI 4 and seccomp are required");
+  }
+  return { landlockAbi: payload.landlock_abi, seccomp: true };
+};
+
+const buildSandboxEnvironment = (workspace: string): NodeJS.ProcessEnv => ({
+  HOME: workspace,
+  LANG: process.env.LANG ?? "C.UTF-8",
+  LC_ALL: process.env.LC_ALL ?? "C.UTF-8",
+  MPLCONFIGDIR: resolve(workspace, ".cache/matplotlib"),
+  PATH: "/opt/venv/bin:/usr/local/bin:/usr/bin:/bin",
+  PYTHONDONTWRITEBYTECODE: "1",
+  TMPDIR: resolve(workspace, "tmp"),
+  TZ: process.env.TZ ?? "Asia/Shanghai",
+  VIRTUAL_ENV: "/opt/venv",
+  XDG_CACHE_HOME: resolve(workspace, ".cache"),
+});
+
+export const executeSandboxCommand = async (
+  workspaceDirectory: string,
+  command: string,
+  timeoutSeconds: number,
+): Promise<SandboxExecutionResult> => {
+  const workspace = await resolveConversationWorkspace(
+    workspaceDirectory,
+    config.RESULT_REF_IMPORT_DIR,
+  );
+  const normalizedTimeout = Math.min(
+    maxSandboxTimeoutSeconds,
+    Math.max(1, Math.trunc(timeoutSeconds)),
+  );
+  for (const path of [resolve(workspace, "tmp"), resolve(workspace, ".cache")]) {
+    await mkdir(path, { mode: 0o700, recursive: true });
+    await setSandboxOwnership(path);
+  }
+
+  const readOnlyPaths = [
+    "/usr",
+    "/bin",
+    "/lib",
+    "/lib64",
+    "/opt/venv",
+    "/etc/ld.so.cache",
+    "/etc/localtime",
+    "/etc/group",
+    "/etc/nsswitch.conf",
+    "/etc/passwd",
+    resolve(config.OPENCODE_SKILLS_ROOT_DIR),
+  ].filter(existsSync);
+  const args = [
+    "--workspace",
+    workspace,
+    "--uid",
+    String(SANDBOX_UID),
+    "--gid",
+    String(SANDBOX_GID),
+    ...readOnlyPaths.flatMap((path) => ["--read-only", path]),
+    "--command",
+    command,
+  ];
+  return collectProcess(args, {
+    cwd: workspace,
+    env: buildSandboxEnvironment(workspace),
+    maxStderrBytes: config.MAX_CLI_STDERR_BYTES,
+    maxStdoutBytes: config.MAX_CLI_OUTPUT_BYTES,
+    timeoutMs: normalizedTimeout * 1000,
+  });
+};
@@ -32,11 +32,20 @@ import {
 import { buildChatRouter } from "./routes/chat.js";
 import { buildAgentPublicRouter } from "./routes/publicApi.js";
 import { opencodeRuntime } from "./runtime/opencode.js";
+import {
+  executeSandboxCommand,
+  probeLandlockSandbox,
+  type SandboxProbe,
+} from "./sandbox/landlockSandbox.js";
 import {
  getRuntimeSessionContext,
  markRuntimeSessionAuthExpired,
  type RuntimeSessionContext,
 } from "./runtime/sessionContext.js";
+import {
+  buildLargeCliResult,
+  stageLargeToolOutput,
+} from "./runtime/toolOutputStaging.js";
 import { ensureDirectory } from "./utils/fileStore.js";
 import { SkillStore } from "./skills/store.js";

@@ -63,6 +72,7 @@ const resultReferenceResolver = new ResultReferenceResolver(
 );
 const internalToken = config.AGENT_INTERNAL_TOKEN ?? randomUUID();
 const credentialRefreshCoordinator = new CredentialRefreshCoordinator();
+let sandboxProbe: SandboxProbe | null = null;

 // 这个 token 只用于 OpenCode 子进程回调本服务的内部工具桥。
 process.env.TJWATER_AGENT_INTERNAL_TOKEN = internalToken;
@@ -78,6 +88,7 @@ app.get("/health", async (_req, res) => {
      ready: true,
      warmed_up: true,
      runtime,
+      sandbox: sandboxProbe,
      sessions: sessionBridge.count(),
    });
  } catch (error) {
@@ -215,6 +226,7 @@ app.post("/internal/tools/tjwater-cli-call", async (req, res) => {

  const timeoutSec =
    typeof req.body?.timeout === "number" && req.body.timeout > 0 ? req.body.timeout : 120;
+  const storeResult = req.body?.store_result === true;

  if (!context.network) {
    res.status(400).json({
@@ -309,6 +321,24 @@ app.post("/internal/tools/tjwater-cli-call", async (req, res) => {
  }

  if (result.stdout.trim()) {
+    try {
+      const dataFile = await stageLargeToolOutput(context, result.stdout, {
+        contentType: "application/json",
+        extension: "json",
+        force: storeResult,
+        prefix: "cli",
+      });
+      if (dataFile) {
+        res.status(200).json(buildLargeCliResult(dataFile));
+        return;
+      }
+    } catch (error) {
+      res.status(409).json({
+        message: "large CLI result could not be staged",
+        detail: error instanceof Error ? error.message : String(error),
+      });
+      return;
+    }
    res.status(200).type("application/json").send(result.stdout);
    return;
  }
@@ -321,6 +351,91 @@ app.post("/internal/tools/tjwater-cli-call", async (req, res) => {
  });
 });

+app.post("/internal/tools/sandbox-shell", async (req, res) => {
+  if (req.header("x-agent-internal-token") !== internalToken) {
+    res.status(403).json({ message: "forbidden" });
+    return;
+  }
+  const sessionId =
+    typeof req.body?.session_id === "string" ? req.body.session_id.trim() : "";
+  const command =
+    typeof req.body?.command === "string" ? req.body.command.trim() : "";
+  const context = sessionId ? getRuntimeSessionContext(sessionId) : null;
+  if (!context) {
+    res.status(404).json({ message: "session context not found", detail: sessionId });
+    return;
+  }
+  if (!context.workspaceDirectory) {
+    res.status(400).json({
+      message: "conversation workspace is required",
+      detail: "create a new conversation before running shell commands",
+    });
+    return;
+  }
+  if (!command) {
+    res.status(400).json({ message: "command is required" });
+    return;
+  }
+  const timeoutSeconds =
+    typeof req.body?.timeout === "number" && Number.isFinite(req.body.timeout)
+      ? req.body.timeout
+      : 120;
+  try {
+    const result = await executeSandboxCommand(
+      context.workspaceDirectory,
+      command,
+      timeoutSeconds,
+    );
+    if (result.outcome === "timeout") {
+      res.status(504).json({
+        ok: false,
+        error: { code: "TIMEOUT", message: "sandbox command timed out" },
+      });
+      return;
+    }
+    if (result.outcome === "output_limit") {
+      res.status(502).json({
+        ok: false,
+        error: {
+          code: "OUTPUT_LIMIT_EXCEEDED",
+          message: `stdout exceeded ${config.MAX_CLI_OUTPUT_BYTES} bytes`,
+        },
+      });
+      return;
+    }
+    if (result.exitCode === 125) {
+      res.status(503).json({
+        ok: false,
+        error: { code: "SANDBOX_UNAVAILABLE", message: result.stderr },
+      });
+      return;
+    }
+    const dataFile = result.stdout
+      ? await stageLargeToolOutput(context, result.stdout, {
+          contentType: "text/plain",
+          extension: "txt",
+          prefix: "shell",
+        })
+      : null;
+    res.json({
+      ok: result.exitCode === 0,
+      exit_code: result.exitCode,
+      signal: result.signal,
+      ...(dataFile ? { data_file: dataFile } : { stdout: result.stdout }),
+      stderr: result.stderr || undefined,
+      stderr_truncated: result.stderrTruncated || undefined,
+    });
+  } catch (error) {
+    res.status(503).json({
+      ok: false,
+      error: {
+        code: "SANDBOX_UNAVAILABLE",
+        message: error instanceof Error ? error.message : String(error),
+      },
+    });
+  }
+});
+
 app.post("/internal/tools/store-render-ref", async (req, res) => {
  if (req.header("x-agent-internal-token") !== internalToken) {
    res.status(403).json({ message: "forbidden" });
@@ -342,6 +457,13 @@ app.post("/internal/tools/store-render-ref", async (req, res) => {
    res.status(400).json({ message: "file_path is required" });
    return;
  }
+  if (!context.workspaceDirectory) {
+    res.status(400).json({
+      message: "conversation workspace is required",
+      detail: "create a new conversation before importing render data",
+    });
+    return;
+  }

  try {
    const record = await resultReferenceResolver.registerRenderPayloadFile(filePath, {
@@ -352,6 +474,7 @@ app.post("/internal/tools/store-render-ref", async (req, res) => {
      sessionId: context.clientSessionId,
      source: RESULT_REFERENCE_SOURCE.agentGenerated,
      traceId: context.traceId,
+      workspaceDirectory: context.workspaceDirectory,
    });
    res.json({
      ok: true,
@@ -618,6 +741,7 @@ app.use(
 );

 const bootstrap = async () => {
+  sandboxProbe = await probeLandlockSandbox();
  await Promise.all([
    sessionMetadataStore.initialize(),
    sessionUiStateStore.initialize(),
@@ -1,5 +1,5 @@
 import { afterEach, beforeEach, describe, expect, it } from "bun:test";
-import { mkdtemp, rm, stat, writeFile } from "node:fs/promises";
+import { mkdir, mkdtemp, rm, stat, symlink, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";

@@ -12,13 +12,18 @@ import {

 describe("ResultReferenceResolver", () => {
  let tempDir: string;
+  let importRoot: string;
+  let conversationWorkspace: string;
  let store: ResultReferenceStore;
  let resolver: ResultReferenceResolver;

  beforeEach(async () => {
    tempDir = await mkdtemp(join(tmpdir(), "tjwater-result-ref-"));
-    store = new ResultReferenceStore(tempDir, 60_000);
-    resolver = new ResultReferenceResolver(store, tempDir, 1024 * 1024);
+    importRoot = join(tempDir, "conversation-workspaces");
+    conversationWorkspace = join(importRoot, "conversation-1");
+    await mkdir(conversationWorkspace, { recursive: true });
+    store = new ResultReferenceStore(join(tempDir, "refs"), 60_000);
+    resolver = new ResultReferenceResolver(store, importRoot, 1024 * 1024);
    await store.initialize();
  });

@@ -127,7 +132,7 @@ describe("ResultReferenceResolver", () => {
  });

  it("registers render refs from local wrapper files and normalizes payloads", async () => {
-    const filePath = join(tempDir, "render-wrapper.json");
+    const filePath = join(conversationWorkspace, "render-wrapper.json");
    await writeFile(
      filePath,
      JSON.stringify(
@@ -166,6 +171,7 @@ describe("ResultReferenceResolver", () => {
      sessionId: "session-3",
      source: RESULT_REFERENCE_SOURCE.agentGenerated,
      traceId: "trace-3",
+      workspaceDirectory: conversationWorkspace,
    });

    expect(record.kind).toBe(RESULT_REFERENCE_KIND.renderJunctionsPayload);
@@ -218,6 +224,7 @@ describe("ResultReferenceResolver", () => {
          sessionId: "session-4",
          source: RESULT_REFERENCE_SOURCE.agentGenerated,
          traceId: "trace-4",
+          workspaceDirectory: outsideDir,
        }),
      ).rejects.toThrow("RESULT_REF_IMPORT_DIR");
    } finally {
@@ -226,9 +233,9 @@ describe("ResultReferenceResolver", () => {
  });

  it("rejects oversized render payload files before parsing", async () => {
-    const filePath = join(tempDir, "oversized.json");
+    const filePath = join(conversationWorkspace, "oversized.json");
    await writeFile(filePath, "x".repeat(128), "utf8");
-    const sizeLimitedResolver = new ResultReferenceResolver(store, tempDir, 64);
+    const sizeLimitedResolver = new ResultReferenceResolver(store, importRoot, 64);

    await expect(
      sizeLimitedResolver.registerRenderPayloadFile(filePath, {
@@ -238,8 +245,65 @@ describe("ResultReferenceResolver", () => {
        sessionId: "session-5",
        source: RESULT_REFERENCE_SOURCE.agentGenerated,
        traceId: "trace-5",
+        workspaceDirectory: conversationWorkspace,
      }),
    ).rejects.toThrow("RESULT_REF_IMPORT_MAX_BYTES");
  });

+  it("rejects render payload files owned by another conversation workspace", async () => {
+    const otherWorkspace = join(importRoot, "conversation-2");
+    await mkdir(otherWorkspace);
+    const filePath = join(otherWorkspace, "render-wrapper.json");
+    await writeFile(
+      filePath,
+      JSON.stringify({
+        metadata: {},
+        location: { file_path: filePath },
+        data: { node_area_map: { J1: "DMA-1" } },
+      }),
+      "utf8",
+    );
+
+    await expect(
+      resolver.registerRenderPayloadFile(filePath, {
+        actorKey: "actor-6",
+        clientSessionId: "client-6",
+        projectKey: "project-key-6",
+        sessionId: "session-6",
+        source: RESULT_REFERENCE_SOURCE.agentGenerated,
+        traceId: "trace-6",
+        workspaceDirectory: conversationWorkspace,
+      }),
+    ).rejects.toThrow("current conversation workspace");
+  });
+
+  it("rejects a conversation workspace that is itself a symbolic link", async () => {
+    const targetWorkspace = join(importRoot, "conversation-target");
+    const linkedWorkspace = join(importRoot, "conversation-linked");
+    await mkdir(targetWorkspace);
+    await symlink(targetWorkspace, linkedWorkspace, "dir");
+    const filePath = join(targetWorkspace, "render-wrapper.json");
+    await writeFile(
+      filePath,
+      JSON.stringify({
+        metadata: {},
+        location: { file_path: filePath },
+        data: { node_area_map: { J1: "DMA-1" } },
+      }),
+      "utf8",
+    );
+
+    await expect(
+      resolver.registerRenderPayloadFile(filePath, {
+        actorKey: "actor-7",
+        clientSessionId: "client-7",
+        projectKey: "project-key-7",
+        sessionId: "session-7",
+        source: RESULT_REFERENCE_SOURCE.agentGenerated,
+        traceId: "trace-7",
+        workspaceDirectory: linkedWorkspace,
+      }),
+    ).rejects.toThrow("symbolic link");
+  });
+
 });
@@ -129,4 +129,72 @@ describe("permission approval policy", () => {
      title: "已按始终允许模式放行",
    });
  });
+
+  it("auto approves sandboxed shell and file access in a conversation workspace", async () => {
+    const root = await mkdtemp(join(tmpdir(), "permission-conversations-"));
+    const conversationRoot = join(root, "conversation-workspaces");
+    const workspaceRoot = join(conversationRoot, "conversation-test");
+    try {
+      await mkdir(workspaceRoot, { recursive: true });
+      expect(
+        resolvePermissionApproval("auto", "bash", {
+          workspaceRoot,
+          metadata: { command: "python3 analysis.py" },
+        }),
+      ).toMatchObject({ autoApprove: true, autoReject: false });
+      expect(
+        canAutoApprovePermission("edit", {
+          workspaceRoot,
+          metadata: { filePath: join(workspaceRoot, "result.json") },
+        }),
+      ).toBe(true);
+      expect(
+        canAutoApprovePermission("glob", {
+          workspaceRoot,
+          metadata: { path: workspaceRoot, pattern: "**/*.json" },
+          patterns: ["**/*.json"],
+        }),
+      ).toBe(true);
+    } finally {
+      await rm(root, { force: true, recursive: true });
+    }
+  });
+
+  it.each([
+    "rm -rf ./target",
+    "rm  -rf ./target",
+    "rm -Rf ./target",
+    "/bin/rm -rf ./target",
+    "command rm --force --recursive ./target",
+    "env LANG=C rm -r -f ./target",
+    "SAFE=1 rm -rf ./target",
+    "env -u HOME rm -rf ./target",
+    "r\"\"m -rf ./target",
+    "(rm -rf ./target)",
+    "! rm -rf ./target",
+    "npm test && rm --recursive --force ./target",
+  ])("rejects direct recursive force removal in always mode: %s", (command) => {
+    expect(
+      resolvePermissionApproval("always", "bash", {
+        metadata: { command },
+        patterns: [command],
+      }),
+    ).toMatchObject({
+      autoApprove: false,
+      autoReject: true,
+      title: "已拒绝递归强制删除",
+    });
+  });
+
+  it.each(["rm tmp.txt", "rm -f tmp.txt", "rm -r tmp-dir", "echo 'rm -rf tmp'"])(
+    "allows non-force-recursive or non-executed removal text in always mode: %s",
+    (command) => {
+      expect(
+        resolvePermissionApproval("always", "bash", {
+          metadata: { command },
+          patterns: [command],
+        }),
+      ).toMatchObject({ autoApprove: true, autoReject: false });
+    },
+  );
 });
@@ -435,6 +435,54 @@ describe("streamPromptResponse", () => {
    expect(events.some((item) => item.event === "permission_request")).toBe(false);
  });

+  it("rejects recursive force removal even in always mode", async () => {
+    const replies: Array<Record<string, unknown>> = [];
+    const runtime = {
+      subscribeEvents: async () =>
+        createEventStream([
+          {
+            type: "permission.asked",
+            properties: {
+              id: "perm-always-rm-rf",
+              sessionID: "runtime-session-1",
+              permission: "bash",
+              patterns: ["/bin/rm -rf ./target"],
+              metadata: { command: "/bin/rm -rf ./target" },
+              always: ["/bin/rm -rf ./target"],
+            },
+          },
+          { type: "session.idle", properties: { sessionID: "runtime-session-1" } },
+        ]),
+      prompt: async () => undefined,
+      messages: async () => [],
+      replyPermission: async (options: Record<string, unknown>) => replies.push(options),
+    } as unknown as OpencodeRuntimeAdapter;
+    const events: Array<{ event: string; data: Record<string, unknown> }> = [];
+
+    await streamPromptResponse({
+      runtime,
+      sessionId: "runtime-session-1",
+      clientSessionId: "client-session-1",
+      message: "delete recursively",
+      approvalMode: "always",
+      write: (event, data) => events.push({ event, data }),
+    });
+
+    expect(replies).toEqual([
+      {
+        requestId: "perm-always-rm-rf",
+        sessionId: "runtime-session-1",
+        reply: "reject",
+      },
+    ]);
+    expect(events.some((item) => item.event === "permission_request")).toBe(false);
+    expect(events.find((item) => item.event === "permission_response")?.data).toEqual({
+      session_id: "client-session-1",
+      request_id: "perm-always-rm-rf",
+      reply: "reject",
+    });
+  });
+
  it("forwards opencode v2 permission requests as SSE payloads", async () => {
    const runtime = {
      subscribeEvents: async () =>
@@ -1,5 +1,8 @@
 import { describe, expect, it } from "bun:test";
 import { type OpencodeClient } from "@opencode-ai/sdk/v2";
+import { mkdtemp, readdir, rm, stat } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join, relative } from "node:path";

 import { config } from "../../src/config.js";
 import { OpencodeRuntimeAdapter } from "../../src/runtime/opencode.js";
@@ -87,6 +90,79 @@ describe("OpencodeRuntimeAdapter.ensureClient", () => {
  });
 });

+describe("OpencodeRuntimeAdapter.createSession", () => {
+  it("creates a real chat session inside a dedicated conversation workspace", async () => {
+    const workspaceRoot = await mkdtemp(join(tmpdir(), "tjwater-conversations-"));
+    const calls: Array<Record<string, unknown>> = [];
+    const client = {
+      session: {
+        create: async (input: Record<string, unknown>) => {
+          calls.push(input);
+          return {
+            data: {
+              id: "runtime-session-1",
+              directory: input.directory,
+            },
+          };
+        },
+      },
+    } as unknown as OpencodeClient;
+    const runtime = Object.assign(Object.create(OpencodeRuntimeAdapter.prototype), {
+      clientPromise: null,
+      closeServer: null,
+      ensureClient: async () => client,
+    }) as OpencodeRuntimeAdapter;
+
+    try {
+      const session = await runtime.createSession("chat", {
+        conversationWorkspace: true,
+        workspaceRoot,
+      });
+      const directory = String(calls[0]?.directory);
+
+      expect(relative(workspaceRoot, directory).startsWith("..")).toBe(false);
+      const workspaceStat = await stat(directory);
+      expect(workspaceStat.isDirectory()).toBe(true);
+      expect(workspaceStat.mode & 0o777).toBe(0o700);
+      expect(session.directory).toBe(directory);
+      expect(calls[0]?.permission).toEqual([
+        { permission: "read", pattern: `${directory}/**`, action: "allow" },
+        { permission: "edit", pattern: `${directory}/**`, action: "ask" },
+      ]);
+    } finally {
+      await rm(workspaceRoot, { force: true, recursive: true });
+    }
+  });
+
+  it("removes an empty conversation workspace when session creation fails", async () => {
+    const workspaceRoot = await mkdtemp(join(tmpdir(), "tjwater-conversations-"));
+    const client = {
+      session: {
+        create: async () => {
+          throw new Error("session creation failed");
+        },
+      },
+    } as unknown as OpencodeClient;
+    const runtime = Object.assign(Object.create(OpencodeRuntimeAdapter.prototype), {
+      clientPromise: null,
+      closeServer: null,
+      ensureClient: async () => client,
+    }) as OpencodeRuntimeAdapter;
+
+    try {
+      await expect(
+        runtime.createSession("chat", {
+          conversationWorkspace: true,
+          workspaceRoot,
+        }),
+      ).rejects.toThrow("session creation failed");
+      expect(await readdir(workspaceRoot)).toEqual([]);
+    } finally {
+      await rm(workspaceRoot, { force: true, recursive: true });
+    }
+  });
+});
+
 describe("OpencodeRuntimeAdapter.warmup", () => {
  it("initializes the project session and model tools before reporting ready", async () => {
    const calls: string[] = [];
@@ -1,7 +1,29 @@
 import { describe, expect, it } from "bun:test";
 import { readFile } from "node:fs/promises";
+import sandboxBash from "../../.opencode/tools/bash.js";
+import tjwaterCli from "../../.opencode/tools/tjwater_cli.js";
+import storeRenderRef, {
+  resolveStoreRenderFilePath,
+} from "../../.opencode/tools/store_render_ref.js";

 describe("internal OpenCode permissions", () => {
+  it("pins the runtime, SDK, plugin, and image to OpenCode 1.18.13", async () => {
+    const [rootPackageText, toolPackageText, dockerfile] = await Promise.all([
+      readFile("package.json", "utf8"),
+      readFile(".opencode/package.json", "utf8"),
+      readFile("Dockerfile", "utf8"),
+    ]);
+    const rootPackage = JSON.parse(rootPackageText) as {
+      dependencies?: Record<string, string>;
+    };
+    const toolPackage = JSON.parse(toolPackageText) as {
+      dependencies?: Record<string, string>;
+    };
+    expect(rootPackage.dependencies?.["@opencode-ai/sdk"]).toBe("1.18.13");
+    expect(toolPackage.dependencies?.["@opencode-ai/plugin"]).toBe("1.18.13");
+    expect(dockerfile.startsWith("FROM smanx/opencode:1.18.13@")).toBe(true);
+  });
+
  it("keeps protected paths denied in every approval mode", async () => {
    const config = JSON.parse(await readFile("opencode.json", "utf8")) as {
      permission?: Record<string, string | Record<string, string>>;
@@ -23,8 +45,140 @@ describe("internal OpenCode permissions", () => {
    expect(edit?.["data/**"]).toBe("deny");
    expect(edit?.["**/logs/**"]).toBe("deny");
    expect(bash?.["*"]).toBe("ask");
+    expect(bash?.["rm *"]).toBe("ask");
+    expect(bash?.["rm -rf *"]).toBe("deny");
+    expect(bash?.["rm -fr *"]).toBe("deny");
+    expect(bash?.["rm -r -f *"]).toBe("deny");
+    expect(bash?.["rm -f -r *"]).toBe("deny");
+    expect(bash?.["rm --recursive --force *"]).toBe("deny");
+    expect(bash?.["rm --force --recursive *"]).toBe("deny");
    expect(bash?.["*.env*"]).toBe("deny");
-    expect(bash?.["*data/*"]).toBe("deny");
-    expect(bash?.["*logs/*"]).toBe("deny");
+    expect(bash?.["*data/*"]).toBeUndefined();
+    expect(bash?.["*logs/*"]).toBeUndefined();
+  });
+});
+
+describe("store_render_ref arguments", () => {
+  it("accepts the observed camelCase alias without changing snake_case precedence", () => {
+    expect(
+      resolveStoreRenderFilePath({
+        filePath: "/app/data/conversation-workspaces/chat-1/partition.json",
+      }),
+    ).toBe("/app/data/conversation-workspaces/chat-1/partition.json");
+
+    expect(
+      resolveStoreRenderFilePath({
+        file_path: "/app/data/conversation-workspaces/chat-1/preferred.json",
+        filePath: "/app/data/conversation-workspaces/chat-1/compatibility.json",
+      }),
+    ).toBe("/app/data/conversation-workspaces/chat-1/preferred.json");
+  });
+
+  it("forwards a camelCase compatibility argument as file_path", async () => {
+    const originalFetch = globalThis.fetch;
+    let requestBody: unknown;
+    globalThis.fetch = (async (
+      _input: RequestInfo | URL,
+      init?: RequestInit,
+    ) => {
+      requestBody = JSON.parse(String(init?.body));
+      return new Response('{"render_ref":"res-test"}');
+    }) as unknown as typeof fetch;
+
+    try {
+      const definition = storeRenderRef as unknown as {
+        args: Record<string, unknown>;
+        execute: (args: unknown, context: unknown) => Promise<unknown>;
+      };
+      expect(definition.args.filePath).toBeDefined();
+      await definition.execute(
+        {
+          reason: "regression test",
+          filePath: "/app/data/conversation-workspaces/chat-1/partition.json",
+        },
+        { sessionID: "session-test" } as never,
+      );
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+
+    expect(requestBody).toEqual({
+      session_id: "session-test",
+      file_path: "/app/data/conversation-workspaces/chat-1/partition.json",
+    });
+  });
+});
+
+describe("sandbox bash tool", () => {
+  it("forwards commands to the authenticated sandbox endpoint", async () => {
+    const originalFetch = globalThis.fetch;
+    const permissionRequests: unknown[] = [];
+    let requestUrl = "";
+    let requestBody: unknown;
+    globalThis.fetch = (async (input: RequestInfo | URL, init?: RequestInit) => {
+      requestUrl = String(input);
+      requestBody = JSON.parse(String(init?.body));
+      return new Response('{"ok":true,"stdout":"done"}');
+    }) as unknown as typeof fetch;
+    try {
+      const definition = sandboxBash as unknown as {
+        execute: (args: unknown, context: unknown) => Promise<unknown>;
+      };
+      await definition.execute(
+        { command: "python3 analysis.py", timeout: 300 },
+        {
+          sessionID: "session-test",
+          ask: async (input: unknown) => permissionRequests.push(input),
+        } as never,
+      );
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+    expect(requestUrl).toEndWith("/internal/tools/sandbox-shell");
+    expect(permissionRequests).toEqual([
+      {
+        permission: "bash",
+        patterns: ["python3 analysis.py"],
+        always: ["python3 analysis.py"],
+        metadata: {
+          command: "python3 analysis.py",
+        },
+      },
+    ]);
+    expect(requestBody).toEqual({
+      session_id: "session-test",
+      command: "python3 analysis.py",
+      timeout: 300,
+    });
+  });
+});
+
+describe("tjwater_cli storage request", () => {
+  it("forwards store_result so small workflow inputs can be staged", async () => {
+    const originalFetch = globalThis.fetch;
+    let requestBody: unknown;
+    globalThis.fetch = (async (_input: RequestInfo | URL, init?: RequestInit) => {
+      requestBody = JSON.parse(String(init?.body));
+      return new Response('{"ok":true,"data_file":{"file_path":"/tmp/test"}}');
+    }) as unknown as typeof fetch;
+    try {
+      const definition = tjwaterCli as unknown as {
+        execute: (args: unknown, context: unknown) => Promise<unknown>;
+      };
+      await definition.execute(
+        {
+          command: "network get-all-reservoirs-properties",
+          reason: "prepare workflow input",
+          store_result: true,
+        },
+        { sessionID: "session-test" } as never,
+      );
+    } finally {
+      globalThis.fetch = originalFetch;
+    }
+    expect(requestBody).toMatchObject({
+      session_id: "session-test",
+      store_result: true,
+    });
  });
 });
@@ -19,6 +19,7 @@ describe("runtime session context", () => {
      projectKey: "project-1",
      sessionId: "runtime-session-1",
      traceId: "trace-1",
+      workspaceDirectory: "/app/data/conversation-workspaces/chat-session-1",
    });

    const runtimeContext = getRuntimeSessionContext("runtime-session-1");
@@ -27,6 +28,9 @@ describe("runtime session context", () => {
    expect(runtimeContext?.clientSessionId).toBe("chat-session-1");
    expect(runtimeContext?.network).toBe("fengyang");
    expect(runtimeContext?.sessionId).toBe("runtime-session-1");
+    expect(runtimeContext?.workspaceDirectory).toBe(
+      "/app/data/conversation-workspaces/chat-session-1",
+    );

    removeRuntimeSessionContext("runtime-session-1");
    expect(getRuntimeSessionContext("runtime-session-1")).toBeNull();
@@ -0,0 +1,98 @@
+import { afterEach, describe, expect, it } from "bun:test";
+import { mkdir, readFile, rm, stat } from "node:fs/promises";
+import { resolve } from "node:path";
+
+import { config } from "../../src/config.js";
+import { type RuntimeSessionContext } from "../../src/runtime/sessionContext.js";
+import {
+  buildLargeCliResult,
+  stageLargeToolOutput,
+} from "../../src/runtime/toolOutputStaging.js";
+
+const createdPaths: string[] = [];
+
+afterEach(async () => {
+  await Promise.all(
+    createdPaths.splice(0).map((path) => rm(path, { force: true, recursive: true })),
+  );
+});
+
+const createContext = async (): Promise<RuntimeSessionContext> => {
+  const importRoot = resolve(config.RESULT_REF_IMPORT_DIR);
+  await mkdir(importRoot, { recursive: true });
+  const workspaceDirectory = resolve(
+    importRoot,
+    `conversation-staging-${crypto.randomUUID()}`,
+  );
+  await mkdir(workspaceDirectory, { mode: 0o700 });
+  createdPaths.push(workspaceDirectory);
+  return {
+    actorKey: "actor-1",
+    clientSessionId: "client-1",
+    projectKey: "project-1",
+    sessionId: "runtime-1",
+    traceId: "trace-1",
+    workspaceDirectory,
+  };
+};
+
+describe("tool output staging", () => {
+  it("keeps small output inline unless storage is forced", async () => {
+    const context = await createContext();
+    await expect(
+      stageLargeToolOutput(context, '{"ok":true}', {
+        contentType: "application/json",
+        extension: "json",
+        prefix: "cli",
+      }),
+    ).resolves.toBeNull();
+
+    const stored = await stageLargeToolOutput(context, '{"ok":true}', {
+      contentType: "application/json",
+      extension: "json",
+      force: true,
+      prefix: "cli",
+    });
+    expect(stored?.file_path).toContain("/tool-data/cli-");
+    expect(await readFile(stored!.file_path, "utf8")).toBe('{"ok":true}');
+    expect((await stat(stored!.file_path)).mode & 0o777).toBe(0o600);
+  });
+
+  it("stages output above the inline threshold and returns a compact descriptor", async () => {
+    const context = await createContext();
+    const content = JSON.stringify({
+      ok: true,
+      schema_version: "tjwater-cli/v1",
+      data: "x".repeat(config.MAX_INLINE_RESULT_BYTES),
+    });
+    const dataFile = await stageLargeToolOutput(context, content, {
+      contentType: "application/json",
+      extension: "json",
+      prefix: "cli",
+    });
+    expect(dataFile?.bytes).toBe(Buffer.byteLength(content));
+    const stored = dataFile!;
+    expect(buildLargeCliResult(stored)).toEqual({
+      ok: true,
+      schema_version: "tjwater-cli/v1",
+      summary: "CLI 结果已保存到当前对话工作区",
+      data_file: stored,
+    });
+  });
+
+  it("does not stage large output for legacy sessions without a workspace", async () => {
+    await expect(
+      stageLargeToolOutput(
+        {
+          actorKey: "actor-1",
+          clientSessionId: "client-1",
+          projectKey: "project-1",
+          sessionId: "runtime-1",
+          traceId: "trace-1",
+        },
+        "x".repeat(config.MAX_INLINE_RESULT_BYTES + 1),
+        { contentType: "text/plain", extension: "txt", prefix: "shell" },
+      ),
+    ).rejects.toThrow("requires a conversation workspace");
+  });
+});
@@ -0,0 +1,94 @@
+import { afterEach, describe, expect, it } from "bun:test";
+import { mkdir, rm } from "node:fs/promises";
+import { resolve } from "node:path";
+
+import { config } from "../../src/config.js";
+import {
+  resolveConversationWorkspace,
+  setSandboxOwnership,
+} from "../../src/runtime/conversationWorkspace.js";
+import {
+  executeSandboxCommand,
+  probeLandlockSandbox,
+} from "../../src/sandbox/landlockSandbox.js";
+
+const createdPaths: string[] = [];
+
+afterEach(async () => {
+  await Promise.all(
+    createdPaths.splice(0).map((path) => rm(path, { force: true, recursive: true })),
+  );
+});
+
+const createWorkspace = async (name: string = crypto.randomUUID()) => {
+  const importRoot = resolve(config.RESULT_REF_IMPORT_DIR);
+  await mkdir(importRoot, { recursive: true });
+  const workspace = resolve(importRoot, `conversation-sandbox-${name}`);
+  await mkdir(workspace, { mode: 0o700 });
+  await setSandboxOwnership(workspace);
+  createdPaths.push(workspace);
+  return workspace;
+};
+
+describe("Landlock sandbox", () => {
+  it("requires Landlock ABI 4 and seccomp", async () => {
+    const probe = await probeLandlockSandbox();
+    expect(probe.landlockAbi).toBeGreaterThanOrEqual(4);
+    expect(probe.seccomp).toBe(true);
+  });
+
+  it("allows local Python analysis while denying filesystem escape, secrets, and network", async () => {
+    const workspace = await createWorkspace("primary");
+    const sibling = await createWorkspace("sibling");
+    await resolveConversationWorkspace(workspace, config.RESULT_REF_IMPORT_DIR);
+
+    const allowed = await executeSandboxCommand(
+      workspace,
+      'python3 -c "import json; open(\'result.json\', \'w\').write(json.__name__)" && cat result.json',
+      10,
+    );
+    expect(allowed).toMatchObject({ exitCode: 0, stdout: "json" });
+
+    const escaped = await executeSandboxCommand(
+      workspace,
+      `cat ${resolve("package.json")}`,
+      10,
+    );
+    expect(escaped.exitCode).not.toBe(0);
+    expect(escaped.stderr).toContain("Permission denied");
+
+    const siblingRead = await executeSandboxCommand(
+      workspace,
+      `ls ${sibling}`,
+      10,
+    );
+    expect(siblingRead.exitCode).not.toBe(0);
+
+    const originalApiKey = process.env.DEEPSEEK_API_KEY;
+    process.env.DEEPSEEK_API_KEY = "must-not-leak";
+    try {
+      const environment = await executeSandboxCommand(
+        workspace,
+        'test -z "$DEEPSEEK_API_KEY"',
+        10,
+      );
+      expect(environment.exitCode).toBe(0);
+    } finally {
+      if (originalApiKey === undefined) {
+        delete process.env.DEEPSEEK_API_KEY;
+      } else {
+        process.env.DEEPSEEK_API_KEY = originalApiKey;
+      }
+    }
+
+    for (const socketType of ["SOCK_STREAM", "SOCK_DGRAM"]) {
+      const network = await executeSandboxCommand(
+        workspace,
+        `python3 -c "import socket; socket.socket(socket.AF_INET, socket.${socketType})"`,
+        10,
+      );
+      expect(network.exitCode).not.toBe(0);
+      expect(network.stderr).toContain("Operation not permitted");
+    }
+  });
+});
@@ -0,0 +1,73 @@
+import { afterEach, describe, expect, it } from "bun:test";
+import { execFile } from "node:child_process";
+import { mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join, resolve } from "node:path";
+import { promisify } from "node:util";
+
+const execFileAsync = promisify(execFile);
+const createdPaths: string[] = [];
+
+afterEach(async () => {
+  await Promise.all(
+    createdPaths.splice(0).map((path) => rm(path, { force: true, recursive: true })),
+  );
+});
+
+describe("service area workflow script", () => {
+  it("writes the wrapped render payload required by store_render_ref", async () => {
+    const directory = await mkdtemp(join(tmpdir(), "service-area-script-"));
+    createdPaths.push(directory);
+    const time = "2026-04-01T08:00:00+08:00";
+    const inputs = {
+      pipes: { data: [{ id: "P1", node1: "R1", node2: "N1" }] },
+      reservoirs: { data: [{ id: "R1" }] },
+      links: { data: [{ id: "P1", flow: 1, time }] },
+      nodes: {
+        data: [
+          { id: "R1", pressure: 30, actual_demand: 0, time },
+          { id: "N1", pressure: 28, actual_demand: 2, time },
+        ],
+      },
+    };
+    const paths = Object.fromEntries(
+      await Promise.all(
+        Object.entries(inputs).map(async ([name, value]) => {
+          const path = join(directory, `${name}.json`);
+          await writeFile(path, JSON.stringify(value));
+          return [name, path] as const;
+        }),
+      ),
+    );
+    const outputPath = join(directory, "service-area-wrapper.json");
+    await execFileAsync(
+      "python3",
+      [
+        resolve(
+          ".opencode/skills/workflow/service-area-analysis/scripts/service_area_partition.py",
+        ),
+        "--pipe-props",
+        paths.pipes!,
+        "--reservoirs",
+        paths.reservoirs!,
+        "--links",
+        paths.links!,
+        "--nodes",
+        paths.nodes!,
+        "--target-time",
+        time,
+        "--output",
+        outputPath,
+      ],
+      { cwd: directory },
+    );
+    const wrapper = JSON.parse(await readFile(outputPath, "utf8")) as {
+      data: { node_area_map: Record<string, string> };
+      location: { file_path: string };
+      metadata: { schema_version: number };
+    };
+    expect(wrapper.location.file_path).toBe(outputPath);
+    expect(wrapper.metadata.schema_version).toBe(1);
+    expect(wrapper.data.node_area_map).toMatchObject({ R1: "R1", N1: "R1" });
+  });
+});
Author	SHA1	Message	Date
jiang	80cfc1f2ab	feat(agent): sandbox conversation analysis Generic Container CI/CD / test-build-publish (push) Successful in 2m44s Details Agent CI/CD v2 / build-test-publish-and-deploy (push) Successful in 2m44s Details	2026-08-25 16:08:33 +08:00
jiang	ce04704af2	fix(agent): isolate conversation workspaces Generic Container CI/CD / test-build-publish (push) Successful in 2m2s Details Agent CI/CD v2 / build-test-publish-and-deploy (push) Successful in 2m2s Details	2026-08-25 13:18:39 +08:00