Update config.yaml

README.md

@@ -35,7 +35,18 @@ CyberStrikeAI is an **AI-native security testing platform** built in Go. It inte
 
 ### System Dashboard Overview
 
-<img src="./images/dashboard.png" alt="System Dashboard" width="100%">
+<table>
+<tr>
+<td width="50%" align="center">
+<strong>Light Mode</strong><br/>
+<img src="./images/dashboard.png" alt="System Dashboard (Light)" width="100%">
+</td>
+<td width="50%" align="center">
+<strong>Dark Mode</strong><br/>
+<img src="./images/dark.png" alt="System Dashboard (Dark)" width="100%">
+</td>
+</tr>
+</table>
 
 *The dashboard provides a comprehensive overview of system runtime status, security vulnerabilities, tool usage, and knowledge base, helping users quickly understand the platform's core features and current state.*
 
@@ -110,7 +121,7 @@ CyberStrikeAI is an **AI-native security testing platform** built in Go. It inte
 - 📄 Large-result pagination, compression, and searchable archives
 - 🔗 Attack-chain graph, risk scoring, and step-by-step replay
 - 🔒 Password-protected web UI, audit logs, and SQLite persistence
-- 📚 Knowledge base (RAG) with embedding-based vector retrieval (cosine similarity), optional **Eino Compose** indexing pipeline, and configurable post-retrieval budgets / reranking hooks
+- 📚 Knowledge base (RAG): **Eino MultiQuery** query rewrite + multi-path vector retrieval + **HTTP rerank** (DashScope `gte-rerank` / Cohere-compatible) + post-processing (dedupe, budget); **Eino Compose** indexing pipeline
 - 📁 Conversation grouping with pinning, rename, and batch management
 - 📂 **Project management**: shared facts (blackboard) across sessions, `upsert_project_fact` + `links` to chain paths; attack-chain and project fact graph views
 - 🛡️ Vulnerability management with CRUD operations, severity tracking, status workflow, and statistics
@@ -455,9 +466,10 @@ A test SSE MCP server is available at `cmd/test-sse-mcp-server/` for validation
 
 ### Knowledge Base
 - **Vector search** – AI agent can automatically search the knowledge base for relevant security knowledge during conversations using the `search_knowledge_base` tool.
-- **Vector retrieval** – cosine similarity over stored embeddings, aligned with Eino `retriever.Retriever` usage.
+- **RAG pipeline (always on)** – **MultiQuery** (LLM query rewrite) → vector prefetch & fusion → **HTTP rerank** (DashScope `gte-rerank` or Cohere-compatible `/v1/rerank`) → post-processing (normalized dedupe, char/token budget, final top_k). Rerank failures degrade to fusion order without breaking search.
-- **Auto-indexing** – scans the `knowledge_base/` directory for Markdown files and automatically indexes them with embeddings.
+- **Vector retrieval** – cosine similarity over stored embeddings with configurable threshold, aligned with Eino `retriever.Retriever` usage.
-- **Web management** – create, update, delete knowledge items through the web UI, with category-based organization.
+- **Auto-indexing** – scans the `knowledge_base/` directory for Markdown files and automatically indexes them with embeddings (Markdown header split + recursive chunking via Eino).
+- **Web management** – create, update, delete knowledge items through the web UI, with category-based organization; settings page exposes MultiQuery / rerank / prefetch options.
 - **Retrieval logs** – tracks all knowledge retrieval operations for audit and debugging.
 
 **Quick Start (Using Pre-built Knowledge Base):**
@@ -479,6 +491,17 @@ A test SSE MCP server is available at `cmd/test-sse-mcp-server/` for validation
      retrieval:
        top_k: 5
        similarity_threshold: 0.7
+       multi_query:
+         max_queries: 4        # LLM rewrite variants (always on)
+       rerank:                 # always on; empty fields inherit openai/embedding credentials
+         provider: ""          # auto: dashscope | cohere from base_url
+         model: ""             # empty: gte-rerank (DashScope) or rerank-multilingual-v3.0 (Cohere)
+         base_url: ""
+         api_key: ""
+       post_retrieve:
+         prefetch_top_k: 20    # vector candidates per MultiQuery variant; 0 = max(top_k×4, 20)
+         max_context_chars: 0
+         max_context_tokens: 0
    ```
 2. **Add knowledge files** – place Markdown files in `knowledge_base/` directory, organized by category (e.g., `knowledge_base/SQL Injection/README.md`).
 3. **Scan and index** – use the web UI to scan the knowledge base directory, which will automatically import files and build vector embeddings.
@@ -539,6 +562,17 @@ knowledge:
   retrieval:
     top_k: 5  # Number of top results to return
     similarity_threshold: 0.7  # Minimum cosine similarity (0-1)
+    multi_query:
+      max_queries: 4  # MultiQuery rewrite variants (always on)
+    rerank:  # HTTP rerank (always on); empty fields inherit openai/embedding credentials
+      provider: ""
+      model: ""
+      base_url: ""
+      api_key: ""
+    post_retrieve:
+      prefetch_top_k: 20  # per MultiQuery variant; 0 = max(top_k×4, 20)
+      max_context_chars: 0
+      max_context_tokens: 0
 roles_dir: "roles"  # Role configuration directory (relative to config file)
 skills_dir: "skills"  # Skills directory (relative to config file)
 agents_dir: "agents"  # Multi-agent Markdown definitions (orchestrator + sub-agents)

README_CN.md

@@ -34,7 +34,18 @@ CyberStrikeAI 是一款 **AI 原生安全测试平台**，基于 Go 构建，集
 
 ### 系统仪表盘概览
 
-<img src="./images/dashboard.png" alt="系统仪表盘" width="100%">
+<table>
+<tr>
+<td width="50%" align="center">
+<strong>浅色模式</strong><br/>
+<img src="./images/dashboard.png" alt="系统仪表盘（浅色）" width="100%">
+</td>
+<td width="50%" align="center">
+<strong>深色模式</strong><br/>
+<img src="./images/dark.png" alt="系统仪表盘（深色）" width="100%">
+</td>
+</tr>
+</table>
 
 *仪表盘提供系统运行状态、安全漏洞、工具使用情况和知识库的全面概览，帮助用户快速了解平台核心功能和当前状态。*
 
@@ -109,7 +120,7 @@ CyberStrikeAI 是一款 **AI 原生安全测试平台**，基于 Go 构建，集
 - 📄 大结果分页、压缩与全文检索
 - 🔗 攻击链可视化、风险打分与步骤回放
 - 🔒 Web 登录保护、审计日志、SQLite 持久化
-- 📚 知识库（RAG）：向量嵌入与余弦相似度检索（与 Eino `retriever.Retriever` 语义一致），可选 **Eino Compose** 索引流水线及检索后处理（预算、重排等配置项）
+- 📚 知识库（RAG）：**Eino MultiQuery** 查询改写 + 多路向量检索 + **HTTP 精排**（DashScope `gte-rerank` / Cohere 兼容）+ 后处理（去重、预算）；索引侧为 **Eino Compose** 流水线
 - 📁 对话分组管理：支持分组创建、置顶、重命名、删除等操作
 - 📂 **项目管理**：共享事实（黑板）跨会话沉淀认知，`upsert_project_fact` + `links` 串联攻击路径；聊天攻击链与项目事实图可视化
 - 🛡️ 漏洞管理功能：完整的漏洞 CRUD 操作，支持严重程度分级、状态流转、按对话/严重程度/状态过滤，以及统计看板
@@ -453,9 +464,10 @@ CyberStrikeAI 支持通过三种传输模式连接外部 MCP 服务器：
 
 ### 知识库功能
 - **向量检索**：AI 智能体在对话过程中可自动调用 `search_knowledge_base` 工具搜索知识库中的安全知识。
-- **向量检索**：基于嵌入余弦相似度与相似度阈值过滤（与 Eino `retriever.Retriever` 语义一致）。
+- **RAG 管线（始终启用）**：**MultiQuery**（LLM 查询改写）→ 向量预取与融合 → **HTTP 精排**（DashScope `gte-rerank` 或 Cohere 兼容 `/v1/rerank`）→ 后处理（规范化去重、字符/token 预算、最终 top_k）。精排失败时自动降级为融合排序，检索仍可用。
-- **自动索引**：扫描 `knowledge_base/` 目录下的 Markdown 文件，自动构建向量嵌入索引。
+- **向量相似度**：基于嵌入余弦相似度与相似度阈值过滤（与 Eino `retriever.Retriever` 语义一致）。
-- **Web 管理**：通过 Web 界面创建、更新、删除知识项，支持分类管理。
+- **自动索引**：扫描 `knowledge_base/` 目录下的 Markdown 文件，自动构建向量嵌入索引（Eino Markdown 标题切分 + 递归分块）。
+- **Web 管理**：通过 Web 界面创建、更新、删除知识项，支持分类管理；设置页可配置 MultiQuery / 精排 / 预取候选数。
 - **检索日志**：记录所有知识检索操作，便于审计与调试。
 
 **快速开始（使用预构建知识库）：**
@@ -477,6 +489,17 @@ CyberStrikeAI 支持通过三种传输模式连接外部 MCP 服务器：
      retrieval:
        top_k: 5
        similarity_threshold: 0.7
+       multi_query:
+         max_queries: 4        # LLM 改写变体上限（始终启用）
+       rerank:                 # 精排始终启用；留空则继承 openai/embedding 凭据
+         provider: ""          # 空=按 base_url 推断 dashscope | cohere
+         model: ""             # 空=DashScope→gte-rerank，Cohere→rerank-multilingual-v3.0
+         base_url: ""
+         api_key: ""
+       post_retrieve:
+         prefetch_top_k: 20    # 每条 MultiQuery 变体的向量候选数；0=max(top_k×4, 20)
+         max_context_chars: 0
+         max_context_tokens: 0
    ```
 2. **添加知识文件**：将 Markdown 文件放入 `knowledge_base/` 目录，按分类组织（如 `knowledge_base/SQL注入/README.md`）。
 3. **扫描索引**：在 Web 界面中点击"扫描知识库"，系统会自动导入文件并构建向量索引。
@@ -537,6 +560,17 @@ knowledge:
   retrieval:
     top_k: 5  # 检索返回的 Top-K 结果数量
     similarity_threshold: 0.7  # 余弦相似度阈值（0-1），低于此值的结果将被过滤
+    multi_query:
+      max_queries: 4  # MultiQuery 改写变体上限（始终启用）
+    rerank:  # HTTP 精排（始终启用）；留空则继承 openai/embedding 凭据
+      provider: ""
+      model: ""
+      base_url: ""
+      api_key: ""
+    post_retrieve:
+      prefetch_top_k: 20  # 每条 MultiQuery 变体；0=max(top_k×4, 20)
+      max_context_chars: 0
+      max_context_tokens: 0
 roles_dir: "roles"  # 角色配置文件目录（相对于配置文件所在目录）
 skills_dir: "skills"  # Skills 目录（相对于配置文件所在目录）
 agents_dir: "agents"  # 多代理 Markdown（主代理 orchestrator.md + 子代理 *.md）

config.yaml

@@ -10,7 +10,7 @@
 # ============================================
 
 # 前端显示的版本号（可选，不填则显示默认版本）
-version: "v1.6.46"
+version: "v1.6.49"
 # 服务器配置
 server:
   host: 0.0.0.0 # 监听地址，0.0.0.0 表示监听所有网络接口
@@ -102,9 +102,69 @@ agent:
 
   system_prompt_path: ""
 # 人机协同（HITL）全局白名单：此处列出的工具始终免审批，与对话页「白名单工具（免审批，逗号分隔）」合并为并集；侧栏「应用」可合并写入本列表并立即生效。
+# 非白名单工具在审批方=审计 Agent 时，按会话 HITL 模式选用提示词：
+#   approval → audit_agent_prompt
+#   review_edit → audit_agent_prompt_review_edit（可改参后放行）
 hitl:
+  # 已决策审计日志保留天数（与 MCP 监控一致；省略默认 90；0 表示不自动清理）
+  retention_days: 90
   # 按你环境里的真实工具名增删（与侧栏一致、小写不敏感）；不需要全局免审批可改为 []
-  tool_whitelist: [read_file, list_dir, glob, grep]
+  tool_whitelist: [read_file, list_dir, glob, grep, tool_search]
+  # audit_agent_prompt: |  # 审批模式；留空使用内置默认，可在「人机协同」页编辑
+  # audit_agent_prompt_review_edit: |  # 审查编辑模式；留空使用内置默认
+
+  audit_agent_prompt: |-
+    你是 CyberStrikeAI 人机协同审计 Agent。审查 Agent 即将执行的工具调用是否会对系统造成实质性损害。
+
+    你会收到 JSON，包含 hitlMode、toolName、arguments/argumentsObj、userMessage、thinking、reasoningChain、planning 等字段。
+
+    裁决基调（默认放行）：
+    - 常规、低风险的渗透测试操作 → approve（如信息收集、端口/服务扫描、目录枚举、只读查询、无害探测命令）
+    - 与用户授权、当前任务目标一致，且未见明确高危迹象 → approve
+    - 仅在「可能对系统造成实质影响」时 → reject
+
+    必须 reject 的高危情形（示例，非穷举）：
+    - 删库、清表、批量删除数据、格式化磁盘、不可逆破坏
+    - 修改/重置密码、创建或篡改管理员账号、持久化后门、开机自启
+    - 向生产环境写入恶意载荷、勒索加密、停止关键服务、修改系统核心配置
+    - 明显越权：与任务/授权目标无关的破坏性操作
+
+    不应单独作为 reject 理由的情形：
+    - 常规 nmap/curl/grep/读文件/枚举类命令本身
+    - 参数略显宽泛但无明确破坏意图
+    - 仅因「信息不足」——若无上述高危迹象，应 approve 并可在 comment 中提示注意点
+
+    仅输出一行 JSON，不要 markdown 代码块：
+    {"decision":"approve"|"reject","comment":"简要理由"}
+  audit_agent_prompt_review_edit: |-
+    你是 CyberStrikeAI 人机协同审计 Agent。审查 Agent 即将执行的工具调用是否会对系统造成实质性损害。
+
+    你会收到 JSON，包含 hitlMode、toolName、arguments/argumentsObj、userMessage、thinking、reasoningChain、planning 等字段。
+
+    裁决基调（默认放行）：
+    - 常规、低风险的渗透测试操作 → approve（如信息收集、端口/服务扫描、目录枚举、只读查询、无害探测命令）
+    - 与用户授权、当前任务目标一致，且未见明确高危迹象 → approve
+    - 仅在「可能对系统造成实质影响」时 → reject；参数可安全收窄时优先 approve + editedArguments
+
+    必须 reject 的高危情形（示例，非穷举）：
+    - 删库、清表、批量删除数据、格式化磁盘、不可逆破坏
+    - 修改/重置密码、创建或篡改管理员账号、持久化后门、开机自启
+    - 向生产环境写入恶意载荷、勒索加密、停止关键服务、修改系统核心配置
+    - 明显越权：与任务/授权目标无关的破坏性操作
+
+    不应单独作为 reject 理由的情形：
+    - 常规 nmap/curl/grep/读文件/枚举类命令本身
+    - 参数略显宽泛但无明确破坏意图（应收窄参数后 approve）
+    - 仅因「信息不足」——若无上述高危迹象，应 approve 并可在 comment 中提示注意点
+
+    仅输出一行 JSON，不要 markdown 代码块：
+    {"decision":"approve"|"reject","comment":"简要理由","editedArguments":{...}}
+
+    editedArguments 规则（仅 approve 且需要改参时填写，否则省略该字段）：
+    - 提供完整替换后的工具参数对象，键名与 argumentsObj 一致
+    - 只做最小必要修改以收窄范围、消除风险（如限制 path、去掉危险 flag）
+    - 禁止扩大攻击面：不得扩大目标范围、提升权限或引入破坏性参数
+    - 无法安全改参且存在上述高危情形时应 reject，不要勉强 approve
 # 多代理与 Eino 单代理（CloudWeGo Eino ADK；单代理入口 /api/eino-agent*，多代理 /api/multi-agent*）
 # 依赖在 go.mod 中拉取；若下载失败可设置: go env -w GOPROXY=https://goproxy.cn,direct
 # Deep / Plan-Execute / Supervisor 由对话页与 WebShell 所选模式在请求体 orchestration 中指定；机器人按 robot_default_agent_mode
@@ -114,7 +174,7 @@ multi_agent:
   batch_use_multi_agent: false # true 时「批量任务」队列中每个子任务也走 Eino 多代理（成本更高）
   # plan_execute 专用：execute↔replan 外层循环上限，0 表示 Eino 默认 10。主/子代理 ReAct 轮次见 agent.max_iterations。
   plan_execute_loop_max_iterations: 0
-  sub_agent_user_context_max_runes: 0 # 子代理 task 描述中自动注入用户原始请求的字符上限；0=默认2000，负数=禁用
+  sub_agent_user_context_max_runes: 0 # 子代理 task 描述中注入用户原文；0=不截断（默认），>0=总字符上限，负数=禁用
   without_general_sub_agent: false # false 时保留 Deep 内置 general-purpose 子代理
   without_write_todos: false
   orchestrator_instruction: "" # Deep 主代理：agents/orchestrator.md（或 kind: orchestrator 的单个 .md）正文优先；正文为空时用此处；皆空则 Eino 默认
@@ -125,7 +185,8 @@ multi_agent:
     disable: false # true：不注册 skill 渐进式披露中间件，也不挂本机 FS/Shell 工具；false：按下方开关加载
     filesystem_tools: true # true：注册 read_file/glob/grep/write/edit/execute（授权环境慎用）；false：仅 skill，不暴露本机读写与 Shell
     skill_tool_name: skill # 模型侧可调用的「加载技能」工具名，一般保持 skill；与技能包文档中的调用名一致即可
-  # Eino ADK 中间件与 Deep/Supervisor 调参（结构体见 internal/config/config.go → MultiAgentEinoMiddlewareConfig）
+  # Eino ADK 中间件与 Deep/Supervisor/plan_execute Executor 调参（结构体见 internal/config/config.go → MultiAgentEinoMiddlewareConfig）
+  # plan_execute：下列 patch/reduction/tool_search/plantask 等同样作用于 Executor（经 ExecPreMiddlewares）；Planner/Replanner 不挂 MCP 前置中间件。
   eino_middleware:
     patch_tool_calls: true # true：修补历史中无 tool_result 的悬空 tool_call（流式中断/重试后更稳）；false：关闭；字段省略时默认等同 true
     tool_search_enable: true # true：工具数 ≥ min 时启用 tool_search，仅前 N 个工具常驻，其余按正则按需解锁，省 token、减误选；false：全量工具进上下文
@@ -149,6 +210,7 @@ multi_agent:
     checkpoint_dir: data/eino-checkpoints # P0：进程崩溃/OOM 后同会话自动 ADK Resume；正常结束会删 .ckpt；与「中断并继续」(last_react_*) 是两套机制
     run_retry_max_attempts: 0 # 429/5xx/网络抖动时可退避重试次数（run loop + summarization 共用 isEinoTransientRunError）；0=默认 10
     run_retry_max_backoff_sec: 0 # 单次退避上限秒数；0=默认 30
+    empty_response_continue_max_attempts: 0 # Run 成功但未捕获助手正文（含流式中断）时 Handler 退避续跑次数；0=默认 5
     deep_output_key: final_answer # P0：Eino session 写入最终助手结论（框架内部；Deep/Supervisor 主/eino_single）
     deep_model_retry_max_retries: 0 # 已废弃，请用 run_retry_max_attempts；保留字段仅为兼容旧配置
     task_tool_description_prefix: "" # 非空：仅 Deep 的 task 工具使用自定义描述前缀，运行时会拼接子代理名称；空则走 Eino 默认生成逻辑
@@ -221,9 +283,17 @@ knowledge:
   retrieval:
     top_k: 5 # 检索返回的Top-K结果数量
     similarity_threshold: 0.4 # 余弦相似度阈值（0-1），低于此值的结果将被过滤
-    # 检索后处理：固定正文规范化去重；上下文预算；可选代码注入 DocumentReranker 做重排
+    # Eino MultiQuery：LLM 改写查询后多路向量检索再融合（始终启用）
+    multi_query:
+      max_queries: 4 # 改写变体上限（含语义覆盖）；建议 3~4
+    # 精排（始终启用）：dashscope 用 gte-rerank；其他 OpenAI 兼容端点走 /v1/rerank
+    rerank:
+      provider: "" # 空=按 base_url 推断：dashscope | cohere
+      model: "" # 空=dashscope→gte-rerank，cohere→rerank-multilingual-v3.0
+      base_url: "" # 留空则用 embedding / openai 的 base_url
+      api_key: "" # 留空则用 embedding / openai 的 api_key
     post_retrieve:
-      prefetch_top_k: 0 # 0 与 top_k 相同；可设为 15~30 以便去重后仍填满 top_k
+      prefetch_top_k: 20 # 每条 MultiQuery 变体的向量候选数；0=内置 max(top_k*4,20)
       max_context_chars: 0 # 0 不限制；否则返回的正文总 Unicode 字符上限（整段 chunk）
       max_context_tokens: 0 # 0 不限制；tiktoken 总 token 上限
     sub_index_filter: ""

docs/MULTI_AGENT_EINO.md

@@ -26,7 +26,7 @@
 | OpenAPI | 多代理路径说明已更新（流式未启用为 SSE 错误事件）。 |
 | 机器人 | `ProcessMessageForRobot` 按 `robot_default_agent_mode`（默认 `eino_single`）调用 `RunEinoSingleChatModelAgent` 或 `RunDeepAgent`。 |
 | 预置编排 | 聊天 / WebShell：`POST /api/multi-agent*` 请求体 `orchestration`：`deep` \| `plan_execute` \| `supervisor`（缺省 `deep`）。`plan_execute` 不构建 YAML/Markdown 子代理；`plan_execute_loop_max_iterations` 仍来自配置。`supervisor` 至少需一个子代理。 |
-| Eino 中间件 | `multi_agent.eino_middleware`（可选）：`patchtoolcalls`（默认开）、`toolsearch`（按阈值拆分 MCP 工具列表）、`plantask`（需 `eino_skills`）、`reduction`（大工具输出截断/落盘）、`checkpoint_dir`（Runner 断点）、`deep_output_key` / `deep_model_retry_max_retries` / `task_tool_description_prefix`（Deep 与 supervisor 主代理共享其中模型重试与 OutputKey）。`plan_execute` 的 Executor 无 Handlers：仅继承 **ToolsConfig** 侧效果（如 `tool_search` 列表拆分），不挂载 patch/plantask/reduction 中间件。 |
+| Eino 中间件 | `multi_agent.eino_middleware`（可选）：`patchtoolcalls`（默认开）、`toolsearch`（按阈值拆分 MCP 工具列表）、`plantask`（需 `eino_skills`）、`reduction`（大工具输出截断/落盘）、`checkpoint_dir`（Runner 断点）、`deep_output_key` / `deep_model_retry_max_retries` / `task_tool_description_prefix`（Deep 与 supervisor 主代理共享其中模型重试与 OutputKey）。**`plan_execute`**：`runner.go` 将 `prependEinoMiddlewares(einoMWMain)` 产物作为 `ExecPreMiddlewares` 挂到 **Executor**（与 Deep/Supervisor 主代理同序：patch → reduction → toolsearch → plantask → filesystem → skill → summarization tail）；Planner/Replanner 仅 summarization tail + prompt 预算截断，不跑 MCP 工具链。 |
 
 ## 进行中 / 待办（ backlog ）
 
@@ -37,7 +37,8 @@
 
 ## 关键文件索引
 
-- `internal/multiagent/runner.go` — DeepAgent 组装与事件循环  
+- `internal/multiagent/runner.go` — DeepAgent / plan_execute / supervisor 组装与事件循环  
+- `internal/multiagent/eino_orchestration.go` — PlanExecute 根节点与 Executor 中间件栈（`buildPlanExecuteExecutorHandlers`）  
 - `internal/handler/multi_agent.go` — SSE 与（同步）HTTP  
 - `internal/handler/multi_agent_prepare.go` — 会话准备（含 WebShell）  
 - `internal/einomcp/` — MCP → Eino Tool  
@@ -59,4 +60,5 @@
 | 2026-03-22 | `orchestrator.md` / `kind: orchestrator` 主代理、列表主/子标记、与 `orchestrator_instruction` 优先级。 |
 | 2026-04-19 | 主聊天「对话模式」：原生 ReAct 与 Deep / Plan-Execute / Supervisor；`POST /api/multi-agent*` 请求体 `orchestration` 与界面一致；`config.yaml` / 设置页不再维护预置编排字段（机器人/批量默认 `deep`）。 |
 | 2026-04-21 | 移除角色 `skills` 与 `/api/roles/skills/list`；`bind_role` 仅继承 tools；Skills 仅通过 Eino `skill` 工具按需加载。 |
+| 2026-07-02 | **plan_execute Executor 中间件对齐**：`ExecPreMiddlewares` 与 Deep 主代理同源；`buildPlanExecuteExecutorHandlers` + 回归测试；文档更正。 |
 | 2026-06-02 | **移除原生 ReAct**：删除 `/api/agent-loop*` 执行入口与 `AgentLoopWithProgress`；统一 Eino ADK（单代理 `/api/eino-agent*`，多代理 `/api/multi-agent*`）；任务 cancel/tasks API 保留。 |

internal/app/app.go

@@ -21,11 +21,13 @@ import (
 	"cyberstrike-ai/internal/database"
 	"cyberstrike-ai/internal/einoobserve"
 	"cyberstrike-ai/internal/handler"
+	"cyberstrike-ai/internal/hitl"
 	"cyberstrike-ai/internal/knowledge"
 	"cyberstrike-ai/internal/logger"
 	"cyberstrike-ai/internal/mcp"
 	"cyberstrike-ai/internal/mcp/builtin"
 	"cyberstrike-ai/internal/monitor"
+	"cyberstrike-ai/internal/multiagent"
 	"cyberstrike-ai/internal/robot"
 	"cyberstrike-ai/internal/security"
 	"cyberstrike-ai/internal/skillpackage"
@@ -67,6 +69,10 @@ type App struct {
 
 // New 创建新应用
 func New(cfg *config.Config, log *logger.Logger, configPath string) (*App, error) {
+	if err := multiagent.InitADK(); err != nil {
+		return nil, fmt.Errorf("初始化 Eino ADK: %w", err)
+	}
+
 	gin.SetMode(gin.ReleaseMode)
 	router := gin.Default()
 
@@ -104,6 +110,10 @@ func New(cfg *config.Config, log *logger.Logger, configPath string) (*App, error
 	monitorRetention.PurgeExpired()
 	monitor.StartRetentionLoop(monitorRetention, log.Logger)
 
+	hitlRetention := hitl.NewService(db, cfg, log.Logger)
+	hitlRetention.PurgeExpired()
+	hitl.StartRetentionLoop(hitlRetention, log.Logger)
+
 	// 创建MCP服务器（带数据库持久化）
 	mcpServer := mcp.NewServerWithStorage(log.Logger, db)
 	mcpServer.ConfigureHTTPToolCallTimeoutFromAgentMinutes(cfg.Agent.ToolTimeoutMinutes)
@@ -135,6 +145,10 @@ func New(cfg *config.Config, log *logger.Logger, configPath string) (*App, error
 		externalMCPMgr.StartAllEnabled()
 	}
 
+	execReconciler := monitor.NewExecutionReconciler(db, mcpServer, externalMCPMgr, log.Logger)
+	execReconciler.ReconcileOnStartup()
+	monitor.StartStaleRunningReconcileLoop(execReconciler, log.Logger)
+
 	// 创建Agent
 	maxIterations := cfg.Agent.MaxIterations
 	if maxIterations <= 0 {
@@ -193,14 +207,12 @@ func New(cfg *config.Config, log *logger.Logger, configPath string) (*App, error
 			return nil, fmt.Errorf("初始化知识库嵌入器失败: %w", err)
 		}
 
-		// 创建检索器
+		// 创建检索器（Eino MultiQuery + 重排流水线）
-		retrievalConfig := &knowledge.RetrievalConfig{
+		retrievalConfig := knowledge.RetrievalConfigFromYAML(cfg.Knowledge.Retrieval)
-			TopK:                cfg.Knowledge.Retrieval.TopK,
-			SimilarityThreshold: cfg.Knowledge.Retrieval.SimilarityThreshold,
-			SubIndexFilter:      cfg.Knowledge.Retrieval.SubIndexFilter,
-			PostRetrieve:        cfg.Knowledge.Retrieval.PostRetrieve,
-		}
 		knowledgeRetriever = knowledge.NewRetriever(knowledgeDB, embedder, retrievalConfig, log.Logger)
+		if err := knowledge.WireRetrieverPipeline(context.Background(), knowledgeRetriever, &cfg.OpenAI); err != nil {
+			return nil, fmt.Errorf("初始化知识库检索流水线失败: %w", err)
+		}
 
 		// 创建索引器（Eino Compose 链）
 		knowledgeIndexer, err = knowledge.NewIndexer(context.Background(), knowledgeDB, embedder, log.Logger, &cfg.Knowledge)
@@ -354,6 +366,7 @@ func New(cfg *config.Config, log *logger.Logger, configPath string) (*App, error
 	configHandler := handler.NewConfigHandler(configPath, cfg, mcpServer, executor, agent, attackChainHandler, externalMCPMgr, log.Logger)
 	configHandler.SetAudit(auditSvc)
 	agentHandler.SetHitlToolWhitelistSaver(configHandler)
+	agentHandler.SetHitlAuditStrategySaver(configHandler)
 	externalMCPHandler := handler.NewExternalMCPHandler(externalMCPMgr, cfg, configPath, log.Logger)
 	externalMCPHandler.SetAudit(auditSvc)
 	roleHandler := handler.NewRoleHandler(cfg, configPath, log.Logger)
@@ -803,11 +816,18 @@ func setupRoutes(
 		protected.POST("/eino-agent", agentHandler.EinoSingleAgentLoop)
 		protected.POST("/eino-agent/stream", agentHandler.EinoSingleAgentLoopStream)
 		protected.GET("/hitl/pending", agentHandler.ListHITLPending)
+		protected.GET("/hitl/logs", agentHandler.ListHITLLogs)
+		protected.DELETE("/hitl/logs", agentHandler.DeleteHITLLogs)
+		protected.GET("/hitl/logs/:id", agentHandler.GetHITLLog)
 		protected.POST("/hitl/decision", agentHandler.DecideHITLInterrupt)
 		protected.POST("/hitl/dismiss", agentHandler.DismissHITLInterrupt)
 		protected.GET("/hitl/config/:conversationId", agentHandler.GetHITLConversationConfig)
 		protected.PUT("/hitl/config", agentHandler.UpsertHITLConversationConfig)
+		protected.GET("/hitl/tool-whitelist", agentHandler.GetHITLGlobalToolWhitelist)
+		protected.PUT("/hitl/tool-whitelist", agentHandler.SetHITLGlobalToolWhitelist)
 		protected.POST("/hitl/tool-whitelist", agentHandler.MergeHITLGlobalToolWhitelist)
+		protected.GET("/hitl/audit-strategy", agentHandler.GetHITLAuditStrategy)
+		protected.PUT("/hitl/audit-strategy", agentHandler.UpdateHITLAuditStrategy)
 		// Agent Loop 取消与任务列表
 		protected.POST("/agent-loop/cancel", agentHandler.CancelAgentLoop)
 		protected.GET("/agent-loop/tasks", agentHandler.ListAgentTasks)
@@ -1778,14 +1798,12 @@ func initializeKnowledge(
 		return nil, fmt.Errorf("初始化知识库嵌入器失败: %w", err)
 	}
 
-	// 创建检索器
+	// 创建检索器（Eino MultiQuery + 重排流水线）
-	retrievalConfig := &knowledge.RetrievalConfig{
+	retrievalConfig := knowledge.RetrievalConfigFromYAML(cfg.Knowledge.Retrieval)
-		TopK:                cfg.Knowledge.Retrieval.TopK,
-		SimilarityThreshold: cfg.Knowledge.Retrieval.SimilarityThreshold,
-		SubIndexFilter:      cfg.Knowledge.Retrieval.SubIndexFilter,
-		PostRetrieve:        cfg.Knowledge.Retrieval.PostRetrieve,
-	}
 	knowledgeRetriever := knowledge.NewRetriever(knowledgeDB, embedder, retrievalConfig, logger)
+	if err := knowledge.WireRetrieverPipeline(context.Background(), knowledgeRetriever, &cfg.OpenAI); err != nil {
+		return nil, fmt.Errorf("初始化知识库检索流水线失败: %w", err)
+	}
 
 	// 创建索引器（Eino Compose 链）
 	knowledgeIndexer, err := knowledge.NewIndexer(context.Background(), knowledgeDB, embedder, logger, &cfg.Knowledge)

internal/config/config.go

@@ -96,8 +96,8 @@ type MultiAgentConfig struct {
 	// OrchestratorInstructionSupervisor supervisor 主代理系统提示（transfer/exit 说明仍由运行追加）；非空且 agents/orchestrator-supervisor.md 正文为空或未存在时生效。
 	OrchestratorInstructionSupervisor string                `yaml:"orchestrator_instruction_supervisor,omitempty" json:"orchestrator_instruction_supervisor,omitempty"`
 	SubAgents                         []MultiAgentSubConfig `yaml:"sub_agents" json:"sub_agents"`
-	// SubAgentUserContextMaxRunes caps the user-context supplement appended to task descriptions for sub-agents.
+	// SubAgentUserContextMaxRunes caps user-context supplement for sub-agent task descriptions.
-	// 0 (default) uses the built-in default of 2000 runes; negative value disables injection entirely.
+	// 0 (default) preserves all user turns verbatim; >0 caps total runes; negative disables injection.
 	SubAgentUserContextMaxRunes int `yaml:"sub_agent_user_context_max_runes,omitempty" json:"sub_agent_user_context_max_runes,omitempty"`
 	// EinoSkills configures CloudWeGo Eino ADK skill middleware + optional local filesystem/execute on DeepAgent.
 	EinoSkills MultiAgentEinoSkillsConfig `yaml:"eino_skills,omitempty" json:"eino_skills,omitempty"`
@@ -107,6 +107,11 @@ type MultiAgentConfig struct {
 	EinoCallbacks MultiAgentEinoCallbacksConfig `yaml:"eino_callbacks,omitempty" json:"eino_callbacks,omitempty"`
 }
 
+// SubAgentUserContextMaxRunesEffective returns max runes for sub-agent task supplement; 0 = unlimited; negative = disabled.
+func (c MultiAgentConfig) SubAgentUserContextMaxRunesEffective() int {
+	return c.SubAgentUserContextMaxRunes
+}
+
 // MultiAgentEinoCallbacksConfig enables Eino unified callbacks on each ADK agent run (deep / plan_execute / supervisor / eino_single).
 // Modes: log_only (zap + optional OTel; no SSE to browser), sse (adds client SSE eino_trace_* when sse_trace_to_client), full (sse rules + stream callback copies closed).
 type MultiAgentEinoCallbacksConfig struct {
@@ -270,6 +275,8 @@ type MultiAgentEinoMiddlewareConfig struct {
 	RunRetryMaxAttempts int `yaml:"run_retry_max_attempts,omitempty" json:"run_retry_max_attempts,omitempty"`
 	// RunRetryMaxBackoffSec 单次退避上限秒数；0=默认 30。
 	RunRetryMaxBackoffSec int `yaml:"run_retry_max_backoff_sec,omitempty" json:"run_retry_max_backoff_sec,omitempty"`
+	// EmptyResponseContinueMaxAttempts Run 成功但未捕获助手正文时 Handler 层退避续跑次数；0=默认 5。
+	EmptyResponseContinueMaxAttempts int `yaml:"empty_response_continue_max_attempts,omitempty" json:"empty_response_continue_max_attempts,omitempty"`
 	// TaskToolDescriptionPrefix when non-empty sets deep.Config TaskToolDescriptionGenerator (sub-agent names appended).
 	TaskToolDescriptionPrefix string `yaml:"task_tool_description_prefix,omitempty" json:"task_tool_description_prefix,omitempty"`
 }
@@ -490,6 +497,17 @@ type RobotWecomConfig struct {
 	AgentID        int64  `yaml:"agent_id" json:"agent_id"`                 // 应用 AgentId
 }
 
+// ValidateWecomConfig 校验企业微信机器人配置；启用时必须配置 token，否则回调无法防伪造。
+func ValidateWecomConfig(w RobotWecomConfig) error {
+	if !w.Enabled {
+		return nil
+	}
+	if strings.TrimSpace(w.Token) == "" {
+		return fmt.Errorf("robots.wecom.enabled 为 true 时必须配置 robots.wecom.token")
+	}
+	return nil
+}
+
 // RobotDingtalkConfig 钉钉机器人配置
 type RobotDingtalkConfig struct {
 	Enabled                    bool   `yaml:"enabled" json:"enabled"`
@@ -614,10 +632,100 @@ type AgentConfig struct {
 }
 
 // HitlConfig 人机协同全局选项；与会话侧栏/API 中的白名单合并为并集后参与判定。
-// tool_whitelist 可在侧栏「应用」时合并写入 config.yaml 并立即生效；其他字段若仅改文件仍需重启。
+// tool_whitelist 可在侧栏「应用」时合并写入 config.yaml 并立即生效。
+// audit_agent_prompt / audit_agent_prompt_review_edit 可在人机协同页编辑并立即生效；空则使用内置默认。
 type HitlConfig struct {
-	// ToolWhitelist 全局免审批工具名（与每条会话配置的 sensitiveTools 语义相同：白名单内工具不触发 HITL）。
+	// ToolWhitelist 全局免审批工具名（与白名单内工具不触发 HITL 审批）。
 	ToolWhitelist []string `yaml:"tool_whitelist,omitempty" json:"tool_whitelist,omitempty"`
+	// AuditAgentPrompt 审批模式（approval）下审计 Agent 系统提示词。
+	AuditAgentPrompt string `yaml:"audit_agent_prompt,omitempty" json:"audit_agent_prompt,omitempty"`
+	// AuditAgentPromptReviewEdit 审查编辑模式（review_edit）下审计 Agent 系统提示词。
+	AuditAgentPromptReviewEdit string `yaml:"audit_agent_prompt_review_edit,omitempty" json:"audit_agent_prompt_review_edit,omitempty"`
+	// RetentionDays 已决策审计日志（hitl_interrupts 非 pending）保留天数；省略时默认 90；0 表示不自动清理。
+	RetentionDays *int `yaml:"retention_days,omitempty" json:"retention_days,omitempty"`
+}
+
+// RetentionDaysEffective returns retention; 0 means keep forever; omitted defaults to 90.
+func (h HitlConfig) RetentionDaysEffective() int {
+	if h.RetentionDays == nil {
+		return 90
+	}
+	if *h.RetentionDays < 0 {
+		return 0
+	}
+	return *h.RetentionDays
+}
+
+const hitlAuditAgentPromptBase = `你是 CyberStrikeAI 人机协同审计 Agent。审查 Agent 即将执行的工具调用是否会对系统造成实质性损害。
+
+你会收到 JSON，包含 hitlMode、toolName、arguments/argumentsObj、userMessage、thinking、reasoningChain、planning 等字段。
+
+裁决基调（默认放行）：
+- 常规、低风险的渗透测试操作 → approve（如信息收集、端口/服务扫描、目录枚举、只读查询、无害探测命令）
+- 与用户授权、当前任务目标一致，且未见明确高危迹象 → approve
+- 仅在「可能对系统造成实质影响」时 → reject
+
+必须 reject 的高危情形（示例，非穷举）：
+- 删库、清表、批量删除数据、格式化磁盘、不可逆破坏
+- 修改/重置密码、创建或篡改管理员账号、持久化后门、开机自启
+- 向生产环境写入恶意载荷、勒索加密、停止关键服务、修改系统核心配置
+- 明显越权：与任务/授权目标无关的破坏性操作
+
+不应单独作为 reject 理由的情形：
+- 常规 nmap/curl/grep/读文件/枚举类命令本身
+- 参数略显宽泛但无明确破坏意图（审查编辑模式可收窄参数后 approve）
+- 仅因「信息不足」——若无上述高危迹象，应 approve 并可在 comment 中提示注意点`
+
+const hitlAuditAgentPromptApprovalOutput = `
+仅输出一行 JSON，不要 markdown 代码块：
+{"decision":"approve"|"reject","comment":"简要理由"}`
+
+const hitlAuditAgentPromptReviewEditOutput = `
+仅输出一行 JSON，不要 markdown 代码块：
+{"decision":"approve"|"reject","comment":"简要理由","editedArguments":{...}}
+
+editedArguments 规则（仅 approve 且需要改参时填写，否则省略该字段）：
+- 提供完整替换后的工具参数对象，键名与 argumentsObj 一致
+- 只做最小必要修改以收窄范围、消除风险（如限制 path、去掉危险 flag）
+- 禁止扩大攻击面：不得扩大目标范围、提升权限或引入破坏性参数
+- 无法安全改参时应 reject，不要勉强 approve`
+
+// DefaultHitlAuditAgentPrompt 内置审批模式审计 Agent 提示词。
+func DefaultHitlAuditAgentPrompt() string {
+	return hitlAuditAgentPromptBase + hitlAuditAgentPromptApprovalOutput
+}
+
+// DefaultHitlAuditAgentPromptReviewEdit 内置审查编辑模式审计 Agent 提示词。
+func DefaultHitlAuditAgentPromptReviewEdit() string {
+	return hitlAuditAgentPromptBase + hitlAuditAgentPromptReviewEditOutput
+}
+
+// EffectiveAuditAgentPrompt 返回审批模式生效的审计 Agent 提示词。
+func (c HitlConfig) EffectiveAuditAgentPrompt() string {
+	return c.EffectiveAuditAgentPromptForMode("approval")
+}
+
+// EffectiveAuditAgentPromptForMode 按 HITL 模式返回生效的审计 Agent 提示词。
+func (c HitlConfig) EffectiveAuditAgentPromptForMode(mode string) string {
+	if normalizeHitlModeForPrompt(mode) == "review_edit" {
+		if s := strings.TrimSpace(c.AuditAgentPromptReviewEdit); s != "" {
+			return s
+		}
+		return DefaultHitlAuditAgentPromptReviewEdit()
+	}
+	if s := strings.TrimSpace(c.AuditAgentPrompt); s != "" {
+		return s
+	}
+	return DefaultHitlAuditAgentPrompt()
+}
+
+func normalizeHitlModeForPrompt(mode string) string {
+	switch strings.ToLower(strings.TrimSpace(mode)) {
+	case "review_edit":
+		return "review_edit"
+	default:
+		return "approval"
+	}
 }
 
 type AuthConfig struct {
@@ -804,33 +912,13 @@ func Load(path string) (*Config, error) {
 
 	// 如果配置了工具目录，从目录加载工具配置
 	if cfg.Security.ToolsDir != "" {
-		configDir := filepath.Dir(path)
+		inlineTools := append([]ToolConfig(nil), cfg.Security.Tools...)
-		toolsDir := cfg.Security.ToolsDir
+		toolsDir := ResolveToolsDir(cfg.Security.ToolsDir, path)
-
+		merged, err := MergeToolsFromDir(toolsDir, inlineTools)
-		// 如果是相对路径，相对于配置文件所在目录
-		if !filepath.IsAbs(toolsDir) {
-			toolsDir = filepath.Join(configDir, toolsDir)
-		}
-
-		tools, err := LoadToolsFromDir(toolsDir)
 		if err != nil {
 			return nil, fmt.Errorf("从工具目录加载工具配置失败: %w", err)
 		}
-
+		cfg.Security.Tools = merged
-		// 合并工具配置：目录中的工具优先，主配置中的工具作为补充
-		existingTools := make(map[string]bool)
-		for _, tool := range tools {
-			existingTools[tool.Name] = true
-		}
-
-		// 添加主配置中不存在于目录中的工具（向后兼容）
-		for _, tool := range cfg.Security.Tools {
-			if !existingTools[tool.Name] {
-				tools = append(tools, tool)
-			}
-		}
-
-		cfg.Security.Tools = tools
 	}
 
 	// 外部 MCP：迁移 + 环境变量展开
@@ -874,6 +962,10 @@ func Load(path string) (*Config, error) {
 		}
 	}
 
+	if err := ValidateWecomConfig(cfg.Robots.Wecom); err != nil {
+		return nil, err
+	}
+
 	return &cfg, nil
 }
 
@@ -1098,6 +1190,75 @@ func PrintMCPConfigJSON(mcp MCPConfig) {
 	fmt.Println("----------------------------------------------------------------")
 }
 
+// ResolveToolsDir 将 tools_dir 解析为绝对路径（相对路径相对于 configPath 所在目录）。
+func ResolveToolsDir(toolsDir, configPath string) string {
+	toolsDir = strings.TrimSpace(toolsDir)
+	if toolsDir == "" {
+		return ""
+	}
+	if filepath.IsAbs(toolsDir) {
+		return toolsDir
+	}
+	return filepath.Join(filepath.Dir(configPath), toolsDir)
+}
+
+// MergeToolsFromDir 从目录加载工具并与 inline 列表合并：目录中的工具优先，主配置中的工具作为补充。
+func MergeToolsFromDir(toolsDir string, inlineTools []ToolConfig) ([]ToolConfig, error) {
+	dirTools, err := LoadToolsFromDir(toolsDir)
+	if err != nil {
+		return nil, err
+	}
+	existing := make(map[string]bool, len(dirTools))
+	for _, tool := range dirTools {
+		existing[tool.Name] = true
+	}
+	merged := append([]ToolConfig(nil), dirTools...)
+	for _, tool := range inlineTools {
+		if !existing[tool.Name] {
+			merged = append(merged, tool)
+		}
+	}
+	return merged, nil
+}
+
+// loadInlineSecurityToolsFromYAML 读取 config.yaml 中 security.tools（不含 tools_dir 扫描结果）。
+func loadInlineSecurityToolsFromYAML(configPath string) ([]ToolConfig, error) {
+	data, err := os.ReadFile(configPath)
+	if err != nil {
+		return nil, fmt.Errorf("读取配置文件失败: %w", err)
+	}
+	var partial struct {
+		Security struct {
+			Tools []ToolConfig `yaml:"tools"`
+		} `yaml:"security"`
+	}
+	if err := yaml.Unmarshal(data, &partial); err != nil {
+		return nil, fmt.Errorf("解析配置文件失败: %w", err)
+	}
+	if partial.Security.Tools == nil {
+		return []ToolConfig{}, nil
+	}
+	return partial.Security.Tools, nil
+}
+
+// ReloadSecurityToolsFromDir 从 tools_dir 重新加载工具并更新 cfg.Security.Tools（ApplyConfig 热重载用）。
+func ReloadSecurityToolsFromDir(cfg *Config, configPath string) error {
+	if cfg == nil || strings.TrimSpace(cfg.Security.ToolsDir) == "" {
+		return nil
+	}
+	inlineTools, err := loadInlineSecurityToolsFromYAML(configPath)
+	if err != nil {
+		return err
+	}
+	toolsDir := ResolveToolsDir(cfg.Security.ToolsDir, configPath)
+	merged, err := MergeToolsFromDir(toolsDir, inlineTools)
+	if err != nil {
+		return fmt.Errorf("从工具目录加载工具配置失败: %w", err)
+	}
+	cfg.Security.Tools = merged
+	return nil
+}
+
 // LoadToolsFromDir 从目录加载所有工具配置文件
 func LoadToolsFromDir(dir string) ([]ToolConfig, error) {
 	var tools []ToolConfig
@@ -1316,7 +1477,12 @@ func Default() *Config {
 			},
 			Retrieval: RetrievalConfig{
 				TopK:                5,
-				SimilarityThreshold: 0.65, // 降低阈值到 0.65，减少漏检
+				SimilarityThreshold: 0.65,
+				MultiQuery:          MultiQueryConfig{MaxQueries: 4},
+				Rerank:              RerankConfig{},
+				PostRetrieve: PostRetrieveConfig{
+					PrefetchTopK: 20,
+				},
 			},
 			Indexing: IndexingConfig{
 				ChunkStrategy:         "markdown_then_recursive",
@@ -1412,7 +1578,7 @@ type EmbeddingConfig struct {
 
 // PostRetrieveConfig 检索后处理：固定对正文做规范化去重（最佳实践）、上下文预算截断；PrefetchTopK 用于多取候选再收敛到 top_k。
 type PostRetrieveConfig struct {
-	// PrefetchTopK 向量检索阶段最多保留的候选数（余弦序），应 ≥ top_k，0 表示与 top_k 相同；上限见知识库包内常量。
+	// PrefetchTopK 向量检索阶段每条 MultiQuery 变体最多保留的候选数；0 表示使用内置默认 max(top_k*4, 20)。
 	PrefetchTopK int `yaml:"prefetch_top_k,omitempty" json:"prefetch_top_k,omitempty"`
 	// MaxContextChars 返回文档内容总 Unicode 字符数上限（整段 chunk，不截断半段）；0 表示不限制。
 	MaxContextChars int `yaml:"max_context_chars,omitempty" json:"max_context_chars,omitempty"`
@@ -1420,13 +1586,62 @@ type PostRetrieveConfig struct {
 	MaxContextTokens int `yaml:"max_context_tokens,omitempty" json:"max_context_tokens,omitempty"`
 }
 
+// MultiQueryConfig Eino MultiQuery 查询改写（始终启用，无关闭开关）。
+type MultiQueryConfig struct {
+	// MaxQueries LLM 生成的检索变体上限（含原问语义覆盖）；0 表示默认 4。
+	MaxQueries int `yaml:"max_queries,omitempty" json:"max_queries,omitempty"`
+}
+
+func (c MultiQueryConfig) MaxQueriesEffective() int {
+	if c.MaxQueries <= 0 {
+		return 4
+	}
+	if c.MaxQueries > 8 {
+		return 8
+	}
+	return c.MaxQueries
+}
+
+// RerankConfig 检索精排（始终启用）；支持 dashscope 与 Cohere 兼容 HTTP API。
+type RerankConfig struct {
+	// Provider: dashscope | cohere；空则按 base_url 自动推断。
+	Provider string `yaml:"provider,omitempty" json:"provider,omitempty"`
+	Model    string `yaml:"model,omitempty" json:"model,omitempty"`
+	BaseURL  string `yaml:"base_url,omitempty" json:"base_url,omitempty"`
+	APIKey   string `yaml:"api_key,omitempty" json:"api_key,omitempty"`
+}
+
+func (c RerankConfig) ProviderEffective(baseURL string) string {
+	p := strings.TrimSpace(strings.ToLower(c.Provider))
+	if p != "" {
+		return p
+	}
+	u := strings.ToLower(baseURL)
+	if strings.Contains(u, "dashscope") {
+		return "dashscope"
+	}
+	return "cohere"
+}
+
+func (c RerankConfig) ModelEffective(provider string) string {
+	if m := strings.TrimSpace(c.Model); m != "" {
+		return m
+	}
+	if provider == "dashscope" {
+		return "gte-rerank"
+	}
+	return "rerank-multilingual-v3.0"
+}
+
 // RetrievalConfig 检索配置
 type RetrievalConfig struct {
 	TopK                int     `yaml:"top_k" json:"top_k"`                               // 检索Top-K
 	SimilarityThreshold float64 `yaml:"similarity_threshold" json:"similarity_threshold"` // 余弦相似度阈值
 	// SubIndexFilter 非空时仅保留 sub_indexes 含该标签（逗号分隔之一）的行；sub_indexes 为空的旧行仍返回。
 	SubIndexFilter string `yaml:"sub_index_filter,omitempty" json:"sub_index_filter,omitempty"`
-	// PostRetrieve 检索后处理（去重、预算截断）；重排通过代码注入 [knowledge.DocumentReranker]。
+	MultiQuery     MultiQueryConfig `yaml:"multi_query" json:"multi_query"`
+	Rerank         RerankConfig     `yaml:"rerank" json:"rerank"`
+	// PostRetrieve 检索后处理（去重、预算截断）；精排在 MultiQuery 融合后执行。
 	PostRetrieve PostRetrieveConfig `yaml:"post_retrieve,omitempty" json:"post_retrieve,omitempty"`
 }
 

internal/config/robots_validate_test.go

@@ -0,0 +1,45 @@
+package config
+
+import "testing"
+
+func TestValidateWecomConfig(t *testing.T) {
+	t.Parallel()
+
+	tests := []struct {
+		name    string
+		cfg     RobotWecomConfig
+		wantErr bool
+	}{
+		{
+			name:    "disabled without token",
+			cfg:     RobotWecomConfig{Enabled: false, Token: ""},
+			wantErr: false,
+		},
+		{
+			name:    "enabled with token",
+			cfg:     RobotWecomConfig{Enabled: true, Token: "secret"},
+			wantErr: false,
+		},
+		{
+			name:    "enabled without token",
+			cfg:     RobotWecomConfig{Enabled: true, Token: ""},
+			wantErr: true,
+		},
+		{
+			name:    "enabled with whitespace token",
+			cfg:     RobotWecomConfig{Enabled: true, Token: "   "},
+			wantErr: true,
+		},
+	}
+
+	for _, tt := range tests {
+		tt := tt
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			err := ValidateWecomConfig(tt.cfg)
+			if (err != nil) != tt.wantErr {
+				t.Fatalf("ValidateWecomConfig() error = %v, wantErr %v", err, tt.wantErr)
+			}
+		})
+	}
+}

internal/config/tools_reload_test.go

@@ -0,0 +1,111 @@
+package config
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+func TestReloadSecurityToolsFromDir(t *testing.T) {
+	root := t.TempDir()
+	toolsDir := filepath.Join(root, "tools")
+	if err := os.MkdirAll(toolsDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+
+	configPath := filepath.Join(root, "config.yaml")
+	if err := os.WriteFile(configPath, []byte(`security:
+  tools_dir: tools
+  tools:
+    - name: inline-only
+      command: inline-cmd
+      enabled: true
+      description: inline tool
+`), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	writeTool := func(name, command string) {
+		t.Helper()
+		content := "name: " + name + "\ncommand: " + command + "\nenabled: true\ndescription: test\n"
+		if err := os.WriteFile(filepath.Join(toolsDir, name+".yaml"), []byte(content), 0644); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	writeTool("alpha", "alpha-cmd")
+
+	cfg := &Config{
+		Security: SecurityConfig{
+			ToolsDir: "tools",
+			Tools: []ToolConfig{
+				{Name: "stale", Command: "stale-cmd", Enabled: true, Description: "should be removed"},
+			},
+		},
+	}
+
+	if err := ReloadSecurityToolsFromDir(cfg, configPath); err != nil {
+		t.Fatalf("reload: %v", err)
+	}
+	if len(cfg.Security.Tools) != 2 {
+		t.Fatalf("expected 2 tools, got %d", len(cfg.Security.Tools))
+	}
+
+	names := map[string]string{}
+	for _, tool := range cfg.Security.Tools {
+		names[tool.Name] = tool.Command
+	}
+	if names["alpha"] != "alpha-cmd" {
+		t.Fatalf("alpha tool missing or wrong command: %#v", names)
+	}
+	if names["inline-only"] != "inline-cmd" {
+		t.Fatalf("inline-only tool missing: %#v", names)
+	}
+	if _, ok := names["stale"]; ok {
+		t.Fatal("stale in-memory tool should not survive reload")
+	}
+
+	writeTool("beta", "beta-cmd")
+	if err := ReloadSecurityToolsFromDir(cfg, configPath); err != nil {
+		t.Fatalf("second reload: %v", err)
+	}
+	if len(cfg.Security.Tools) != 3 {
+		t.Fatalf("expected 3 tools after add, got %d", len(cfg.Security.Tools))
+	}
+	foundBeta := false
+	for _, tool := range cfg.Security.Tools {
+		if tool.Name == "beta" {
+			foundBeta = true
+			break
+		}
+	}
+	if !foundBeta {
+		t.Fatal("beta tool not found after second reload")
+	}
+}
+
+func TestMergeToolsFromDir_DirOverridesInline(t *testing.T) {
+	root := t.TempDir()
+	toolsDir := filepath.Join(root, "tools")
+	if err := os.MkdirAll(toolsDir, 0755); err != nil {
+		t.Fatal(err)
+	}
+	content := "name: shared\ncommand: dir-cmd\nenabled: true\ndescription: from dir\n"
+	if err := os.WriteFile(filepath.Join(toolsDir, "shared.yaml"), []byte(content), 0644); err != nil {
+		t.Fatal(err)
+	}
+
+	inline := []ToolConfig{
+		{Name: "shared", Command: "inline-cmd", Enabled: true, Description: "from inline"},
+	}
+	merged, err := MergeToolsFromDir(toolsDir, inline)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(merged) != 1 {
+		t.Fatalf("expected 1 tool, got %d", len(merged))
+	}
+	if merged[0].Command != "dir-cmd" {
+		t.Fatalf("dir tool should win, got command %q", merged[0].Command)
+	}
+}

internal/database/conversation.go

@@ -3,6 +3,7 @@ package database
 import (
 	"database/sql"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"os"
 	"path/filepath"
@@ -13,6 +14,9 @@ import (
 	"go.uber.org/zap"
 )
 
+// ProjectFilterUnbound 列表 API 中 project_id=__none__ 表示仅未绑定项目的对话。
+const ProjectFilterUnbound = "__none__"
+
 // Conversation 对话
 type Conversation struct {
 	ID        string    `json:"id"`
@@ -361,20 +365,44 @@ func (db *DB) GetConversationLite(id string) (*Conversation, error) {
 	return &conv, nil
 }
 
+func conversationProjectIDColumn(alias string) string {
+	if alias != "" {
+		return alias + ".project_id"
+	}
+	return "project_id"
+}
+
+func appendConversationProjectFilter(where string, args []interface{}, projectID, alias string) (string, []interface{}) {
+	pid := strings.TrimSpace(projectID)
+	if pid == "" {
+		return where, args
+	}
+	col := conversationProjectIDColumn(alias)
+	if pid == ProjectFilterUnbound {
+		return where + fmt.Sprintf(" AND (%s IS NULL OR TRIM(COALESCE(%s, '')) = '')", col, col), args
+	}
+	return where + fmt.Sprintf(" AND %s = ?", col), append(args, pid)
+}
+
 // CountConversations 统计对话数量。
-func (db *DB) CountConversations(search string) (int, error) {
+func (db *DB) CountConversations(search, projectID string) (int, error) {
 	var count int
 	var err error
 	if search != "" {
 		searchPattern := "%" + search + "%"
-		err = db.QueryRow(
+		where := ` WHERE (c.title LIKE ?
-			`SELECT COUNT(*) FROM conversations c
+			    OR EXISTS (SELECT 1 FROM messages m WHERE m.conversation_id = c.id AND m.content LIKE ?))`
-			 WHERE c.title LIKE ?
+		args := []interface{}{searchPattern, searchPattern}
-			    OR EXISTS (SELECT 1 FROM messages m WHERE m.conversation_id = c.id AND m.content LIKE ?)`,
+		where, args = appendConversationProjectFilter(where, args, projectID, "c")
-			searchPattern, searchPattern,
+		err = db.QueryRow(`SELECT COUNT(*) FROM conversations c`+where, args...).Scan(&count)
-		).Scan(&count)
 	} else {
-		err = db.QueryRow(`SELECT COUNT(*) FROM conversations`).Scan(&count)
+		where := ""
+		args := []interface{}{}
+		where, args = appendConversationProjectFilter(where, args, projectID, "")
+		if where != "" {
+			where = " WHERE" + strings.TrimPrefix(where, " AND")
+		}
+		err = db.QueryRow(`SELECT COUNT(*) FROM conversations`+where, args...).Scan(&count)
 	}
 	if err != nil {
 		return 0, fmt.Errorf("统计对话失败: %w", err)
@@ -395,7 +423,7 @@ func conversationOrderClause(sortBy, tableAlias string) string {
 }
 
 // ListConversations 列出所有对话
-func (db *DB) ListConversations(limit, offset int, search, sortBy string) ([]*Conversation, error) {
+func (db *DB) ListConversations(limit, offset int, search, sortBy, projectID string) ([]*Conversation, error) {
 	var rows *sql.Rows
 	var err error
 
@@ -403,20 +431,30 @@ func (db *DB) ListConversations(limit, offset int, search, sortBy string) ([]*Co
 		// 使用 EXISTS 子查询代替 LEFT JOIN + DISTINCT，避免大表笛卡尔积
 		searchPattern := "%" + search + "%"
 		orderClause := conversationOrderClause(sortBy, "c")
+		where := ` WHERE (c.title LIKE ?
+			    OR EXISTS (SELECT 1 FROM messages m WHERE m.conversation_id = c.id AND m.content LIKE ?))`
+		args := []interface{}{searchPattern, searchPattern}
+		where, args = appendConversationProjectFilter(where, args, projectID, "c")
+		args = append(args, limit, offset)
 		rows, err = db.Query(
 			`SELECT c.id, c.title, COALESCE(c.pinned, 0), c.created_at, c.updated_at, c.project_id
-			 FROM conversations c
+			 FROM conversations c`+where+`
-			 WHERE c.title LIKE ?
-			    OR EXISTS (SELECT 1 FROM messages m WHERE m.conversation_id = c.id AND m.content LIKE ?)
 			 `+orderClause+`
 			 LIMIT ? OFFSET ?`,
-			searchPattern, searchPattern, limit, offset,
+			args...,
 		)
 	} else {
 		orderClause := conversationOrderClause(sortBy, "")
+		where := ""
+		args := []interface{}{}
+		where, args = appendConversationProjectFilter(where, args, projectID, "")
+		if where != "" {
+			where = " WHERE" + strings.TrimPrefix(where, " AND")
+		}
+		args = append(args, limit, offset)
 		rows, err = db.Query(
-			"SELECT id, title, COALESCE(pinned, 0), created_at, updated_at, project_id FROM conversations "+orderClause+" LIMIT ? OFFSET ?",
+			"SELECT id, title, COALESCE(pinned, 0), created_at, updated_at, project_id FROM conversations"+where+" "+orderClause+" LIMIT ? OFFSET ?",
-			limit, offset,
+			args...,
 		)
 	}
 
@@ -472,23 +510,30 @@ const ungroupedConversationsSQL = `
 	)`
 
 // CountUngroupedConversations 统计不在任何分组中的对话数量。
-func (db *DB) CountUngroupedConversations() (int, error) {
+func (db *DB) CountUngroupedConversations(projectID string) (int, error) {
+	where := ungroupedConversationsSQL
+	args := []interface{}{}
+	where, args = appendConversationProjectFilter(where, args, projectID, "c")
 	var count int
-	if err := db.QueryRow(`SELECT COUNT(*) ` + ungroupedConversationsSQL).Scan(&count); err != nil {
+	if err := db.QueryRow(`SELECT COUNT(*) `+where, args...).Scan(&count); err != nil {
 		return 0, fmt.Errorf("统计未分组对话失败: %w", err)
 	}
 	return count, nil
 }
 
 // ListUngroupedConversations 列出不在任何分组中的对话（最近对话侧栏）。
-func (db *DB) ListUngroupedConversations(limit, offset int, sortBy string) ([]*Conversation, error) {
+func (db *DB) ListUngroupedConversations(limit, offset int, sortBy, projectID string) ([]*Conversation, error) {
 	orderClause := conversationOrderClause(sortBy, "c")
+	where := ungroupedConversationsSQL
+	args := []interface{}{}
+	where, args = appendConversationProjectFilter(where, args, projectID, "c")
+	args = append(args, limit, offset)
 	rows, err := db.Query(
 		`SELECT c.id, c.title, COALESCE(c.pinned, 0), c.created_at, c.updated_at, c.project_id `+
-			ungroupedConversationsSQL+`
+			where+`
 		 `+orderClause+`
 		 LIMIT ? OFFSET ?`,
-		limit, offset,
+		args...,
 	)
 	if err != nil {
 		return nil, fmt.Errorf("查询未分组对话失败: %w", err)
@@ -533,6 +578,19 @@ func (db *DB) ListUngroupedConversations(limit, offset int, sortBy string) ([]*C
 	return conversations, rows.Err()
 }
 
+// GetConversationTitle 获取对话标题（轻量查询，不加载消息）
+func (db *DB) GetConversationTitle(id string) (string, error) {
+	var title string
+	err := db.QueryRow("SELECT title FROM conversations WHERE id = ?", id).Scan(&title)
+	if err != nil {
+		if err == sql.ErrNoRows {
+			return "", fmt.Errorf("对话不存在")
+		}
+		return "", fmt.Errorf("查询对话标题失败: %w", err)
+	}
+	return title, nil
+}
+
 // UpdateConversationTitle 更新对话标题
 func (db *DB) UpdateConversationTitle(id, title string) error {
 	// 注意：不更新 updated_at，因为重命名操作不应该改变对话的更新时间
@@ -1013,6 +1071,77 @@ type ProcessDetail struct {
 	CreatedAt      time.Time `json:"createdAt"`
 }
 
+// GetTurnUserMessage 返回锚点消息所在轮次中的用户原文（最近一条 user 消息，不含完整历史）。
+func (db *DB) GetTurnUserMessage(conversationID, anchorMessageID string) (string, error) {
+	conversationID = strings.TrimSpace(conversationID)
+	anchorMessageID = strings.TrimSpace(anchorMessageID)
+	if conversationID == "" || anchorMessageID == "" {
+		return "", nil
+	}
+	var content string
+	err := db.QueryRow(`
+SELECT m.content FROM messages m
+WHERE m.conversation_id = ? AND m.role = 'user'
+  AND m.created_at <= COALESCE((SELECT created_at FROM messages WHERE id = ? AND conversation_id = ?), m.created_at)
+ORDER BY m.created_at DESC, m.rowid DESC
+LIMIT 1`, conversationID, anchorMessageID, conversationID).Scan(&content)
+	if err != nil {
+		if errors.Is(err, sql.ErrNoRows) {
+			return "", nil
+		}
+		return "", fmt.Errorf("query turn user message: %w", err)
+	}
+	return content, nil
+}
+
+// AssistantCognitionTexts 单条助手消息上的思考/推理/规划文本。
+type AssistantCognitionTexts struct {
+	Thinking       string
+	ReasoningChain string
+	Planning       string
+}
+
+// GetAssistantCognitionTexts 聚合助手消息在 process_details 中的 thinking / reasoning_chain / planning。
+func (db *DB) GetAssistantCognitionTexts(assistantMessageID string) (AssistantCognitionTexts, error) {
+	assistantMessageID = strings.TrimSpace(assistantMessageID)
+	if assistantMessageID == "" {
+		return AssistantCognitionTexts{}, nil
+	}
+	rows, err := db.Query(`
+SELECT event_type, message FROM process_details
+WHERE message_id = ? AND event_type IN ('thinking', 'reasoning_chain', 'planning')
+ORDER BY created_at ASC, rowid ASC`, assistantMessageID)
+	if err != nil {
+		return AssistantCognitionTexts{}, fmt.Errorf("query assistant cognition: %w", err)
+	}
+	defer rows.Close()
+
+	var thinkingParts, reasoningParts, planningParts []string
+	for rows.Next() {
+		var eventType, message string
+		if err := rows.Scan(&eventType, &message); err != nil {
+			continue
+		}
+		msg := strings.TrimSpace(message)
+		if msg == "" {
+			continue
+		}
+		switch eventType {
+		case "thinking":
+			thinkingParts = append(thinkingParts, msg)
+		case "reasoning_chain":
+			reasoningParts = append(reasoningParts, msg)
+		case "planning":
+			planningParts = append(planningParts, msg)
+		}
+	}
+	return AssistantCognitionTexts{
+		Thinking:       strings.Join(thinkingParts, "\n\n"),
+		ReasoningChain: strings.Join(reasoningParts, "\n\n"),
+		Planning:       strings.Join(planningParts, "\n\n"),
+	}, nil
+}
+
 // AddProcessDetail 添加过程详情事件
 func (db *DB) AddProcessDetail(messageID, conversationID, eventType, message string, data interface{}) error {
 	id := uuid.New().String()

internal/database/conversation_project_filter_test.go

@@ -0,0 +1,60 @@
+package database
+
+import (
+	"path/filepath"
+	"testing"
+
+	"go.uber.org/zap"
+)
+
+func TestConversationProjectFilter(t *testing.T) {
+	tmp := t.TempDir()
+	dbPath := filepath.Join(tmp, "conversations.db")
+	db, err := NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	defer db.Close()
+
+	p, err := db.CreateProject(&Project{Name: "target-a", Status: "active"})
+	if err != nil {
+		t.Fatalf("CreateProject: %v", err)
+	}
+
+	convNone, err := db.CreateConversation("unbound", ConversationCreateMeta{})
+	if err != nil {
+		t.Fatalf("CreateConversation unbound: %v", err)
+	}
+	convBound, err := db.CreateConversation("bound", ConversationCreateMeta{ProjectID: p.ID})
+	if err != nil {
+		t.Fatalf("CreateConversation bound: %v", err)
+	}
+
+	totalAll, err := db.CountConversations("", "")
+	if err != nil || totalAll < 2 {
+		t.Fatalf("CountConversations all: total=%d err=%v", totalAll, err)
+	}
+
+	totalBound, err := db.CountConversations("", p.ID)
+	if err != nil || totalBound != 1 {
+		t.Fatalf("CountConversations project: total=%d err=%v", totalBound, err)
+	}
+
+	totalUnbound, err := db.CountConversations("", ProjectFilterUnbound)
+	if err != nil || totalUnbound != 1 {
+		t.Fatalf("CountConversations unbound: total=%d err=%v", totalUnbound, err)
+	}
+
+	listBound, err := db.ListConversations(10, 0, "", "", p.ID)
+	if err != nil || len(listBound) != 1 || listBound[0].ID != convBound.ID {
+		t.Fatalf("ListConversations project: %+v err=%v", listBound, err)
+	}
+
+	listUnbound, err := db.ListConversations(10, 0, "", "", ProjectFilterUnbound)
+	if err != nil || len(listUnbound) != 1 || listUnbound[0].ID != convNone.ID {
+		t.Fatalf("ListConversations unbound: %+v err=%v", listUnbound, err)
+	}
+
+	_ = convNone
+	_ = convBound
+}

internal/database/hitl_logs.go

@@ -0,0 +1,75 @@
+package database
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	"go.uber.org/zap"
+)
+
+// DeleteHitlInterruptLogsByIDs deletes decided HITL audit logs by id (pending rows are skipped).
+func (db *DB) DeleteHitlInterruptLogsByIDs(ids []string) (int64, error) {
+	if db == nil {
+		return 0, fmt.Errorf("database is nil")
+	}
+	clean := make([]string, 0, len(ids))
+	for _, id := range ids {
+		id = strings.TrimSpace(id)
+		if id != "" {
+			clean = append(clean, id)
+		}
+	}
+	if len(clean) == 0 {
+		return 0, nil
+	}
+	placeholders := strings.TrimRight(strings.Repeat("?,", len(clean)), ",")
+	q := fmt.Sprintf(`DELETE FROM hitl_interrupts WHERE status != 'pending' AND id IN (%s)`, placeholders)
+	args := make([]interface{}, len(clean))
+	for i, id := range clean {
+		args[i] = id
+	}
+	res, err := db.Exec(q, args...)
+	if err != nil {
+		db.logger.Error("批量删除人机协同审计日志失败", zap.Error(err), zap.Int("count", len(clean)))
+		return 0, fmt.Errorf("批量删除人机协同审计日志失败: %w", err)
+	}
+	n, _ := res.RowsAffected()
+	return n, nil
+}
+
+// DeleteHitlInterruptLogsMatching deletes decided logs matching whereSQL (e.g. "WHERE 1=1 AND status != 'pending' ...").
+func (db *DB) DeleteHitlInterruptLogsMatching(whereSQL string, args []interface{}) (int64, error) {
+	if db == nil {
+		return 0, fmt.Errorf("database is nil")
+	}
+	whereSQL = strings.TrimSpace(whereSQL)
+	if whereSQL == "" {
+		return 0, fmt.Errorf("where clause is required")
+	}
+	q := `DELETE FROM hitl_interrupts ` + whereSQL
+	res, err := db.Exec(q, args...)
+	if err != nil {
+		db.logger.Error("清空人机协同审计日志失败", zap.Error(err))
+		return 0, fmt.Errorf("清空人机协同审计日志失败: %w", err)
+	}
+	n, _ := res.RowsAffected()
+	return n, nil
+}
+
+// PurgeHitlInterruptLogsBefore deletes decided logs with decided/created time before cutoff.
+func (db *DB) PurgeHitlInterruptLogsBefore(cutoff time.Time) (int64, error) {
+	if db == nil {
+		return 0, fmt.Errorf("database is nil")
+	}
+	res, err := db.Exec(
+		`DELETE FROM hitl_interrupts WHERE status != 'pending' AND datetime(COALESCE(decided_at, created_at)) < datetime(?)`,
+		cutoff.UTC().Format(time.RFC3339),
+	)
+	if err != nil {
+		db.logger.Error("清理过期人机协同审计日志失败", zap.Error(err))
+		return 0, fmt.Errorf("清理过期人机协同审计日志失败: %w", err)
+	}
+	n, _ := res.RowsAffected()
+	return n, nil
+}

internal/database/hitl_logs_test.go

@@ -0,0 +1,106 @@
+package database
+
+import (
+	"path/filepath"
+	"testing"
+	"time"
+
+	"go.uber.org/zap"
+)
+
+func ensureHitlInterruptsTable(t *testing.T, db *DB) {
+	t.Helper()
+	if _, err := db.Exec(`
+CREATE TABLE IF NOT EXISTS hitl_interrupts (
+    id TEXT PRIMARY KEY,
+    conversation_id TEXT NOT NULL,
+    message_id TEXT,
+    mode TEXT NOT NULL,
+    tool_name TEXT NOT NULL,
+    tool_call_id TEXT,
+    payload TEXT,
+    status TEXT NOT NULL,
+    decision TEXT,
+    decision_comment TEXT,
+    created_at DATETIME NOT NULL,
+    decided_at DATETIME
+);`); err != nil {
+		t.Fatalf("create hitl_interrupts: %v", err)
+	}
+}
+
+func TestDeleteHitlInterruptLogsByIDs_skipsPending(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "hitl.db")
+	db, err := NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	defer db.Close()
+	ensureHitlInterruptsTable(t, db)
+
+	now := time.Now().UTC().Format(time.RFC3339)
+	if _, err := db.Exec(`INSERT INTO hitl_interrupts
+		(id, conversation_id, mode, tool_name, status, created_at)
+		VALUES ('pending-1', 'c1', 'approval', 'exec', 'pending', ?)`, now); err != nil {
+		t.Fatalf("insert pending: %v", err)
+	}
+	if _, err := db.Exec(`INSERT INTO hitl_interrupts
+		(id, conversation_id, mode, tool_name, status, decision, created_at, decided_at)
+		VALUES ('done-1', 'c1', 'approval', 'exec', 'decided', 'approve', ?, ?)`, now, now); err != nil {
+		t.Fatalf("insert decided: %v", err)
+	}
+
+	deleted, err := db.DeleteHitlInterruptLogsByIDs([]string{"pending-1", "done-1"})
+	if err != nil {
+		t.Fatalf("DeleteHitlInterruptLogsByIDs: %v", err)
+	}
+	if deleted != 1 {
+		t.Fatalf("deleted = %d, want 1", deleted)
+	}
+
+	var status string
+	if err := db.QueryRow(`SELECT status FROM hitl_interrupts WHERE id = 'pending-1'`).Scan(&status); err != nil {
+		t.Fatalf("pending row missing: %v", err)
+	}
+	if err := db.QueryRow(`SELECT id FROM hitl_interrupts WHERE id = 'done-1'`).Scan(new(string)); err == nil {
+		t.Fatal("decided row should be deleted")
+	}
+}
+
+func TestPurgeHitlInterruptLogsBefore(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "hitl.db")
+	db, err := NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	defer db.Close()
+	ensureHitlInterruptsTable(t, db)
+
+	old := time.Now().AddDate(0, 0, -100).UTC().Format(time.RFC3339)
+	recent := time.Now().AddDate(0, 0, -1).UTC().Format(time.RFC3339)
+	for _, row := range []struct{ id, decided string }{
+		{"old-1", old},
+		{"new-1", recent},
+	} {
+		if _, err := db.Exec(`INSERT INTO hitl_interrupts
+			(id, conversation_id, mode, tool_name, status, decision, created_at, decided_at)
+			VALUES (?, 'c1', 'approval', 'exec', 'decided', 'approve', ?, ?)`, row.id, row.decided, row.decided); err != nil {
+			t.Fatalf("insert %s: %v", row.id, err)
+		}
+	}
+
+	cutoff := time.Now().AddDate(0, 0, -90)
+	deleted, err := db.PurgeHitlInterruptLogsBefore(cutoff)
+	if err != nil {
+		t.Fatalf("PurgeHitlInterruptLogsBefore: %v", err)
+	}
+	if deleted != 1 {
+		t.Fatalf("deleted = %d, want 1", deleted)
+	}
+	if err := db.QueryRow(`SELECT id FROM hitl_interrupts WHERE id = 'old-1'`).Scan(new(string)); err == nil {
+		t.Fatal("old row should be purged")
+	}
+	if err := db.QueryRow(`SELECT id FROM hitl_interrupts WHERE id = 'new-1'`).Scan(new(string)); err != nil {
+		t.Fatalf("new row should remain: %v", err)
+	}
+}

internal/database/monitor.go

@@ -3,7 +3,6 @@ package database
 import (
 	"database/sql"
 	"encoding/json"
-	"sort"
 	"strings"
 	"time"
 
@@ -227,6 +226,167 @@ func (db *DB) LoadToolExecutionsWithPagination(offset, limit int, status, toolNa
 	return executions, nil
 }
 
+func toolExecutionsFilterSQL(status, toolName string) (string, []interface{}) {
+	args := []interface{}{}
+	conditions := []string{}
+	if status != "" {
+		conditions = append(conditions, "status = ?")
+		args = append(args, status)
+	}
+	if toolName != "" {
+		conditions = append(conditions, "LOWER(tool_name) LIKE ?")
+		args = append(args, "%"+strings.ToLower(toolName)+"%")
+	}
+	if len(conditions) == 0 {
+		return "", args
+	}
+	return ` WHERE ` + strings.Join(conditions, ` AND `), args
+}
+
+// ToolStatsSummary 工具调用汇总（全量聚合，不含逐工具明细）
+type ToolStatsSummary struct {
+	TotalCalls   int
+	SuccessCalls int
+	FailedCalls  int
+	LastCallTime *time.Time
+	ToolCount    int
+}
+
+// ToolStatsSummaryResult 汇总 + Top N 工具排行
+type ToolStatsSummaryResult struct {
+	Summary  ToolStatsSummary
+	TopTools []*mcp.ToolStats
+}
+
+// LoadToolStatsSummary 聚合统计信息，仅返回汇总与 Top N 工具（避免全量 map 传输）
+func (db *DB) LoadToolStatsSummary(topN int) (*ToolStatsSummaryResult, error) {
+	if topN <= 0 {
+		topN = 6
+	}
+	if topN > 100 {
+		topN = 100
+	}
+
+	result := &ToolStatsSummaryResult{
+		TopTools: make([]*mcp.ToolStats, 0, topN),
+	}
+
+	summaryQuery := `
+		SELECT COUNT(*),
+			COALESCE(SUM(total_calls), 0),
+			COALESCE(SUM(success_calls), 0),
+			COALESCE(SUM(failed_calls), 0),
+			MAX(last_call_time)
+		FROM tool_stats
+	`
+	var lastCallRaw sql.NullString
+	err := db.QueryRow(summaryQuery).Scan(
+		&result.Summary.ToolCount,
+		&result.Summary.TotalCalls,
+		&result.Summary.SuccessCalls,
+		&result.Summary.FailedCalls,
+		&lastCallRaw,
+	)
+	if err != nil {
+		return nil, err
+	}
+	if lastCallRaw.Valid && strings.TrimSpace(lastCallRaw.String) != "" {
+		if t, parseErr := time.Parse(time.RFC3339Nano, lastCallRaw.String); parseErr == nil {
+			result.Summary.LastCallTime = &t
+		} else if t, parseErr := time.Parse("2006-01-02 15:04:05.999999999-07:00", lastCallRaw.String); parseErr == nil {
+			result.Summary.LastCallTime = &t
+		} else if t, parseErr := time.Parse("2006-01-02 15:04:05", lastCallRaw.String); parseErr == nil {
+			result.Summary.LastCallTime = &t
+		}
+	}
+
+	topQuery := `
+		SELECT tool_name, total_calls, success_calls, failed_calls, last_call_time
+		FROM tool_stats
+		WHERE total_calls > 0
+		ORDER BY total_calls DESC, tool_name ASC
+		LIMIT ?
+	`
+	rows, err := db.Query(topQuery, topN)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+
+	for rows.Next() {
+		var stat mcp.ToolStats
+		var lastCallTime sql.NullTime
+		if err := rows.Scan(
+			&stat.ToolName,
+			&stat.TotalCalls,
+			&stat.SuccessCalls,
+			&stat.FailedCalls,
+			&lastCallTime,
+		); err != nil {
+			db.logger.Warn("加载 Top 工具统计失败", zap.Error(err))
+			continue
+		}
+		if lastCallTime.Valid {
+			stat.LastCallTime = &lastCallTime.Time
+		}
+		result.TopTools = append(result.TopTools, &stat)
+	}
+
+	return result, nil
+}
+
+// LoadToolExecutionListPage 分页加载执行记录列表（不含 arguments/result，供监控列表使用）
+func (db *DB) LoadToolExecutionListPage(offset, limit int, status, toolName string) ([]*mcp.ToolExecution, error) {
+	if limit <= 0 {
+		limit = 20
+	}
+	if limit > 100 {
+		limit = 100
+	}
+
+	query := `
+		SELECT id, tool_name, status, start_time, end_time, duration_ms
+		FROM tool_executions
+	`
+	whereSQL, args := toolExecutionsFilterSQL(status, toolName)
+	query += whereSQL + ` ORDER BY start_time DESC LIMIT ? OFFSET ?`
+	args = append(args, limit, offset)
+
+	rows, err := db.Query(query, args...)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+
+	executions := make([]*mcp.ToolExecution, 0, limit)
+	for rows.Next() {
+		var exec mcp.ToolExecution
+		var endTime sql.NullTime
+		var durationMs sql.NullInt64
+
+		if err := rows.Scan(
+			&exec.ID,
+			&exec.ToolName,
+			&exec.Status,
+			&exec.StartTime,
+			&endTime,
+			&durationMs,
+		); err != nil {
+			db.logger.Warn("加载执行记录列表失败", zap.Error(err))
+			continue
+		}
+		if endTime.Valid {
+			exec.EndTime = &endTime.Time
+		}
+		if durationMs.Valid {
+			exec.Duration = time.Duration(durationMs.Int64) * time.Millisecond
+		}
+		executions = append(executions, &exec)
+	}
+
+	return executions, nil
+}
+
 // GetToolExecution 根据ID获取单条工具执行记录
 func (db *DB) GetToolExecution(id string) (*mcp.ToolExecution, error) {
 	query := `
@@ -288,6 +448,93 @@ func (db *DB) GetToolExecution(id string) (*mcp.ToolExecution, error) {
 	return &exec, nil
 }
 
+// CancelOrphanedRunningToolExecutions 将仍为 running 的记录批量标记为 cancelled（如进程重启后无对应执行协程）。
+func (db *DB) CancelOrphanedRunningToolExecutions(endTime time.Time, errMsg string) (int64, error) {
+	errMsg = strings.TrimSpace(errMsg)
+	if errMsg == "" {
+		errMsg = "执行已中断（服务重启或会话结束）"
+	}
+	query := `
+		UPDATE tool_executions
+		SET status = 'cancelled',
+		    error = ?,
+		    end_time = ?,
+		    duration_ms = MAX(0, CAST((julianday(?) - julianday(start_time)) * 86400000 AS INTEGER))
+		WHERE status = 'running'
+	`
+	res, err := db.Exec(query, errMsg, endTime, endTime)
+	if err != nil {
+		return 0, err
+	}
+	return res.RowsAffected()
+}
+
+// FinalizeStaleRunningToolExecutions 将「非活跃且超过 minAge」的 running 记录标记为 cancelled。
+// activeIDs 为当前进程内仍登记 cancel 的 executionId；不在集合内且已超时的视为孤儿记录。
+func (db *DB) FinalizeStaleRunningToolExecutions(endTime time.Time, minAge time.Duration, activeIDs map[string]struct{}, errMsg string) (int64, error) {
+	errMsg = strings.TrimSpace(errMsg)
+	if errMsg == "" {
+		errMsg = "执行已中断（会话已结束）"
+	}
+	if minAge < 0 {
+		minAge = 0
+	}
+	cutoff := endTime.Add(-minAge)
+	rows, err := db.Query(`
+		SELECT id, start_time FROM tool_executions
+		WHERE status = 'running' AND start_time <= ?
+	`, cutoff)
+	if err != nil {
+		return 0, err
+	}
+	defer rows.Close()
+
+	type staleRow struct {
+		id        string
+		startTime time.Time
+	}
+	var stale []staleRow
+	for rows.Next() {
+		var row staleRow
+		if err := rows.Scan(&row.id, &row.startTime); err != nil {
+			db.logger.Warn("读取 stale running 执行记录失败", zap.Error(err))
+			continue
+		}
+		if activeIDs != nil {
+			if _, active := activeIDs[row.id]; active {
+				continue
+			}
+		}
+		stale = append(stale, row)
+	}
+	if err := rows.Err(); err != nil {
+		return 0, err
+	}
+	if len(stale) == 0 {
+		return 0, nil
+	}
+
+	var affected int64
+	for _, row := range stale {
+		durationMs := endTime.Sub(row.startTime).Milliseconds()
+		if durationMs < 0 {
+			durationMs = 0
+		}
+		res, err := db.Exec(`
+			UPDATE tool_executions
+			SET status = 'cancelled', error = ?, end_time = ?, duration_ms = ?
+			WHERE id = ? AND status = 'running'
+		`, errMsg, endTime, durationMs, row.id)
+		if err != nil {
+			db.logger.Warn("更新 stale running 执行记录失败", zap.Error(err), zap.String("executionId", row.id))
+			continue
+		}
+		n, _ := res.RowsAffected()
+		affected += n
+	}
+	return affected, nil
+}
+
 // DeleteToolExecution 删除工具执行记录
 func (db *DB) DeleteToolExecution(id string) error {
 	query := `DELETE FROM tool_executions WHERE id = ?`
@@ -600,13 +847,28 @@ func truncateCallsTimelineBucket(t time.Time, dailyBuckets bool) time.Time {
 
 // LoadCallsTimeline 按时间范围加载调用趋势（since 起至今，含边界）
 func (db *DB) LoadCallsTimeline(since time.Time, dailyBuckets bool) ([]CallsTimelineBucket, error) {
-	// 在 Go 侧按本地时区分桶，避免 SQLite strftime 对 UTC 存储时间分桶后再误当本地时间解析（差 8h 等问题）
+	var query string
-	query := `
+	if dailyBuckets {
-		SELECT start_time,
+		query = `
-			CASE WHEN status IN ('failed', 'cancelled') THEN 1 ELSE 0 END AS failed
+			SELECT date(start_time, 'localtime') AS bucket,
-		FROM tool_executions
+				COUNT(*) AS total,
-		WHERE start_time >= ?
+				SUM(CASE WHEN status IN ('failed', 'cancelled') THEN 1 ELSE 0 END) AS failed
-	`
+			FROM tool_executions
+			WHERE start_time >= ?
+			GROUP BY bucket
+			ORDER BY bucket
+		`
+	} else {
+		query = `
+			SELECT strftime('%Y-%m-%d %H:00:00', start_time, 'localtime') AS bucket,
+				COUNT(*) AS total,
+				SUM(CASE WHEN status IN ('failed', 'cancelled') THEN 1 ELSE 0 END) AS failed
+			FROM tool_executions
+			WHERE start_time >= ?
+			GROUP BY bucket
+			ORDER BY bucket
+		`
+	}
 
 	rows, err := db.Query(query, since)
 	if err != nil {
@@ -614,35 +876,35 @@ func (db *DB) LoadCallsTimeline(since time.Time, dailyBuckets bool) ([]CallsTime
 	}
 	defer rows.Close()
 
-	bucketMap := make(map[time.Time]struct{ total, failed int })
+	buckets := make([]CallsTimelineBucket, 0)
 	for rows.Next() {
-		var startTime time.Time
+		var bucketStr string
-		var failed int
+		var total, failed int
-		if err := rows.Scan(&startTime, &failed); err != nil {
+		if err := rows.Scan(&bucketStr, &total, &failed); err != nil {
 			db.logger.Warn("加载调用趋势失败", zap.Error(err))
 			continue
 		}
-		key := truncateCallsTimelineBucket(startTime, dailyBuckets)
+		bucketTime, err := parseCallsTimelineBucket(bucketStr, dailyBuckets)
-		entry := bucketMap[key]
+		if err != nil {
-		entry.total++
+			db.logger.Warn("解析调用趋势时间桶失败", zap.Error(err), zap.String("bucket", bucketStr))
-		entry.failed += failed
+			continue
-		bucketMap[key] = entry
+		}
-	}
-
-	buckets := make([]CallsTimelineBucket, 0, len(bucketMap))
-	for bucketTime, counts := range bucketMap {
 		buckets = append(buckets, CallsTimelineBucket{
 			BucketTime: bucketTime,
-			Total:      counts.total,
+			Total:      total,
-			Failed:     counts.failed,
+			Failed:     failed,
 		})
 	}
-	sort.Slice(buckets, func(i, j int) bool {
-		return buckets[i].BucketTime.Before(buckets[j].BucketTime)
-	})
 	return buckets, nil
 }
 
+func parseCallsTimelineBucket(bucketStr string, dailyBuckets bool) (time.Time, error) {
+	if dailyBuckets {
+		return time.ParseInLocation("2006-01-02", bucketStr, time.Local)
+	}
+	return time.ParseInLocation("2006-01-02 15:04:05", bucketStr, time.Local)
+}
+
 // DecreaseToolStats 减少工具统计信息（用于删除执行记录时）
 // 如果统计信息变为0，则删除该统计记录
 func (db *DB) DecreaseToolStats(toolName string, totalCalls, successCalls, failedCalls int) error {

internal/database/monitor_reconcile_test.go

@@ -0,0 +1,102 @@
+package database
+
+import (
+	"path/filepath"
+	"testing"
+	"time"
+
+	"cyberstrike-ai/internal/mcp"
+
+	"go.uber.org/zap"
+)
+
+func TestCancelOrphanedRunningToolExecutions(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "monitor.db")
+	db, err := NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	defer db.Close()
+
+	start := time.Now().Add(-2 * time.Hour)
+	exec := &mcp.ToolExecution{
+		ID:        "orphan-hydra",
+		ToolName:  "hydra",
+		Arguments: map[string]interface{}{"target": "127.0.0.1"},
+		Status:    "running",
+		StartTime: start,
+	}
+	if err := db.SaveToolExecution(exec); err != nil {
+		t.Fatalf("SaveToolExecution: %v", err)
+	}
+
+	end := time.Now()
+	n, err := db.CancelOrphanedRunningToolExecutions(end, "执行已中断（服务重启）")
+	if err != nil {
+		t.Fatalf("CancelOrphanedRunningToolExecutions: %v", err)
+	}
+	if n != 1 {
+		t.Fatalf("expected 1 row updated, got %d", n)
+	}
+
+	got, err := db.GetToolExecution("orphan-hydra")
+	if err != nil {
+		t.Fatalf("GetToolExecution: %v", err)
+	}
+	if got.Status != "cancelled" {
+		t.Fatalf("expected cancelled, got %s", got.Status)
+	}
+	if got.EndTime == nil {
+		t.Fatal("expected end_time to be set")
+	}
+	if got.Duration <= 0 {
+		t.Fatalf("expected positive duration, got %v", got.Duration)
+	}
+}
+
+func TestFinalizeStaleRunningToolExecutions_skipsActive(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "monitor.db")
+	db, err := NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	defer db.Close()
+
+	now := time.Now()
+	oldStart := now.Add(-5 * time.Minute)
+	if err := db.SaveToolExecution(&mcp.ToolExecution{
+		ID: "stale", ToolName: "hydra", Status: "running", StartTime: oldStart,
+	}); err != nil {
+		t.Fatalf("SaveToolExecution stale: %v", err)
+	}
+	if err := db.SaveToolExecution(&mcp.ToolExecution{
+		ID: "active", ToolName: "hydra", Status: "running", StartTime: oldStart,
+	}); err != nil {
+		t.Fatalf("SaveToolExecution active: %v", err)
+	}
+
+	active := map[string]struct{}{"active": {}}
+	n, err := db.FinalizeStaleRunningToolExecutions(now, time.Minute, active, "执行已中断（会话已结束）")
+	if err != nil {
+		t.Fatalf("FinalizeStaleRunningToolExecutions: %v", err)
+	}
+	if n != 1 {
+		t.Fatalf("expected 1 stale row updated, got %d", n)
+	}
+
+	stale, err := db.GetToolExecution("stale")
+	if err != nil {
+		t.Fatalf("GetToolExecution stale: %v", err)
+	}
+	if stale.Status != "cancelled" {
+		t.Fatalf("stale expected cancelled, got %s", stale.Status)
+	}
+
+	activeExec, err := db.GetToolExecution("active")
+	if err != nil {
+		t.Fatalf("GetToolExecution active: %v", err)
+	}
+	if activeExec.Status != "running" {
+		t.Fatalf("active expected running, got %s", activeExec.Status)
+	}
+}

internal/database/monitor_summary_test.go

@@ -0,0 +1,86 @@
+package database
+
+import (
+	"fmt"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"cyberstrike-ai/internal/mcp"
+
+	"go.uber.org/zap"
+)
+
+func TestLoadToolStatsSummaryAndListPage(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "monitor-summary.db")
+	db, err := NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	defer db.Close()
+
+	now := time.Now()
+	tools := []struct {
+		name   string
+		calls  int
+		ok     int
+		fail   int
+		result string
+	}{
+		{"alpha::run", 10, 9, 1, `{"content":[{"type":"text","text":"` + string(make([]byte, 64*1024)) + `"}]}`},
+		{"beta::scan", 5, 5, 0, `{"content":[{"type":"text","text":"ok"}]}`},
+		{"gamma::ping", 1, 1, 0, `{"content":[{"type":"text","text":"pong"}]}`},
+	}
+
+	for _, tool := range tools {
+		if err := db.UpdateToolStats(tool.name, tool.calls, tool.ok, tool.fail, &now); err != nil {
+			t.Fatalf("UpdateToolStats(%s): %v", tool.name, err)
+		}
+		for j := 0; j < tool.calls; j++ {
+			exec := &mcp.ToolExecution{
+				ID:        fmt.Sprintf("%s-exec-%d", tool.name, j),
+				ToolName:  tool.name,
+				Arguments: map[string]interface{}{"n": j},
+				Status:    "completed",
+				StartTime: now.Add(-time.Duration(j) * time.Minute),
+				Result:    &mcp.ToolResult{Content: []mcp.Content{{Type: "text", Text: tool.result}}},
+			}
+			end := exec.StartTime.Add(time.Second)
+			exec.EndTime = &end
+			exec.Duration = time.Second
+			if err := db.SaveToolExecution(exec); err != nil {
+				t.Fatalf("SaveToolExecution: %v", err)
+			}
+		}
+	}
+
+	summary, err := db.LoadToolStatsSummary(2)
+	if err != nil {
+		t.Fatalf("LoadToolStatsSummary: %v", err)
+	}
+	if summary.Summary.ToolCount != 3 {
+		t.Fatalf("toolCount = %d, want 3", summary.Summary.ToolCount)
+	}
+	if summary.Summary.TotalCalls != 16 {
+		t.Fatalf("totalCalls = %d, want 16", summary.Summary.TotalCalls)
+	}
+	if len(summary.TopTools) != 2 {
+		t.Fatalf("top tools = %d, want 2", len(summary.TopTools))
+	}
+	if summary.TopTools[0].ToolName != "alpha::run" {
+		t.Fatalf("top tool = %q, want alpha::run", summary.TopTools[0].ToolName)
+	}
+
+	list, err := db.LoadToolExecutionListPage(0, 5, "", "")
+	if err != nil {
+		t.Fatalf("LoadToolExecutionListPage: %v", err)
+	}
+	if len(list) != 5 {
+		t.Fatalf("list len = %d, want 5", len(list))
+	}
+	for _, exec := range list {
+		if exec.Arguments != nil || exec.Result != nil || exec.Error != "" {
+			t.Fatalf("expected lite execution row, got args/result/error on %s", exec.ID)
+		}
+	}
+}

internal/database/project.go

@@ -111,19 +111,43 @@ func (db *DB) GetProject(id string) (*Project, error) {
 	return &p, nil
 }
 
-// CountProjects 统计项目数量。
+func projectListSearchPattern(q string) string {
-func (db *DB) CountProjects(status, search string) (int, error) {
+	q = strings.TrimSpace(q)
-	query := `SELECT COUNT(*) FROM projects WHERE 1=1`
+	if q == "" {
-	args := []interface{}{}
+		return ""
+	}
+	var b strings.Builder
+	b.WriteByte('%')
+	for _, r := range q {
+		switch r {
+		case '%', '_', '\\':
+			b.WriteByte('\\')
+			b.WriteRune(r)
+		default:
+			b.WriteRune(r)
+		}
+	}
+	b.WriteByte('%')
+	return b.String()
+}
+
+func appendProjectListFilters(query string, args []interface{}, status, search string) (string, []interface{}) {
 	if s := strings.TrimSpace(status); s != "" {
 		query += " AND status = ?"
 		args = append(args, s)
 	}
-	if q := strings.TrimSpace(search); q != "" {
+	if pattern := projectListSearchPattern(search); pattern != "" {
-		pattern := "%" + q + "%"
+		query += ` AND (LOWER(name) LIKE LOWER(?) ESCAPE '\' OR LOWER(COALESCE(description,'')) LIKE LOWER(?) ESCAPE '\' OR LOWER(id) LIKE LOWER(?) ESCAPE '\')`
-		query += " AND (name LIKE ? OR COALESCE(description,'') LIKE ?)"
+		args = append(args, pattern, pattern, pattern)
-		args = append(args, pattern, pattern)
 	}
+	return query, args
+}
+
+// CountProjects 统计项目数量。
+func (db *DB) CountProjects(status, search string) (int, error) {
+	query := `SELECT COUNT(*) FROM projects WHERE 1=1`
+	args := []interface{}{}
+	query, args = appendProjectListFilters(query, args, status, search)
 	var count int
 	if err := db.QueryRow(query, args...).Scan(&count); err != nil {
 		return 0, fmt.Errorf("统计项目失败: %w", err)
@@ -139,15 +163,7 @@ func (db *DB) ListProjects(status, search string, limit, offset int) ([]*Project
 	query := `SELECT id, name, COALESCE(description,''), COALESCE(scope_json,''), status, pinned, created_at, updated_at
 		FROM projects WHERE 1=1`
 	args := []interface{}{}
-	if s := strings.TrimSpace(status); s != "" {
+	query, args = appendProjectListFilters(query, args, status, search)
-		query += " AND status = ?"
-		args = append(args, s)
-	}
-	if q := strings.TrimSpace(search); q != "" {
-		pattern := "%" + q + "%"
-		query += " AND (name LIKE ? OR COALESCE(description,'') LIKE ?)"
-		args = append(args, pattern, pattern)
-	}
 	query += " ORDER BY pinned DESC, updated_at DESC LIMIT ? OFFSET ?"
 	args = append(args, limit, offset)
 

internal/database/project_search_test.go

@@ -0,0 +1,82 @@
+package database
+
+import (
+	"path/filepath"
+	"testing"
+
+	"go.uber.org/zap"
+)
+
+func TestListProjectsSearchCaseInsensitive(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "projects-search.db")
+	db, err := NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer db.Close()
+
+	p1, err := db.CreateProject(&Project{Name: "Alpha Security Review", Status: "active"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	p2, err := db.CreateProject(&Project{Name: "beta-scan", Status: "active"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := db.CreateProject(&Project{Name: "Other", Status: "archived"}); err != nil {
+		t.Fatal(err)
+	}
+
+	cases := []struct {
+		name   string
+		search string
+		status string
+		want   []string
+	}{
+		{name: "case insensitive name", search: "alpha", status: "active", want: []string{p1.ID}},
+		{name: "upper query", search: "BETA", status: "active", want: []string{p2.ID}},
+		{name: "search by id substring", search: p1.ID[:8], status: "", want: []string{p1.ID}},
+		{name: "status filter", search: "alpha", status: "archived", want: nil},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			list, err := db.ListProjects(tc.status, tc.search, 50, 0)
+			if err != nil {
+				t.Fatal(err)
+			}
+			got := make([]string, 0, len(list))
+			for _, p := range list {
+				got = append(got, p.ID)
+			}
+			if len(got) != len(tc.want) {
+				t.Fatalf("got %v want %v", got, tc.want)
+			}
+			for i := range got {
+				if got[i] != tc.want[i] {
+					t.Fatalf("got %v want %v", got, tc.want)
+				}
+			}
+		})
+	}
+}
+
+func TestProjectListSearchPatternEscapesWildcards(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "projects-like.db")
+	db, err := NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer db.Close()
+
+	p, err := db.CreateProject(&Project{Name: "100% coverage", Status: "active"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	list, err := db.ListProjects("active", "100%", 50, 0)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(list) != 1 || list[0].ID != p.ID {
+		t.Fatalf("expected exact match for literal %% query, got %#v", list)
+	}
+}

internal/einomcp/tool_invoke_notify.go

@@ -2,8 +2,8 @@ package einomcp
 
 import "sync"
 
-// ToolInvokeNotifyHolder 由 Eino run loop 在迭代开始前 Set 回调；MCP/execute 桥在工具调用结束时 Fire，
+// ToolInvokeNotifyHolder 由 Eino run loop 与 MCP/execute 桥共享；Fire 在工具原始返回时触发。
-// 用于清除 pending tool_call（tool_result 由 ADK schema.Tool 事件推送，含流式工具与 reduction 后正文）。
+// UI 的 tool_result 须等 ADK schema.Tool 事件（reduction 后正文），不在此 holder 的回调里推送。
 type ToolInvokeNotifyHolder struct {
 	mu sync.RWMutex
 	fn func(toolCallID, toolName, einoAgent string, success bool, content string, invokeErr error)

internal/handler/agent.go

@@ -77,6 +77,13 @@ type responsePlanAgg struct {
 	b    strings.Builder
 }
 
+// thinkingBuf aggregates thinking_stream_* / reasoning_chain_stream_* before flush to process_details.
+type thinkingBuf struct {
+	b         strings.Builder
+	meta      map[string]interface{}
+	persistAs string // "thinking" | "reasoning_chain"
+}
+
 func normalizeProcessDetailText(s string) string {
 	s = strings.ReplaceAll(s, "\r\n", "\n")
 	s = strings.ReplaceAll(s, "\r", "\n")
@@ -179,6 +186,8 @@ type AgentHandler struct {
 	batchCronParser   cron.Parser
 	// hitlWhitelistSaver 侧栏「应用」HITL 时将会话增量白名单合并写入 config.yaml（可选）
 	hitlWhitelistSaver HitlToolWhitelistSaver
+	hitlStrategySaver  HitlAuditStrategySaver
+	auditLLM           *openai.Client
 	audit              *audit.Service
 }
 
@@ -200,9 +209,8 @@ func (h *AgentHandler) CancelRunningTaskForConversation(conversationID string) {
 	if h == nil || conversationID == "" || h.tasks == nil {
 		return
 	}
-	if execID := h.tasks.ActiveMCPExecutionID(conversationID); execID != "" {
+	h.cancelActiveMCPToolForConversation(conversationID)
-		h.agent.CancelMCPToolExecutionWithNote(execID, "")
+	h.tasks.AbortActiveEinoExecute(conversationID, "")
-	}
 	if ok, err := h.tasks.CancelTask(conversationID, ErrTaskCancelled); ok {
 		h.logger.Info("已取消会话运行中任务", zap.String("conversationId", conversationID))
 	} else if err != nil {
@@ -210,9 +218,19 @@ func (h *AgentHandler) CancelRunningTaskForConversation(conversationID string) {
 	}
 }
 
-// HitlToolWhitelistSaver 合并 HITL 免审批工具到全局配置并落盘
+func (h *AgentHandler) cancelActiveMCPToolForConversation(conversationID string) {
+	if h == nil || h.tasks == nil || h.agent == nil {
+		return
+	}
+	if execID := h.tasks.ActiveMCPExecutionID(conversationID); execID != "" {
+		h.agent.CancelMCPToolExecutionWithNote(execID, "")
+	}
+}
+
+// HitlToolWhitelistSaver 合并/设置 HITL 免审批工具到全局配置并落盘
 type HitlToolWhitelistSaver interface {
 	MergeHitlToolWhitelistIntoConfig(add []string) error
+	SetHitlToolWhitelist(tools []string) error
 }
 
 // NewAgentHandler 创建新的Agent处理器
@@ -228,6 +246,11 @@ func NewAgentHandler(agent *agent.Agent, db *database.DB, cfg *config.Config, lo
 	bus := NewTaskEventBus()
 	tm := NewAgentTaskManager()
 	tm.SetTaskEventBus(bus)
+	llmHTTP := &http.Client{Timeout: 2 * time.Minute}
+	var llmCfg *config.OpenAIConfig
+	if cfg != nil {
+		llmCfg = &cfg.OpenAI
+	}
 	handler := &AgentHandler{
 		agent:            agent,
 		db:               db,
@@ -238,7 +261,9 @@ func NewAgentHandler(agent *agent.Agent, db *database.DB, cfg *config.Config, lo
 		config:           cfg,
 		hitlManager:      NewHITLManager(db, logger),
 		batchCronParser:  cron.NewParser(cron.Minute | cron.Hour | cron.Dom | cron.Month | cron.Dow | cron.Descriptor),
+		auditLLM:         openai.NewClient(llmCfg, llmHTTP, logger),
 	}
+	tm.SetToolCanceler(handler.cancelActiveMCPToolForConversation)
 	if err := handler.hitlManager.EnsureSchema(); err != nil {
 		logger.Warn("初始化 HITL 表失败", zap.Error(err))
 	}
@@ -311,6 +336,7 @@ func chatReasoningToClientIntent(r *ChatReasoningRequest) *reasoning.ClientInten
 type HITLRequest struct {
 	Enabled        bool     `json:"enabled"`
 	Mode           string   `json:"mode,omitempty"`
+	Reviewer       string   `json:"reviewer,omitempty"` // human | audit_agent
 	SensitiveTools []string `json:"sensitiveTools,omitempty"`
 	TimeoutSeconds int      `json:"timeoutSeconds,omitempty"`
 }
@@ -840,11 +866,6 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 
 	// thinking_stream_*（ReAct 等助手正文流）与 reasoning_chain_stream_*（Eino ReasoningContent）：
 	// 不逐条落库，按 streamId 聚合，flush 时分别落 thinking / reasoning_chain。
-	type thinkingBuf struct {
-		b         strings.Builder
-		meta      map[string]interface{}
-		persistAs string // "thinking" | "reasoning_chain"
-	}
 	thinkingStreams := make(map[string]*thinkingBuf) // streamId -> buf
 	flushedThinking := make(map[string]bool)         // streamId -> flushed
 	seenToolCallSigs := make(map[string]string)      // toolCallId -> payload signature
@@ -857,6 +878,12 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 	// response_start + response_delta：前端时间线显示为「📝 规划中」（monitor.js），不落逐条 delta；
 	// 聚合为一条 planning 写入 process_details，刷新后与线上一致。
 	var respPlan responsePlanAgg
+	if assistantMessageID != "" {
+		h.tasks.SetHitlAssistantMessageID(conversationID, assistantMessageID)
+	}
+	syncHitlCognition := func() {
+		h.syncHitlCognitionFromProgress(conversationID, assistantMessageID, thinkingStreams, &respPlan)
+	}
 	flushResponsePlan := func() {
 		if assistantMessageID == "" {
 			return
@@ -876,6 +903,7 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 		if err := h.db.AddProcessDetail(assistantMessageID, conversationID, "planning", content, data); err != nil {
 			h.logger.Warn("保存过程详情失败", zap.Error(err), zap.String("eventType", "planning"))
 		}
+		syncHitlCognition()
 		respPlan.meta = nil
 		respPlan.b.Reset()
 	}
@@ -912,6 +940,7 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 			}
 			flushedThinking[sid] = true
 		}
+		syncHitlCognition()
 	}
 
 	return func(eventType, message string, data interface{}) {
@@ -972,6 +1001,25 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 			}
 		}
 
+		if eventType == "tool_result" {
+			if dataMap, ok := data.(map[string]interface{}); ok {
+				toolName, _ := dataMap["toolName"].(string)
+				toolCallID, _ := dataMap["toolCallId"].(string)
+				success := true
+				if v, ok := dataMap["success"].(bool); ok {
+					success = v
+				}
+				resultText := ""
+				if r, ok := dataMap["result"].(string); ok {
+					resultText = r
+				}
+				if strings.TrimSpace(resultText) == "" {
+					resultText = message
+				}
+				h.recordHitlToolExecutionResult(conversationID, toolCallID, toolName, success, resultText)
+			}
+		}
+
 		// 处理知识检索日志记录
 		if eventType == "tool_result" && h.knowledgeManager != nil {
 			if dataMap, ok := data.(map[string]interface{}); ok {
@@ -1179,6 +1227,7 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 					respPlan.meta[k] = v
 				}
 			}
+			syncHitlCognition()
 			return
 		}
 		if eventType == "response" {
@@ -1248,6 +1297,7 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 					}
 				}
 			}
+			syncHitlCognition()
 			return
 		}
 
@@ -1352,6 +1402,7 @@ func (h *AgentHandler) cancelToolContinueAfter(conversationID, preferredExecID,
 func (h *AgentHandler) CancelAgentLoop(c *gin.Context) {
 	var req struct {
 		ConversationID string `json:"conversationId" binding:"required"`
+		ExecutionID    string `json:"executionId,omitempty"`
 		Reason         string `json:"reason,omitempty"`
 		ContinueAfter  bool   `json:"continueAfter,omitempty"`
 	}
@@ -1368,7 +1419,7 @@ func (h *AgentHandler) CancelAgentLoop(c *gin.Context) {
 		}
 		note := strings.TrimSpace(req.Reason)
 		activeExec := strings.TrimSpace(h.tasks.ActiveMCPExecutionID(req.ConversationID))
-		if ok, payload := h.cancelToolContinueAfter(req.ConversationID, "", note); ok {
+		if ok, payload := h.cancelToolContinueAfter(req.ConversationID, strings.TrimSpace(req.ExecutionID), note); ok {
 			execID, _ := payload["executionId"].(string)
 			h.logger.Info("对话页仅终止当前工具",
 				zap.String("conversationId", req.ConversationID),
@@ -1411,6 +1462,8 @@ func (h *AgentHandler) CancelAgentLoop(c *gin.Context) {
 
 	var cause error = ErrTaskCancelled
 	msg := "已提交取消请求，任务将在当前步骤完成后停止。"
+	h.cancelActiveMCPToolForConversation(req.ConversationID)
+	h.tasks.AbortActiveEinoExecute(req.ConversationID, "")
 	ok, err := h.tasks.CancelTask(req.ConversationID, cause)
 	if err != nil {
 		h.logger.Error("取消任务失败", zap.Error(err))
@@ -1477,17 +1530,51 @@ func (h *AgentHandler) SubscribeAgentTaskEvents(c *gin.Context) {
 	}
 }
 
+// enrichAgentTasksWithConversationTitles 为任务列表附加当前会话标题（供顶栏/任务页展示，重命名后自动同步）
+func (h *AgentHandler) enrichAgentTasksWithConversationTitles(tasks []*AgentTask) {
+	if h == nil || h.db == nil {
+		return
+	}
+	for _, task := range tasks {
+		if task == nil || strings.TrimSpace(task.ConversationID) == "" {
+			continue
+		}
+		if title, err := h.db.GetConversationTitle(task.ConversationID); err == nil {
+			task.Title = strings.TrimSpace(title)
+		}
+	}
+}
+
+// enrichCompletedTasksWithConversationTitles 为已完成任务附加当前会话标题
+func (h *AgentHandler) enrichCompletedTasksWithConversationTitles(tasks []*CompletedTask) {
+	if h == nil || h.db == nil {
+		return
+	}
+	for _, task := range tasks {
+		if task == nil || strings.TrimSpace(task.ConversationID) == "" {
+			continue
+		}
+		if title, err := h.db.GetConversationTitle(task.ConversationID); err == nil {
+			task.Title = strings.TrimSpace(title)
+		}
+	}
+}
+
 // ListAgentTasks 列出所有运行中的任务
 func (h *AgentHandler) ListAgentTasks(c *gin.Context) {
+	tasks := h.tasks.GetActiveTasks()
+	h.enrichAgentTasksWithConversationTitles(tasks)
 	c.JSON(http.StatusOK, gin.H{
-		"tasks": h.tasks.GetActiveTasks(),
+		"tasks": tasks,
 	})
 }
 
 // ListCompletedTasks 列出最近完成的任务历史
 func (h *AgentHandler) ListCompletedTasks(c *gin.Context) {
+	tasks := h.tasks.GetCompletedTasks()
+	h.enrichCompletedTasksWithConversationTitles(tasks)
 	c.JSON(http.StatusOK, gin.H{
-		"tasks": h.tasks.GetCompletedTasks(),
+		"tasks": tasks,
 	})
 }
 

internal/handler/config.go

@@ -798,6 +798,10 @@ func (h *ConfigHandler) UpdateConfig(c *gin.Context) {
 
 	// 更新机器人配置
 	if req.Robots != nil {
+		if err := config.ValidateWecomConfig(req.Robots.Wecom); err != nil {
+			c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+			return
+		}
 		h.config.Robots = *req.Robots
 		h.logger.Info("更新机器人配置",
 			zap.Bool("wechat_enabled", h.config.Robots.Wechat.Enabled),
@@ -1329,6 +1333,17 @@ func (h *ConfigHandler) ApplyConfig(c *gin.Context) {
 		h.logger.Info("已更新嵌入模型配置记录")
 	}
 
+	// 从 tools 目录重新加载工具配置（新增/修改/删除 yaml 后无需重启）
+	if err := config.ReloadSecurityToolsFromDir(h.config, h.configPath); err != nil {
+		h.logger.Error("重新加载工具配置失败", zap.Error(err))
+		if h.audit != nil {
+			h.audit.RecordFail(c, "config", "apply", "应用配置失败：重新加载工具", map[string]interface{}{"error": err.Error()})
+		}
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "重新加载工具配置失败: " + err.Error()})
+		return
+	}
+	h.logger.Info("已从 tools 目录重新加载工具配置", zap.Int("tools_count", len(h.config.Security.Tools)))
+
 	// 重新注册工具（根据新的启用状态）
 	h.logger.Info("重新注册工具")
 
@@ -1417,12 +1432,7 @@ func (h *ConfigHandler) ApplyConfig(c *gin.Context) {
 
 	// 更新检索器配置（如果知识库启用）
 	if h.config.Knowledge.Enabled && h.retrieverUpdater != nil {
-		retrievalConfig := &knowledge.RetrievalConfig{
+		retrievalConfig := knowledge.RetrievalConfigFromYAML(h.config.Knowledge.Retrieval)
-			TopK:                h.config.Knowledge.Retrieval.TopK,
-			SimilarityThreshold: h.config.Knowledge.Retrieval.SimilarityThreshold,
-			SubIndexFilter:      h.config.Knowledge.Retrieval.SubIndexFilter,
-			PostRetrieve:        h.config.Knowledge.Retrieval.PostRetrieve,
-		}
 		h.retrieverUpdater.UpdateConfig(retrievalConfig)
 		h.logger.Info("检索器配置已更新",
 			zap.Int("top_k", retrievalConfig.TopK),
@@ -1705,6 +1715,13 @@ func updateKnowledgeConfig(doc *yaml.Node, cfg config.KnowledgeConfig) {
 	setIntInMap(retrievalNode, "top_k", cfg.Retrieval.TopK)
 	setFloatInMap(retrievalNode, "similarity_threshold", cfg.Retrieval.SimilarityThreshold)
 	setStringInMap(retrievalNode, "sub_index_filter", cfg.Retrieval.SubIndexFilter)
+	mqNode := ensureMap(retrievalNode, "multi_query")
+	setIntInMap(mqNode, "max_queries", cfg.Retrieval.MultiQuery.MaxQueries)
+	rerankNode := ensureMap(retrievalNode, "rerank")
+	setStringInMap(rerankNode, "provider", cfg.Retrieval.Rerank.Provider)
+	setStringInMap(rerankNode, "model", cfg.Retrieval.Rerank.Model)
+	setStringInMap(rerankNode, "base_url", cfg.Retrieval.Rerank.BaseURL)
+	setStringInMap(rerankNode, "api_key", cfg.Retrieval.Rerank.APIKey)
 	postNode := ensureMap(retrievalNode, "post_retrieve")
 	setIntInMap(postNode, "prefetch_top_k", cfg.Retrieval.PostRetrieve.PrefetchTopK)
 	setIntInMap(postNode, "max_context_chars", cfg.Retrieval.PostRetrieve.MaxContextChars)
@@ -1751,6 +1768,20 @@ func mergeHitlToolWhitelistSlice(existing, add []string) []string {
 	return out
 }
 
+// SetHitlToolWhitelist 将全局免审批工具白名单整表写入 config.yaml（替换，非合并）。
+func (h *ConfigHandler) SetHitlToolWhitelist(tools []string) error {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	h.config.Hitl.ToolWhitelist = mergeHitlToolWhitelistSlice(nil, tools)
+	if err := h.saveConfig(); err != nil {
+		return err
+	}
+	h.logger.Info("HITL 全局工具白名单已写入配置文件",
+		zap.Int("count", len(h.config.Hitl.ToolWhitelist)),
+	)
+	return nil
+}
+
 // MergeHitlToolWhitelistIntoConfig 将会话侧栏提交的免审批工具名合并进内存配置并写入 config.yaml（与全局白名单去重规则一致：小写键、保留首次出现的原始大小写）。
 func (h *ConfigHandler) MergeHitlToolWhitelistIntoConfig(add []string) error {
 	h.mu.Lock()
@@ -1771,6 +1802,21 @@ func updateHitlConfig(doc *yaml.Node, cfg config.HitlConfig) {
 	hitlNode := ensureMap(root, "hitl")
 	// flow 样式 [a, b, c] 单行展示，工具多时比块序列省行数
 	setFlowStringSliceInMap(hitlNode, "tool_whitelist", cfg.ToolWhitelist)
+	setStringInMap(hitlNode, "audit_agent_prompt", cfg.AuditAgentPrompt)
+	setStringInMap(hitlNode, "audit_agent_prompt_review_edit", cfg.AuditAgentPromptReviewEdit)
+}
+
+// UpdateHitlAuditAgentStrategy 更新审批/审查编辑两套审计 Agent 提示词并写入 config.yaml。
+func (h *ConfigHandler) UpdateHitlAuditAgentStrategy(approvalPrompt, reviewEditPrompt string) error {
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	h.config.Hitl.AuditAgentPrompt = strings.TrimSpace(approvalPrompt)
+	h.config.Hitl.AuditAgentPromptReviewEdit = strings.TrimSpace(reviewEditPrompt)
+	if err := h.saveConfig(); err != nil {
+		return err
+	}
+	h.logger.Info("HITL 审计 Agent 提示词已写入配置文件")
+	return nil
 }
 
 func updateRobotsConfig(doc *yaml.Node, cfg config.RobotsConfig) {

internal/handler/conversation.go

@@ -103,6 +103,7 @@ func (h *ConversationHandler) ListConversations(c *gin.Context) {
 	limitStr := c.DefaultQuery("limit", "50")
 	offsetStr := c.DefaultQuery("offset", "0")
 	search := c.Query("search") // 获取搜索参数
+	projectID := strings.TrimSpace(c.Query("project_id"))
 
 	limit, _ := strconv.Atoi(limitStr)
 	offset, _ := strconv.Atoi(offsetStr)
@@ -114,7 +115,7 @@ func (h *ConversationHandler) ListConversations(c *gin.Context) {
 		limit = 1000
 	}
 
-	excludeGrouped := strings.TrimSpace(search) == "" &&
+	excludeGrouped := strings.TrimSpace(search) == "" && projectID == "" &&
 		(c.Query("exclude_grouped") == "true" || c.Query("exclude_grouped") == "1")
 	sortBy := strings.TrimSpace(c.Query("sort_by"))
 
@@ -122,14 +123,14 @@ func (h *ConversationHandler) ListConversations(c *gin.Context) {
 	var total int
 	var err error
 	if excludeGrouped {
-		conversations, err = h.db.ListUngroupedConversations(limit, offset, sortBy)
+		conversations, err = h.db.ListUngroupedConversations(limit, offset, sortBy, projectID)
 		if err == nil {
-			total, err = h.db.CountUngroupedConversations()
+			total, err = h.db.CountUngroupedConversations(projectID)
 		}
 	} else {
-		conversations, err = h.db.ListConversations(limit, offset, search, sortBy)
+		conversations, err = h.db.ListConversations(limit, offset, search, sortBy, projectID)
 		if err == nil {
-			total, err = h.db.CountConversations(search)
+			total, err = h.db.CountConversations(search, projectID)
 		}
 	}
 	if err != nil {

internal/handler/eino_empty_response_continue.go

@@ -0,0 +1,83 @@
+package handler
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	"cyberstrike-ai/internal/agent"
+	"cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/multiagent"
+
+	"go.uber.org/zap"
+)
+
+// rebindEinoRunningTask 中断并继续 / 空正文续跑：重建 cancel 链与超时 ctx，保持任务 running。
+func (h *AgentHandler) rebindEinoRunningTask(conversationID string, timeoutCancel context.CancelFunc) (context.Context, context.CancelCauseFunc, context.Context, context.CancelFunc) {
+	if timeoutCancel != nil {
+		timeoutCancel()
+	}
+	baseCtx, cancelWithCause := context.WithCancelCause(context.Background())
+	h.tasks.BindTaskCancel(conversationID, cancelWithCause)
+	taskCtx, newTimeoutCancel := context.WithTimeout(baseCtx, 600*time.Minute)
+	h.tasks.UpdateTaskStatus(conversationID, "running")
+	return baseCtx, cancelWithCause, taskCtx, newTimeoutCancel
+}
+
+// tryContinueOnEinoEmptyResponse Run 成功但 Response 为 emptyHint 时退避续跑；true 表示已准备下一段 Run。
+func (h *AgentHandler) tryContinueOnEinoEmptyResponse(
+	taskCtx context.Context,
+	mw *config.MultiAgentEinoMiddlewareConfig,
+	conversationID string,
+	result *multiagent.RunResult,
+	attempt *int,
+	curHistory *[]agent.ChatMessage,
+	curFinalMessage *string,
+	progressCallback func(eventType, message string, data interface{}),
+) bool {
+	if result == nil || !multiagent.IsEinoEmptyResponseResult(result) || !multiagent.HasEinoResumeTrace(result) {
+		return false
+	}
+	maxAttempts := multiagent.EmptyResponseContinueMaxAttemptsFromConfig(mw)
+	if *attempt >= maxAttempts {
+		if h.logger != nil {
+			h.logger.Warn("eino empty response continue exhausted",
+				zap.String("conversationId", conversationID),
+				zap.Int("maxAttempts", maxAttempts))
+		}
+		return false
+	}
+	*attempt++
+	h.persistEinoAgentTraceForResume(conversationID, result)
+
+	backoff := multiagent.EmptyResponseContinueBackoff(*attempt-1, mw)
+	waitMsg := fmt.Sprintf("会话已结束但未捕获到助手正文，%d 秒后第 %d/%d 次自动续跑…",
+		int(backoff.Seconds()), *attempt, maxAttempts)
+	if progressCallback != nil {
+		progressCallback("eino_empty_response_continue", waitMsg, map[string]interface{}{
+			"conversationId": conversationID,
+			"source":         "eino",
+			"attempt":        *attempt,
+			"maxAttempts":    maxAttempts,
+			"backoffSec":     int(backoff.Seconds()),
+		})
+	}
+	select {
+	case <-taskCtx.Done():
+		return false
+	case <-time.After(backoff):
+	}
+
+	inject := multiagent.FormatEmptyResponseContinueUserMessage()
+	h.applyEinoTraceResumeSegment(conversationID, result, curHistory, curFinalMessage, inject)
+	if progressCallback != nil {
+		progressCallback("eino_empty_response_continue", "已恢复上下文，正在续跑…", map[string]interface{}{
+			"conversationId": conversationID,
+			"source":         "eino",
+			"attempt":        *attempt,
+			"maxAttempts":    maxAttempts,
+			"contextSource":  "empty_response_continue",
+		})
+	}
+	return true
+}

internal/handler/eino_single_agent.go

@@ -178,6 +178,7 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 	var cumulativeMCPExecutionIDs []string
 	// 同一请求内分段续跑时，主代理 iteration 事件按偏移累计，避免 UI 出现「第3轮 → 第1轮」回跳。
 	var mainIterationOffset int
+	var emptyResponseContinueAttempt int
 
 	for {
 		segmentMainIterationMax := 0
@@ -239,6 +240,13 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 		}
 
 		if runErr == nil {
+			mw := &h.config.MultiAgent.EinoMiddleware
+			if h.tryContinueOnEinoEmptyResponse(taskCtx, mw, conversationID, result, &emptyResponseContinueAttempt, &curHistory, &curFinalMessage, progressCallback) {
+				mainIterationOffset += segmentMainIterationMax
+				timeoutCancel()
+				baseCtx, cancelWithCause, taskCtx, timeoutCancel = h.rebindEinoRunningTask(conversationID, timeoutCancel)
+				continue
+			}
 			timeoutCancel()
 			break
 		}

internal/handler/hitl.go

@@ -23,6 +23,7 @@ import (
 type hitlRuntimeConfig struct {
 	Enabled        bool
 	Mode           string
+	Reviewer       string
 	SensitiveTools map[string]struct{}
 	Timeout        time.Duration
 }
@@ -49,6 +50,8 @@ type HITLManager struct {
 	mu      sync.RWMutex
 	runtime map[string]hitlRuntimeConfig
 	pending map[string]*pendingInterrupt
+	// approvedExec 审批通过、待回写 tool_result 的队列（按会话 FIFO）
+	approvedExec map[string][]hitlApprovedExecTrack
 }
 
 func NewHITLManager(db *database.DB, logger *zap.Logger) *HITLManager {
@@ -90,6 +93,7 @@ CREATE TABLE IF NOT EXISTS hitl_conversation_configs (
 	if err != nil {
 		return err
 	}
+	m.migrateHitlSchemaColumns()
 
 	// On startup, cancel all orphaned pending interrupts from previous process.
 	// Their in-memory channels are gone, so they can never be resolved.
@@ -141,6 +145,7 @@ func (m *HITLManager) ActivateConversation(conversationID string, req *HITLReque
 	m.runtime[conversationID] = hitlRuntimeConfig{
 		Enabled:        true,
 		Mode:           normalizeHitlMode(req.Mode),
+		Reviewer:       normalizeHitlReviewer(req.Reviewer),
 		SensitiveTools: tools,
 		Timeout:        timeout,
 	}
@@ -153,17 +158,14 @@ func (m *HITLManager) DeactivateConversation(conversationID string) {
 	m.mu.Unlock()
 }
 
-// hitlConfigGlobalToolWhitelist 来自 config.yaml hitl.tool_whitelist（去重、去空）。
+// hitlConfigGlobalToolWhitelist 来自 config.yaml hitl.tool_whitelist（去重、去空），并合并内置元工具免审批项。
 func (h *AgentHandler) hitlConfigGlobalToolWhitelist() []string {
 	if h == nil || h.config == nil {
-		return nil
+		return multiagent.MergeHitlExemptMetaTools(nil)
 	}
 	raw := h.config.Hitl.ToolWhitelist
-	if len(raw) == 0 {
-		return nil
-	}
 	seen := make(map[string]struct{})
-	out := make([]string, 0, len(raw))
+	out := make([]string, 0, len(raw)+len(multiagent.HitlExemptMetaTools))
 	for _, t := range raw {
 		n := strings.ToLower(strings.TrimSpace(t))
 		if n == "" {
@@ -175,44 +177,35 @@ func (h *AgentHandler) hitlConfigGlobalToolWhitelist() []string {
 		seen[n] = struct{}{}
 		out = append(out, strings.TrimSpace(t))
 	}
-	return out
+	return multiagent.MergeHitlExemptMetaTools(out)
 }
 
-// hitlRequestWithMergedConfigWhitelist 将会话/API 中的白名单与 config.yaml 全局白名单合并（并集），仅用于运行时 Activate；不写入数据库。
+// hitlRequestWithMergedConfigWhitelist 将会话/API 中的白名单与 config.yaml 全局白名单及内置元工具免审批项合并（并集），仅用于运行时 Activate；不写入数据库。
 func (h *AgentHandler) hitlRequestWithMergedConfigWhitelist(req *HITLRequest) *HITLRequest {
-	gw := h.hitlConfigGlobalToolWhitelist()
-	if len(gw) == 0 {
-		return req
-	}
 	if req == nil {
 		return nil
 	}
 	seen := make(map[string]struct{})
-	union := make([]string, 0, len(gw)+len(req.SensitiveTools))
+	union := make([]string, 0, len(req.SensitiveTools)+16)
-	for _, t := range gw {
+	add := func(t string) {
 		n := strings.ToLower(strings.TrimSpace(t))
 		if n == "" {
-			continue
+			return
 		}
 		if _, ok := seen[n]; ok {
-			continue
+			return
 		}
 		seen[n] = struct{}{}
 		union = append(union, strings.TrimSpace(t))
 	}
+	for _, t := range h.hitlConfigGlobalToolWhitelist() {
+		add(t)
+	}
 	for _, t := range req.SensitiveTools {
-		n := strings.ToLower(strings.TrimSpace(t))
+		add(t)
-		if n == "" {
-			continue
-		}
-		if _, ok := seen[n]; ok {
-			continue
-		}
-		seen[n] = struct{}{}
-		union = append(union, strings.TrimSpace(t))
 	}
 	out := *req
-	out.SensitiveTools = union
+	out.SensitiveTools = multiagent.MergeHitlExemptMetaTools(union)
 	return &out
 }
 
@@ -362,22 +355,22 @@ func (m *HITLManager) SaveConversationConfig(conversationID string, req *HITLReq
 		timeout = 0
 	}
 	_, err := m.db.Exec(`INSERT INTO hitl_conversation_configs
-		(conversation_id, enabled, mode, sensitive_tools, timeout_seconds, updated_at)
+		(conversation_id, enabled, mode, reviewer, sensitive_tools, timeout_seconds, updated_at)
-		VALUES (?, ?, ?, ?, ?, ?)
+		VALUES (?, ?, ?, ?, ?, ?, ?)
 		ON CONFLICT(conversation_id) DO UPDATE SET
-		enabled=excluded.enabled, mode=excluded.mode, sensitive_tools=excluded.sensitive_tools, timeout_seconds=excluded.timeout_seconds, updated_at=excluded.updated_at`,
+		enabled=excluded.enabled, mode=excluded.mode, reviewer=excluded.reviewer, sensitive_tools=excluded.sensitive_tools, timeout_seconds=excluded.timeout_seconds, updated_at=excluded.updated_at`,
-		conversationID, boolToInt(req.Enabled), mode, string(tools), timeout, time.Now())
+		conversationID, boolToInt(req.Enabled), mode, normalizeHitlReviewer(req.Reviewer), string(tools), timeout, time.Now())
 	return err
 }
 
 func (m *HITLManager) LoadConversationConfig(conversationID string) (*HITLRequest, error) {
 	var enabledInt int
-	var mode, toolsJSON string
+	var mode, reviewer, toolsJSON string
 	var timeout int
-	err := m.db.QueryRow(`SELECT enabled, mode, sensitive_tools, timeout_seconds FROM hitl_conversation_configs WHERE conversation_id = ?`, conversationID).
+	err := m.db.QueryRow(`SELECT enabled, mode, COALESCE(reviewer,'human'), sensitive_tools, timeout_seconds FROM hitl_conversation_configs WHERE conversation_id = ?`, conversationID).
-		Scan(&enabledInt, &mode, &toolsJSON, &timeout)
+		Scan(&enabledInt, &mode, &reviewer, &toolsJSON, &timeout)
 	if errors.Is(err, sql.ErrNoRows) {
-		return &HITLRequest{Enabled: false, Mode: "off", SensitiveTools: []string{}, TimeoutSeconds: 0}, nil
+		return &HITLRequest{Enabled: false, Mode: "off", Reviewer: "human", SensitiveTools: []string{}, TimeoutSeconds: 0}, nil
 	}
 	if err != nil {
 		return nil, err
@@ -390,6 +383,7 @@ func (m *HITLManager) LoadConversationConfig(conversationID string) (*HITLReques
 	return &HITLRequest{
 		Enabled:        enabledInt == 1,
 		Mode:           mode,
+		Reviewer:       normalizeHitlReviewer(reviewer),
 		SensitiveTools: tools,
 		TimeoutSeconds: timeout,
 	}, nil
@@ -413,15 +407,16 @@ func (m *HITLManager) waitDecision(ctx context.Context, p *pendingInterrupt, tim
 		if p.Mode != "review_edit" && len(d.EditedArguments) > 0 {
 			d.EditedArguments = nil
 		}
-		_, _ = m.db.Exec(`UPDATE hitl_interrupts SET status='decided', decision=?, decision_comment=?, decided_at=? WHERE id=?`,
+		_, _ = m.db.Exec(`UPDATE hitl_interrupts SET status='decided', decision=?, decision_comment=?, decided_at=?, decided_by='human' WHERE id=?`,
 			d.Decision, d.Comment, time.Now(), p.InterruptID)
 		return d, nil
 	case <-timeoutCh:
-		_, _ = m.db.Exec(`UPDATE hitl_interrupts SET status='timeout', decision='approve', decision_comment='timeout auto approve', decided_at=? WHERE id=?`,
+		comment := "HITL timeout auto-reject for safety"
-			time.Now(), p.InterruptID)
+		_, _ = m.db.Exec(`UPDATE hitl_interrupts SET status='timeout', decision='reject', decision_comment=?, decided_at=?, decided_by='system' WHERE id=?`,
-		return hitlDecision{Decision: "approve", Comment: "timeout auto approve"}, nil
+			comment, time.Now(), p.InterruptID)
+		return hitlDecision{Decision: "reject", Comment: comment}, nil
 	case <-ctx.Done():
-		_, _ = m.db.Exec(`UPDATE hitl_interrupts SET status='cancelled', decision='reject', decision_comment='task cancelled', decided_at=? WHERE id=?`,
+		_, _ = m.db.Exec(`UPDATE hitl_interrupts SET status='cancelled', decision='reject', decision_comment='task cancelled', decided_at=?, decided_by='system' WHERE id=?`,
 			time.Now(), p.InterruptID)
 		return hitlDecision{Decision: "reject", Comment: "task cancelled"}, ctx.Err()
 	}
@@ -445,12 +440,57 @@ func (h *AgentHandler) waitHITLApproval(runCtx context.Context, cancelRun contex
 	if !need {
 		return nil, nil
 	}
+	h.enrichHitlApprovalPayload(conversationID, assistantMessageID, payload)
 	payloadRaw, _ := json.Marshal(payload)
 	p, err := h.hitlManager.CreatePendingInterrupt(conversationID, assistantMessageID, cfg.Mode, toolName, toolCallID, string(payloadRaw))
 	if err != nil {
 		h.logger.Warn("创建 HITL 中断失败", zap.Error(err))
 		return nil, err
 	}
+
+	if cfg.Reviewer == "audit_agent" {
+		ad := h.auditAgentReview(runCtx, cfg.Mode, toolName, payload)
+		now := time.Now()
+		_, _ = h.db.Exec(`UPDATE hitl_interrupts SET status='decided', decision=?, decision_comment=?, decided_at=?, decided_by='audit_agent' WHERE id=?`,
+			ad.Decision, ad.Comment, now, p.InterruptID)
+		if sendEventFunc != nil {
+			sendEventFunc("hitl_audit_agent", "审计 Agent 已裁决", map[string]interface{}{
+				"conversationId": conversationID,
+				"interruptId":    p.InterruptID,
+				"toolName":       toolName,
+				"mode":           cfg.Mode,
+				"decision":       ad.Decision,
+				"comment":        ad.Comment,
+				"editedArgs":     ad.EditedArguments,
+				"decidedBy":      "audit_agent",
+			})
+		}
+		if ad.Decision == "reject" {
+			if sendEventFunc != nil {
+				sendEventFunc("hitl_rejected", "审计 Agent 拒绝本次工具调用", map[string]interface{}{
+					"conversationId": conversationID,
+					"interruptId":    p.InterruptID,
+					"toolName":       toolName,
+					"comment":        ad.Comment,
+					"decidedBy":      "audit_agent",
+				})
+			}
+			return &ad, nil
+		}
+		if sendEventFunc != nil {
+			sendEventFunc("hitl_resumed", "审计 Agent 已通过，继续执行", map[string]interface{}{
+				"conversationId": conversationID,
+				"interruptId":    p.InterruptID,
+				"toolName":       toolName,
+				"comment":        ad.Comment,
+				"editedArgs":     ad.EditedArguments,
+				"decidedBy":      "audit_agent",
+			})
+		}
+		h.hitlManager.TrackApprovedHitlExecution(p.InterruptID, conversationID, toolName, toolCallID)
+		return &ad, nil
+	}
+
 	if sendEventFunc != nil {
 		sendEventFunc("hitl_interrupt", "命中人机协同审批", map[string]interface{}{
 			"conversationId": conversationID,
@@ -479,8 +519,12 @@ func (h *AgentHandler) waitHITLApproval(runCtx context.Context, cancelRun contex
 		return nil, waitErr
 	}
 	if d.Decision == "reject" {
+		rejectMsg := "人工拒绝本次工具调用，模型将基于反馈继续迭代"
+		if strings.Contains(strings.ToLower(strings.TrimSpace(d.Comment)), "timeout") {
+			rejectMsg = "审批超时，安全起见已自动拒绝，模型将基于反馈继续迭代"
+		}
 		if sendEventFunc != nil {
-			sendEventFunc("hitl_rejected", "人工拒绝本次工具调用，模型将基于反馈继续迭代", map[string]interface{}{
+			sendEventFunc("hitl_rejected", rejectMsg, map[string]interface{}{
 				"conversationId": conversationID,
 				"interruptId":    p.InterruptID,
 				"toolName":       toolName,
@@ -498,6 +542,7 @@ func (h *AgentHandler) waitHITLApproval(runCtx context.Context, cancelRun contex
 			"editedArgs":     d.EditedArguments,
 		})
 	}
+	h.hitlManager.TrackApprovedHitlExecution(p.InterruptID, conversationID, toolName, toolCallID)
 	return &d, nil
 }
 
@@ -527,11 +572,6 @@ func (h *AgentHandler) handleHITLToolCall(runCtx context.Context, cancelRun cont
 }
 
 func (h *AgentHandler) ListHITLPending(c *gin.Context) {
-	conversationID := strings.TrimSpace(c.Query("conversationId"))
-	status := strings.TrimSpace(c.Query("status"))
-	if status == "" {
-		status = "pending"
-	}
 	page, _ := strconv.Atoi(c.DefaultQuery("page", "1"))
 	if page < 1 {
 		page = 1
@@ -539,15 +579,12 @@ func (h *AgentHandler) ListHITLPending(c *gin.Context) {
 	pageSize, _ := strconv.Atoi(c.DefaultQuery("pageSize", "20"))
 	pageSize = int(math.Max(1, math.Min(float64(pageSize), 200)))
 	offset := (page - 1) * pageSize
-	q := `SELECT id, conversation_id, message_id, mode, tool_name, tool_call_id, payload, status, decision, decision_comment, created_at, decided_at FROM hitl_interrupts WHERE 1=1`
+	q, args := h.buildHitlListQuery(false)
-	args := []interface{}{}
+	q, args = h.appendHitlListFilters(q, args, c)
-	if conversationID != "" {
+	total, err := h.countHitlQuery(q, args)
-		q += " AND conversation_id = ?"
+	if err != nil {
-		args = append(args, conversationID)
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-	}
+		return
-	if status != "all" {
-		q += " AND status = ?"
-		args = append(args, status)
 	}
 	q += " ORDER BY created_at DESC LIMIT ? OFFSET ?"
 	args = append(args, pageSize, offset)
@@ -557,41 +594,12 @@ func (h *AgentHandler) ListHITLPending(c *gin.Context) {
 		return
 	}
 	defer rows.Close()
-	items := make([]map[string]interface{}, 0)
+	items, err := h.scanHitlInterruptRows(rows)
-	for rows.Next() {
+	if err != nil {
-		var id, cid, mode, toolName, toolCallID, payload, rowStatus string
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		var messageID sql.NullString
+		return
-		var decision, comment sql.NullString
-		var createdAt time.Time
-		var decidedAt sql.NullTime
-		if err := rows.Scan(&id, &cid, &messageID, &mode, &toolName, &toolCallID, &payload, &rowStatus, &decision, &comment, &createdAt, &decidedAt); err != nil {
-			continue
-		}
-		msgID := ""
-		if messageID.Valid {
-			msgID = messageID.String
-		}
-		items = append(items, map[string]interface{}{
-			"id":             id,
-			"conversationId": cid,
-			"messageId":      msgID,
-			"mode":           mode,
-			"toolName":       toolName,
-			"toolCallId":     toolCallID,
-			"payload":        payload,
-			"status":         rowStatus,
-			"decision":       decision.String,
-			"comment":        comment.String,
-			"createdAt":      createdAt,
-			"decidedAt": func() interface{} {
-				if decidedAt.Valid {
-					return decidedAt.Time
-				}
-				return nil
-			}(),
-		})
 	}
-	c.JSON(http.StatusOK, gin.H{"items": items, "page": page, "pageSize": pageSize})
+	c.JSON(http.StatusOK, gin.H{"items": items, "page": page, "pageSize": pageSize, "total": total})
 }
 
 type hitlDecisionReq struct {
@@ -636,7 +644,7 @@ func (h *AgentHandler) DismissHITLInterrupt(c *gin.Context) {
 		return
 	}
 	res, err := h.db.Exec(`UPDATE hitl_interrupts SET status='cancelled', decision='reject',
-		decision_comment='dismissed by user', decided_at=CURRENT_TIMESTAMP
+		decision_comment='dismissed by user', decided_at=CURRENT_TIMESTAMP, decided_by='human'
 		WHERE id=? AND status='pending'`, req.InterruptID)
 	if err != nil {
 		c.JSON(500, gin.H{"error": err.Error()})
@@ -732,6 +740,7 @@ func (h *AgentHandler) UpsertHITLConversationConfig(c *gin.Context) {
 		return
 	}
 	req.Mode = normalizeHitlMode(req.Mode)
+	req.Reviewer = normalizeHitlReviewer(req.Reviewer)
 	if err := h.hitlManager.SaveConversationConfig(req.ConversationID, &req.HITLRequest); err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
@@ -753,6 +762,44 @@ type mergeHitlGlobalWhitelistReq struct {
 	SensitiveTools []string `json:"sensitiveTools"`
 }
 
+type setHitlGlobalWhitelistReq struct {
+	ToolWhitelist []string `json:"toolWhitelist"`
+}
+
+// GetHITLGlobalToolWhitelist 返回 config.yaml 中的全局免审批工具白名单。
+func (h *AgentHandler) GetHITLGlobalToolWhitelist(c *gin.Context) {
+	c.JSON(http.StatusOK, gin.H{
+		"toolWhitelist": h.hitlConfigGlobalToolWhitelist(),
+	})
+}
+
+// SetHITLGlobalToolWhitelist 整表替换 config.yaml 中的全局免审批工具白名单。
+func (h *AgentHandler) SetHITLGlobalToolWhitelist(c *gin.Context) {
+	if h.hitlWhitelistSaver == nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "HITL 配置持久化不可用"})
+		return
+	}
+	var req setHitlGlobalWhitelistReq
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+	if err := h.hitlWhitelistSaver.SetHitlToolWhitelist(req.ToolWhitelist); err != nil {
+		h.logger.Warn("写入 HITL 工具白名单到 config.yaml 失败", zap.Error(err))
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	if h.audit != nil {
+		h.audit.RecordOK(c, "hitl", "tool_whitelist_update", "HITL 全局白名单更新", "hitl_config", "tool_whitelist", nil)
+	}
+	c.JSON(http.StatusOK, gin.H{
+		"ok":                      true,
+		"toolWhitelist":           h.hitlConfigGlobalToolWhitelist(),
+		"hitlGlobalToolWhitelist":   h.hitlConfigGlobalToolWhitelist(),
+		"hitlGlobalWhitelistMerged": false,
+	})
+}
+
 // MergeHITLGlobalToolWhitelist 无会话 ID 时将侧栏提交的免审批工具合并进 config.yaml（与 PUT /hitl/config 中白名单落盘规则一致）。
 func (h *AgentHandler) MergeHITLGlobalToolWhitelist(c *gin.Context) {
 	if h.hitlWhitelistSaver == nil {

internal/handler/hitl_audit_agent.go

@@ -0,0 +1,357 @@
+package handler
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"cyberstrike-ai/internal/config"
+
+	"github.com/gin-gonic/gin"
+	"go.uber.org/zap"
+)
+
+// auditAgentReview 在 reviewer=audit_agent 时由 LLM 代行审批。
+// 白名单工具在 shouldInterrupt 阶段已跳过，到达此处的一律需要裁决。
+func (h *AgentHandler) auditAgentReview(ctx context.Context, hitlMode, toolName string, payload map[string]interface{}) hitlDecision {
+	if h == nil {
+		return hitlDecision{Decision: "reject", Comment: "audit agent: handler unavailable"}
+	}
+	mode := normalizeHitlMode(hitlMode)
+	prompt := config.DefaultHitlAuditAgentPrompt()
+	if h.config != nil {
+		prompt = h.config.Hitl.EffectiveAuditAgentPromptForMode(mode)
+	}
+	if h.auditLLM == nil {
+		return hitlDecision{Decision: "reject", Comment: "audit agent: LLM 未配置"}
+	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	callCtx, cancel := context.WithTimeout(ctx, 90*time.Second)
+	defer cancel()
+
+	userContent := buildAuditAgentReviewInput(mode, toolName, payload)
+	requestBody := map[string]interface{}{
+		"model": h.auditLLMModel(),
+		"messages": []map[string]interface{}{
+			{"role": "system", "content": prompt},
+			{"role": "user", "content": userContent},
+		},
+		"temperature":           0.1,
+		"max_completion_tokens": 1024,
+		// 审计裁决需要结构化 JSON；关闭 thinking 避免 Qwen 等把正文放进 reasoning_content 导致解析失败。
+		"thinking": map[string]interface{}{"type": "disabled"},
+	}
+
+	var apiResponse struct {
+		Choices []struct {
+			Message struct {
+				Content          string `json:"content"`
+				ReasoningContent string `json:"reasoning_content"`
+			} `json:"message"`
+		} `json:"choices"`
+	}
+	if err := h.auditLLM.ChatCompletion(callCtx, requestBody, &apiResponse); err != nil {
+		h.logger.Warn("审计 Agent LLM 调用失败", zap.Error(err), zap.String("tool", toolName))
+		return hitlDecision{
+			Decision: "reject",
+			Comment:  "audit agent: LLM 调用失败，保守拒绝",
+		}
+	}
+	if len(apiResponse.Choices) == 0 {
+		return hitlDecision{Decision: "reject", Comment: "audit agent: LLM 无有效响应，保守拒绝"}
+	}
+	msg := apiResponse.Choices[0].Message
+	raw := strings.TrimSpace(msg.Content)
+	if raw == "" {
+		raw = strings.TrimSpace(msg.ReasoningContent)
+	}
+	dec, err := parseAuditAgentLLMContent(raw)
+	if err != nil {
+		snippet := raw
+		if len(snippet) > 240 {
+			snippet = snippet[:240] + "..."
+		}
+		h.logger.Warn("审计 Agent 响应解析失败",
+			zap.Error(err),
+			zap.String("tool", toolName),
+			zap.String("mode", mode),
+			zap.String("snippet", snippet),
+		)
+		return hitlDecision{Decision: "reject", Comment: "audit agent: 响应无法解析，保守拒绝"}
+	}
+	if mode != "review_edit" && len(dec.EditedArguments) > 0 {
+		h.logger.Warn("审计 Agent 在审批模式下返回 editedArguments，已忽略",
+			zap.String("tool", toolName),
+		)
+		dec.EditedArguments = nil
+	}
+	if dec.Comment == "" {
+		dec.Comment = "audit agent: " + dec.Decision
+	} else if !strings.HasPrefix(strings.ToLower(dec.Comment), "audit agent") {
+		dec.Comment = "audit agent: " + dec.Comment
+	}
+	return dec
+}
+
+func (h *AgentHandler) auditLLMModel() string {
+	if h.config != nil && strings.TrimSpace(h.config.OpenAI.Model) != "" {
+		return strings.TrimSpace(h.config.OpenAI.Model)
+	}
+	return ""
+}
+
+func buildAuditAgentReviewInput(hitlMode, toolName string, payload map[string]interface{}) string {
+	review := map[string]interface{}{
+		"hitlMode": normalizeHitlMode(hitlMode),
+		"toolName": strings.TrimSpace(toolName),
+	}
+	if payload != nil {
+		for _, k := range []string{"arguments", "argumentsObj", "command", hitlPayloadUserMessage, hitlPayloadThinking, hitlPayloadReasoningChain, hitlPayloadPlanning} {
+			if v, ok := payload[k]; ok && v != nil && fmt.Sprint(v) != "" {
+				review[k] = v
+			}
+		}
+	}
+	b, err := json.MarshalIndent(review, "", "  ")
+	if err != nil {
+		return fmt.Sprintf(`{"hitlMode":%q,"toolName":%q}`, normalizeHitlMode(hitlMode), toolName)
+	}
+	return string(b)
+}
+
+func parseAuditAgentLLMContent(content string) (hitlDecision, error) {
+	s := strings.TrimSpace(content)
+	if s == "" {
+		return hitlDecision{}, errors.New("empty content")
+	}
+	for _, candidate := range auditAgentJSONCandidates(s) {
+		dec, comment, editedArgs, err := parseAuditAgentDecisionObject(candidate)
+		if err == nil {
+			return hitlDecision{
+				Decision:        dec,
+				Comment:         comment,
+				EditedArguments: editedArgs,
+			}, nil
+		}
+	}
+	return hitlDecision{}, fmt.Errorf("no valid decision json in response")
+}
+
+func auditAgentJSONCandidates(s string) []string {
+	out := make([]string, 0, 4)
+	seen := make(map[string]struct{})
+	add := func(c string) {
+		c = strings.TrimSpace(c)
+		if c == "" {
+			return
+		}
+		if _, ok := seen[c]; ok {
+			return
+		}
+		seen[c] = struct{}{}
+		out = append(out, c)
+	}
+	add(s)
+	add(stripMarkdownCodeFence(s))
+	if obj := extractFirstJSONObject(s); obj != "" {
+		add(obj)
+	}
+	if obj := extractFirstJSONObject(stripMarkdownCodeFence(s)); obj != "" {
+		add(obj)
+	}
+	return out
+}
+
+func stripMarkdownCodeFence(s string) string {
+	s = strings.TrimSpace(s)
+	for _, fence := range []string{"```json", "```JSON", "```"} {
+		if strings.HasPrefix(s, fence) {
+			s = strings.TrimPrefix(s, fence)
+		}
+	}
+	s = strings.TrimSuffix(s, "```")
+	return strings.TrimSpace(s)
+}
+
+func extractFirstJSONObject(s string) string {
+	start := strings.Index(s, "{")
+	if start < 0 {
+		return ""
+	}
+	depth := 0
+	inStr := false
+	esc := false
+	for i := start; i < len(s); i++ {
+		ch := s[i]
+		if inStr {
+			if esc {
+				esc = false
+				continue
+			}
+			if ch == '\\' {
+				esc = true
+				continue
+			}
+			if ch == '"' {
+				inStr = false
+			}
+			continue
+		}
+		switch ch {
+		case '"':
+			inStr = true
+		case '{':
+			depth++
+		case '}':
+			depth--
+			if depth == 0 {
+				return s[start : i+1]
+			}
+		}
+	}
+	return ""
+}
+
+func parseAuditAgentDecisionObject(jsonText string) (decision, comment string, editedArgs map[string]interface{}, err error) {
+	var parsed map[string]interface{}
+	if err := json.Unmarshal([]byte(jsonText), &parsed); err != nil {
+		return "", "", nil, err
+	}
+	rawDecision := auditAgentPickString(parsed, "decision", "Decision", "result", "action", "verdict", "决策", "决定")
+	decision = normalizeAuditAgentDecision(rawDecision)
+	if decision == "" {
+		return "", "", nil, fmt.Errorf("missing decision")
+	}
+	comment = auditAgentPickString(parsed, "comment", "Comment", "reason", "message", "rationale", "备注", "理由", "说明")
+	editedArgs = auditAgentPickObject(parsed, "editedArguments", "edited_arguments", "editedArgs")
+	return decision, strings.TrimSpace(comment), editedArgs, nil
+}
+
+func auditAgentPickString(m map[string]interface{}, keys ...string) string {
+	for _, k := range keys {
+		if v, ok := m[k]; ok && v != nil {
+			s := strings.TrimSpace(fmt.Sprint(v))
+			if s != "" {
+				return s
+			}
+		}
+	}
+	return ""
+}
+
+func auditAgentPickObject(m map[string]interface{}, keys ...string) map[string]interface{} {
+	for _, k := range keys {
+		v, ok := m[k]
+		if !ok || v == nil {
+			continue
+		}
+		switch t := v.(type) {
+		case map[string]interface{}:
+			if len(t) > 0 {
+				return t
+			}
+		case string:
+			s := strings.TrimSpace(t)
+			if s == "" || s == "{}" {
+				continue
+			}
+			var obj map[string]interface{}
+			if err := json.Unmarshal([]byte(s), &obj); err == nil && len(obj) > 0 {
+				return obj
+			}
+		}
+	}
+	return nil
+}
+
+func normalizeAuditAgentDecision(v string) string {
+	d := strings.ToLower(strings.TrimSpace(v))
+	switch d {
+	case "approve", "approved", "pass", "passed", "allow", "allowed", "yes", "ok", "accept", "accepted":
+		return "approve"
+	case "reject", "rejected", "deny", "denied", "no", "block", "blocked", "refuse", "refused":
+		return "reject"
+	}
+	switch strings.TrimSpace(v) {
+	case "通过", "批准", "允许", "同意", "放行":
+		return "approve"
+	case "拒绝", "驳回", "禁止", "否决":
+		return "reject"
+	}
+	return ""
+}
+
+type hitlAuditStrategyReq struct {
+	AuditAgentPrompt           string `json:"auditAgentPrompt"`
+	AuditAgentPromptReviewEdit string `json:"auditAgentPromptReviewEdit"`
+}
+
+func (h *AgentHandler) GetHITLAuditStrategy(c *gin.Context) {
+	approvalPrompt := config.DefaultHitlAuditAgentPrompt()
+	reviewEditPrompt := config.DefaultHitlAuditAgentPromptReviewEdit()
+	approvalCustom := false
+	reviewEditCustom := false
+	if h.config != nil {
+		approvalPrompt = h.config.Hitl.EffectiveAuditAgentPromptForMode("approval")
+		reviewEditPrompt = h.config.Hitl.EffectiveAuditAgentPromptForMode("review_edit")
+		approvalCustom = strings.TrimSpace(h.config.Hitl.AuditAgentPrompt) != ""
+		reviewEditCustom = strings.TrimSpace(h.config.Hitl.AuditAgentPromptReviewEdit) != ""
+	}
+	c.JSON(http.StatusOK, gin.H{
+		"auditAgentPrompt":                  approvalPrompt,
+		"auditAgentPromptCustom":            approvalCustom,
+		"auditAgentPromptReviewEdit":        reviewEditPrompt,
+		"auditAgentPromptReviewEditCustom":  reviewEditCustom,
+		"defaultAuditAgentPrompt":           config.DefaultHitlAuditAgentPrompt(),
+		"defaultAuditAgentPromptReviewEdit": config.DefaultHitlAuditAgentPromptReviewEdit(),
+	})
+}
+
+func (h *AgentHandler) UpdateHITLAuditStrategy(c *gin.Context) {
+	if h.hitlStrategySaver == nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "HITL 策略持久化不可用"})
+		return
+	}
+	var req hitlAuditStrategyReq
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+	approvalPrompt := strings.TrimSpace(req.AuditAgentPrompt)
+	reviewEditPrompt := strings.TrimSpace(req.AuditAgentPromptReviewEdit)
+	if err := h.hitlStrategySaver.UpdateHitlAuditAgentStrategy(approvalPrompt, reviewEditPrompt); err != nil {
+		h.logger.Warn("保存审计 Agent 提示词失败", zap.Error(err))
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	if h.audit != nil {
+		h.audit.RecordOK(c, "hitl", "audit_strategy_update", "HITL 审计策略更新", "hitl_config", "audit_agent_prompt", nil)
+	}
+	if h.config != nil {
+		h.config.Hitl.AuditAgentPrompt = approvalPrompt
+		h.config.Hitl.AuditAgentPromptReviewEdit = reviewEditPrompt
+	}
+	c.JSON(http.StatusOK, gin.H{
+		"ok":                                true,
+		"auditAgentPrompt":                  config.HitlConfig{AuditAgentPrompt: approvalPrompt}.EffectiveAuditAgentPromptForMode("approval"),
+		"auditAgentPromptCustom":            approvalPrompt != "",
+		"auditAgentPromptReviewEdit":        config.HitlConfig{AuditAgentPromptReviewEdit: reviewEditPrompt}.EffectiveAuditAgentPromptForMode("review_edit"),
+		"auditAgentPromptReviewEditCustom":  reviewEditPrompt != "",
+	})
+}
+
+// HitlAuditStrategySaver 持久化审计 Agent 提示词到 config.yaml。
+type HitlAuditStrategySaver interface {
+	UpdateHitlAuditAgentStrategy(approvalPrompt, reviewEditPrompt string) error
+}
+
+// SetHitlAuditStrategySaver 设置审计策略落盘。
+func (h *AgentHandler) SetHitlAuditStrategySaver(s HitlAuditStrategySaver) {
+	h.hitlStrategySaver = s
+}

internal/handler/hitl_audit_agent_test.go

@@ -0,0 +1,88 @@
+package handler
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestParseAuditAgentLLMContentApprove(t *testing.T) {
+	d, err := parseAuditAgentLLMContent(`{"decision":"approve","comment":"与任务一致"}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if d.Decision != "approve" || d.Comment != "与任务一致" {
+		t.Fatalf("unexpected %+v", d)
+	}
+}
+
+func TestParseAuditAgentLLMContentReject(t *testing.T) {
+	d, err := parseAuditAgentLLMContent("```json\n{\"decision\":\"reject\",\"comment\":\"风险过高\"}\n```")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if d.Decision != "reject" {
+		t.Fatalf("expected reject, got %s", d.Decision)
+	}
+}
+
+func TestParseAuditAgentLLMContentInvalid(t *testing.T) {
+	_, err := parseAuditAgentLLMContent(`{"decision":"maybe"}`)
+	if err == nil {
+		t.Fatal("expected error for invalid decision")
+	}
+}
+
+func TestParseAuditAgentLLMContentProseWrapped(t *testing.T) {
+	d, err := parseAuditAgentLLMContent("好的，裁决如下：\n```json\n{\"decision\":\"approve\",\"comment\":\"只读 ls\"}\n```\n以上。")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if d.Decision != "approve" {
+		t.Fatalf("expected approve, got %s", d.Decision)
+	}
+}
+
+func TestParseAuditAgentLLMContentChineseDecision(t *testing.T) {
+	d, err := parseAuditAgentLLMContent(`{"decision":"通过","comment":"风险低"}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if d.Decision != "approve" {
+		t.Fatalf("expected approve, got %s", d.Decision)
+	}
+}
+
+func TestParseAuditAgentLLMContentWithEditedArguments(t *testing.T) {
+	d, err := parseAuditAgentLLMContent(`{"decision":"approve","comment":"收窄路径","editedArguments":{"path":"/safe"}}`)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if d.Decision != "approve" {
+		t.Fatalf("expected approve, got %s", d.Decision)
+	}
+	if d.EditedArguments == nil || d.EditedArguments["path"] != "/safe" {
+		t.Fatalf("unexpected edited args: %+v", d.EditedArguments)
+	}
+}
+
+func TestBuildAuditAgentReviewInputIncludesMode(t *testing.T) {
+	s := buildAuditAgentReviewInput("review_edit", "execute", map[string]interface{}{
+		"arguments": `{"command":"pwd"}`,
+	})
+	if !strings.Contains(s, "review_edit") || !strings.Contains(s, "execute") {
+		t.Fatalf("unexpected input: %s", s)
+	}
+}
+
+func TestBuildAuditAgentReviewInput(t *testing.T) {
+	s := buildAuditAgentReviewInput("approval", "nmap", map[string]interface{}{
+		"arguments":   `{"target":"10.0.0.1"}`,
+		"userMessage": "扫描内网",
+	})
+	if s == "" {
+		t.Fatal("expected non-empty input")
+	}
+	if !strings.Contains(s, "nmap") || !strings.Contains(s, "10.0.0.1") || !strings.Contains(s, "扫描内网") {
+		t.Fatalf("unexpected input: %s", s)
+	}
+}

internal/handler/hitl_cognition.go

@@ -0,0 +1,97 @@
+package handler
+
+import (
+	"strings"
+)
+
+type hitlCognitionState struct {
+	AssistantMessageID string
+	UserMessage        string
+	Thinking           string
+	ReasoningChain     string
+	Planning           string
+}
+
+// GetHitlCognition 返回当前运行任务上缓存的本轮 HITL 上下文（不含会话历史）。
+func (m *AgentTaskManager) GetHitlCognition(conversationID string) hitlCognitionFields {
+	conversationID = strings.TrimSpace(conversationID)
+	if m == nil || conversationID == "" {
+		return hitlCognitionFields{}
+	}
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	t, ok := m.tasks[conversationID]
+	if !ok || t == nil || t.hitlCognition == nil {
+		return hitlCognitionFields{}
+	}
+	c := t.hitlCognition
+	return hitlCognitionFields{
+		UserMessage:    c.UserMessage,
+		Thinking:       c.Thinking,
+		ReasoningChain: c.ReasoningChain,
+		Planning:       c.Planning,
+	}
+}
+
+// ResetHitlCognition 新任务开始时重置本轮 HITL 上下文。
+func (m *AgentTaskManager) ResetHitlCognition(conversationID, userMessage string) {
+	conversationID = strings.TrimSpace(conversationID)
+	if m == nil || conversationID == "" {
+		return
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	t, ok := m.tasks[conversationID]
+	if !ok || t == nil {
+		return
+	}
+	t.hitlCognition = &hitlCognitionState{UserMessage: strings.TrimSpace(userMessage)}
+}
+
+// SetHitlAssistantMessageID 记录当前助手消息 ID，供 HITL 与 DB 回退对齐。
+func (m *AgentTaskManager) SetHitlAssistantMessageID(conversationID, assistantMessageID string) {
+	conversationID = strings.TrimSpace(conversationID)
+	assistantMessageID = strings.TrimSpace(assistantMessageID)
+	if m == nil || conversationID == "" || assistantMessageID == "" {
+		return
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	t, ok := m.tasks[conversationID]
+	if !ok || t == nil {
+		return
+	}
+	if t.hitlCognition == nil {
+		t.hitlCognition = &hitlCognitionState{}
+	}
+	t.hitlCognition.AssistantMessageID = assistantMessageID
+}
+
+// UpdateHitlCognitionSnapshot 从进行中的进度流快照更新 thinking / reasoning / planning。
+func (m *AgentTaskManager) UpdateHitlCognitionSnapshot(conversationID, assistantMessageID, thinking, reasoningChain, planning string) {
+	conversationID = strings.TrimSpace(conversationID)
+	if m == nil || conversationID == "" {
+		return
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	t, ok := m.tasks[conversationID]
+	if !ok || t == nil {
+		return
+	}
+	if t.hitlCognition == nil {
+		t.hitlCognition = &hitlCognitionState{}
+	}
+	if id := strings.TrimSpace(assistantMessageID); id != "" {
+		t.hitlCognition.AssistantMessageID = id
+	}
+	if s := strings.TrimSpace(thinking); s != "" {
+		t.hitlCognition.Thinking = s
+	}
+	if s := strings.TrimSpace(reasoningChain); s != "" {
+		t.hitlCognition.ReasoningChain = s
+	}
+	if s := strings.TrimSpace(planning); s != "" {
+		t.hitlCognition.Planning = s
+	}
+}

internal/handler/hitl_context.go

@@ -0,0 +1,102 @@
+package handler
+
+import (
+	"strings"
+)
+
+const (
+	hitlPayloadUserMessage    = "userMessage"
+	hitlPayloadThinking       = "thinking"
+	hitlPayloadReasoningChain = "reasoningChain"
+	hitlPayloadPlanning       = "planning"
+)
+
+type hitlCognitionFields struct {
+	UserMessage    string
+	Thinking       string
+	ReasoningChain string
+	Planning       string
+}
+
+func (h *AgentHandler) enrichHitlApprovalPayload(conversationID, assistantMessageID string, payload map[string]interface{}) {
+	if h == nil || payload == nil {
+		return
+	}
+	cog := h.collectHitlCognition(conversationID, assistantMessageID)
+	if s := strings.TrimSpace(cog.UserMessage); s != "" {
+		payload[hitlPayloadUserMessage] = s
+	}
+	if s := strings.TrimSpace(cog.Thinking); s != "" {
+		payload[hitlPayloadThinking] = s
+	}
+	if s := strings.TrimSpace(cog.ReasoningChain); s != "" {
+		payload[hitlPayloadReasoningChain] = s
+	}
+	if s := strings.TrimSpace(cog.Planning); s != "" {
+		payload[hitlPayloadPlanning] = s
+	}
+}
+
+func (h *AgentHandler) collectHitlCognition(conversationID, assistantMessageID string) hitlCognitionFields {
+	var out hitlCognitionFields
+	if h.tasks != nil {
+		out = h.tasks.GetHitlCognition(conversationID)
+	}
+	if strings.TrimSpace(out.UserMessage) == "" && h.db != nil {
+		if msg, err := h.db.GetTurnUserMessage(conversationID, assistantMessageID); err == nil {
+			out.UserMessage = msg
+		}
+	}
+	if h.db != nil && assistantMessageID != "" {
+		dbCog, err := h.db.GetAssistantCognitionTexts(assistantMessageID)
+		if err == nil {
+			if strings.TrimSpace(out.Thinking) == "" {
+				out.Thinking = dbCog.Thinking
+			}
+			if strings.TrimSpace(out.ReasoningChain) == "" {
+				out.ReasoningChain = dbCog.ReasoningChain
+			}
+			if strings.TrimSpace(out.Planning) == "" {
+				out.Planning = dbCog.Planning
+			}
+		}
+	}
+	return out
+}
+
+func snapshotHitlCognitionFromStreams(thinkingStreams map[string]*thinkingBuf, respPlan *responsePlanAgg) (thinking, reasoningChain, planning string) {
+	if len(thinkingStreams) > 0 {
+		var thinkingParts, reasoningParts []string
+		for _, tb := range thinkingStreams {
+			if tb == nil {
+				continue
+			}
+			content := strings.TrimSpace(tb.b.String())
+			if content == "" {
+				continue
+			}
+			if tb.persistAs == "reasoning_chain" {
+				reasoningParts = append(reasoningParts, content)
+			} else {
+				thinkingParts = append(thinkingParts, content)
+			}
+		}
+		thinking = strings.Join(thinkingParts, "\n\n")
+		reasoningChain = strings.Join(reasoningParts, "\n\n")
+	}
+	if respPlan != nil {
+		planning = strings.TrimSpace(respPlan.b.String())
+	}
+	return thinking, reasoningChain, planning
+}
+
+func (h *AgentHandler) syncHitlCognitionFromProgress(conversationID, assistantMessageID string, thinkingStreams map[string]*thinkingBuf, respPlan *responsePlanAgg) {
+	if h == nil || h.tasks == nil {
+		return
+	}
+	thinking, reasoning, planning := snapshotHitlCognitionFromStreams(thinkingStreams, respPlan)
+	if thinking == "" && reasoning == "" && planning == "" {
+		return
+	}
+	h.tasks.UpdateHitlCognitionSnapshot(conversationID, assistantMessageID, thinking, reasoning, planning)
+}

internal/handler/hitl_context_test.go

@@ -0,0 +1,46 @@
+package handler
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"cyberstrike-ai/internal/database"
+
+	"go.uber.org/zap"
+)
+
+func TestEnrichHitlApprovalPayload(t *testing.T) {
+	tmp := t.TempDir()
+	db, err := database.NewDB(filepath.Join(tmp, "test.sqlite"), zap.NewNop())
+	if err != nil {
+		t.Fatalf("db: %v", err)
+	}
+	defer os.RemoveAll(tmp)
+
+	conv, err := db.CreateConversation("hitl ctx", database.ConversationCreateMeta{})
+	if err != nil {
+		t.Fatalf("conv: %v", err)
+	}
+	if _, err := db.AddMessage(conv.ID, "user", "scan 10.0.0.1 please", nil); err != nil {
+		t.Fatalf("user msg: %v", err)
+	}
+	asst, err := db.AddMessage(conv.ID, "assistant", "", nil)
+	if err != nil {
+		t.Fatalf("asst msg: %v", err)
+	}
+	if err := db.AddProcessDetail(asst.ID, conv.ID, "thinking", "need port scan first", nil); err != nil {
+		t.Fatalf("detail: %v", err)
+	}
+
+	h := &AgentHandler{db: db, tasks: NewAgentTaskManager()}
+	payload := map[string]interface{}{"toolName": "nmap", "arguments": "{}"}
+	h.enrichHitlApprovalPayload(conv.ID, asst.ID, payload)
+
+	if got := payload["userMessage"]; got != "scan 10.0.0.1 please" {
+		t.Fatalf("userMessage=%v", got)
+	}
+	if got := payload["thinking"]; got != "need port scan first" {
+		t.Fatalf("thinking=%v", got)
+	}
+}

internal/handler/hitl_execution.go

@@ -0,0 +1,132 @@
+package handler
+
+import (
+	"encoding/json"
+	"strings"
+	"time"
+)
+
+const hitlPayloadExecutionResult = "executionResult"
+
+type hitlExecutionResult struct {
+	Success    bool      `json:"success"`
+	Result     string    `json:"result,omitempty"`
+	ToolName   string    `json:"toolName,omitempty"`
+	ToolCallID string    `json:"toolCallId,omitempty"`
+	RecordedAt time.Time `json:"recordedAt"`
+}
+
+type hitlApprovedExecTrack struct {
+	InterruptID    string
+	ConversationID string
+	ToolName       string
+	ToolCallID     string
+}
+
+// TrackApprovedHitlExecution 审批通过后登记，待 tool_result 回写执行结果。
+func (m *HITLManager) TrackApprovedHitlExecution(interruptID, conversationID, toolName, toolCallID string) {
+	if m == nil {
+		return
+	}
+	interruptID = strings.TrimSpace(interruptID)
+	conversationID = strings.TrimSpace(conversationID)
+	if interruptID == "" || conversationID == "" {
+		return
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if m.approvedExec == nil {
+		m.approvedExec = make(map[string][]hitlApprovedExecTrack)
+	}
+	m.approvedExec[conversationID] = append(m.approvedExec[conversationID], hitlApprovedExecTrack{
+		InterruptID:    interruptID,
+		ConversationID: conversationID,
+		ToolName:       strings.TrimSpace(toolName),
+		ToolCallID:     strings.TrimSpace(toolCallID),
+	})
+}
+
+func (m *HITLManager) popApprovedInterruptForTool(conversationID, toolCallID, toolName string) string {
+	if m == nil {
+		return ""
+	}
+	conversationID = strings.TrimSpace(conversationID)
+	toolCallID = strings.TrimSpace(toolCallID)
+	toolName = strings.TrimSpace(toolName)
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	queue := m.approvedExec[conversationID]
+	if len(queue) == 0 {
+		return ""
+	}
+	idx := -1
+	if toolCallID != "" {
+		for i, t := range queue {
+			if t.ToolCallID == toolCallID {
+				idx = i
+				break
+			}
+		}
+	}
+	if idx < 0 && toolName != "" {
+		for i, t := range queue {
+			if strings.EqualFold(t.ToolName, toolName) {
+				idx = i
+				break
+			}
+		}
+	}
+	if idx < 0 {
+		return ""
+	}
+	id := queue[idx].InterruptID
+	queue = append(queue[:idx], queue[idx+1:]...)
+	if len(queue) == 0 {
+		delete(m.approvedExec, conversationID)
+	} else {
+		m.approvedExec[conversationID] = queue
+	}
+	return id
+}
+
+func mergeHitlPayloadExecutionResult(payloadJSON string, exec hitlExecutionResult) (string, error) {
+	root := make(map[string]interface{})
+	if strings.TrimSpace(payloadJSON) != "" {
+		_ = json.Unmarshal([]byte(payloadJSON), &root)
+	}
+	if root == nil {
+		root = make(map[string]interface{})
+	}
+	root[hitlPayloadExecutionResult] = exec
+	out, err := json.Marshal(root)
+	if err != nil {
+		return payloadJSON, err
+	}
+	return string(out), nil
+}
+
+func (h *AgentHandler) recordHitlToolExecutionResult(conversationID, toolCallID, toolName string, success bool, result string) {
+	if h == nil || h.hitlManager == nil || h.db == nil {
+		return
+	}
+	interruptID := h.hitlManager.popApprovedInterruptForTool(conversationID, toolCallID, toolName)
+	if interruptID == "" {
+		return
+	}
+	var payloadJSON string
+	err := h.db.QueryRow(`SELECT payload FROM hitl_interrupts WHERE id = ?`, interruptID).Scan(&payloadJSON)
+	if err != nil {
+		return
+	}
+	merged, err := mergeHitlPayloadExecutionResult(payloadJSON, hitlExecutionResult{
+		Success:    success,
+		Result:     strings.TrimSpace(result),
+		ToolName:   strings.TrimSpace(toolName),
+		ToolCallID: strings.TrimSpace(toolCallID),
+		RecordedAt: time.Now(),
+	})
+	if err != nil {
+		return
+	}
+	_, _ = h.db.Exec(`UPDATE hitl_interrupts SET payload = ? WHERE id = ?`, merged, interruptID)
+}

internal/handler/hitl_execution_test.go

@@ -0,0 +1,39 @@
+package handler
+
+import (
+	"encoding/json"
+	"testing"
+)
+
+func TestMergeHitlPayloadExecutionResult(t *testing.T) {
+	merged, err := mergeHitlPayloadExecutionResult(`{"userMessage":"hi","toolName":"nmap"}`, hitlExecutionResult{
+		Success: true,
+		Result:  "open ports: 80",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var root map[string]interface{}
+	if err := json.Unmarshal([]byte(merged), &root); err != nil {
+		t.Fatal(err)
+	}
+	if root["userMessage"] != "hi" {
+		t.Fatalf("userMessage lost: %v", root["userMessage"])
+	}
+	exec, ok := root["executionResult"].(map[string]interface{})
+	if !ok || exec["success"] != true {
+		t.Fatalf("executionResult missing: %v", root["executionResult"])
+	}
+}
+
+func TestPopApprovedInterruptForTool(t *testing.T) {
+	m := NewHITLManager(nil, nil)
+	m.TrackApprovedHitlExecution("hitl_a", "conv1", "nmap", "tc1")
+	m.TrackApprovedHitlExecution("hitl_b", "conv1", "exec", "")
+	if id := m.popApprovedInterruptForTool("conv1", "tc1", "nmap"); id != "hitl_a" {
+		t.Fatalf("tc1 match=%q", id)
+	}
+	if id := m.popApprovedInterruptForTool("conv1", "", "exec"); id != "hitl_b" {
+		t.Fatalf("tool name match=%q", id)
+	}
+}

internal/handler/hitl_logs.go

@@ -0,0 +1,263 @@
+package handler
+
+import (
+	"database/sql"
+	"errors"
+	"math"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"cyberstrike-ai/internal/config"
+
+	"github.com/gin-gonic/gin"
+)
+
+func normalizeHitlReviewer(v string) string {
+	switch strings.ToLower(strings.TrimSpace(v)) {
+	case "audit_agent", "agent", "ai":
+		return "audit_agent"
+	default:
+		return "human"
+	}
+}
+
+func normalizeHitlDecidedBy(v string) string {
+	switch strings.ToLower(strings.TrimSpace(v)) {
+	case "audit_agent", "agent", "ai":
+		return "audit_agent"
+	case "system", "timeout":
+		return "system"
+	case "manual":
+		return "manual"
+	default:
+		return "human"
+	}
+}
+
+func (m *HITLManager) migrateHitlSchemaColumns() {
+	_, _ = m.db.Exec(`ALTER TABLE hitl_interrupts ADD COLUMN decided_by TEXT NOT NULL DEFAULT 'human'`)
+	_, _ = m.db.Exec(`ALTER TABLE hitl_conversation_configs ADD COLUMN reviewer TEXT NOT NULL DEFAULT 'human'`)
+}
+
+func hitlInterruptRowToMap(
+	id, cid, mode, toolName, toolCallID, payload, rowStatus, decidedBy string,
+	messageID sql.NullString,
+	decision, comment sql.NullString,
+	createdAt time.Time,
+	decidedAt sql.NullTime,
+) map[string]interface{} {
+	msgID := ""
+	if messageID.Valid {
+		msgID = messageID.String
+	}
+	return map[string]interface{}{
+		"id":             id,
+		"conversationId": cid,
+		"messageId":      msgID,
+		"mode":           mode,
+		"toolName":       toolName,
+		"toolCallId":     toolCallID,
+		"payload":        payload,
+		"status":         rowStatus,
+		"decision":       decision.String,
+		"comment":        comment.String,
+		"decidedBy":      decidedBy,
+		"createdAt":      createdAt,
+		"decidedAt": func() interface{} {
+			if decidedAt.Valid {
+				return decidedAt.Time
+			}
+			return nil
+		}(),
+	}
+}
+
+func (h *AgentHandler) buildHitlListQuery(logs bool) (string, []interface{}) {
+	where, args := h.buildHitlLogsWhere(logs)
+	q := `SELECT id, conversation_id, message_id, mode, tool_name, tool_call_id, payload, status, decision, decision_comment, COALESCE(decided_by,'human'), created_at, decided_at FROM hitl_interrupts` + where
+	return q, args
+}
+
+func (h *AgentHandler) buildHitlLogsWhere(logs bool) (string, []interface{}) {
+	q := " WHERE 1=1"
+	args := []interface{}{}
+	if logs {
+		q += " AND status != 'pending'"
+	} else {
+		q += " AND status = 'pending'"
+	}
+	return q, args
+}
+
+func (h *AgentHandler) appendHitlListFilters(q string, args []interface{}, c *gin.Context) (string, []interface{}) {
+	conversationID := strings.TrimSpace(c.Query("conversationId"))
+	toolName := strings.TrimSpace(c.Query("toolName"))
+	decision := strings.TrimSpace(c.Query("decision"))
+	decidedBy := strings.TrimSpace(c.Query("decidedBy"))
+	status := strings.TrimSpace(c.Query("status"))
+	search := strings.TrimSpace(c.Query("q"))
+
+	if conversationID != "" {
+		q += " AND conversation_id = ?"
+		args = append(args, conversationID)
+	}
+	if toolName != "" {
+		q += " AND tool_name LIKE ?"
+		args = append(args, "%"+toolName+"%")
+	}
+	if decision != "" && decision != "all" {
+		q += " AND decision = ?"
+		args = append(args, decision)
+	}
+	if decidedBy != "" && decidedBy != "all" {
+		q += " AND COALESCE(decided_by,'human') = ?"
+		args = append(args, normalizeHitlDecidedBy(decidedBy))
+	}
+	if status != "" && status != "all" {
+		q += " AND status = ?"
+		args = append(args, status)
+	}
+	if search != "" {
+		like := "%" + search + "%"
+		q += " AND (id LIKE ? OR conversation_id LIKE ? OR tool_name LIKE ? OR payload LIKE ? OR COALESCE(decision_comment,'') LIKE ?)"
+		args = append(args, like, like, like, like, like)
+	}
+	return q, args
+}
+
+func (h *AgentHandler) scanHitlInterruptRows(rows *sql.Rows) ([]map[string]interface{}, error) {
+	items := make([]map[string]interface{}, 0)
+	for rows.Next() {
+		var id, cid, mode, toolName, toolCallID, payload, rowStatus, decidedBy string
+		var messageID sql.NullString
+		var decision, comment sql.NullString
+		var createdAt time.Time
+		var decidedAt sql.NullTime
+		if err := rows.Scan(&id, &cid, &messageID, &mode, &toolName, &toolCallID, &payload, &rowStatus, &decision, &comment, &decidedBy, &createdAt, &decidedAt); err != nil {
+			continue
+		}
+		items = append(items, hitlInterruptRowToMap(id, cid, mode, toolName, toolCallID, payload, rowStatus, decidedBy, messageID, decision, comment, createdAt, decidedAt))
+	}
+	return items, nil
+}
+
+func (h *AgentHandler) countHitlQuery(baseQ string, args []interface{}) (int, error) {
+	countQ := "SELECT COUNT(*) FROM (" + baseQ + ") AS hitl_cnt"
+	var total int
+	if err := h.db.QueryRow(countQ, args...).Scan(&total); err != nil {
+		return 0, err
+	}
+	return total, nil
+}
+
+func (h *AgentHandler) ListHITLLogs(c *gin.Context) {
+	page, _ := strconv.Atoi(c.DefaultQuery("page", "1"))
+	if page < 1 {
+		page = 1
+	}
+	pageSize, _ := strconv.Atoi(c.DefaultQuery("pageSize", "20"))
+	pageSize = int(math.Max(1, math.Min(float64(pageSize), 200)))
+	offset := (page - 1) * pageSize
+
+	q, args := h.buildHitlListQuery(true)
+	q, args = h.appendHitlListFilters(q, args, c)
+	total, err := h.countHitlQuery(q, args)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	q += " ORDER BY COALESCE(decided_at, created_at) DESC LIMIT ? OFFSET ?"
+	args = append(args, pageSize, offset)
+	rows, err := h.db.Query(q, args...)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	defer rows.Close()
+	items, err := h.scanHitlInterruptRows(rows)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, gin.H{"items": items, "page": page, "pageSize": pageSize, "total": total, "retentionDays": h.hitlRetentionDays()})
+}
+
+func (h *AgentHandler) hitlRetentionDays() int {
+	if h.config != nil {
+		return h.config.Hitl.RetentionDaysEffective()
+	}
+	return config.HitlConfig{}.RetentionDaysEffective()
+}
+
+// DeleteHITLLogs 批量删除或按筛选清空已决策的人机协同审计日志（不删除 pending）。
+func (h *AgentHandler) DeleteHITLLogs(c *gin.Context) {
+	var request struct {
+		IDs []string `json:"ids"`
+		All bool     `json:"all"`
+	}
+	if err := c.ShouldBindJSON(&request); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "请求参数无效: " + err.Error()})
+		return
+	}
+
+	var deleted int64
+	var err error
+	if request.All {
+		where, args := h.buildHitlLogsWhere(true)
+		where, args = h.appendHitlListFilters(where, args, c)
+		deleted, err = h.db.DeleteHitlInterruptLogsMatching(where, args)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+		if h.audit != nil {
+			h.audit.RecordOK(c, "hitl", "logs_clear", "清空人机协同审计日志", "hitl_interrupt", "", map[string]interface{}{
+				"deleted": deleted,
+			})
+		}
+	} else {
+		if len(request.IDs) == 0 {
+			c.JSON(http.StatusBadRequest, gin.H{"error": "审计日志 ID 列表不能为空"})
+			return
+		}
+		deleted, err = h.db.DeleteHitlInterruptLogsByIDs(request.IDs)
+		if err != nil {
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+		if h.audit != nil {
+			h.audit.RecordOK(c, "hitl", "logs_delete_batch", "批量删除人机协同审计日志", "hitl_interrupt", "", map[string]interface{}{
+				"count":   len(request.IDs),
+				"deleted": deleted,
+			})
+		}
+	}
+
+	c.JSON(http.StatusOK, gin.H{"message": "删除成功", "deleted": deleted})
+}
+
+func (h *AgentHandler) GetHITLLog(c *gin.Context) {
+	id := strings.TrimSpace(c.Param("id"))
+	if id == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "id is required"})
+		return
+	}
+	q := `SELECT id, conversation_id, message_id, mode, tool_name, tool_call_id, payload, status, decision, decision_comment, COALESCE(decided_by,'human'), created_at, decided_at FROM hitl_interrupts WHERE id = ?`
+	var rowID, cid, mode, toolName, toolCallID, payload, rowStatus, decidedBy string
+	var messageID sql.NullString
+	var decision, comment sql.NullString
+	var createdAt time.Time
+	var decidedAt sql.NullTime
+	err := h.db.QueryRow(q, id).Scan(&rowID, &cid, &messageID, &mode, &toolName, &toolCallID, &payload, &rowStatus, &decision, &comment, &decidedBy, &createdAt, &decidedAt)
+	if errors.Is(err, sql.ErrNoRows) {
+		c.JSON(http.StatusNotFound, gin.H{"error": "not found"})
+		return
+	}
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, hitlInterruptRowToMap(rowID, cid, mode, toolName, toolCallID, payload, rowStatus, decidedBy, messageID, decision, comment, createdAt, decidedAt))
+}

internal/handler/monitor.go

@@ -5,6 +5,7 @@ import (
 	"errors"
 	"io"
 	"net/http"
+	"sort"
 	"strconv"
 	"strings"
 	"time"
@@ -68,16 +69,34 @@ func (h *MonitorHandler) SetAgentHandler(ah *AgentHandler) {
 	h.agentHandler = ah
 }
 
+const monitorPageTopTools = 6
+
+// MonitorStatsSummary 工具调用汇总
+type MonitorStatsSummary struct {
+	TotalCalls   int        `json:"totalCalls"`
+	SuccessCalls int        `json:"successCalls"`
+	FailedCalls  int        `json:"failedCalls"`
+	LastCallTime *time.Time `json:"lastCallTime,omitempty"`
+	ToolCount    int        `json:"toolCount"`
+}
+
 // MonitorResponse 监控响应
 type MonitorResponse struct {
-	Executions    []*mcp.ToolExecution      `json:"executions"`
+	Executions    []*mcp.ToolExecution `json:"executions"`
-	Stats         map[string]*mcp.ToolStats `json:"stats"`
+	Summary       *MonitorStatsSummary `json:"summary"`
-	Timestamp     time.Time                 `json:"timestamp"`
+	TopTools      []*mcp.ToolStats     `json:"topTools"`
-	Total         int                       `json:"total,omitempty"`
+	Timestamp     time.Time            `json:"timestamp"`
-	Page          int                       `json:"page,omitempty"`
+	Total         int                  `json:"total"`
-	PageSize      int                       `json:"page_size,omitempty"`
+	Page          int                  `json:"page"`
-	TotalPages    int                       `json:"total_pages,omitempty"`
+	PageSize      int                  `json:"pageSize"`
-	RetentionDays int                       `json:"retention_days,omitempty"`
+	TotalPages    int                  `json:"totalPages"`
+	RetentionDays int                  `json:"retentionDays"`
+}
+
+// StatsResponse 统计信息响应（Dashboard 等）
+type StatsResponse struct {
+	Summary  *MonitorStatsSummary `json:"summary"`
+	TopTools []*mcp.ToolStats     `json:"topTools"`
 }
 
 // Monitor 获取监控信息
@@ -101,9 +120,9 @@ func (h *MonitorHandler) Monitor(c *gin.Context) {
 	// 解析工具筛选参数（兼容 mcp__tool 与内部 mcp::tool）
 	toolName := normalizeToolNameFilter(c.Query("tool"))
 
-	executions, total := h.loadExecutionsWithPagination(page, pageSize, status, toolName)
+	executions, total := h.loadExecutionListWithPagination(page, pageSize, status, toolName)
 	h.enrichExecutionsConversationID(executions)
-	stats := h.loadStats()
+	summary, topTools := h.loadStatsSummary(monitorPageTopTools)
 
 	totalPages := (total + pageSize - 1) / pageSize
 	if totalPages == 0 {
@@ -112,7 +131,8 @@ func (h *MonitorHandler) Monitor(c *gin.Context) {
 
 	c.JSON(http.StatusOK, MonitorResponse{
 		Executions:    executions,
-		Stats:         stats,
+		Summary:       summary,
+		TopTools:      topTools,
 		Timestamp:     time.Now(),
 		Total:         total,
 		Page:          page,
@@ -134,6 +154,112 @@ func (h *MonitorHandler) loadExecutions() []*mcp.ToolExecution {
 	return executions
 }
 
+func (h *MonitorHandler) loadExecutionListWithPagination(page, pageSize int, status, toolName string) ([]*mcp.ToolExecution, int) {
+	if h.db == nil {
+		allExecutions := h.mcpServer.GetAllExecutions()
+		if status != "" || toolName != "" {
+			filtered := make([]*mcp.ToolExecution, 0)
+			for _, exec := range allExecutions {
+				matchStatus := status == "" || exec.Status == status
+				matchTool := toolNameFilterMatches(exec.ToolName, toolName)
+				if matchStatus && matchTool {
+					filtered = append(filtered, exec)
+				}
+			}
+			allExecutions = filtered
+		}
+		total := len(allExecutions)
+		offset := (page - 1) * pageSize
+		end := offset + pageSize
+		if end > total {
+			end = total
+		}
+		if offset >= total {
+			return []*mcp.ToolExecution{}, total
+		}
+		pageSlice := allExecutions[offset:end]
+		out := make([]*mcp.ToolExecution, 0, len(pageSlice))
+		for _, exec := range pageSlice {
+			if exec == nil {
+				continue
+			}
+			out = append(out, slimToolExecution(exec))
+		}
+		return out, total
+	}
+
+	offset := (page - 1) * pageSize
+	executions, err := h.db.LoadToolExecutionListPage(offset, pageSize, status, toolName)
+	if err != nil {
+		h.logger.Warn("从数据库加载执行记录列表失败，回退到内存数据", zap.Error(err))
+		return h.loadExecutionListWithPaginationFromMemory(page, pageSize, status, toolName)
+	}
+
+	total, err := h.db.CountToolExecutions(status, toolName)
+	if err != nil {
+		h.logger.Warn("获取执行记录总数失败", zap.Error(err))
+		total = offset + len(executions)
+		if len(executions) == pageSize {
+			total = offset + len(executions) + 1
+		}
+	}
+
+	return executions, total
+}
+
+func (h *MonitorHandler) loadExecutionListWithPaginationFromMemory(page, pageSize int, status, toolName string) ([]*mcp.ToolExecution, int) {
+	allExecutions := h.mcpServer.GetAllExecutions()
+	if status != "" || toolName != "" {
+		filtered := make([]*mcp.ToolExecution, 0)
+		for _, exec := range allExecutions {
+			matchStatus := status == "" || exec.Status == status
+			matchTool := toolNameFilterMatches(exec.ToolName, toolName)
+			if matchStatus && matchTool {
+				filtered = append(filtered, exec)
+			}
+		}
+		allExecutions = filtered
+	}
+	total := len(allExecutions)
+	offset := (page - 1) * pageSize
+	end := offset + pageSize
+	if end > total {
+		end = total
+	}
+	if offset >= total {
+		return []*mcp.ToolExecution{}, total
+	}
+	pageSlice := allExecutions[offset:end]
+	out := make([]*mcp.ToolExecution, 0, len(pageSlice))
+	for _, exec := range pageSlice {
+		if exec == nil {
+			continue
+		}
+		out = append(out, slimToolExecution(exec))
+	}
+	return out, total
+}
+
+func slimToolExecution(exec *mcp.ToolExecution) *mcp.ToolExecution {
+	if exec == nil {
+		return nil
+	}
+	slim := &mcp.ToolExecution{
+		ID:        exec.ID,
+		ToolName:  exec.ToolName,
+		Status:    exec.Status,
+		StartTime: exec.StartTime,
+	}
+	if exec.EndTime != nil {
+		end := *exec.EndTime
+		slim.EndTime = &end
+	}
+	if exec.Duration > 0 {
+		slim.Duration = exec.Duration
+	}
+	return slim
+}
+
 func (h *MonitorHandler) loadExecutionsWithPagination(page, pageSize int, status, toolName string) ([]*mcp.ToolExecution, int) {
 	if h.db == nil {
 		allExecutions := h.mcpServer.GetAllExecutions()
@@ -206,7 +332,78 @@ func (h *MonitorHandler) loadExecutionsWithPagination(page, pageSize int, status
 	return executions, total
 }
 
-func (h *MonitorHandler) loadStats() map[string]*mcp.ToolStats {
+func (h *MonitorHandler) loadStatsSummary(topN int) (*MonitorStatsSummary, []*mcp.ToolStats) {
+	if topN <= 0 {
+		topN = monitorPageTopTools
+	}
+
+	if h.db != nil {
+		result, err := h.db.LoadToolStatsSummary(topN)
+		if err == nil {
+			return dbStatsSummaryToMonitor(result), result.TopTools
+		}
+		h.logger.Warn("从数据库加载统计汇总失败，回退到内存数据", zap.Error(err))
+	}
+
+	stats := h.loadStatsMap()
+	return summarizeToolStats(stats, topN)
+}
+
+func dbStatsSummaryToMonitor(result *database.ToolStatsSummaryResult) *MonitorStatsSummary {
+	if result == nil {
+		return &MonitorStatsSummary{}
+	}
+	summary := &MonitorStatsSummary{
+		TotalCalls:   result.Summary.TotalCalls,
+		SuccessCalls: result.Summary.SuccessCalls,
+		FailedCalls:  result.Summary.FailedCalls,
+		ToolCount:    result.Summary.ToolCount,
+	}
+	if result.Summary.LastCallTime != nil {
+		t := *result.Summary.LastCallTime
+		summary.LastCallTime = &t
+	}
+	return summary
+}
+
+func summarizeToolStats(stats map[string]*mcp.ToolStats, topN int) (*MonitorStatsSummary, []*mcp.ToolStats) {
+	summary := &MonitorStatsSummary{}
+	if len(stats) == 0 {
+		return summary, nil
+	}
+
+	all := make([]*mcp.ToolStats, 0, len(stats))
+	for _, stat := range stats {
+		if stat == nil {
+			continue
+		}
+		summary.ToolCount++
+		summary.TotalCalls += stat.TotalCalls
+		summary.SuccessCalls += stat.SuccessCalls
+		summary.FailedCalls += stat.FailedCalls
+		if stat.LastCallTime != nil && (summary.LastCallTime == nil || stat.LastCallTime.After(*summary.LastCallTime)) {
+			t := *stat.LastCallTime
+			summary.LastCallTime = &t
+		}
+		if stat.TotalCalls > 0 {
+			statCopy := *stat
+			all = append(all, &statCopy)
+		}
+	}
+
+	sort.Slice(all, func(i, j int) bool {
+		if all[i].TotalCalls == all[j].TotalCalls {
+			return all[i].ToolName < all[j].ToolName
+		}
+		return all[i].TotalCalls > all[j].TotalCalls
+	})
+	if len(all) > topN {
+		all = all[:topN]
+	}
+	return summary, all
+}
+
+func (h *MonitorHandler) loadStatsMap() map[string]*mcp.ToolStats {
 	// 合并内部MCP服务器和外部MCP管理器的统计信息
 	stats := make(map[string]*mcp.ToolStats)
 
@@ -334,7 +531,7 @@ func (h *MonitorHandler) CancelExecution(c *gin.Context) {
 
 func (h *MonitorHandler) enrichExecutionsConversationID(executions []*mcp.ToolExecution) {
 	for _, exec := range executions {
-		if exec == nil {
+		if exec == nil || exec.Status != "running" {
 			continue
 		}
 		exec.ConversationID = h.conversationIDForRunningExecution(exec.ID)
@@ -415,8 +612,17 @@ func (h *MonitorHandler) BatchGetToolNames(c *gin.Context) {
 
 // GetStats 获取统计信息
 func (h *MonitorHandler) GetStats(c *gin.Context) {
-	stats := h.loadStats()
+	topN := 30
-	c.JSON(http.StatusOK, stats)
+	if topStr := c.Query("top"); topStr != "" {
+		if t, err := strconv.Atoi(topStr); err == nil && t > 0 && t <= 100 {
+			topN = t
+		}
+	}
+	summary, topTools := h.loadStatsSummary(topN)
+	c.JSON(http.StatusOK, StatsResponse{
+		Summary:  summary,
+		TopTools: topTools,
+	})
 }
 
 // CallsTimelinePoint 调用趋势数据点

internal/handler/multi_agent.go

@@ -188,6 +188,7 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 	var cumulativeMCPExecutionIDs []string
 	// 同一请求内分段续跑时，主代理 iteration 事件按偏移累计，避免 UI 出现「第3轮 → 第1轮」回跳。
 	var mainIterationOffset int
+	var emptyResponseContinueAttempt int
 
 	for {
 		segmentMainIterationMax := 0
@@ -251,6 +252,13 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 		}
 
 		if runErr == nil {
+			mw := &h.config.MultiAgent.EinoMiddleware
+			if h.tryContinueOnEinoEmptyResponse(taskCtx, mw, conversationID, result, &emptyResponseContinueAttempt, &curHistory, &curFinalMessage, progressCallback) {
+				mainIterationOffset += segmentMainIterationMax
+				timeoutCancel()
+				baseCtx, cancelWithCause, taskCtx, timeoutCancel = h.rebindEinoRunningTask(conversationID, timeoutCancel)
+				continue
+			}
 			timeoutCancel()
 			break
 		}

internal/handler/openapi.go

@@ -740,14 +740,21 @@ func (h *OpenAPIHandler) GetOpenAPISpec(c *gin.Context) {
 					"properties": map[string]interface{}{
 						"executions": map[string]interface{}{
 							"type":        "array",
-							"description": "执行记录列表",
+							"description": "执行记录列表（轻量字段，不含 arguments/result）",
 							"items": map[string]interface{}{
 								"$ref": "#/components/schemas/ToolExecution",
 							},
 						},
-						"stats": map[string]interface{}{
+						"summary": map[string]interface{}{
 							"type":        "object",
-							"description": "统计信息",
+							"description": "工具调用汇总",
+						},
+						"topTools": map[string]interface{}{
+							"type":        "array",
+							"description": "调用量 Top N 工具",
+							"items": map[string]interface{}{
+								"type": "object",
+							},
 						},
 						"timestamp": map[string]interface{}{
 							"type":        "string",
@@ -756,20 +763,24 @@ func (h *OpenAPIHandler) GetOpenAPISpec(c *gin.Context) {
 						},
 						"total": map[string]interface{}{
 							"type":        "integer",
-							"description": "总数",
+							"description": "执行记录总数",
 						},
 						"page": map[string]interface{}{
 							"type":        "integer",
 							"description": "当前页",
 						},
-						"page_size": map[string]interface{}{
+						"pageSize": map[string]interface{}{
 							"type":        "integer",
 							"description": "每页数量",
 						},
-						"total_pages": map[string]interface{}{
+						"totalPages": map[string]interface{}{
 							"type":        "integer",
 							"description": "总页数",
 						},
+						"retentionDays": map[string]interface{}{
+							"type":        "integer",
+							"description": "执行记录保留天数",
+						},
 					},
 				},
 				"ConfigResponse": map[string]interface{}{
@@ -1232,6 +1243,34 @@ func (h *OpenAPIHandler) GetOpenAPISpec(c *gin.Context) {
 								"type": "string",
 							},
 						},
+						{
+							"name":        "project_id",
+							"in":          "query",
+							"required":    false,
+							"description": "按项目筛选；传 __none__ 表示仅未绑定项目的对话",
+							"schema": map[string]interface{}{
+								"type": "string",
+							},
+						},
+						{
+							"name":        "exclude_grouped",
+							"in":          "query",
+							"required":    false,
+							"description": "为 true 时排除已加入分组的对话（默认在未搜索且未按项目筛选时启用）",
+							"schema": map[string]interface{}{
+								"type": "boolean",
+							},
+						},
+						{
+							"name":        "sort_by",
+							"in":          "query",
+							"required":    false,
+							"description": "排序字段：updated_at（默认）或 created_at",
+							"schema": map[string]interface{}{
+								"type": "string",
+								"enum": []string{"updated_at", "created_at"},
+							},
+						},
 					},
 					"responses": map[string]interface{}{
 						"200": map[string]interface{}{

internal/handler/project_context.go

@@ -8,6 +8,7 @@ import (
 )
 
 // agentSessionContextBlock 注入会话工作目录与项目黑板（用于 system prompt 追加块）。
+// 用户输入由 message history 承载；压缩后由 summarization 摘要指令保留关键约束。
 func (h *AgentHandler) agentSessionContextBlock(conversationID string) string {
 	var parts []string
 	if ws := h.buildWorkspaceBlock(conversationID); ws != "" {

internal/handler/robot.go

@@ -447,7 +447,7 @@ func (h *RobotHandler) cmdUnbindProject(platform, userID string) string {
 }
 
 func (h *RobotHandler) cmdList() string {
-	convs, err := h.db.ListConversations(50, 0, "", "")
+	convs, err := h.db.ListConversations(50, 0, "", "", "")
 	if err != nil {
 		return "获取对话列表失败: " + err.Error()
 	}
@@ -711,12 +711,27 @@ type wecomReplyXML struct {
 	Content      string   `xml:"Content"`
 }
 
+// wecomRequireToken 企业微信回调必须配置 Token；未配置时拒绝请求，防止未授权触发 Agent。
+func (h *RobotHandler) wecomRequireToken(c *gin.Context) (string, bool) {
+	token := strings.TrimSpace(h.config.Robots.Wecom.Token)
+	if token == "" {
+		h.logger.Warn("企业微信已启用但未配置 token，已拒绝回调（请在配置中设置 robots.wecom.token）")
+		c.String(http.StatusForbidden, "")
+		return "", false
+	}
+	return token, true
+}
+
 // HandleWecomGET 企业微信 URL 校验（GET）
 func (h *RobotHandler) HandleWecomGET(c *gin.Context) {
 	if !h.config.Robots.Wecom.Enabled {
 		c.String(http.StatusNotFound, "")
 		return
 	}
+	token, ok := h.wecomRequireToken(c)
+	if !ok {
+		return
+	}
 	// Gin 的 Query() 会自动 URL 解码，拿到的就是正确的 base64 字符串
 	echostr := c.Query("echostr")
 	msgSignature := c.Query("msg_signature")
@@ -724,7 +739,7 @@ func (h *RobotHandler) HandleWecomGET(c *gin.Context) {
 	nonce := c.Query("nonce")
 
 	// 验证签名：将 token、timestamp、nonce、echostr 四个参数排序后拼接计算 SHA1
-	signature := h.signWecomRequest(h.config.Robots.Wecom.Token, timestamp, nonce, echostr)
+	signature := h.signWecomRequest(token, timestamp, nonce, echostr)
 	if signature != msgSignature {
 		h.logger.Warn("企业微信 URL 验证签名失败", zap.String("expected", msgSignature), zap.String("got", signature))
 		c.String(http.StatusBadRequest, "invalid signature")
@@ -865,27 +880,28 @@ func (h *RobotHandler) HandleWecomPOST(c *gin.Context) {
 	}
 	h.logger.Debug("企业微信 POST 收到请求", zap.String("body", string(bodyRaw)))
 
-	// 验证请求签名防止伪造。企业微信签名算法同 URL 验证，使用 token、timestamp、nonce、 Encrypt 四个字段
+	// 验证请求签名防止伪造。企业微信签名算法同 URL 验证，使用 token、timestamp、nonce、 Encrypt 四个字段。
-	// 若配置了 Token 则必须校验签名，避免未授权请求触发 Agent（防止平台被接管）
+	// 启用企业微信时必须配置 token 并校验签名，避免未授权请求触发 Agent。
-	token := h.config.Robots.Wecom.Token
+	token, ok := h.wecomRequireToken(c)
-	if token != "" {
+	if !ok {
-		if msgSignature == "" {
+		return
-			h.logger.Warn("企业微信 POST 缺少签名，已拒绝（需配置 token 并确保回调携带 msg_signature）")
+	}
-			c.String(http.StatusOK, "")
+	if msgSignature == "" {
-			return
+		h.logger.Warn("企业微信 POST 缺少签名，已拒绝（需确保回调携带 msg_signature）")
-		}
+		c.String(http.StatusOK, "")
-		var tmp wecomXML
+		return
-		if err := xml.Unmarshal(bodyRaw, &tmp); err != nil {
+	}
-			h.logger.Warn("企业微信 POST 签名验证前解析 XML 失败", zap.Error(err))
+	var tmp wecomXML
-			c.String(http.StatusOK, "")
+	if err := xml.Unmarshal(bodyRaw, &tmp); err != nil {
-			return
+		h.logger.Warn("企业微信 POST 签名验证前解析 XML 失败", zap.Error(err))
-		}
+		c.String(http.StatusOK, "")
-		expected := h.signWecomRequest(token, timestamp, nonce, tmp.Encrypt)
+		return
-		if expected != msgSignature {
+	}
-			h.logger.Warn("企业微信 POST 签名验证失败", zap.String("expected", expected), zap.String("got", msgSignature))
+	expected := h.signWecomRequest(token, timestamp, nonce, tmp.Encrypt)
-			c.String(http.StatusOK, "")
+	if expected != msgSignature {
-			return
+		h.logger.Warn("企业微信 POST 签名验证失败", zap.String("expected", expected), zap.String("got", msgSignature))
-		}
+		c.String(http.StatusOK, "")
+		return
 	}
 
 	var body wecomXML
@@ -899,6 +915,13 @@ func (h *RobotHandler) HandleWecomPOST(c *gin.Context) {
 	// 保存企业 ID（用于明文模式回复）
 	enterpriseID := body.ToUserName
 
+	// 配置了 EncodingAESKey 时必须走加密消息，拒绝明文 XML 绕过
+	if strings.TrimSpace(h.config.Robots.Wecom.EncodingAESKey) != "" && strings.TrimSpace(body.Encrypt) == "" {
+		h.logger.Warn("企业微信已配置加密模式但收到明文消息，已拒绝")
+		c.String(http.StatusOK, "")
+		return
+	}
+
 	// 加密模式：先解密再解析内层 XML
 	if body.Encrypt != "" && h.config.Robots.Wecom.EncodingAESKey != "" {
 		h.logger.Debug("企业微信进入加密模式解密流程")

internal/handler/robot_wecom_test.go

@@ -0,0 +1,78 @@
+package handler
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"cyberstrike-ai/internal/config"
+
+	"github.com/gin-gonic/gin"
+	"go.uber.org/zap"
+)
+
+func newWecomTestHandler(token string, aesKey string) *RobotHandler {
+	return &RobotHandler{
+		config: &config.Config{
+			Robots: config.RobotsConfig{
+				Wecom: config.RobotWecomConfig{
+					Enabled:        true,
+					Token:          token,
+					EncodingAESKey: aesKey,
+				},
+			},
+		},
+		logger: zap.NewNop(),
+	}
+}
+
+func TestHandleWecomPOST_rejectsWhenTokenEmpty(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	h := newWecomTestHandler("", "")
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	body := `<?xml version="1.0"?><xml><FromUserName>attacker</FromUserName><MsgType>text</MsgType><Content>hi</Content></xml>`
+	c.Request = httptest.NewRequest(http.MethodPost, "/api/robot/wecom", strings.NewReader(body))
+
+	h.HandleWecomPOST(c)
+
+	if w.Code != http.StatusForbidden {
+		t.Fatalf("status = %d, want %d", w.Code, http.StatusForbidden)
+	}
+	if w.Body.String() == "success" {
+		t.Fatal("expected rejection, got success")
+	}
+}
+
+func TestHandleWecomPOST_rejectsPlaintextWhenEncryptionConfigured(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	h := newWecomTestHandler("secret-token", "abcdefghijklmnopqrstuvwxyz0123456789ABCD")
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	body := `<?xml version="1.0"?><xml><FromUserName>attacker</FromUserName><MsgType>text</MsgType><Content>hi</Content></xml>`
+	c.Request = httptest.NewRequest(http.MethodPost, "/api/robot/wecom?timestamp=1&nonce=2&msg_signature=fake", strings.NewReader(body))
+
+	h.HandleWecomPOST(c)
+
+	if w.Body.String() == "success" {
+		t.Fatal("expected rejection for plaintext in encryption mode, got success")
+	}
+}
+
+func TestHandleWecomGET_rejectsWhenTokenEmpty(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+
+	h := newWecomTestHandler("", "")
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest(http.MethodGet, "/api/robot/wecom?msg_signature=x&timestamp=1&nonce=2&echostr=abc", nil)
+
+	h.HandleWecomGET(c)
+
+	if w.Code != http.StatusForbidden {
+		t.Fatalf("status = %d, want %d", w.Code, http.StatusForbidden)
+	}
+}

internal/handler/task_manager.go

@@ -26,6 +26,7 @@ func shouldPersistEinoAgentTraceAfterRunError(baseCtx context.Context) bool {
 // AgentTask 描述正在运行的Agent任务
 type AgentTask struct {
 	ConversationID string    `json:"conversationId"`
+	Title          string    `json:"title,omitempty"`
 	Message        string    `json:"message,omitempty"`
 	StartedAt      time.Time `json:"startedAt"`
 	Status         string    `json:"status"`
@@ -42,6 +43,9 @@ type AgentTask struct {
 	// activeEinoExecuteAbortNote AbortActiveEinoExecute 写入的用户说明，由 execute 收尾时合并进工具结果
 	activeEinoExecuteAbortNote string
 
+	// hitlCognition 本轮运行中供 HITL/审计 Agent 读取的上下文（用户原话 + 思考，不含会话历史）
+	hitlCognition *hitlCognitionState
+
 	cancel func(error)
 }
 
@@ -233,6 +237,7 @@ func (m *AgentTaskManager) ActiveMCPExecutionID(conversationID string) string {
 // CompletedTask 已完成的任务（用于历史记录）
 type CompletedTask struct {
 	ConversationID string    `json:"conversationId"`
+	Title          string    `json:"title,omitempty"`
 	Message        string    `json:"message,omitempty"`
 	StartedAt      time.Time `json:"startedAt"`
 	CompletedAt    time.Time `json:"completedAt"`
@@ -247,6 +252,8 @@ type AgentTaskManager struct {
 	maxHistorySize   int              // 最大历史记录数
 	historyRetention time.Duration    // 历史记录保留时间
 	eventBus         *TaskEventBus    // 可选：任务结束时关闭镜像 SSE 订阅
+	// toolCanceler 在用户整轮停止任务时终止当前 MCP 工具（非「中断并继续」）。
+	toolCanceler func(conversationID string)
 }
 
 const (
@@ -277,6 +284,13 @@ func (m *AgentTaskManager) SetTaskEventBus(b *TaskEventBus) {
 	m.eventBus = b
 }
 
+// SetToolCanceler 设置整轮停止任务时终止当前 MCP 工具的回调（由 AgentHandler 注入）。
+func (m *AgentTaskManager) SetToolCanceler(fn func(conversationID string)) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	m.toolCanceler = fn
+}
+
 // GetTask 返回运行中任务（无则 nil）。
 func (m *AgentTaskManager) GetTask(conversationID string) *AgentTask {
 	m.mu.RLock()
@@ -343,6 +357,7 @@ func (m *AgentTaskManager) StartTask(conversationID, message string, cancel cont
 	}
 
 	m.tasks[conversationID] = task
+	task.hitlCognition = &hitlCognitionState{UserMessage: strings.TrimSpace(message)}
 	return task, nil
 }
 
@@ -372,14 +387,21 @@ func (m *AgentTaskManager) CancelTask(conversationID string, cause error) (bool,
 		task.InterruptContinueNote = ""
 	}
 	cancel := task.cancel
-	m.mu.Unlock()
-
 	if cause == nil {
 		cause = ErrTaskCancelled
 	}
+	var toolCanceler func(string)
+	if errors.Is(cause, ErrTaskCancelled) {
+		toolCanceler = m.toolCanceler
+	}
+	m.mu.Unlock()
+
 	if cancel != nil {
 		cancel(cause)
 	}
+	if toolCanceler != nil {
+		toolCanceler(conversationID)
+	}
 	return true, nil
 }
 

internal/handler/task_manager_tool_cancel_test.go

@@ -0,0 +1,80 @@
+package handler
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"cyberstrike-ai/internal/multiagent"
+)
+
+func TestCancelTaskInvokesToolCancelerOnFullStop(t *testing.T) {
+	tm := NewAgentTaskManager()
+	called := false
+	tm.SetToolCanceler(func(conversationID string) {
+		if conversationID == "conv-1" {
+			called = true
+		}
+	})
+
+	_, cancel := context.WithCancelCause(context.Background())
+	_, err := tm.StartTask("conv-1", "hello", cancel)
+	if err != nil {
+		t.Fatalf("StartTask: %v", err)
+	}
+
+	ok, err := tm.CancelTask("conv-1", ErrTaskCancelled)
+	if err != nil || !ok {
+		t.Fatalf("CancelTask: ok=%v err=%v", ok, err)
+	}
+	if !called {
+		t.Fatal("expected tool canceler to be invoked on full task cancel")
+	}
+}
+
+func TestCancelTaskSkipsToolCancelerOnInterruptContinue(t *testing.T) {
+	tm := NewAgentTaskManager()
+	called := false
+	tm.SetToolCanceler(func(conversationID string) {
+		called = true
+	})
+
+	_, cancel := context.WithCancelCause(context.Background())
+	_, err := tm.StartTask("conv-1", "hello", cancel)
+	if err != nil {
+		t.Fatalf("StartTask: %v", err)
+	}
+
+	ok, err := tm.CancelTask("conv-1", multiagent.ErrInterruptContinue)
+	if err != nil || !ok {
+		t.Fatalf("CancelTask: ok=%v err=%v", ok, err)
+	}
+	if called {
+		t.Fatal("tool canceler must not run for interrupt-continue")
+	}
+}
+
+func TestCancelTaskDefaultCauseIsTaskCancelled(t *testing.T) {
+	tm := NewAgentTaskManager()
+	var gotCause error
+	tm.SetToolCanceler(func(conversationID string) {
+		if conversationID == "conv-2" {
+			gotCause = ErrTaskCancelled
+		}
+	})
+
+	ctx, cancel := context.WithCancelCause(context.Background())
+	if _, err := tm.StartTask("conv-2", "hello", cancel); err != nil {
+		t.Fatalf("StartTask: %v", err)
+	}
+
+	if _, err := tm.CancelTask("conv-2", nil); err != nil {
+		t.Fatalf("CancelTask: %v", err)
+	}
+	if !errors.Is(context.Cause(ctx), ErrTaskCancelled) {
+		t.Fatalf("expected ErrTaskCancelled cause, got %v", context.Cause(ctx))
+	}
+	if gotCause != ErrTaskCancelled {
+		t.Fatalf("expected tool canceler path for default cancel cause")
+	}
+}

internal/handler/terminal_ws_windows.go

@@ -0,0 +1,16 @@
+//go:build windows
+
+package handler
+
+import (
+	"net/http"
+
+	"github.com/gin-gonic/gin"
+)
+
+// RunCommandWS 交互式 PTY 终端依赖 Unix PTY（见 terminal_ws_unix.go）；Windows 暂不支持。
+func (h *TerminalHandler) RunCommandWS(c *gin.Context) {
+	c.JSON(http.StatusNotImplemented, gin.H{
+		"error": "Interactive WebSocket terminal is not supported on Windows; use POST /terminal/run or /terminal/run/stream instead.",
+	})
+}

internal/hitl/retention.go

@@ -0,0 +1,71 @@
+package hitl
+
+import (
+	"time"
+
+	"cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/database"
+
+	"go.uber.org/zap"
+)
+
+const retentionPurgeInterval = time.Hour
+
+// Service manages HITL audit log retention (decided hitl_interrupts rows).
+type Service struct {
+	db     *database.DB
+	cfg    *config.Config
+	logger *zap.Logger
+}
+
+// NewService creates a HITL audit log retention service.
+func NewService(db *database.DB, cfg *config.Config, logger *zap.Logger) *Service {
+	return &Service{db: db, cfg: cfg, logger: logger}
+}
+
+// RetentionDays returns configured retention; 0 means keep forever.
+func (s *Service) RetentionDays() int {
+	if s == nil || s.cfg == nil {
+		return config.HitlConfig{}.RetentionDaysEffective()
+	}
+	return s.cfg.Hitl.RetentionDaysEffective()
+}
+
+// PurgeExpired deletes decided HITL log rows older than retention_days when configured.
+func (s *Service) PurgeExpired() {
+	if s == nil || s.db == nil || s.cfg == nil {
+		return
+	}
+	days := s.cfg.Hitl.RetentionDaysEffective()
+	if days <= 0 {
+		return
+	}
+	cutoff := time.Now().AddDate(0, 0, -days)
+	n, err := s.db.PurgeHitlInterruptLogsBefore(cutoff)
+	if err != nil {
+		if s.logger != nil {
+			s.logger.Warn("清理过期人机协同审计日志失败", zap.Error(err))
+		}
+		return
+	}
+	if n > 0 && s.logger != nil {
+		s.logger.Info("已清理过期人机协同审计日志", zap.Int64("deleted", n), zap.Int("retention_days", days))
+	}
+}
+
+// StartRetentionLoop periodically purges expired HITL audit log rows.
+func StartRetentionLoop(s *Service, logger *zap.Logger) {
+	if s == nil {
+		return
+	}
+	go func() {
+		ticker := time.NewTicker(retentionPurgeInterval)
+		defer ticker.Stop()
+		for range ticker.C {
+			s.PurgeExpired()
+			if logger != nil {
+				logger.Debug("hitl audit log retention tick completed")
+			}
+		}
+	}()
+}

internal/hitl/retention_test.go

@@ -0,0 +1,50 @@
+package hitl
+
+import (
+	"path/filepath"
+	"testing"
+	"time"
+
+	appconfig "cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/database"
+
+	"go.uber.org/zap"
+)
+
+func TestServicePurgeExpired_respectsZeroRetention(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "hitl.db")
+	db, err := database.NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	defer db.Close()
+	if _, err := db.Exec(`CREATE TABLE IF NOT EXISTS hitl_interrupts (
+		id TEXT PRIMARY KEY,
+		conversation_id TEXT NOT NULL,
+		mode TEXT NOT NULL,
+		tool_name TEXT NOT NULL,
+		status TEXT NOT NULL,
+		decision TEXT,
+		created_at DATETIME NOT NULL,
+		decided_at DATETIME
+	)`); err != nil {
+		t.Fatalf("create table: %v", err)
+	}
+
+	old := time.Now().AddDate(0, 0, -100).UTC().Format(time.RFC3339)
+	if _, err := db.Exec(`INSERT INTO hitl_interrupts
+		(id, conversation_id, mode, tool_name, status, decision, created_at, decided_at)
+		VALUES ('old-1', 'c1', 'approval', 'exec', 'decided', 'approve', ?, ?)`, old, old); err != nil {
+		t.Fatalf("insert: %v", err)
+	}
+
+	zero := 0
+	svc := NewService(db, &appconfig.Config{
+		Hitl: appconfig.HitlConfig{RetentionDays: &zero},
+	}, zap.NewNop())
+	svc.PurgeExpired()
+
+	if err := db.QueryRow(`SELECT id FROM hitl_interrupts WHERE id = 'old-1'`).Scan(new(string)); err != nil {
+		t.Fatalf("record should remain when retention_days=0: %v", err)
+	}
+}

internal/knowledge/eino_pipeline_retriever.go

@@ -0,0 +1,96 @@
+package knowledge
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"cyberstrike-ai/internal/config"
+
+	"github.com/cloudwego/eino/callbacks"
+	"github.com/cloudwego/eino/components"
+	"github.com/cloudwego/eino/components/retriever"
+	"github.com/cloudwego/eino/schema"
+	"go.uber.org/zap"
+)
+
+// knowledgePipelineRetriever: MultiQuery → vector candidates → rerank → post-process.
+type knowledgePipelineRetriever struct {
+	inner retriever.Retriever
+	base  *Retriever
+}
+
+func newKnowledgePipelineRetriever(inner retriever.Retriever, base *Retriever) *knowledgePipelineRetriever {
+	if inner == nil || base == nil {
+		return nil
+	}
+	return &knowledgePipelineRetriever{inner: inner, base: base}
+}
+
+func (p *knowledgePipelineRetriever) GetType() string {
+	return "KnowledgeRAGPipeline"
+}
+
+func (p *knowledgePipelineRetriever) Retrieve(ctx context.Context, query string, opts ...retriever.Option) (out []*schema.Document, err error) {
+	if p == nil || p.inner == nil || p.base == nil {
+		return nil, fmt.Errorf("knowledge pipeline retriever: nil")
+	}
+	q := strings.TrimSpace(query)
+	if q == "" {
+		return nil, fmt.Errorf("查询不能为空")
+	}
+
+	ro := retriever.GetCommonOptions(nil, opts...)
+	finalTopK := p.base.config.TopK
+	if finalTopK <= 0 {
+		finalTopK = 5
+	}
+	if ro.TopK != nil && *ro.TopK > 0 {
+		finalTopK = *ro.TopK
+	}
+
+	ctx = callbacks.EnsureRunInfo(ctx, p.GetType(), components.ComponentOfRetriever)
+	ctx = callbacks.OnStart(ctx, &retriever.CallbackInput{Query: q, TopK: finalTopK, Extra: ro.DSLInfo})
+	defer func() {
+		if err != nil {
+			_ = callbacks.OnError(ctx, err)
+			return
+		}
+		_ = callbacks.OnEnd(ctx, &retriever.CallbackOutput{Docs: out})
+	}()
+
+	out, err = p.inner.Retrieve(ctx, q, opts...)
+	if err != nil {
+		return nil, err
+	}
+	if len(out) == 0 {
+		return out, nil
+	}
+
+	if rr := p.base.documentReranker(); rr != nil && len(out) > 1 {
+		reranked, rerr := rr.Rerank(ctx, q, out)
+		if rerr != nil {
+			if p.base.logger != nil {
+				p.base.logger.Warn("知识检索重排失败，已使用融合序", zap.Error(rerr))
+			}
+		} else if len(reranked) > 0 {
+			out = reranked
+		}
+	}
+
+	tokenModel := ""
+	if p.base.embedder != nil {
+		tokenModel = p.base.embedder.EmbeddingModelName()
+	}
+	var postPO *config.PostRetrieveConfig
+	if p.base.config != nil {
+		postPO = &p.base.config.PostRetrieve
+	}
+	out, err = ApplyPostRetrieve(out, postPO, tokenModel, finalTopK)
+	if err != nil {
+		return nil, err
+	}
+	return out, nil
+}
+
+var _ retriever.Retriever = (*knowledgePipelineRetriever)(nil)

internal/knowledge/eino_retrieve_chain.go

@@ -8,8 +8,7 @@ import (
 	"github.com/cloudwego/eino/schema"
 )
 
-// BuildKnowledgeRetrieveChain 编译「查询字符串 → 文档列表」的 Eino Chain，底层为 SQLite 向量检索（[VectorEinoRetriever]）。
+// BuildKnowledgeRetrieveChain 编译「查询字符串 → 文档列表」的 Eino Chain（MultiQuery → 向量 → 重排 → 后处理）。
-// 去重、上下文预算截断与最终 Top-K 均在 [VectorEinoRetriever.Retrieve] 内完成，与 HTTP/MCP 检索路径一致。
 func BuildKnowledgeRetrieveChain(ctx context.Context, r *Retriever) (compose.Runnable[string, []*schema.Document], error) {
 	if r == nil {
 		return nil, fmt.Errorf("retriever is nil")

internal/knowledge/eino_retriever_adapter.go

@@ -11,19 +11,10 @@ import (
 	"github.com/cloudwego/eino/components"
 	"github.com/cloudwego/eino/components/retriever"
 	"github.com/cloudwego/eino/schema"
-	"go.uber.org/zap"
 )
 
 // VectorEinoRetriever implements [retriever.Retriever] on top of SQLite-stored embeddings + cosine similarity.
-//
+// It returns prefetch-sized vector candidates only; rerank and post-process run in [knowledgePipelineRetriever].
-// Options:
-//   - [retriever.WithTopK]
-//   - [retriever.WithDSLInfo] with [DSLRiskType] (string), [DSLSimilarityThreshold] (float, cosine 0–1), [DSLSubIndexFilter] (string)
-//
-// Document scores are cosine similarity; [retriever.WithScoreThreshold] is not mapped to a different metric.
-//
-// After vector search: optional [DocumentReranker] (see [Retriever.SetDocumentReranker]), then
-// [ApplyPostRetrieve] (normalized-text dedupe, context budget, final Top-K) using [config.PostRetrieveConfig].
 type VectorEinoRetriever struct {
 	inner *Retriever
 }
@@ -119,26 +110,6 @@ func (h *VectorEinoRetriever) Retrieve(ctx context.Context, query string, opts .
 		return nil, err
 	}
 	out = retrievalResultsToDocuments(results)
-
-	if rr := h.inner.documentReranker(); rr != nil && len(out) > 1 {
-		reranked, rerr := rr.Rerank(ctx, q, out)
-		if rerr != nil {
-			if h.inner.logger != nil {
-				h.inner.logger.Warn("知识检索重排失败，已使用向量序", zap.Error(rerr))
-			}
-		} else if len(reranked) > 0 {
-			out = reranked
-		}
-	}
-
-	tokenModel := ""
-	if h.inner.embedder != nil {
-		tokenModel = h.inner.embedder.EmbeddingModelName()
-	}
-	out, err = ApplyPostRetrieve(out, postPO, tokenModel, finalTopK)
-	if err != nil {
-		return nil, err
-	}
 	return out, nil
 }
 

internal/knowledge/rerank_http.go

@@ -0,0 +1,226 @@
+package knowledge
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"cyberstrike-ai/internal/config"
+
+	"github.com/cloudwego/eino/schema"
+	"go.uber.org/zap"
+)
+
+// HTTPReranker calls a hosted rerank API (DashScope or Cohere-compatible).
+type HTTPReranker struct {
+	provider string
+	model    string
+	baseURL  string
+	apiKey   string
+	client   *http.Client
+	logger   *zap.Logger
+}
+
+// NewHTTPReranker builds a rerank client from knowledge retrieval config; openAI supplies fallback credentials.
+func NewHTTPReranker(rc *config.RerankConfig, openAI *config.OpenAIConfig, logger *zap.Logger) (*HTTPReranker, error) {
+	if rc == nil {
+		return nil, fmt.Errorf("rerank config is nil")
+	}
+	baseURL := strings.TrimSpace(rc.BaseURL)
+	apiKey := strings.TrimSpace(rc.APIKey)
+	if openAI != nil {
+		if baseURL == "" {
+			baseURL = strings.TrimSpace(openAI.BaseURL)
+		}
+		if apiKey == "" {
+			apiKey = strings.TrimSpace(openAI.APIKey)
+		}
+	}
+	if apiKey == "" {
+		return nil, fmt.Errorf("rerank api_key is required")
+	}
+	provider := rc.ProviderEffective(baseURL)
+	model := rc.ModelEffective(provider)
+	return &HTTPReranker{
+		provider: provider,
+		model:    model,
+		baseURL:  strings.TrimSuffix(baseURL, "/"),
+		apiKey:   apiKey,
+		client:   &http.Client{Timeout: 60 * time.Second},
+		logger:   logger,
+	}, nil
+}
+
+func (r *HTTPReranker) Rerank(ctx context.Context, query string, docs []*schema.Document) ([]*schema.Document, error) {
+	if r == nil {
+		return docs, nil
+	}
+	q := strings.TrimSpace(query)
+	if q == "" || len(docs) == 0 {
+		return docs, nil
+	}
+	if len(docs) == 1 {
+		return docs, nil
+	}
+	texts := make([]string, 0, len(docs))
+	for _, d := range docs {
+		if d == nil {
+			texts = append(texts, "")
+			continue
+		}
+		texts = append(texts, d.Content)
+	}
+	var order []int
+	var err error
+	switch r.provider {
+	case "dashscope":
+		order, err = r.rerankDashScope(ctx, q, texts, len(docs))
+	default:
+		order, err = r.rerankCohere(ctx, q, texts, len(docs))
+	}
+	if err != nil {
+		return nil, err
+	}
+	out := make([]*schema.Document, 0, len(order))
+	for _, idx := range order {
+		if idx < 0 || idx >= len(docs) || docs[idx] == nil {
+			continue
+		}
+		out = append(out, docs[idx])
+	}
+	if len(out) == 0 {
+		return docs, nil
+	}
+	return out, nil
+}
+
+func (r *HTTPReranker) rerankCohere(ctx context.Context, query string, documents []string, topN int) ([]int, error) {
+	url := r.cohereRerankURL()
+	body := map[string]any{
+		"model":     r.model,
+		"query":     query,
+		"documents": documents,
+		"top_n":     topN,
+	}
+	raw, err := json.Marshal(body)
+	if err != nil {
+		return nil, err
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(raw))
+	if err != nil {
+		return nil, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "Bearer "+r.apiKey)
+	resp, err := r.client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("rerank request: %w", err)
+	}
+	defer resp.Body.Close()
+	respBody, _ := io.ReadAll(resp.Body)
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return nil, fmt.Errorf("rerank http %d: %s", resp.StatusCode, truncateForRerankLog(string(respBody)))
+	}
+	var parsed struct {
+		Results []struct {
+			Index int `json:"index"`
+		} `json:"results"`
+	}
+	if err := json.Unmarshal(respBody, &parsed); err != nil {
+		return nil, fmt.Errorf("rerank decode: %w", err)
+	}
+	order := make([]int, 0, len(parsed.Results))
+	for _, row := range parsed.Results {
+		order = append(order, row.Index)
+	}
+	return order, nil
+}
+
+func (r *HTTPReranker) rerankDashScope(ctx context.Context, query string, documents []string, topN int) ([]int, error) {
+	url := r.dashscopeRerankURL()
+	body := map[string]any{
+		"model": r.model,
+		"input": map[string]any{
+			"query":     query,
+			"documents": documents,
+		},
+		"parameters": map[string]any{
+			"return_documents": false,
+			"top_n":            topN,
+		},
+	}
+	raw, err := json.Marshal(body)
+	if err != nil {
+		return nil, err
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(raw))
+	if err != nil {
+		return nil, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Authorization", "Bearer "+r.apiKey)
+	resp, err := r.client.Do(req)
+	if err != nil {
+		return nil, fmt.Errorf("dashscope rerank: %w", err)
+	}
+	defer resp.Body.Close()
+	respBody, _ := io.ReadAll(resp.Body)
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return nil, fmt.Errorf("dashscope rerank http %d: %s", resp.StatusCode, truncateForRerankLog(string(respBody)))
+	}
+	var parsed struct {
+		Output struct {
+			Results []struct {
+				Index int `json:"index"`
+			} `json:"results"`
+		} `json:"output"`
+	}
+	if err := json.Unmarshal(respBody, &parsed); err != nil {
+		return nil, fmt.Errorf("dashscope rerank decode: %w", err)
+	}
+	order := make([]int, 0, len(parsed.Output.Results))
+	for _, row := range parsed.Output.Results {
+		order = append(order, row.Index)
+	}
+	return order, nil
+}
+
+func (r *HTTPReranker) cohereRerankURL() string {
+	base := r.baseURL
+	if base == "" {
+		base = "https://api.cohere.com"
+	}
+	if strings.HasSuffix(base, "/v1") {
+		return base + "/rerank"
+	}
+	return base + "/v1/rerank"
+}
+
+func (r *HTTPReranker) dashscopeRerankURL() string {
+	base := strings.TrimSpace(r.baseURL)
+	if base == "" {
+		return "https://dashscope.aliyuncs.com/api/v1/services/rerank/text-rerank/text-rerank"
+	}
+	if strings.Contains(base, "/api/v1/services/rerank") {
+		return base
+	}
+	if strings.Contains(base, "dashscope.aliyuncs.com") || strings.Contains(base, "compatible-mode") {
+		return "https://dashscope.aliyuncs.com/api/v1/services/rerank/text-rerank/text-rerank"
+	}
+	return strings.TrimSuffix(base, "/")
+}
+
+func truncateForRerankLog(s string) string {
+	s = strings.TrimSpace(s)
+	if len(s) > 512 {
+		return s[:512] + "..."
+	}
+	return s
+}
+
+var _ DocumentReranker = (*HTTPReranker)(nil)

internal/knowledge/rerank_http_test.go

@@ -0,0 +1,97 @@
+package knowledge
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"cyberstrike-ai/internal/config"
+
+	"github.com/cloudwego/eino/schema"
+)
+
+func TestHTTPReranker_CohereOrder(t *testing.T) {
+	t.Parallel()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/rerank" {
+			t.Fatalf("path %s", r.URL.Path)
+		}
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"results": []map[string]any{
+				{"index": 2, "relevance_score": 0.9},
+				{"index": 0, "relevance_score": 0.5},
+			},
+		})
+	}))
+	defer srv.Close()
+
+	rr, err := NewHTTPReranker(&config.RerankConfig{
+		Provider: "cohere",
+		Model:    "rerank-multilingual-v3.0",
+		BaseURL:  srv.URL,
+		APIKey:   "test-key",
+	}, nil, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	docs := []*schema.Document{
+		{ID: "a", Content: "alpha"},
+		{ID: "b", Content: "beta"},
+		{ID: "c", Content: "gamma"},
+	}
+	out, err := rr.Rerank(context.Background(), "query", docs)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(out) != 2 || out[0].ID != "c" || out[1].ID != "a" {
+		t.Fatalf("order wrong: %#v", out)
+	}
+}
+
+func TestHTTPReranker_DashScopeOrder(t *testing.T) {
+	t.Parallel()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"output": map[string]any{
+				"results": []map[string]any{
+					{"index": 1, "relevance_score": 0.88},
+				},
+			},
+		})
+	}))
+	defer srv.Close()
+
+	rr, err := NewHTTPReranker(&config.RerankConfig{
+		Provider: "dashscope",
+		Model:    "gte-rerank",
+		BaseURL:  srv.URL,
+		APIKey:   "test-key",
+	}, nil, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	docs := []*schema.Document{{ID: "a", Content: "a"}, {ID: "b", Content: "b"}}
+	out, err := rr.Rerank(context.Background(), "q", docs)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(out) != 1 || out[0].ID != "b" {
+		t.Fatalf("got %#v", out)
+	}
+}
+
+func TestRerankConfigDefaults(t *testing.T) {
+	t.Parallel()
+	rc := config.RerankConfig{}
+	if rc.ProviderEffective("https://dashscope.aliyuncs.com/x") != "dashscope" {
+		t.Fatal("dashscope detect")
+	}
+	if rc.ModelEffective("dashscope") != "gte-rerank" {
+		t.Fatal("dashscope model")
+	}
+	if rc.ModelEffective("cohere") != "rerank-multilingual-v3.0" {
+		t.Fatal("cohere model")
+	}
+}

internal/knowledge/retrieval_postprocess.go

@@ -19,7 +19,7 @@ import (
 // postRetrieveMaxPrefetchCap 限制单次向量候选上限，避免误配置导致全表扫压力过大。
 const postRetrieveMaxPrefetchCap = 200
 
-// DocumentReranker 可选重排（如交叉编码器 / 第三方 Rerank API），由 [Retriever.SetDocumentReranker] 注入；失败时在适配层降级为向量序。
+// DocumentReranker 精排（HTTP dashscope / Cohere 兼容 API），由 [WireRetrieverPipeline] 注入。
 type DocumentReranker interface {
 	Rerank(ctx context.Context, query string, docs []*schema.Document) ([]*schema.Document, error)
 }
@@ -167,13 +167,16 @@ func truncateDocumentsByBudget(docs []*schema.Document, maxRunes, maxTokens int,
 	return out, nil
 }
 
-// EffectivePrefetchTopK 计算向量检索应拉取的候选条数（供粗排 / 去重 / 重排）。
+// EffectivePrefetchTopK 计算每条 MultiQuery 变体在向量阶段的候选条数（供融合 / 重排 / 后处理）。
 func EffectivePrefetchTopK(topK int, po *config.PostRetrieveConfig) int {
 	if topK < 1 {
 		topK = 5
 	}
-	fetch := topK
+	fetch := topK * 4
-	if po != nil && po.PrefetchTopK > fetch {
+	if fetch < 20 {
+		fetch = 20
+	}
+	if po != nil && po.PrefetchTopK > 0 {
 		fetch = po.PrefetchTopK
 	}
 	if fetch > postRetrieveMaxPrefetchCap {
@@ -182,7 +185,7 @@ func EffectivePrefetchTopK(topK int, po *config.PostRetrieveConfig) int {
 	return fetch
 }
 
-// ApplyPostRetrieve 检索后处理：规范化正文去重 → 预算截断 → 最终 TopK。重排在 [VectorEinoRetriever] 中单独调用以便失败时降级。
+// ApplyPostRetrieve 检索后处理：规范化正文去重 → 预算截断 → 最终 TopK（精排已在流水线中完成）。
 func ApplyPostRetrieve(docs []*schema.Document, po *config.PostRetrieveConfig, tokenModel string, finalTopK int) ([]*schema.Document, error) {
 	if finalTopK < 1 {
 		finalTopK = 5

internal/knowledge/retrieval_postprocess_test.go

@@ -28,8 +28,8 @@ func TestDedupeByNormalizedContent(t *testing.T) {
 }
 
 func TestEffectivePrefetchTopK(t *testing.T) {
-	if g := EffectivePrefetchTopK(5, nil); g != 5 {
+	if g := EffectivePrefetchTopK(5, nil); g != 20 {
-		t.Fatalf("got %d", g)
+		t.Fatalf("default prefetch got %d want 20", g)
 	}
 	if g := EffectivePrefetchTopK(5, &config.PostRetrieveConfig{PrefetchTopK: 50}); g != 50 {
 		t.Fatalf("got %d", g)

internal/knowledge/retriever.go

@@ -27,15 +27,19 @@ type Retriever struct {
 
 	rerankMu sync.RWMutex
 	reranker DocumentReranker
+
+	pipeline retriever.Retriever
+	wireOpenAI *config.OpenAIConfig
 }
 
 // RetrievalConfig 检索配置
 type RetrievalConfig struct {
 	TopK                int
 	SimilarityThreshold float64
-	// SubIndexFilter 非空时仅检索 sub_indexes 包含该标签（逗号分隔之一）的行；空 sub_indexes 的旧行仍保留以兼容。
+	SubIndexFilter      string
-	SubIndexFilter string
+	MultiQuery          config.MultiQueryConfig
-	PostRetrieve   config.PostRetrieveConfig
+	Rerank              config.RerankConfig
+	PostRetrieve        config.PostRetrieveConfig
 }
 
 // NewRetriever 创建新的检索器
@@ -48,7 +52,7 @@ func NewRetriever(db *sql.DB, embedder *Embedder, config *RetrievalConfig, logge
 	}
 }
 
-// UpdateConfig 更新检索配置
+// UpdateConfig 更新检索配置并重建 Eino MultiQuery + 重排流水线。
 func (r *Retriever) UpdateConfig(cfg *RetrievalConfig) {
 	if cfg != nil {
 		r.config = cfg
@@ -57,12 +61,18 @@ func (r *Retriever) UpdateConfig(cfg *RetrievalConfig) {
 				zap.Int("top_k", cfg.TopK),
 				zap.Float64("similarity_threshold", cfg.SimilarityThreshold),
 				zap.String("sub_index_filter", cfg.SubIndexFilter),
+				zap.Int("multi_query_max", cfg.MultiQuery.MaxQueriesEffective()),
 				zap.Int("post_retrieve_prefetch_top_k", cfg.PostRetrieve.PrefetchTopK),
 				zap.Int("post_retrieve_max_context_chars", cfg.PostRetrieve.MaxContextChars),
 				zap.Int("post_retrieve_max_context_tokens", cfg.PostRetrieve.MaxContextTokens),
 			)
 		}
 	}
+	if r.wireOpenAI != nil {
+		if err := WireRetrieverPipeline(context.Background(), r, r.wireOpenAI); err != nil && r.logger != nil {
+			r.logger.Warn("检索流水线重建失败", zap.Error(err))
+		}
+	}
 }
 
 // SetDocumentReranker 注入可选重排器（并发安全）；nil 表示禁用。
@@ -103,7 +113,7 @@ func cosineSimilarity(a, b []float32) float64 {
 	return dotProduct / (math.Sqrt(normA) * math.Sqrt(normB))
 }
 
-// Search 搜索知识库。统一经 [VectorEinoRetriever]（Eino retriever.Retriever 边界）。
+// Search 搜索知识库（Eino MultiQuery → 向量检索 → 重排 → 后处理）。
 func (r *Retriever) Search(ctx context.Context, req *SearchRequest) ([]*RetrievalResult, error) {
 	if req == nil {
 		return nil, fmt.Errorf("请求不能为空")
@@ -113,7 +123,7 @@ func (r *Retriever) Search(ctx context.Context, req *SearchRequest) ([]*Retrieva
 		return nil, fmt.Errorf("查询不能为空")
 	}
 	opts := r.einoRetrieverOptions(req)
-	docs, err := NewVectorEinoRetriever(r).Retrieve(ctx, q, opts...)
+	docs, err := r.activeEinoRetriever().Retrieve(ctx, q, opts...)
 	if err != nil {
 		return nil, err
 	}
@@ -143,7 +153,19 @@ func (r *Retriever) einoRetrieverOptions(req *SearchRequest) []retriever.Option
 
 // EinoRetrieve 直接返回 [schema.Document]，供 Eino Graph / Chain 使用。
 func (r *Retriever) EinoRetrieve(ctx context.Context, query string, opts ...retriever.Option) ([]*schema.Document, error) {
-	return NewVectorEinoRetriever(r).Retrieve(ctx, query, opts...)
+	return r.activeEinoRetriever().Retrieve(ctx, query, opts...)
+}
+
+func (r *Retriever) activeEinoRetriever() retriever.Retriever {
+	if r != nil && r.pipeline != nil {
+		return r.pipeline
+	}
+	return NewVectorEinoRetriever(r)
+}
+
+// AsEinoRetriever 将知识库检索流水线暴露为 Eino [retriever.Retriever]。
+func (r *Retriever) AsEinoRetriever() retriever.Retriever {
+	return r.activeEinoRetriever()
 }
 
 func (r *Retriever) knowledgeEmbeddingSelectSQL(riskType, subIndexFilter string) (string, []interface{}) {
@@ -299,7 +321,14 @@ func (r *Retriever) vectorSearch(ctx context.Context, req *SearchRequest) ([]*Re
 	return results, nil
 }
 
-// AsEinoRetriever 将纯向量检索暴露为 Eino [retriever.Retriever]。
+// RetrievalConfigFromYAML maps API/YAML retrieval settings into the knowledge package.
-func (r *Retriever) AsEinoRetriever() retriever.Retriever {
+func RetrievalConfigFromYAML(r config.RetrievalConfig) *RetrievalConfig {
-	return NewVectorEinoRetriever(r)
+	return &RetrievalConfig{
+		TopK:                r.TopK,
+		SimilarityThreshold: r.SimilarityThreshold,
+		SubIndexFilter:      r.SubIndexFilter,
+		MultiQuery:          r.MultiQuery,
+		Rerank:              r.Rerank,
+		PostRetrieve:        r.PostRetrieve,
+	}
 }

internal/knowledge/wire_retriever.go

@@ -0,0 +1,74 @@
+package knowledge
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/openai"
+
+	einoopenai "github.com/cloudwego/eino-ext/components/model/openai"
+	"github.com/cloudwego/eino/flow/retriever/multiquery"
+	"go.uber.org/zap"
+)
+
+// WireRetrieverPipeline builds Eino MultiQuery + HTTP rerank + post-process pipeline on r.
+// Call once after NewRetriever; UpdateConfig re-invokes when wireOpenAI is set.
+func WireRetrieverPipeline(ctx context.Context, r *Retriever, openAI *config.OpenAIConfig) error {
+	if r == nil {
+		return fmt.Errorf("retriever is nil")
+	}
+	if openAI == nil {
+		return fmt.Errorf("openai config is nil")
+	}
+	if r.config == nil {
+		return fmt.Errorf("retrieval config is nil")
+	}
+	r.wireOpenAI = openAI
+
+	httpClient := openai.NewEinoHTTPClient(openAI, &http.Client{Timeout: 120 * time.Second})
+	chatCfg := &einoopenai.ChatModelConfig{
+		APIKey:     strings.TrimSpace(openAI.APIKey),
+		BaseURL:    strings.TrimSuffix(strings.TrimSpace(openAI.BaseURL), "/"),
+		Model:      strings.TrimSpace(openAI.Model),
+		HTTPClient: httpClient,
+	}
+	if chatCfg.Model == "" {
+		chatCfg.Model = "gpt-4o"
+	}
+	rewriteLLM, err := einoopenai.NewChatModel(ctx, chatCfg)
+	if err != nil {
+		return fmt.Errorf("multi_query rewrite model: %w", err)
+	}
+
+	reranker, err := NewHTTPReranker(&r.config.Rerank, openAI, r.logger)
+	if err != nil {
+		return fmt.Errorf("reranker: %w", err)
+	}
+	r.SetDocumentReranker(reranker)
+
+	vec := NewVectorEinoRetriever(r)
+	mq, err := multiquery.NewRetriever(ctx, &multiquery.Config{
+		RewriteLLM:    rewriteLLM,
+		MaxQueriesNum: r.config.MultiQuery.MaxQueriesEffective(),
+		OrigRetriever: vec,
+	})
+	if err != nil {
+		return fmt.Errorf("multi_query: %w", err)
+	}
+
+	r.pipeline = newKnowledgePipelineRetriever(mq, r)
+	if r.logger != nil {
+		provider := r.config.Rerank.ProviderEffective(strings.TrimSpace(openAI.BaseURL))
+		r.logger.Info("知识库检索流水线已启用",
+			zap.String("pipeline", "MultiQuery→Vector→Rerank→PostRetrieve"),
+			zap.Int("multi_query_max", r.config.MultiQuery.MaxQueriesEffective()),
+			zap.String("rerank_provider", provider),
+			zap.String("rerank_model", r.config.Rerank.ModelEffective(provider)),
+		)
+	}
+	return nil
+}

internal/mcp/external_manager.go

@@ -814,6 +814,23 @@ func (m *ExternalMCPManager) CancelToolExecution(id string) bool {
 	return m.CancelToolExecutionWithNote(id, "")
 }
 
+// ActiveRunningExecutionIDs 返回当前进程内仍登记 cancel 的外部 MCP executionId 快照。
+func (m *ExternalMCPManager) ActiveRunningExecutionIDs() map[string]struct{} {
+	if m == nil {
+		return nil
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if len(m.runningCancels) == 0 {
+		return nil
+	}
+	out := make(map[string]struct{}, len(m.runningCancels))
+	for id := range m.runningCancels {
+		out[id] = struct{}{}
+	}
+	return out
+}
+
 // updateStats 更新统计信息
 func (m *ExternalMCPManager) updateStats(toolName string, failed bool) {
 	now := time.Now()

internal/mcp/server.go

@@ -1170,6 +1170,23 @@ func (s *Server) CancelToolExecution(id string) bool {
 	return s.CancelToolExecutionWithNote(id, "")
 }
 
+// ActiveRunningExecutionIDs 返回当前进程内仍登记 cancel 的 executionId 快照。
+func (s *Server) ActiveRunningExecutionIDs() map[string]struct{} {
+	if s == nil {
+		return nil
+	}
+	s.runningCancelsMu.Lock()
+	defer s.runningCancelsMu.Unlock()
+	if len(s.runningCancels) == 0 {
+		return nil
+	}
+	out := make(map[string]struct{}, len(s.runningCancels))
+	for id := range s.runningCancels {
+		out[id] = struct{}{}
+	}
+	return out
+}
+
 // initDefaultPrompts 初始化默认提示词模板
 func (s *Server) initDefaultPrompts() {
 	s.mu.Lock()

internal/monitor/reconcile.go

@@ -0,0 +1,101 @@
+package monitor
+
+import (
+	"time"
+
+	"cyberstrike-ai/internal/database"
+	"cyberstrike-ai/internal/mcp"
+
+	"go.uber.org/zap"
+)
+
+const (
+	staleRunningMinAge       = 45 * time.Second
+	staleRunningReconcileGap = 2 * time.Minute
+)
+
+// ExecutionReconciler 在启动或运行期将无对应协程的 running 执行记录收尾为 cancelled。
+type ExecutionReconciler struct {
+	db          *database.DB
+	mcpServer   *mcp.Server
+	externalMgr *mcp.ExternalMCPManager
+	logger      *zap.Logger
+}
+
+// NewExecutionReconciler creates a reconciler for orphaned MCP tool executions.
+func NewExecutionReconciler(db *database.DB, mcpServer *mcp.Server, externalMgr *mcp.ExternalMCPManager, logger *zap.Logger) *ExecutionReconciler {
+	return &ExecutionReconciler{
+		db:          db,
+		mcpServer:   mcpServer,
+		externalMgr: externalMgr,
+		logger:      logger,
+	}
+}
+
+// ReconcileOnStartup marks every persisted running row as cancelled (safe right after process start).
+func (r *ExecutionReconciler) ReconcileOnStartup() {
+	if r == nil || r.db == nil {
+		return
+	}
+	now := time.Now()
+	n, err := r.db.CancelOrphanedRunningToolExecutions(now, "执行已中断（服务重启）")
+	if err != nil {
+		if r.logger != nil {
+			r.logger.Warn("启动时清理孤儿 running 工具执行记录失败", zap.Error(err))
+		}
+		return
+	}
+	if n > 0 && r.logger != nil {
+		r.logger.Info("启动时已收尾孤儿 running 工具执行记录", zap.Int64("count", n))
+	}
+}
+
+func (r *ExecutionReconciler) activeExecutionIDs() map[string]struct{} {
+	ids := make(map[string]struct{})
+	if r.mcpServer != nil {
+		for id := range r.mcpServer.ActiveRunningExecutionIDs() {
+			ids[id] = struct{}{}
+		}
+	}
+	if r.externalMgr != nil {
+		for id := range r.externalMgr.ActiveRunningExecutionIDs() {
+			ids[id] = struct{}{}
+		}
+	}
+	return ids
+}
+
+// ReconcileStaleRunning finalizes running rows that are not tracked in-memory and older than staleRunningMinAge.
+func (r *ExecutionReconciler) ReconcileStaleRunning() {
+	if r == nil || r.db == nil {
+		return
+	}
+	now := time.Now()
+	n, err := r.db.FinalizeStaleRunningToolExecutions(now, staleRunningMinAge, r.activeExecutionIDs(), "执行已中断（会话已结束）")
+	if err != nil {
+		if r.logger != nil {
+			r.logger.Warn("定期收尾 stale running 工具执行记录失败", zap.Error(err))
+		}
+		return
+	}
+	if n > 0 && r.logger != nil {
+		r.logger.Info("已收尾 stale running 工具执行记录", zap.Int64("count", n))
+	}
+}
+
+// StartStaleRunningReconcileLoop periodically reconciles orphaned running tool executions.
+func StartStaleRunningReconcileLoop(r *ExecutionReconciler, logger *zap.Logger) {
+	if r == nil {
+		return
+	}
+	go func() {
+		ticker := time.NewTicker(staleRunningReconcileGap)
+		defer ticker.Stop()
+		for range ticker.C {
+			r.ReconcileStaleRunning()
+			if logger != nil {
+				logger.Debug("monitor stale running reconcile tick completed")
+			}
+		}
+	}()
+}

internal/monitor/reconcile_test.go

@@ -0,0 +1,38 @@
+package monitor
+
+import (
+	"path/filepath"
+	"testing"
+	"time"
+
+	"cyberstrike-ai/internal/database"
+	"cyberstrike-ai/internal/mcp"
+
+	"go.uber.org/zap"
+)
+
+func TestExecutionReconciler_ReconcileOnStartup(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "monitor.db")
+	db, err := database.NewDB(dbPath, zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	defer db.Close()
+
+	if err := db.SaveToolExecution(&mcp.ToolExecution{
+		ID: "run-1", ToolName: "hydra", Status: "running", StartTime: time.Now().Add(-time.Hour),
+	}); err != nil {
+		t.Fatalf("SaveToolExecution: %v", err)
+	}
+
+	r := NewExecutionReconciler(db, mcp.NewServer(zap.NewNop()), nil, zap.NewNop())
+	r.ReconcileOnStartup()
+
+	got, err := db.GetToolExecution("run-1")
+	if err != nil {
+		t.Fatalf("GetToolExecution: %v", err)
+	}
+	if got.Status != "cancelled" {
+		t.Fatalf("expected cancelled after startup reconcile, got %s", got.Status)
+	}
+}

internal/multiagent/adk_init.go

@@ -0,0 +1,16 @@
+package multiagent
+
+import (
+	"fmt"
+
+	"github.com/cloudwego/eino/adk"
+)
+
+// InitADK configures global Eino ADK settings. Call once at process startup before
+// any ADK middleware or agents are created.
+func InitADK() error {
+	if err := adk.SetLanguage(adk.LanguageChinese); err != nil {
+		return fmt.Errorf("adk set language: %w", err)
+	}
+	return nil
+}

internal/multiagent/eino_adk_run_loop.go

@@ -299,6 +299,8 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 
 	var toolResultSent sync.Map // toolCallID -> struct{}；ADK Tool 事件去重（权威正文来自 reduction 处理后的 agent 上下文）
 	tryEmitToolResultProgress := func(toolName, content, toolCallID string, isErr bool, agentName string) {
+		// 仅由 ADK schema.Tool 事件调用；MCP/execute 桥在 reduction 前的 ToolInvokeNotify 不得推送 tool_result，
+		// 否则全量输出会先占位并触发 toolResultSent 去重，导致 UI/监控展示与 agent 实际收到的截断正文不一致。
 		if progress == nil {
 			return
 		}
@@ -316,6 +318,7 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 			"isError":        isErr,
 			"result":         content,
 			"resultPreview":  preview,
+			"agentFacing":    true, // 与 reduction 后送入 ChatModel 的正文一致，供前端展示
 			"conversationId": conversationID,
 			"einoAgent":      agentName,
 			"einoRole":       einoRoleTag(agentName),
@@ -350,25 +353,6 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 		}
 		progress("tool_result", fmt.Sprintf("工具结果 (%s)", toolName), data)
 	}
-	if args.ToolInvokeNotify != nil {
-		args.ToolInvokeNotify.Set(func(toolCallID, toolName, einoAgent string, success bool, content string, invokeErr error) {
-			// Eino execute / MCP 桥在工具返回时 Fire；若 ADK schema.Tool 事件迟迟不到，此处立即推送
-			// tool_result 解除 UI「执行中」。tryEmitToolResultProgress 经 toolResultSent 去重，ADK 晚到不重复。
-			isErr := !success || invokeErr != nil
-			body := einoToolResultBody(content)
-			if einoToolResultIsError(toolName, content) {
-				isErr = true
-			}
-			if tail := friendlyEinoExecuteInvokeTail(invokeErr); tail != "" {
-				if body == "" {
-					body = tail
-				} else if !strings.Contains(body, tail) {
-					body = strings.TrimSpace(body) + "\n\n" + tail
-				}
-			}
-			tryEmitToolResultProgress(toolName, body, toolCallID, isErr, einoAgent)
-		})
-	}
 
 	if args.EinoCallbacks != nil {
 		ctx = einoobserve.AttachAgentRunCallbacks(ctx, args.EinoCallbacks, einoobserve.Params{

internal/multiagent/eino_empty_response.go

@@ -0,0 +1,59 @@
+package multiagent
+
+import (
+	"strings"
+	"time"
+
+	"cyberstrike-ai/internal/config"
+)
+
+const defaultEmptyResponseContinueMaxAttempts = 5
+
+// IsEinoEmptyResponseResult 判断 Run 是否以「未捕获助手正文」占位结束（非真实用户可见回复）。
+func IsEinoEmptyResponseResult(result *RunResult) bool {
+	if result == nil {
+		return false
+	}
+	return isEinoEmptyResponseText(result.Response)
+}
+
+func isEinoEmptyResponseText(s string) bool {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return false
+	}
+	return strings.Contains(s, "no assistant text was captured") ||
+		strings.Contains(s, "未捕获到助手文本输出")
+}
+
+// HasEinoResumeTrace 轨迹非空，续跑才有上下文可恢复。
+func HasEinoResumeTrace(result *RunResult) bool {
+	if result == nil {
+		return false
+	}
+	s := strings.TrimSpace(result.LastAgentTraceInput)
+	return s != "" && s != "[]" && s != "null"
+}
+
+// EmptyResponseContinueMaxAttemptsFromConfig 无助手正文时 Handler 层退避续跑上限；0=默认 5。
+func EmptyResponseContinueMaxAttemptsFromConfig(mw *config.MultiAgentEinoMiddlewareConfig) int {
+	if mw != nil && mw.EmptyResponseContinueMaxAttempts > 0 {
+		return mw.EmptyResponseContinueMaxAttempts
+	}
+	return defaultEmptyResponseContinueMaxAttempts
+}
+
+// EmptyResponseContinueBackoff 与 run_retry 相同指数退避（2s, 4s, 8s… capped）。
+func EmptyResponseContinueBackoff(attempt int, mw *config.MultiAgentEinoMiddlewareConfig) time.Duration {
+	maxBackoff := defaultEinoRunRetryMaxBackoff
+	if mw != nil && mw.RunRetryMaxBackoffSec > 0 {
+		maxBackoff = time.Duration(mw.RunRetryMaxBackoffSec) * time.Second
+	}
+	return einoTransientRetryBackoff(attempt, maxBackoff)
+}
+
+// FormatEmptyResponseContinueUserMessage 系统自动续跑时注入的 user 轮次（不写入 messages 表气泡）。
+func FormatEmptyResponseContinueUserMessage() string {
+	return strings.TrimSpace(`【系统自动续跑 / Auto resume】
+上一轮 Eino 会话未产出可见助手正文（可能流式中断或仅完成工具调用）。请基于已有轨迹与工具结果继续推进，并给出阶段性总结；勿重复已完成步骤。`)
+}

internal/multiagent/eino_empty_response_test.go

@@ -0,0 +1,38 @@
+package multiagent
+
+import "testing"
+
+func TestIsEinoEmptyResponseResult(t *testing.T) {
+	empty := &RunResult{
+		Response: "(Eino ADK single-agent session completed but no assistant text was captured. Check process details or logs.) " +
+			"（Eino ADK 单代理会话已完成，但未捕获到助手文本输出。请查看过程详情或日志。）",
+	}
+	if !IsEinoEmptyResponseResult(empty) {
+		t.Fatal("expected empty placeholder response")
+	}
+	ok := &RunResult{Response: "扫描完成，发现 2 个开放端口。"}
+	if IsEinoEmptyResponseResult(ok) {
+		t.Fatalf("expected real response, got placeholder match")
+	}
+	if IsEinoEmptyResponseResult(nil) {
+		t.Fatal("nil result should be false")
+	}
+}
+
+func TestHasEinoResumeTrace(t *testing.T) {
+	if HasEinoResumeTrace(nil) {
+		t.Fatal("nil")
+	}
+	if HasEinoResumeTrace(&RunResult{LastAgentTraceInput: "[]"}) {
+		t.Fatal("enable resume on empty trace")
+	}
+	if !HasEinoResumeTrace(&RunResult{LastAgentTraceInput: `[{"role":"user","content":"hi"}]`}) {
+		t.Fatal("expected resume trace")
+	}
+}
+
+func TestEmptyResponseContinueMaxAttemptsFromConfig(t *testing.T) {
+	if got := EmptyResponseContinueMaxAttemptsFromConfig(nil); got != defaultEmptyResponseContinueMaxAttempts {
+		t.Fatalf("default: got %d want %d", got, defaultEmptyResponseContinueMaxAttempts)
+	}
+}

internal/multiagent/eino_execute_streaming_wrap.go

@@ -84,7 +84,7 @@ func (w *einoStreamingShellWrap) ExecuteStreaming(ctx context.Context, input *fi
 	if security.IsBackgroundShellCommand(req.Command) && !req.RunInBackendGround {
 		req.RunInBackendGround = true
 	}
-	req.Command = security.PrepareNonInteractiveShellCommand(prependPythonUnbufferedEnv(req.Command))
+	req.Command = prependPythonUnbufferedEnv(req.Command)
 	convID := mcp.MCPConversationIDFromContext(ctx)
 	execReg := mcp.EinoExecuteRunRegistryFromContext(ctx)
 

internal/multiagent/eino_execute_streaming_wrap_test.go

@@ -61,12 +61,6 @@ func TestEinoStreamingShellWrap_PreparesNonInteractiveCommand(t *testing.T) {
 			t.Fatalf("recv: %v", rerr)
 		}
 	}
-	if !strings.Contains(inner.lastCommand, "exec </dev/null") {
-		t.Fatalf("missing stdin redirect in inner command: %q", inner.lastCommand)
-	}
-	if !strings.Contains(inner.lastCommand, "GIT_PAGER=cat") {
-		t.Fatalf("missing pager export in inner command: %q", inner.lastCommand)
-	}
 	if !strings.Contains(inner.lastCommand, "PYTHONUNBUFFERED=1") {
 		t.Fatalf("missing python unbuffer in inner command: %q", inner.lastCommand)
 	}

internal/multiagent/eino_orchestration.go

@@ -80,34 +80,9 @@ func NewPlanExecuteRoot(ctx context.Context, a *PlanExecuteRootArgs) (adk.Resuma
 		return nil, fmt.Errorf("plan_execute replanner: %w", err)
 	}
 
-	// 组装 executor handler 栈，顺序与 Deep/Supervisor 主代理一致（outermost first）。
+	execHandlers, err := buildPlanExecuteExecutorHandlers(ctx, a)
-	var execHandlers []adk.ChatModelAgentMiddleware
+	if err != nil {
-	// 1. patchtoolcalls, reduction, toolsearch, plantask（来自 prependEinoMiddlewares）
+		return nil, err
-	if len(a.ExecPreMiddlewares) > 0 {
-		execHandlers = append(execHandlers, a.ExecPreMiddlewares...)
-	}
-	// 2. filesystem 中间件（可选）
-	if a.FilesystemMiddleware != nil {
-		execHandlers = append(execHandlers, a.FilesystemMiddleware)
-	}
-	// 3. skill 中间件（可选）
-	if a.SkillMiddleware != nil {
-		execHandlers = append(execHandlers, a.SkillMiddleware)
-	}
-	// 4. pre-summarization normalize + continuation dedup, then summarization (与 Deep/Supervisor 一致)
-	if a.AppCfg != nil {
-		sumMw, sumErr := newEinoSummarizationMiddleware(ctx, a.ExecModel, a.AppCfg, a.MwCfg, a.ConversationID, a.DB, a.ProjectID, a.Logger)
-		if sumErr != nil {
-			return nil, fmt.Errorf("plan_execute executor summarization: %w", sumErr)
-		}
-		execHandlers = appendEinoChatModelTailMiddlewares(execHandlers, einoChatModelTailConfig{
-			logger:         a.Logger,
-			phase:          "plan_execute_executor",
-			summarization:  sumMw,
-			modelName:      a.ModelName,
-			conversationID: a.ConversationID,
-			trace:          a.ModelFacingTrace,
-		})
 	}
 	executor, err := newPlanExecuteExecutor(ctx, &planexecute.ExecutorConfig{
 		Model:         a.ExecModel,
@@ -130,6 +105,39 @@ func NewPlanExecuteRoot(ctx context.Context, a *PlanExecuteRootArgs) (adk.Resuma
 	})
 }
 
+// buildPlanExecuteExecutorHandlers 组装 Executor 中间件栈（outermost first），与 Deep/Supervisor 主代理对齐：
+// ExecPreMiddlewares（patch / reduction / toolsearch / plantask）→ filesystem → skill → summarization tail。
+func buildPlanExecuteExecutorHandlers(ctx context.Context, a *PlanExecuteRootArgs) ([]adk.ChatModelAgentMiddleware, error) {
+	if a == nil {
+		return nil, fmt.Errorf("plan_execute: args 为空")
+	}
+	var execHandlers []adk.ChatModelAgentMiddleware
+	if len(a.ExecPreMiddlewares) > 0 {
+		execHandlers = append(execHandlers, a.ExecPreMiddlewares...)
+	}
+	if a.FilesystemMiddleware != nil {
+		execHandlers = append(execHandlers, a.FilesystemMiddleware)
+	}
+	if a.SkillMiddleware != nil {
+		execHandlers = append(execHandlers, a.SkillMiddleware)
+	}
+	if a.AppCfg != nil {
+		sumMw, sumErr := newEinoSummarizationMiddleware(ctx, a.ExecModel, a.AppCfg, a.MwCfg, a.ConversationID, a.DB, a.ProjectID, a.Logger)
+		if sumErr != nil {
+			return nil, fmt.Errorf("plan_execute executor summarization: %w", sumErr)
+		}
+		execHandlers = appendEinoChatModelTailMiddlewares(execHandlers, einoChatModelTailConfig{
+			logger:         a.Logger,
+			phase:          "plan_execute_executor",
+			summarization:  sumMw,
+			modelName:      a.ModelName,
+			conversationID: a.ConversationID,
+			trace:          a.ModelFacingTrace,
+		})
+	}
+	return execHandlers, nil
+}
+
 // planExecutePlannerGenInput 将 orchestrator instruction 作为 SystemMessage 注入 planner 输入。
 // 返回 nil 时 Eino 使用内置默认 planner prompt。
 func planExecutePlannerGenInput(

internal/multiagent/eino_single_runner.go

@@ -184,6 +184,7 @@ func RunEinoSingleChatModelAgent(
 		Name:          einoSingleAgentName,
 		Description:   "Eino ADK ChatModelAgent with MCP tools for authorized security testing.",
 		Instruction:   ins,
+		GenModelInput: literalInstructionGenModelInput,
 		Model:         mainModel,
 		ToolsConfig:   mainToolsCfg,
 		MaxIterations: maxIter,

internal/multiagent/eino_summarize.go

@@ -22,15 +22,60 @@ import (
 	"go.uber.org/zap"
 )
 
-// einoSummarizeUserInstruction：压缩历史时保留渗透测试关键信息。
+// einoSummarizeUserInstruction：压缩历史时保留渗透测试与用户约束关键信息。
-const einoSummarizeUserInstruction = `在保持所有关键安全测试信息完整的前提下压缩对话历史。
+// 结构对齐 Eino 最佳实践（禁止工具、<analysis>+<summary>、<all_user_messages>），章节为安全测试领域化。
+const einoSummarizeUserInstruction = `关键：仅以纯文本响应。禁止调用任何工具（read_file、exec、grep、glob、write、edit 等）。
+上述对话中已包含全部待压缩上下文；不要要求用户粘贴历史，不要输出「请提供待压缩的对话历史」等占位/meta 回复。
+工具调用将被拒绝并浪费唯一一次摘要机会。
 
-必须保留：已确认漏洞与攻击路径、工具输出中的核心发现、凭证与认证细节、架构与薄弱点、当前进度、失败尝试与死路、策略决策。
+你的任务：在保持所有关键安全测试信息完整的前提下压缩对话历史，使后续代理能无缝继续同一授权测试任务。
-保留精确技术细节（URL、路径、参数、Payload、版本号、报错原文可摘要但要点不丢）。
-将冗长扫描输出概括为结论；重复发现合并表述。
-已枚举资产须保留**可继承的摘要**：主域、关键子域/主机短表（或数量+代表样例）、高价值目标与已识别服务/端口要点，避免后续子代理因「看不见清单」而重复全量枚举。
 
-输出须使后续代理能无缝继续同一授权测试任务。`
+压缩原则：
+- 必须保留：已确认漏洞与攻击路径、工具输出核心发现、凭证与认证细节、架构与薄弱点、当前进度、失败尝试与死路、策略决策
+- 保留精确技术细节（URL、路径、参数、Payload、版本号；报错原文可摘要但要点不丢）
+- 冗长扫描输出概括为结论；重复发现合并表述
+- 已枚举资产须保留可继承摘要：主域、关键子域/主机短表（或数量+代表样例）、高价值目标、已识别服务/端口要点
+
+输出格式（严格遵循，仅一轮回复）：
+1. 先输出 <analysis> 块：按时间顺序梳理对话，检查是否涵盖下方各章节要点；analysis 仅供自检，保持简洁（建议 ≤400 字）
+2. 再输出 <summary> 块：按以下章节写入可继承的压缩报告（无信息处写「无」，禁止留空模板占位符）
+
+<summary>
+## 1. 授权范围与约束
+- 目标/范围/禁止项（域名、路径、IP、环境）
+- 凭证/认证信息（账号、Token、Cookie；敏感值原文保留）
+- 用户指定的方法、工具、优先级与待办
+- 否定约束（不测什么、不用什么手法）
+
+## 2. 资产与服务枚举摘要
+- 主域/核心资产、关键子域或主机短表（或数量+代表样例）
+- 高价值目标、已识别服务/端口要点
+- 资产状态（存活/可攻/已排除/待验证）
+
+## 3. 架构与已知薄弱点
+- 技术栈/部署拓扑/信任边界
+- 已识别薄弱点列表
+
+## 4. 已确认漏洞与攻击路径
+- 漏洞名/CVE、URL/路径、参数/Payload、PoC 要点、影响等级
+- 攻击链/利用路径（步骤化）
+
+## 5. 工具核心发现与扫描结论
+- 各工具结论（概括核心输出，非冗长日志）
+- 重复发现合并表述
+
+## 6. 所有用户消息
+<all_user_messages>
+- [逐条列出非 tool 结果的用户消息要点；敏感约束与原文措辞尽量保留]
+</all_user_messages>
+
+## 7. 当前进度、策略决策与下一步
+- 当前位置（已完成/进行中/卡点）
+- 失败尝试与死路（方法、现象/报错摘要、结论）
+- 策略决策与下一步具体操作（须与最近用户请求及未完成任务一致）
+</summary>
+
+提醒：不要调用任何工具；必须基于上文已有对话直接输出 <analysis> 与 <summary>，勿输出 analysis 以外的正文。`
 
 // newEinoSummarizationMiddleware 使用 Eino ADK Summarization 中间件（见 https://www.cloudwego.io/zh/docs/eino/core_modules/eino_adk/eino_adk_chatmodelagentmiddleware/middleware_summarization/）。
 // 触发阈值：估算 token 超过 openai.max_total_tokens * summarization_trigger_ratio（默认 0.8）时摘要。
@@ -144,6 +189,7 @@ func newEinoSummarizationMiddleware(
 			},
 		},
 		Finalize: func(ctx context.Context, originalMessages []adk.Message, summary adk.Message) ([]adk.Message, error) {
+			summary = stripAnalysisFromSummarizationMessage(summary)
 			out, ferr := summarizeFinalizeWithRecentAssistantToolTrail(ctx, originalMessages, summary, tokenCounter, recentTrailMax)
 			if ferr != nil {
 				return nil, ferr

internal/multiagent/eino_summarize_output.go

@@ -0,0 +1,73 @@
+package multiagent
+
+import (
+	"regexp"
+	"strings"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+)
+
+var (
+	summarizationAnalysisBlockRegex = regexp.MustCompile(`(?is)<analysis>\s*.*?\s*</analysis>`)
+	summarizationSummaryBlockRegex  = regexp.MustCompile(`(?is)<summary>\s*(.*?)\s*</summary>`)
+)
+
+// stripAnalysisFromSummarizationMessage removes the <analysis> block from a post-processed
+// Eino summary user message. Analysis helps one-shot generation quality but should not
+// occupy continuation context after compaction.
+func stripAnalysisFromSummarizationMessage(msg adk.Message) adk.Message {
+	if msg == nil {
+		return msg
+	}
+	cloned := *msg
+	if cloned.Content != "" {
+		cloned.Content = stripAnalysisFromSummarizationText(cloned.Content)
+	}
+	if len(cloned.UserInputMultiContent) > 0 {
+		parts := make([]schema.MessageInputPart, len(cloned.UserInputMultiContent))
+		copy(parts, cloned.UserInputMultiContent)
+		// Only the first text part carries model output plus Eino preamble/transcript path.
+		for i := range parts {
+			if parts[i].Type != schema.ChatMessagePartTypeText || parts[i].Text == "" {
+				continue
+			}
+			if i == 0 {
+				parts[i].Text = stripAnalysisFromSummarizationText(parts[i].Text)
+			}
+			break
+		}
+		cloned.UserInputMultiContent = parts
+	}
+	return &cloned
+}
+
+func stripAnalysisFromSummarizationText(text string) string {
+	text = strings.TrimSpace(text)
+	if text == "" {
+		return text
+	}
+	stripped := strings.TrimSpace(summarizationAnalysisBlockRegex.ReplaceAllString(text, ""))
+	if stripped == "" {
+		return text
+	}
+	return stripped
+}
+
+// extractSummarizationSummaryBody returns the inner text of the last <summary> block when present.
+// Used by tests and optional strict compaction paths.
+func extractSummarizationSummaryBody(text string) (string, bool) {
+	text = strings.TrimSpace(text)
+	if text == "" {
+		return "", false
+	}
+	all := summarizationSummaryBlockRegex.FindAllStringSubmatch(text, -1)
+	if len(all) == 0 || len(all[len(all)-1]) < 2 {
+		return "", false
+	}
+	body := strings.TrimSpace(all[len(all)-1][1])
+	if body == "" {
+		return "", false
+	}
+	return body, true
+}

internal/multiagent/eino_summarize_output_test.go

@@ -0,0 +1,67 @@
+package multiagent
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/cloudwego/eino/schema"
+)
+
+func TestStripAnalysisFromSummarizationText(t *testing.T) {
+	in := "<analysis>internal notes</analysis>\n\n<summary>\n## 1. 授权\n- example.com\n</summary>"
+	got := stripAnalysisFromSummarizationText(in)
+	if strings.Contains(got, "<analysis>") {
+		t.Fatalf("analysis block should be removed: %q", got)
+	}
+	if !strings.Contains(got, "## 1. 授权") {
+		t.Fatalf("summary body should remain: %q", got)
+	}
+}
+
+func TestStripAnalysisFromSummarizationMessage_UserInputMultiContent(t *testing.T) {
+	msg := &schema.Message{
+		Role: schema.User,
+		UserInputMultiContent: []schema.MessageInputPart{
+			{
+				Type: schema.ChatMessagePartTypeText,
+				Text: "此会话延续自此前一段因上下文耗尽而终止的对话。\n\n<analysis>draft</analysis>\n<summary>body</summary>\n\n完整记录位于：/tmp/transcript.txt",
+			},
+			{
+				Type: schema.ChatMessagePartTypeText,
+				Text: "请从我们中断的地方继续对话，无需向用户提出任何进一步的问题。",
+			},
+		},
+	}
+	out := stripAnalysisFromSummarizationMessage(msg)
+	if len(out.UserInputMultiContent) != 2 {
+		t.Fatalf("expected 2 parts, got %d", len(out.UserInputMultiContent))
+	}
+	if strings.Contains(out.UserInputMultiContent[0].Text, "<analysis>") {
+		t.Fatalf("part 0 should drop analysis: %q", out.UserInputMultiContent[0].Text)
+	}
+	if !strings.Contains(out.UserInputMultiContent[0].Text, "<summary>body</summary>") {
+		t.Fatalf("part 0 should keep summary: %q", out.UserInputMultiContent[0].Text)
+	}
+	if out.UserInputMultiContent[1].Text != "请从我们中断的地方继续对话，无需向用户提出任何进一步的问题。" {
+		t.Fatalf("continue instruction part should be unchanged: %q", out.UserInputMultiContent[1].Text)
+	}
+}
+
+func TestExtractSummarizationSummaryBody(t *testing.T) {
+	body, ok := extractSummarizationSummaryBody("<analysis>x</analysis><summary>  kept  </summary>")
+	if !ok || body != "kept" {
+		t.Fatalf("extract summary body: ok=%v body=%q", ok, body)
+	}
+	_, ok = extractSummarizationSummaryBody("plain text only")
+	if ok {
+		t.Fatal("expected false for plain text")
+	}
+}
+
+func TestStripAnalysisFromSummarizationText_NoAnalysisUnchanged(t *testing.T) {
+	in := "<summary>only summary</summary>"
+	got := stripAnalysisFromSummarizationText(in)
+	if got != in {
+		t.Fatalf("expected unchanged text, got %q", got)
+	}
+}

internal/multiagent/eino_summarize_payload.go

@@ -1,35 +1,12 @@
 package multiagent
 
 import (
-	"github.com/bytedance/sonic"
+	copenai "cyberstrike-ai/internal/openai"
 )
 
 // stripReasoningFromSummarizationPayload removes thinking / reasoning fields from a
 // chat-completions JSON body. Applied only to summarization Generate calls via
 // model.ModelOptions on the shared ChatModel — main-agent requests are unchanged.
 func stripReasoningFromSummarizationPayload(rawBody []byte) ([]byte, error) {
-	var payload map[string]any
+	return copenai.StripReasoningFromChatCompletionBody(rawBody)
-	if err := sonic.Unmarshal(rawBody, &payload); err != nil {
-		return rawBody, nil
-	}
-	changed := false
-	for _, key := range []string{
-		"thinking",
-		"reasoning_effort",
-		"output_config",
-		"reasoning",
-	} {
-		if _, ok := payload[key]; ok {
-			delete(payload, key)
-			changed = true
-		}
-	}
-	if !changed {
-		return rawBody, nil
-	}
-	out, err := sonic.Marshal(payload)
-	if err != nil {
-		return rawBody, err
-	}
-	return out, nil
 }

internal/multiagent/eino_summarize_test.go

@@ -409,9 +409,9 @@ func TestSanitizeSystemContentForTranscript_BestPractice(t *testing.T) {
 		"需要写入请使用 upsert_project_fact。",
 		project.FactIndexSectionEndMarker,
 		"",
-		"# Skills System",
+		transcriptSkillsSystemMarker,
-		"**How to Use Skills**",
+		"**如何使用 Skill（技能）（渐进式展示）：**",
-		"Remember: Skills make you more capable",
+		"记住：Skill 让你更加强大和稳定",
 	}, "\n")
 
 	out := sanitizeSystemContentForTranscript(system)
@@ -421,7 +421,7 @@ func TestSanitizeSystemContentForTranscript_BestPractice(t *testing.T) {
 	if strings.Contains(out, "- nmap") || strings.Contains(out, "高强度扫描要求") {
 		t.Fatalf("static persona should be stripped: %q", out)
 	}
-	if strings.Contains(out, "# Skills System") || strings.Contains(out, "How to Use Skills") {
+	if strings.Contains(out, transcriptSkillsSystemMarker) || strings.Contains(out, "如何使用 Skill") {
 		t.Fatalf("skills boilerplate should be stripped: %q", out)
 	}
 	if !strings.Contains(out, transcriptStaticSystemOmitNote) {
@@ -435,7 +435,7 @@ func TestSanitizeSystemContentForTranscript_BestPractice(t *testing.T) {
 func TestFormatSummarizationTranscript_OmitsBloatedSystem(t *testing.T) {
 	t.Parallel()
 	msgs := []adk.Message{
-		schema.SystemMessage("以下是当前会话绑定的工具名称索引\n- nmap\n\n你是CyberStrikeAI\n" + project.FactIndexSectionStartMarker + "\n## 项目黑板索引（project: p1, id: x）\n（暂无事实）\n" + project.FactIndexSectionEndMarker + "\n# Skills System\nboiler"),
+		schema.SystemMessage("以下是当前会话绑定的工具名称索引\n- nmap\n\n你是CyberStrikeAI\n" + project.FactIndexSectionStartMarker + "\n## 项目黑板索引（project: p1, id: x）\n（暂无事实）\n" + project.FactIndexSectionEndMarker + "\n" + transcriptSkillsSystemMarker + "\nboiler"),
 		schema.UserMessage("hello"),
 		schema.AssistantMessage("reply", nil),
 	}

internal/multiagent/eino_summarize_transcript.go

@@ -20,7 +20,9 @@ const (
 	transcriptStaticSystemOmitNote = "[static system prompt omitted — unchanged in live context after compaction]"
 	transcriptToolIndexStartMarker   = "以下是当前会话绑定的工具名称索引"
 	transcriptPersonaStartMarker     = "你是CyberStrikeAI"
-	transcriptSkillsSystemMarker     = "# Skills System"
+	// ADK LanguageChinese injects skill middleware prompt with this header (see eino adk/middlewares/skill/prompt.go).
+	transcriptSkillsSystemMarker        = "# Skill 系统"
+	transcriptSkillsSystemMarkerEnglish = "# Skills System"
 )
 
 type transcriptToolCall struct {
@@ -86,13 +88,23 @@ func stripToolNamesIndexFromSystem(s string) string {
 }
 
 func stripSkillsSystemBoilerplate(s string) string {
-	idx := strings.Index(s, transcriptSkillsSystemMarker)
+	idx := indexFirstSubstring(s, transcriptSkillsSystemMarker, transcriptSkillsSystemMarkerEnglish)
 	if idx < 0 {
 		return strings.TrimSpace(s)
 	}
 	return strings.TrimSpace(s[:idx])
 }
 
+func indexFirstSubstring(s string, markers ...string) int {
+	first := -1
+	for _, m := range markers {
+		if i := strings.Index(s, m); i >= 0 && (first < 0 || i < first) {
+			first = i
+		}
+	}
+	return first
+}
+
 func extractProjectBlackboardSection(s string) string {
 	start := strings.Index(s, project.FactIndexSectionStartMarker)
 	if start < 0 {

internal/multiagent/eino_transient_retry.go

@@ -190,29 +190,26 @@ func einoMessagesForRunRestart(args *einoADKRunLoopArgs, baseMsgs, accumulated [
 	return append([]adk.Message(nil), baseMsgs...), einoRestartContextInitial
 }
 
-// adkMessagesHasUserContent 从尾部向前查找，是否已有与 want 相同的 user 消息（避免重复 append）。
+// adkMessagesHasUserContent reports whether the conversation tail is already a user turn
+// with the given content. Only the last message counts: matching text in an earlier round
+// (e.g. user repeats the same prompt after an assistant reply) must not suppress appending
+// the new user turn — Claude 4.6+ rejects requests whose final message is assistant.
 func adkMessagesHasUserContent(msgs []adk.Message, want string) bool {
 	want = strings.TrimSpace(want)
 	if want == "" {
 		return true
 	}
-	for i := len(msgs) - 1; i >= 0; i-- {
+	if len(msgs) == 0 {
-		m := msgs[i]
+		return false
-		if m == nil {
-			continue
-		}
-		if m.Role == schema.User {
-			return strings.TrimSpace(m.Content) == want
-		}
-		if m.Role == schema.Assistant || m.Role == schema.Tool {
-			continue
-		}
-		break
 	}
-	return false
+	last := msgs[len(msgs)-1]
+	if last == nil || last.Role != schema.User {
+		return false
+	}
+	return strings.TrimSpace(last.Content) == want
 }
 
-// appendUserMessageIfNeeded 在 history 轨迹之后追加本轮 user 消息（仅当轨迹中尚未包含该句）。
+// appendUserMessageIfNeeded 在 history 轨迹之后追加本轮 user 消息（仅当尾部已是相同 user 句）。
 func appendUserMessageIfNeeded(msgs []adk.Message, userMessage string) []adk.Message {
 	if strings.TrimSpace(userMessage) == "" || adkMessagesHasUserContent(msgs, userMessage) {
 		return msgs

internal/multiagent/eino_transient_retry_test.go

@@ -143,3 +143,18 @@ func TestAppendUserMessageIfNeeded(t *testing.T) {
 		t.Fatalf("should not duplicate user message: len=%d", len(dup))
 	}
 }
+
+func TestAppendUserMessageIfNeeded_repeatPromptAfterAssistant(t *testing.T) {
+	t.Parallel()
+	msgs := []adk.Message{
+		schema.UserMessage("扫描 example.com"),
+		schema.AssistantMessage("开始扫描...", nil),
+	}
+	out := appendUserMessageIfNeeded(msgs, "扫描 example.com")
+	if len(out) != 3 {
+		t.Fatalf("should append new user turn after assistant reply: len=%d", len(out))
+	}
+	if out[2].Role != schema.User || out[2].Content != "扫描 example.com" {
+		t.Fatalf("tail should be repeated user prompt, got role=%s content=%q", out[2].Role, out[2].Content)
+	}
+}

internal/multiagent/gen_model_input.go

@@ -0,0 +1,23 @@
+package multiagent
+
+import (
+	"context"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+)
+
+// literalInstructionGenModelInput passes Instruction through as a system message without
+// FString template formatting. Eino defaultGenModelInput formats instruction whenever
+// SessionValues exist; prompts with literal curly braces (project blackboard "{关系边: ...}",
+// JSON examples, link syntax) then fail with "could not find key".
+//
+// Matches eino/adk/prebuilt/deep genModelInput — the supported fix per Eino docs.
+func literalInstructionGenModelInput(ctx context.Context, instruction string, input *adk.AgentInput) ([]adk.Message, error) {
+	msgs := make([]adk.Message, 0, len(input.Messages)+1)
+	if instruction != "" {
+		msgs = append(msgs, schema.SystemMessage(instruction))
+	}
+	msgs = append(msgs, input.Messages...)
+	return msgs, nil
+}

internal/multiagent/gen_model_input_test.go

@@ -0,0 +1,33 @@
+package multiagent
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+)
+
+func TestLiteralInstructionGenModelInput_PreservesLiteralCurlyBraces(t *testing.T) {
+	t.Parallel()
+	instruction := "- [finding/x] summary {关系边: discovered_on←target/dev}\n" +
+		"如 finding 上 {from:target/*, type:discovered_on}"
+	msgs, err := literalInstructionGenModelInput(context.Background(), instruction, &adk.AgentInput{
+		Messages: []adk.Message{schema.UserMessage("继续")},
+	})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(msgs) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(msgs))
+	}
+	if msgs[0].Role != schema.System {
+		t.Fatalf("first message must be system, got %s", msgs[0].Role)
+	}
+	for _, want := range []string{"{关系边:", "{from:target/*, type:discovered_on}"} {
+		if !strings.Contains(msgs[0].Content, want) {
+			t.Fatalf("system content missing %q: %q", want, msgs[0].Content)
+		}
+	}
+}

internal/multiagent/hitl_middleware.go

@@ -3,7 +3,6 @@ package multiagent
 import (
 	"context"
 	"errors"
-	"fmt"
 	"strings"
 
 	"github.com/cloudwego/eino/adk"
@@ -75,8 +74,8 @@ func hitlInvokableToolCallMiddleware() compose.InvokableToolMiddleware {
 					if err != nil {
 						if IsHumanRejectError(err) {
 							// Human rejection should be a soft tool result so the model can continue iterating.
-							msg := fmt.Sprintf("[HITL Reject] Tool '%s' was rejected by human reviewer. Reason: %s\nPlease adjust parameters/plan and continue without this call.",
+							// tool_search 须保持 JSON，否则 Eino toolsearch 中间件解析历史时会硬崩 ChatModel。
-								input.Name, strings.TrimSpace(err.Error()))
+							msg := HitlRejectToolResult(input.Name, err.Error())
 							// transfer_to_agent 在 Eino 中标记为 returnDirectly：工具成功后 ReAct 子图会直接 END，
 							// 并依赖真实工具内的 SendToolGenAction 触发移交。HITL 拒绝时不会执行真实工具，
 							// 若仍走 returnDirectly 分支，监督者会在无 Transfer 动作的情况下结束，模型不再迭代。
@@ -103,8 +102,7 @@ func hitlStreamableToolCallMiddleware() compose.StreamableToolMiddleware {
 					edited, err := fn(ctx, input.Name, input.Arguments)
 					if err != nil {
 						if IsHumanRejectError(err) {
-							msg := fmt.Sprintf("[HITL Reject] Tool '%s' was rejected by human reviewer. Reason: %s\nPlease adjust parameters/plan and continue without this call.",
+							msg := HitlRejectToolResult(input.Name, err.Error())
-								input.Name, strings.TrimSpace(err.Error()))
 							hitlClearReturnDirectlyIfTransfer(ctx, input.Name)
 							return &compose.StreamToolOutput{
 								Result: schema.StreamReaderFromArray([]string{msg}),

internal/multiagent/hitl_toolsearch_compat.go

@@ -0,0 +1,85 @@
+package multiagent
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+)
+
+const toolSearchToolName = "tool_search"
+
+// HitlExemptMetaTools 为编排/元工具：不直接执行攻击动作，但会阻塞 agent 控制流。
+// tool_search 必须免审批，否则其 HITL 拒绝结果与 Eino toolsearch 中间件不兼容（会硬崩 ChatModel）。
+var HitlExemptMetaTools = []string{
+	toolSearchToolName,
+	"skill",
+	"task",
+	"write_todos",
+	"transfer_to_agent",
+	"exit",
+	"TaskCreate",
+	"TaskGet",
+	"TaskUpdate",
+	"TaskList",
+}
+
+// IsToolSearchTool reports whether name is the Eino dynamictool tool_search meta-tool.
+func IsToolSearchTool(name string) bool {
+	return strings.EqualFold(strings.TrimSpace(name), toolSearchToolName)
+}
+
+// MergeHitlExemptMetaTools unions configured whitelist with built-in meta-tool exemptions.
+func MergeHitlExemptMetaTools(configured []string) []string {
+	merged := make([]string, 0, len(configured)+len(HitlExemptMetaTools))
+	seen := make(map[string]struct{}, len(configured)+len(HitlExemptMetaTools))
+	add := func(name string) {
+		n := strings.ToLower(strings.TrimSpace(name))
+		if n == "" {
+			return
+		}
+		if _, ok := seen[n]; ok {
+			return
+		}
+		seen[n] = struct{}{}
+		merged = append(merged, strings.TrimSpace(name))
+	}
+	for _, t := range configured {
+		add(t)
+	}
+	for _, t := range HitlExemptMetaTools {
+		add(t)
+	}
+	return merged
+}
+
+type toolSearchHitlRejectPayload struct {
+	SelectedTools []string `json:"selectedTools"`
+	HitlRejected  bool     `json:"_hitlRejected"`
+	Reason        string   `json:"reason"`
+}
+
+// HitlRejectToolResult returns a tool result body safe for downstream consumers.
+// tool_search must stay JSON-shaped so toolsearch.extractSelectedTools does not terminate the graph.
+func HitlRejectToolResult(toolName, reason string) string {
+	reason = strings.TrimSpace(reason)
+	if !IsToolSearchTool(toolName) {
+		if reason == "" {
+			reason = "rejected by reviewer"
+		}
+		return fmt.Sprintf("[HITL Reject] Tool '%s' was rejected by reviewer. Reason: %s\nPlease adjust parameters/plan and continue without this call.",
+			strings.TrimSpace(toolName), reason)
+	}
+	payload := toolSearchHitlRejectPayload{
+		SelectedTools: []string{},
+		HitlRejected:  true,
+		Reason:        reason,
+	}
+	if payload.Reason == "" {
+		payload.Reason = "tool_search rejected by reviewer; no dynamic tools unlocked"
+	}
+	out, err := json.Marshal(payload)
+	if err != nil {
+		return `{"selectedTools":[],"_hitlRejected":true,"reason":"tool_search rejected by reviewer"}`
+	}
+	return string(out)
+}

internal/multiagent/hitl_toolsearch_compat_test.go

@@ -0,0 +1,48 @@
+package multiagent
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+func TestHitlRejectToolResult_toolSearchIsJSON(t *testing.T) {
+	raw := HitlRejectToolResult("tool_search", "rejected by user: timeout")
+	var payload toolSearchHitlRejectPayload
+	if err := json.Unmarshal([]byte(raw), &payload); err != nil {
+		t.Fatalf("unmarshal: %v", err)
+	}
+	if len(payload.SelectedTools) != 0 {
+		t.Fatalf("expected empty selectedTools, got %v", payload.SelectedTools)
+	}
+	if !payload.HitlRejected {
+		t.Fatal("expected _hitlRejected true")
+	}
+	if !strings.Contains(payload.Reason, "timeout") {
+		t.Fatalf("reason=%q", payload.Reason)
+	}
+}
+
+func TestHitlRejectToolResult_otherToolKeepsLegacyText(t *testing.T) {
+	raw := HitlRejectToolResult("nmap", "too risky")
+	if strings.HasPrefix(raw, "{") {
+		t.Fatalf("expected legacy text, got %q", raw)
+	}
+	if !strings.HasPrefix(raw, "[HITL Reject]") {
+		t.Fatalf("expected [HITL Reject] prefix, got %q", raw)
+	}
+}
+
+func TestMergeHitlExemptMetaTools_includesToolSearch(t *testing.T) {
+	merged := MergeHitlExemptMetaTools([]string{"read_file"})
+	found := false
+	for _, name := range merged {
+		if IsToolSearchTool(name) {
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Fatalf("tool_search missing from %v", merged)
+	}
+}

internal/multiagent/plan_execute_executor_handlers_test.go

@@ -0,0 +1,85 @@
+package multiagent
+
+import (
+	"context"
+	"fmt"
+	"testing"
+
+	"cyberstrike-ai/internal/config"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/components/tool"
+)
+
+type stubChatModelAgentMiddleware struct {
+	adk.BaseChatModelAgentMiddleware
+	tag string
+}
+
+func stubMW(tag string) adk.ChatModelAgentMiddleware {
+	return &stubChatModelAgentMiddleware{tag: tag}
+}
+
+func TestBuildPlanExecuteExecutorHandlers_IncludesExecPreMiddlewares(t *testing.T) {
+	t.Parallel()
+	pre := []adk.ChatModelAgentMiddleware{
+		stubMW("patch"),
+		stubMW("reduction"),
+	}
+
+	got, err := buildPlanExecuteExecutorHandlers(context.Background(), &PlanExecuteRootArgs{
+		ExecPreMiddlewares:   pre,
+		FilesystemMiddleware: stubMW("filesystem"),
+		SkillMiddleware:      stubMW("skill"),
+	})
+	if err != nil {
+		t.Fatalf("buildPlanExecuteExecutorHandlers: %v", err)
+	}
+	if len(got) != 4 {
+		t.Fatalf("expected 4 pre-tail handlers (2 pre + fs + skill), got %d", len(got))
+	}
+	for i, want := range []string{"patch", "reduction", "filesystem", "skill"} {
+		st, ok := got[i].(*stubChatModelAgentMiddleware)
+		if !ok || st.tag != want {
+			t.Fatalf("handler[%d]: got %#v want tag %q", i, got[i], want)
+		}
+	}
+}
+
+func stubTools(n int) []tool.BaseTool {
+	out := make([]tool.BaseTool, n)
+	for i := 0; i < n; i++ {
+		out[i] = stubTool{name: fmt.Sprintf("t%d", i)}
+	}
+	return out
+}
+
+func TestBuildPlanExecuteExecutorHandlers_NilArgs(t *testing.T) {
+	t.Parallel()
+	if _, err := buildPlanExecuteExecutorHandlers(context.Background(), nil); err == nil {
+		t.Fatal("expected error for nil args")
+	}
+}
+
+func TestPrependEinoMiddlewares_Main_IncludesPatch(t *testing.T) {
+	t.Parallel()
+	ctx := context.Background()
+	mw := configMultiAgentEinoMiddlewareForTest()
+	mw.ReductionEnable = false
+	mw.ToolSearchEnable = false
+	mw.PlantaskEnable = false
+	_, extra, _, err := prependEinoMiddlewares(ctx, mw, einoMWMain, stubTools(25), nil, "", "conv-test", "", nil)
+	if err != nil {
+		t.Fatalf("prependEinoMiddlewares: %v", err)
+	}
+	if len(extra) == 0 {
+		t.Fatal("expected patch middleware on einoMWMain when patch_tool_calls enabled")
+	}
+}
+
+func configMultiAgentEinoMiddlewareForTest() *config.MultiAgentEinoMiddlewareConfig {
+	patch := true
+	return &config.MultiAgentEinoMiddlewareConfig{
+		PatchToolCalls: &patch,
+	}
+}

internal/multiagent/runner.go

@@ -254,10 +254,11 @@ func RunDeepAgent(
 				)
 			}
 			sa, err := adk.NewChatModelAgent(ctx, &adk.ChatModelAgentConfig{
-				Name:        id,
+				Name:          id,
-				Description: desc,
+				Description:   desc,
-				Instruction: subInstrFinal,
+				Instruction:   subInstrFinal,
-				Model:       subModel,
+				GenModelInput: literalInstructionGenModelInput,
+				Model:         subModel,
 				ToolsConfig: adk.ToolsConfig{
 					ToolsNodeConfig: compose.ToolsNodeConfig{
 						Tools:               subToolsForCfg,
@@ -372,8 +373,15 @@ func RunDeepAgent(
 
 	// noNestedTaskMiddleware 必须在最外层（最先拦截），防止 skill 或其他中间件内部触发 task 调用绕过检测。
 	deepHandlers := []adk.ChatModelAgentMiddleware{newNoNestedTaskMiddleware()}
-	taskEnrichExtra := systemPromptExtra
+	var taskBlackboardSupplement string
-	if mw := newTaskContextEnrichMiddleware(userMessage, history, ma.SubAgentUserContextMaxRunes, taskEnrichExtra); mw != nil {
+	if appCfg.Project.Enabled && db != nil {
+		if pid := strings.TrimSpace(projectID); pid != "" {
+			if block, err := project.BuildFactIndexBlock(db, pid, appCfg.Project); err == nil {
+				taskBlackboardSupplement = strings.TrimSpace(block)
+			}
+		}
+	}
+	if mw := newTaskContextEnrichMiddleware(userMessage, history, ma.SubAgentUserContextMaxRunesEffective(), taskBlackboardSupplement); mw != nil {
 		deepHandlers = append(deepHandlers, mw)
 	}
 	if len(mainOrchestratorPre) > 0 {
@@ -424,6 +432,22 @@ func RunDeepAgent(
 	var da adk.Agent
 	switch orchMode {
 	case "plan_execute":
+		plannerModelCfg := &einoopenai.ChatModelConfig{
+			APIKey:     appCfg.OpenAI.APIKey,
+			BaseURL:    strings.TrimSuffix(appCfg.OpenAI.BaseURL, "/"),
+			Model:      appCfg.OpenAI.Model,
+			HTTPClient: httpClient,
+		}
+		reasoning.ApplyPlanExecutePlannerModelConfig(plannerModelCfg, &appCfg.OpenAI)
+		peMainModel, perr := einoopenai.NewChatModel(ctx, plannerModelCfg)
+		if perr != nil {
+			return nil, fmt.Errorf("plan_execute 规划模型: %w", perr)
+		}
+		if logger != nil {
+			logger.Info("plan_execute: planner/replanner 使用无 reasoning 的独立 ChatModel（ToolChoiceForced 兼容）",
+				zap.String("model", appCfg.OpenAI.Model),
+			)
+		}
 		execModel, perr := einoopenai.NewChatModel(ctx, baseModelCfg)
 		if perr != nil {
 			return nil, fmt.Errorf("plan_execute 执行器模型: %w", perr)
@@ -437,7 +461,7 @@ func RunDeepAgent(
 			}
 		}
 		peRoot, perr := NewPlanExecuteRoot(ctx, &PlanExecuteRootArgs{
-			MainToolCallingModel: mainModel,
+			MainToolCallingModel: peMainModel,
 			ExecModel:            execModel,
 			OrchInstruction:      orchInstruction,
 			ToolsCfg:             mainToolsCfg,
@@ -450,6 +474,7 @@ func RunDeepAgent(
 			ProjectID:            projectID,
 			Logger:               logger,
 			ModelName:            appCfg.OpenAI.Model,
+			// 与 Deep/Supervisor 主代理同源：patch / reduction / toolsearch / plantask（见 buildPlanExecuteExecutorHandlers）。
 			ExecPreMiddlewares:   mainOrchestratorPre,
 			SkillMiddleware:      einoSkillMW,
 			FilesystemMiddleware: peFsMw,
@@ -472,6 +497,7 @@ func RunDeepAgent(
 			Name:          orchestratorName,
 			Description:   orchDescription,
 			Instruction:   supInstr,
+			GenModelInput: literalInstructionGenModelInput,
 			Model:         mainModel,
 			ToolsConfig:   mainToolsCfg,
 			MaxIterations: deepMaxIter,

internal/multiagent/sub_agent_context.go

@@ -3,6 +3,7 @@ package multiagent
 import (
 	"context"
 	"encoding/json"
+	"fmt"
 	"strings"
 
 	"cyberstrike-ai/internal/agent"
@@ -11,7 +12,7 @@ import (
 	"github.com/cloudwego/eino/components/tool"
 )
 
-const defaultSubAgentUserContextMaxRunes = 2000
+const userContextSupplementHeader = "\n\n## 用户历史输入（原文，子代理必读）\n"
 
 // taskContextEnrichMiddleware intercepts "task" tool calls on the orchestrator
 // and appends the user's original conversation messages to the task description.
@@ -30,13 +31,14 @@ type taskContextEnrichMiddleware struct {
 // newTaskContextEnrichMiddleware returns a middleware that enriches task
 // descriptions with user conversation context. Returns nil if disabled
 // (maxRunes < 0) or no user messages exist.
+// projectBlackboard 仅传项目黑板索引块（BuildFactIndexBlock）；勿传完整 systemPromptExtra。
 func newTaskContextEnrichMiddleware(userMessage string, history []agent.ChatMessage, maxRunes int, projectBlackboard string) adk.ChatModelAgentMiddleware {
 	supplement := buildUserContextSupplement(userMessage, history, maxRunes)
 	if bb := strings.TrimSpace(projectBlackboard); bb != "" {
 		if supplement != "" {
-			supplement += "\n\n## 项目黑板索引\n" + bb
+			supplement += "\n\n" + bb
 		} else {
-			supplement = "\n\n## 项目黑板索引\n" + bb
+			supplement = "\n\n" + bb
 		}
 	}
 	if supplement == "" {
@@ -86,9 +88,6 @@ func buildUserContextSupplement(userMessage string, history []agent.ChatMessage,
 	if maxRunes < 0 {
 		return ""
 	}
-	if maxRunes == 0 {
-		maxRunes = defaultSubAgentUserContextMaxRunes
-	}
 
 	var userMsgs []string
 	for _, h := range history {
@@ -107,12 +106,16 @@ func buildUserContextSupplement(userMessage string, history []agent.ChatMessage,
 		return ""
 	}
 
-	joined := strings.Join(userMsgs, "\n---\n")
+	lines := make([]string, 0, len(userMsgs))
-	if len([]rune(joined)) > maxRunes {
+	for i, msg := range userMsgs {
+		lines = append(lines, fmt.Sprintf("[第%d轮] %s", i+1, msg))
+	}
+	joined := strings.Join(lines, "\n")
+	if maxRunes > 0 && len([]rune(joined)) > maxRunes {
 		joined = truncateKeepFirstLast(userMsgs, maxRunes)
 	}
 
-	return "\n\n## 会话上下文（自动补充，确保你了解用户完整意图）\n" + joined
+	return userContextSupplementHeader + joined
 }
 
 // truncateKeepFirstLast keeps the first and last user messages, giving each

internal/multiagent/sub_agent_context_test.go

@@ -74,7 +74,7 @@ func TestBuildUserContextSupplement_DisabledByNegative(t *testing.T) {
 func TestBuildUserContextSupplement_CustomMaxRunes(t *testing.T) {
 	msg := strings.Repeat("A", 200)
 	result := buildUserContextSupplement(msg, nil, 50)
-	header := "\n\n## 会话上下文（自动补充，确保你了解用户完整意图）\n"
+	header := userContextSupplementHeader
 	body := strings.TrimPrefix(result, header)
 	if len([]rune(body)) > 50 {
 		t.Errorf("body should be capped at 50 runes, got %d", len([]rune(body)))
@@ -89,7 +89,7 @@ func TestBuildUserContextSupplement_TruncateKeepsFirstAndLast(t *testing.T) {
 		history = append(history, agent.ChatMessage{Role: "user", Content: strings.Repeat("B", 500)})
 	}
 	last := "最后一条指令"
-	result := buildUserContextSupplement(last, history, 0)
+	result := buildUserContextSupplement(last, history, 800)
 	if !strings.Contains(result, "http://target.com") {
 		t.Error("first message (target URL) should survive truncation")
 	}

internal/openai/claude_bridge.go

@@ -806,10 +806,12 @@ func isClaudeProvider(cfg *config.OpenAIConfig) bool {
 // Eino HTTP Client Bridge
 // ============================================================
 
-// NewEinoHTTPClient 为 einoopenai.ChatModelConfig 返回一个 http.Client，包含两层 transport 包装：
+// NewEinoHTTPClient 为 einoopenai.ChatModelConfig 返回一个 http.Client，包含多层 transport 包装：
-//  1. 当 cfg.Provider 为 claude 时，最内层套 claudeRoundTripper，把 OpenAI /chat/completions 透明
+//  1. 当 cfg.Provider 为 claude 时，套 claudeRoundTripper，把 OpenAI /chat/completions 透明
 //     桥接为 Anthropic /v1/messages（并把 Claude SSE 翻译回 OpenAI SSE 格式）。
-//  2. 最外层无条件套 einoSSESanitizingRoundTripper，吞掉中转站发的 SSE 心跳/注释/控制行
+//  2. reasoningToolChoiceCompatRoundTripper：tool_choice=required/object 时剥离 thinking 字段，避免
+//     plan_execute replanner 等强制工具调用与推理模式冲突（部分网关返回 400）。
+//  3. 最外层无条件套 einoSSESanitizingRoundTripper，吞掉中转站发的 SSE 心跳/注释/控制行
 //     (": keepalive" / "event: ping" / "retry: 3000" 等)，避免 Eino 用的 meguminnnnnnnnn/go-openai
 //     SDK 在累计超过 300 个非 "data:" 行后抛 "stream has sent too many empty messages"。
 //
@@ -825,6 +827,7 @@ func NewEinoHTTPClient(cfg *config.OpenAIConfig, base *http.Client) *http.Client
 	if transport == nil {
 		transport = http.DefaultTransport
 	}
+	transport = &reasoningToolChoiceCompatRoundTripper{base: transport}
 	if isClaudeProvider(cfg) {
 		transport = &claudeRoundTripper{
 			base:   transport,

internal/openai/reasoning_payload.go

@@ -0,0 +1,79 @@
+package openai
+
+import (
+	"github.com/bytedance/sonic"
+)
+
+// reasoningPayloadKeys are OpenAI-compatible root fields that enable "thinking" /
+// extended-reasoning modes on gateways such as DashScope/Qwen and MiniMax.
+var reasoningPayloadKeys = []string{
+	"thinking",
+	"reasoning_effort",
+	"output_config",
+	"reasoning",
+}
+
+// StripReasoningFromChatCompletionBody removes thinking / reasoning fields from a
+// chat-completions JSON body.
+func StripReasoningFromChatCompletionBody(rawBody []byte) ([]byte, error) {
+	var payload map[string]any
+	if err := sonic.Unmarshal(rawBody, &payload); err != nil {
+		return rawBody, nil
+	}
+	if !stripReasoningFields(payload) {
+		return rawBody, nil
+	}
+	out, err := sonic.Marshal(payload)
+	if err != nil {
+		return rawBody, err
+	}
+	return out, nil
+}
+
+// StripReasoningIfForcedToolChoice removes thinking / reasoning fields when the
+// request sets tool_choice to "required" or an object. Several providers reject
+// that combination (e.g. DashScope: "tool_choice does not support being set to
+// required or object in thinking mode").
+func StripReasoningIfForcedToolChoice(rawBody []byte) ([]byte, error) {
+	var payload map[string]any
+	if err := sonic.Unmarshal(rawBody, &payload); err != nil {
+		return rawBody, nil
+	}
+	if !forcedToolChoiceIncompatibleWithThinking(payload) {
+		return rawBody, nil
+	}
+	if !stripReasoningFields(payload) {
+		return rawBody, nil
+	}
+	out, err := sonic.Marshal(payload)
+	if err != nil {
+		return rawBody, err
+	}
+	return out, nil
+}
+
+func stripReasoningFields(payload map[string]any) bool {
+	changed := false
+	for _, key := range reasoningPayloadKeys {
+		if _, ok := payload[key]; ok {
+			delete(payload, key)
+			changed = true
+		}
+	}
+	return changed
+}
+
+func forcedToolChoiceIncompatibleWithThinking(payload map[string]any) bool {
+	tc, ok := payload["tool_choice"]
+	if !ok || tc == nil {
+		return false
+	}
+	switch v := tc.(type) {
+	case string:
+		return v == "required"
+	case map[string]any:
+		return true
+	default:
+		return false
+	}
+}

internal/openai/reasoning_payload_test.go

@@ -0,0 +1,120 @@
+package openai
+
+import (
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+)
+
+func TestStripReasoningFromChatCompletionBody(t *testing.T) {
+	in := []byte(`{"model":"deepseek-chat","messages":[],"thinking":{"type":"enabled"},"reasoning_effort":"high"}`)
+	out, err := StripReasoningFromChatCompletionBody(in)
+	if err != nil {
+		t.Fatal(err)
+	}
+	s := string(out)
+	if strings.Contains(s, "thinking") || strings.Contains(s, "reasoning_effort") {
+		t.Fatalf("expected reasoning fields stripped, got %s", s)
+	}
+	if !strings.Contains(s, `"model":"deepseek-chat"`) {
+		t.Fatalf("expected model preserved, got %s", s)
+	}
+
+	plain := []byte(`{"model":"gpt-4o","messages":[]}`)
+	out2, err := StripReasoningFromChatCompletionBody(plain)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(out2) != string(plain) {
+		t.Fatalf("expected unchanged payload, got %s", out2)
+	}
+}
+
+func TestStripReasoningIfForcedToolChoice(t *testing.T) {
+	cases := []struct {
+		name    string
+		in      string
+		strip   bool
+		contain string
+	}{
+		{
+			name:  "required strips thinking",
+			in:    `{"model":"minimax","messages":[],"thinking":{"type":"enabled"},"tool_choice":"required","tools":[]}`,
+			strip: true,
+		},
+		{
+			name:  "object tool_choice strips thinking",
+			in:    `{"model":"qwen","messages":[],"thinking":{"type":"enabled"},"tool_choice":{"type":"function","function":{"name":"respond"}}}`,
+			strip: true,
+		},
+		{
+			name:    "auto keeps thinking",
+			in:      `{"model":"qwen","messages":[],"thinking":{"type":"enabled"},"tool_choice":"auto"}`,
+			strip:   false,
+			contain: "thinking",
+		},
+		{
+			name:    "no tool_choice keeps thinking",
+			in:      `{"model":"qwen","messages":[],"thinking":{"type":"enabled"}}`,
+			strip:   false,
+			contain: "thinking",
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			out, err := StripReasoningIfForcedToolChoice([]byte(tc.in))
+			if err != nil {
+				t.Fatal(err)
+			}
+			s := string(out)
+			hasThinking := strings.Contains(s, "thinking")
+			if tc.strip && hasThinking {
+				t.Fatalf("expected thinking stripped, got %s", s)
+			}
+			if !tc.strip && tc.contain != "" && !strings.Contains(s, tc.contain) {
+				t.Fatalf("expected %q in %s", tc.contain, s)
+			}
+			if !tc.strip && string(out) != tc.in {
+				t.Fatalf("expected unchanged payload, got %s", s)
+			}
+		})
+	}
+}
+
+func TestReasoningToolChoiceCompatRoundTripper(t *testing.T) {
+	var gotBody string
+	rt := &reasoningToolChoiceCompatRoundTripper{
+		base: roundTripperFunc(func(req *http.Request) (*http.Response, error) {
+			b, _ := io.ReadAll(req.Body)
+			gotBody = string(b)
+			return &http.Response{
+				StatusCode: 200,
+				Body:       io.NopCloser(strings.NewReader(`{"choices":[{"message":{"content":"ok"}}]}`)),
+				Header:     http.Header{"Content-Type": []string{"application/json"}},
+			}, nil
+		}),
+	}
+	req, err := http.NewRequest(http.MethodPost, "https://example.com/v1/chat/completions", strings.NewReader(
+		`{"model":"m","thinking":{"type":"enabled"},"tool_choice":"required","messages":[]}`,
+	))
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = rt.RoundTrip(req)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(gotBody, "thinking") {
+		t.Fatalf("expected thinking stripped in transit, got %s", gotBody)
+	}
+	if !strings.Contains(gotBody, `"tool_choice":"required"`) {
+		t.Fatalf("expected tool_choice preserved, got %s", gotBody)
+	}
+}
+
+type roundTripperFunc func(*http.Request) (*http.Response, error)
+
+func (f roundTripperFunc) RoundTrip(req *http.Request) (*http.Response, error) {
+	return f(req)
+}

internal/openai/reasoning_tool_choice_compat.go

@@ -0,0 +1,43 @@
+package openai
+
+import (
+	"bytes"
+	"io"
+	"net/http"
+	"strconv"
+	"strings"
+)
+
+// reasoningToolChoiceCompatRoundTripper strips thinking/reasoning fields from
+// chat/completions requests that force tool_choice, which some gateways reject
+// when thinking mode is enabled on the same request.
+type reasoningToolChoiceCompatRoundTripper struct {
+	base http.RoundTripper
+}
+
+func (rt *reasoningToolChoiceCompatRoundTripper) RoundTrip(req *http.Request) (*http.Response, error) {
+	if rt == nil || rt.base == nil || req == nil || req.Body == nil {
+		if rt != nil && rt.base != nil {
+			return rt.base.RoundTrip(req)
+		}
+		return http.DefaultTransport.RoundTrip(req)
+	}
+	if req.Method != http.MethodPost || !strings.HasSuffix(req.URL.Path, "/chat/completions") {
+		return rt.base.RoundTrip(req)
+	}
+
+	body, err := io.ReadAll(req.Body)
+	_ = req.Body.Close()
+	if err != nil {
+		return nil, err
+	}
+
+	patched, perr := StripReasoningIfForcedToolChoice(body)
+	if perr != nil {
+		patched = body
+	}
+	req.Body = io.NopCloser(bytes.NewReader(patched))
+	req.ContentLength = int64(len(patched))
+	req.Header.Set("Content-Length", strconv.Itoa(len(patched)))
+	return rt.base.RoundTrip(req)
+}

internal/reasoning/eino.go

@@ -26,6 +26,35 @@ const (
 	wireOutputConfig
 )
 
+// ApplyPlanExecutePlannerModelConfig configures the plan_execute planner/replanner
+// ChatModel. Those Eino agents call WithToolChoice(Forced); several gateways reject
+// thinking / reasoning fields on the same request (tool_choice required/object).
+// Executor should keep the normal ApplyToEinoChatModelConfig path.
+func ApplyPlanExecutePlannerModelConfig(cfg *einoopenai.ChatModelConfig, oa *config.OpenAIConfig) {
+	if cfg == nil || oa == nil {
+		return
+	}
+	offOA := *oa
+	offReasoning := oa.Reasoning
+	offReasoning.Mode = "off"
+	offOA.Reasoning = offReasoning
+	ApplyToEinoChatModelConfig(cfg, &offOA, nil)
+	clearReasoningFromChatModelConfig(cfg)
+}
+
+func clearReasoningFromChatModelConfig(cfg *einoopenai.ChatModelConfig) {
+	if cfg == nil {
+		return
+	}
+	cfg.ReasoningEffort = ""
+	if cfg.ExtraFields != nil {
+		for _, key := range []string{"thinking", "reasoning_effort", "output_config", "reasoning"} {
+			delete(cfg.ExtraFields, key)
+		}
+	}
+	applyThinkingDisabled(cfg)
+}
+
 // ApplyToEinoChatModelConfig merges reasoning-related options into cfg.
 // Precondition: cfg already has APIKey, BaseURL, Model, HTTPClient set.
 func ApplyToEinoChatModelConfig(cfg *einoopenai.ChatModelConfig, oa *config.OpenAIConfig, client *ClientIntent) {

internal/reasoning/eino_test.go

@@ -49,6 +49,30 @@ func TestApplyOpenAICompat_xhighExtraField(t *testing.T) {
 	}
 }
 
+func TestApplyPlanExecutePlannerModelConfig_stripsReasoningWhenGlobalOn(t *testing.T) {
+	cfg := &einoopenai.ChatModelConfig{}
+	oa := &config.OpenAIConfig{
+		BaseURL: "https://antchat.example.com/v1",
+		Model:   "minimax-m3",
+		Reasoning: config.OpenAIReasoningConfig{
+			Profile: "openai_compat",
+			Mode:    "on",
+			Effort:  "high",
+		},
+	}
+	ApplyPlanExecutePlannerModelConfig(cfg, oa)
+	if cfg.ReasoningEffort != "" {
+		t.Fatalf("expected ReasoningEffort cleared, got %q", cfg.ReasoningEffort)
+	}
+	th, ok := cfg.ExtraFields["thinking"].(map[string]any)
+	if !ok || th["type"] != "disabled" {
+		t.Fatalf("expected thinking disabled, got %#v", cfg.ExtraFields)
+	}
+	if _, ok := cfg.ExtraFields["reasoning_effort"]; ok {
+		t.Fatalf("expected reasoning_effort stripped, got %#v", cfg.ExtraFields)
+	}
+}
+
 func TestApplyReasoningOff_disablesThinking(t *testing.T) {
 	cfg := &einoopenai.ChatModelConfig{}
 	oa := &config.OpenAIConfig{

internal/security/executor.go

@@ -162,9 +162,8 @@ func (e *Executor) ExecuteTool(ctx context.Context, toolName string, args map[st
 			output, err = runCommandWithPTY(ctx, cmd2, cb)
 		}
 	} else {
-		outputBytes, err2 := cmd.CombinedOutput()
+		// 非流式：内存缓冲 + ctx 取消杀进程组；行为对齐原 CombinedOutput，避免双流管道 fan-in 死锁。
-		output = string(outputBytes)
+		output, err = combinedOutputCancellable(ctx, cmd)
-		err = err2
 		if err != nil && shouldRetryWithPTY(output) {
 			e.logger.Info("检测到工具需要 TTY，使用 PTY 重试",
 				zap.String("tool", toolName),
@@ -692,83 +691,21 @@ func (e *Executor) formatParamValue(param config.ParameterConfig, value interfac
 // IsBackgroundShellCommand 检测命令是否为完全后台命令（末尾有独立 &，且不在引号内）。
 // command1 & command2 不算完全后台（command2 仍在前台执行）。
 func IsBackgroundShellCommand(command string) bool {
-	// 移除首尾空格
 	command = strings.TrimSpace(command)
 	if command == "" {
 		return false
 	}
-
+	positions := findStandaloneAmpersandPositions(command)
-	// 检查命令中所有不在引号内的 & 符号
+	if len(positions) == 0 {
-	// 找到最后一个 & 符号，检查它是否在命令末尾
-	inSingleQuote := false
-	inDoubleQuote := false
-	escaped := false
-	lastAmpersandPos := -1
-
-	for i, r := range command {
-		if escaped {
-			escaped = false
-			continue
-		}
-		if r == '\\' {
-			escaped = true
-			continue
-		}
-		if r == '\'' && !inDoubleQuote {
-			inSingleQuote = !inSingleQuote
-			continue
-		}
-		if r == '"' && !inSingleQuote {
-			inDoubleQuote = !inDoubleQuote
-			continue
-		}
-		if r == '&' && !inSingleQuote && !inDoubleQuote {
-			// 检查 & 前后是否有空格或换行（确保是独立的 &，而不是变量名的一部分）
-			isStandalone := false
-
-			// 检查前面：空格、制表符、换行符，或者是命令开头
-			if i == 0 {
-				isStandalone = true
-			} else {
-				prev := command[i-1]
-				if prev == ' ' || prev == '\t' || prev == '\n' || prev == '\r' {
-					isStandalone = true
-				}
-			}
-
-			// 检查后面：空格、制表符、换行符，或者是命令末尾
-			if isStandalone {
-				if i == len(command)-1 {
-					// 在末尾，肯定是独立的 &
-					lastAmpersandPos = i
-				} else {
-					next := command[i+1]
-					if next == ' ' || next == '\t' || next == '\n' || next == '\r' {
-						// 后面有空格，是独立的 &
-						lastAmpersandPos = i
-					}
-				}
-			}
-		}
-	}
-
-	// 如果没有找到 & 符号，不是后台命令
-	if lastAmpersandPos == -1 {
 		return false
 	}
-
+	last := positions[len(positions)-1]
-	// 检查最后一个 & 后面是否还有非空内容
+	afterAmpersand := strings.TrimSpace(command[last+1:])
-	afterAmpersand := strings.TrimSpace(command[lastAmpersandPos+1:])
+	if afterAmpersand != "" {
-	if afterAmpersand == "" {
+		return false
-		// & 在末尾或后面只有空白字符，这是完全后台命令
-		// 检查 & 前面是否有内容
-		beforeAmpersand := strings.TrimSpace(command[:lastAmpersandPos])
-		return beforeAmpersand != ""
 	}
-
+	beforeAmpersand := strings.TrimSpace(command[:last])
-	// 如果 & 后面还有非空内容，说明是 command1 & command2 的情况
+	return beforeAmpersand != ""
-	// 这种情况下，command2会在前台执行，所以不算完全后台命令
-	return false
 }
 
 // executeSystemCommand 执行系统命令
@@ -804,7 +741,7 @@ func (e *Executor) executeSystemCommand(ctx context.Context, args map[string]int
 		zap.String("command", command),
 	)
 
-	command = PrepareNonInteractiveShellCommand(command)
+	command = PrepareShellCommandForExecute(command)
 
 	// 获取shell类型（可选，默认为sh）
 	shell := "sh"
@@ -845,10 +782,8 @@ func (e *Executor) executeSystemCommand(ctx context.Context, args map[string]int
 		commandWithoutAmpersand := strings.TrimSuffix(strings.TrimSpace(command), "&")
 		commandWithoutAmpersand = strings.TrimSpace(commandWithoutAmpersand)
 
-		// 构建新命令：将用户命令置于独立重定向的后台作业，再 echo $pid。
+		// 构建新命令：后台作业重定向标准流后 echo $pid（与 RedirectBackgroundJobStdio 一致）。
-		// 若子进程与 echo 共享同一 stdout 管道，且长时间不向 stdout 写入换行，
+		pidCommand := RedirectBackgroundJobStdio(commandWithoutAmpersand+" &") + " pid=$!; echo $pid"
-		// bufio.ReadString('\n') 会永久阻塞（例如 beacon 持续写二进制/单行日志）。
-		pidCommand := fmt.Sprintf("%s </dev/null >/dev/null 2>&1 & pid=$!; echo $pid", commandWithoutAmpersand)
 
 		// 创建新命令来获取PID
 		var pidCmd *exec.Cmd
@@ -981,9 +916,7 @@ func (e *Executor) executeSystemCommand(ctx context.Context, args map[string]int
 			output, err = runCommandWithPTY(ctx, cmd2, cb)
 		}
 	} else {
-		outputBytes, err2 := cmd.CombinedOutput()
+		output, err = combinedOutputCancellable(ctx, cmd)
-		output = string(outputBytes)
-		err = err2
 		if err != nil && shouldRetryWithPTY(output) {
 			e.logger.Info("检测到系统命令需要 TTY，使用 PTY 重试")
 			cmd2 := exec.CommandContext(ctx, shell, "-c", command)
@@ -1027,12 +960,58 @@ func (e *Executor) executeSystemCommand(ctx context.Context, args map[string]int
 	}, nil
 }
 
+// combinedOutputCancellable 行为对齐 cmd.CombinedOutput（stdout/stderr 写入内存缓冲），
+// 但在 ctx 取消时 terminateCmdTree 终止整棵进程树。
+// 非流式路径不使用双流管道 fan-in，避免 stderr 撑满管道缓冲区时与 stdout 互相阻塞导致死锁。
+// 无输出空闲检测由上层 agent.tool_timeout_minutes 兜底，不改变原 CombinedOutput 语义。
+func combinedOutputCancellable(ctx context.Context, cmd *exec.Cmd) (string, error) {
+	var stdoutBuf, stderrBuf strings.Builder
+	cmd.Stdout = &stdoutBuf
+	cmd.Stderr = &stderrBuf
+
+	session, err := StartShellSession(cmd)
+	if err != nil {
+		return "", err
+	}
+
+	done := make(chan error, 1)
+	go func() {
+		done <- session.Wait()
+	}()
+
+	stopWatch := make(chan struct{})
+	go func() {
+		select {
+		case <-ctx.Done():
+			TerminateShellCmdSession(session)
+		case <-stopWatch:
+		}
+	}()
+	defer close(stopWatch)
+
+	var waitErr error
+	select {
+	case waitErr = <-done:
+	case <-ctx.Done():
+		waitErr = <-done
+		return joinCommandOutput(stdoutBuf.String(), stderrBuf.String()), ctx.Err()
+	}
+	return joinCommandOutput(stdoutBuf.String(), stderrBuf.String()), waitErr
+}
+
+func joinCommandOutput(stdout, stderr string) string {
+	if stderr == "" {
+		return stdout
+	}
+	if stdout == "" {
+		return stderr
+	}
+	return stdout + stderr
+}
+
 // streamCommandOutput 以“边读边回调”的方式读取命令 stdout/stderr。
 // 使用定长块读取，避免按行读取在无换行输出时永久阻塞；ctx 取消时终止进程树。
 func streamCommandOutput(ctx context.Context, cmd *exec.Cmd, cb ToolOutputCallback, noOutputSec int) (string, error) {
-	if err := prepareShellCmdSession(cmd); err != nil {
-		return "", err
-	}
 	stdoutPipe, err := cmd.StdoutPipe()
 	if err != nil {
 		return "", err
@@ -1042,7 +1021,8 @@ func streamCommandOutput(ctx context.Context, cmd *exec.Cmd, cb ToolOutputCallba
 		_ = stdoutPipe.Close()
 		return "", err
 	}
-	if err := cmd.Start(); err != nil {
+	session, err := StartShellSession(cmd)
+	if err != nil {
 		_ = stdoutPipe.Close()
 		_ = stderrPipe.Close()
 		return "", err
@@ -1052,7 +1032,7 @@ func streamCommandOutput(ctx context.Context, cmd *exec.Cmd, cb ToolOutputCallba
 	go func() {
 		select {
 		case <-ctx.Done():
-			terminateCmdTree(cmd)
+			TerminateShellCmdSession(session)
 		case <-stopWatch:
 		}
 	}()
@@ -1091,7 +1071,9 @@ func streamCommandOutput(ctx context.Context, cmd *exec.Cmd, cb ToolOutputCallba
 		if deltaBuilder.Len() == 0 {
 			return
 		}
-		cb(deltaBuilder.String())
+		if cb != nil {
+			cb(deltaBuilder.String())
+		}
 		deltaBuilder.Reset()
 		lastFlush = time.Now()
 	}
@@ -1102,13 +1084,13 @@ func streamCommandOutput(ctx context.Context, cmd *exec.Cmd, cb ToolOutputCallba
 	}
 
 	fireInactivity := func() {
-		terminateCmdTree(cmd)
+		TerminateShellCmdSession(session)
 		msg := ShellNoOutputTimeoutMessage(idleWatch.Sec)
 		outBuilder.WriteString(msg)
 		if cb != nil {
 			cb(msg)
 		}
-		_ = cmd.Wait()
+		_ = session.Wait()
 	}
 
 chunksLoop:
@@ -1118,6 +1100,11 @@ chunksLoop:
 			idleCh = idleWatch.Expired
 		}
 		select {
+		case <-ctx.Done():
+			TerminateShellCmdSession(session)
+			flush()
+			_ = session.Wait()
+			return outBuilder.String(), ctx.Err()
 		case <-idleCh:
 			fireInactivity()
 			return outBuilder.String(), fmt.Errorf("shell inactivity timeout (%ds)", idleWatch.Sec)
@@ -1138,7 +1125,7 @@ chunksLoop:
 	flush()
 
 	// 等待命令结束，返回最终退出状态
-	waitErr := cmd.Wait()
+	waitErr := session.Wait()
 	return outBuilder.String(), waitErr
 }
 
@@ -1210,13 +1197,18 @@ func runCommandWithPTY(ctx context.Context, cmd *exec.Cmd, cb ToolOutputCallback
 	}
 	defer func() { _ = ptmx.Close() }()
 
+	rootPID := 0
+	if cmd.Process != nil {
+		rootPID = cmd.Process.Pid
+	}
+
 	// ctx 取消时尽快终止子进程
 	done := make(chan struct{})
 	go func() {
 		select {
 		case <-ctx.Done():
 			_ = ptmx.Close() // 触发读退出
-			terminateCmdTree(cmd)
+			terminateProcessGroup(rootPID, cmd)
 		case <-done:
 		}
 	}()

internal/security/executor_test.go

@@ -2,6 +2,8 @@ package security
 
 import (
 	"context"
+	"os/exec"
+	"runtime"
 	"strings"
 	"testing"
 	"time"
@@ -147,3 +149,33 @@ func indexOf(slice []string, s string) int {
 	}
 	return -1
 }
+
+// TestCombinedOutputCancellable_ContextCancelKillsTree 验证 ctx 取消时能在数秒内结束（杀进程组，非挂死）。
+func TestCombinedOutputCancellable_ContextCancelKillsTree(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("unix process group kill")
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	cmd := exec.CommandContext(ctx, "sh", "-c", "sleep 300")
+	ConfigureShellCmdForAgentExecute(cmd)
+
+	done := make(chan error, 1)
+	go func() {
+		_, err := combinedOutputCancellable(ctx, cmd)
+		done <- err
+	}()
+
+	time.Sleep(150 * time.Millisecond)
+	cancel()
+
+	select {
+	case err := <-done:
+		if err == nil {
+			t.Fatal("expected context cancel error")
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("combinedOutputCancellable did not return within 5s after context cancel")
+	}
+}

internal/security/procattr_unix.go

@@ -19,13 +19,23 @@ func prepareShellCmdSession(cmd *exec.Cmd) error {
 	return nil
 }
 
-// terminateCmdTree 尽力终止 cmd 及其进程组（Unix 下 Setsid 后 PGID == 首进程 PID）。
+// terminateProcessGroup 对 rootPID 对应进程组发 SIGKILL；rootPID 为 0 时回退到 cmd.Process.Pid。
-func terminateCmdTree(cmd *exec.Cmd) {
+func terminateProcessGroup(rootPID int, cmd *exec.Cmd) {
-	if cmd == nil || cmd.Process == nil {
+	pid := rootPID
+	if pid <= 0 && cmd != nil && cmd.Process != nil {
+		pid = cmd.Process.Pid
+	}
+	if pid <= 0 {
 		return
 	}
-	pid := cmd.Process.Pid
 	if err := syscall.Kill(-pid, syscall.SIGKILL); err != nil {
-		_ = cmd.Process.Kill()
+		if cmd != nil && cmd.Process != nil {
+			_ = cmd.Process.Kill()
+		}
 	}
 }
+
+// terminateCmdTree 尽力终止 cmd 及其进程组（Unix 下 Setsid 后 PGID == 首进程 PID）。
+func terminateCmdTree(cmd *exec.Cmd) {
+	terminateProcessGroup(0, cmd)
+}

internal/security/procattr_windows.go

@@ -2,16 +2,42 @@
 
 package security
 
-import "os/exec"
+import (
+	"os/exec"
+	"strconv"
+	"syscall"
+)
 
 func prepareShellCmdSession(cmd *exec.Cmd) error {
-	_ = cmd
+	if cmd == nil {
+		return nil
+	}
+	// 独立进程组，便于 taskkill /T 终止整棵子进程树。
+	if cmd.SysProcAttr == nil {
+		cmd.SysProcAttr = &syscall.SysProcAttr{}
+	}
+	cmd.SysProcAttr.CreationFlags = syscall.CREATE_NEW_PROCESS_GROUP
 	return nil
 }
 
-func terminateCmdTree(cmd *exec.Cmd) {
+// terminateProcessGroup 使用 taskkill /F /T 终止进程及其子进程；rootPID 为 0 时回退到 cmd.Process.Pid。
-	if cmd == nil || cmd.Process == nil {
+func terminateProcessGroup(rootPID int, cmd *exec.Cmd) {
+	pid := rootPID
+	if pid <= 0 && cmd != nil && cmd.Process != nil {
+		pid = cmd.Process.Pid
+	}
+	if pid <= 0 {
 		return
 	}
-	_ = cmd.Process.Kill()
+	tk := exec.Command("taskkill", "/F", "/T", "/PID", strconv.Itoa(pid))
+	if err := tk.Run(); err != nil {
+		if cmd != nil && cmd.Process != nil {
+			_ = cmd.Process.Kill()
+		}
+	}
+}
+
+// terminateCmdTree 使用 taskkill /F /T 终止进程及其子进程（Windows 上 Process.Kill 无法保证杀掉 python 等孙进程）。
+func terminateCmdTree(cmd *exec.Cmd) {
+	terminateProcessGroup(0, cmd)
 }

internal/security/shell_background_io.go

@@ -0,0 +1,111 @@
+package security
+
+import "strings"
+
+const backgroundJobStdioRedirect = " </dev/null >/dev/null 2>&1"
+
+// findStandaloneAmpersandPositions 返回不在引号内的独立 & 下标（排除 &&）。
+func findStandaloneAmpersandPositions(command string) []int {
+	command = strings.TrimSpace(command)
+	if command == "" {
+		return nil
+	}
+
+	var positions []int
+	inSingleQuote := false
+	inDoubleQuote := false
+	escaped := false
+
+	for i := 0; i < len(command); i++ {
+		r := command[i]
+		if escaped {
+			escaped = false
+			continue
+		}
+		if r == '\\' {
+			escaped = true
+			continue
+		}
+		if r == '\'' && !inDoubleQuote {
+			inSingleQuote = !inSingleQuote
+			continue
+		}
+		if r == '"' && !inSingleQuote {
+			inDoubleQuote = !inDoubleQuote
+			continue
+		}
+		if r != '&' || inSingleQuote || inDoubleQuote {
+			continue
+		}
+		if i+1 < len(command) && command[i+1] == '&' {
+			continue
+		}
+		if i > 0 && command[i-1] == '&' {
+			continue
+		}
+
+		isStandalone := i == 0
+		if !isStandalone {
+			prev := command[i-1]
+			isStandalone = prev == ' ' || prev == '\t' || prev == '\n' || prev == '\r'
+		}
+		if !isStandalone {
+			continue
+		}
+		if i == len(command)-1 {
+			positions = append(positions, i)
+			continue
+		}
+		next := command[i+1]
+		if next == ' ' || next == '\t' || next == '\n' || next == '\r' {
+			positions = append(positions, i)
+		}
+	}
+	return positions
+}
+
+func segmentHasStdioRedirect(segment string) bool {
+	lower := strings.ToLower(strings.TrimSpace(segment))
+	if lower == "" {
+		return false
+	}
+	if strings.Contains(lower, ">/dev/null") || strings.Contains(lower, "2>/dev/null") {
+		return true
+	}
+	if strings.Contains(lower, "&>") || strings.Contains(lower, "&>>") {
+		return true
+	}
+	if strings.Contains(lower, "2>&1") && strings.Contains(lower, "/dev/null") {
+		return true
+	}
+	return false
+}
+
+// RedirectBackgroundJobStdio 为每个独立 & 前的后台段注入 </dev/null >/dev/null 2>&1，
+// 避免后台子进程占用 execute/exec 管道导致挂死。
+func RedirectBackgroundJobStdio(command string) string {
+	positions := findStandaloneAmpersandPositions(command)
+	if len(positions) == 0 {
+		return command
+	}
+
+	out := command
+	for j := len(positions) - 1; j >= 0; j-- {
+		i := positions[j]
+		before := out[:i]
+		after := out[i:]
+		trimmed := strings.TrimRight(before, " \t\r\n")
+		if segmentHasStdioRedirect(trimmed) {
+			continue
+		}
+		trailing := before[len(trimmed):]
+		out = trimmed + backgroundJobStdioRedirect + trailing + after
+	}
+	return out
+}
+
+// PrepareShellCommandForExecute 组合 execute/exec 用的非交互包装与后台 IO 重定向。
+// 须先注入 exec </dev/null，再改写 & 后台段，否则段内 </dev/null 会使 stdin 重定向被误判为已存在。
+func PrepareShellCommandForExecute(shellCommand string) string {
+	return RedirectBackgroundJobStdio(PrepareNonInteractiveShellCommand(shellCommand))
+}

internal/security/shell_background_io_test.go

@@ -0,0 +1,64 @@
+package security
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestRedirectBackgroundJobStdio_mixedCommand(t *testing.T) {
+	in := "java -jar app.jar & JRMP_PID=$!; echo started"
+	out := RedirectBackgroundJobStdio(in)
+	if !strings.Contains(out, "java -jar app.jar </dev/null >/dev/null 2>&1 &") {
+		t.Fatalf("expected redirect before &: %q", out)
+	}
+	if !strings.Contains(out, "echo started") {
+		t.Fatalf("foreground tail preserved: %q", out)
+	}
+}
+
+func TestRedirectBackgroundJobStdio_trailingOnly(t *testing.T) {
+	in := "sleep 120 &"
+	out := RedirectBackgroundJobStdio(in)
+	want := "sleep 120 </dev/null >/dev/null 2>&1 &"
+	if strings.TrimSpace(out) != want {
+		t.Fatalf("got %q want %q", out, want)
+	}
+}
+
+func TestRedirectBackgroundJobStdio_skipsAlreadyRedirected(t *testing.T) {
+	in := "sleep 1 >/dev/null 2>&1 & echo ok"
+	out := RedirectBackgroundJobStdio(in)
+	if out != in {
+		t.Fatalf("should not double-redirect: %q", out)
+	}
+}
+
+func TestRedirectBackgroundJobStdio_skipsAndAnd(t *testing.T) {
+	in := "test -f /etc/passwd && echo ok"
+	out := RedirectBackgroundJobStdio(in)
+	if out != in {
+		t.Fatalf("&& must not be treated as background &: %q", out)
+	}
+}
+
+func TestPrepareShellCommandForExecute(t *testing.T) {
+	out := PrepareShellCommandForExecute("java -jar x & echo hi")
+	if !strings.Contains(out, "exec </dev/null") {
+		t.Fatalf("missing stdin redirect: %q", out)
+	}
+	if !strings.Contains(out, "GIT_PAGER=cat") {
+		t.Fatalf("missing pager export: %q", out)
+	}
+	if !strings.Contains(out, "java -jar x </dev/null >/dev/null 2>&1 &") {
+		t.Fatalf("missing background redirect: %q", out)
+	}
+}
+
+func TestIsBackgroundShellCommand_usesSharedParser(t *testing.T) {
+	if !IsBackgroundShellCommand("sleep 1 &") {
+		t.Fatal("trailing & should be background")
+	}
+	if IsBackgroundShellCommand("sleep 1 & echo hi") {
+		t.Fatal("mixed should not be fully background")
+	}
+}

internal/security/shell_execute_stream.go

@@ -27,6 +27,11 @@ func TerminateShellCmdTree(cmd *exec.Cmd) {
 	terminateCmdTree(cmd)
 }
 
+// TerminateShellCmdSession 使用 Start 时缓存的进程组 ID 终止（shell 已退出时仍有效）。
+func TerminateShellCmdSession(session *ShellSession) {
+	TerminateShellSession(session)
+}
+
 // EinoStreamingShell 为 Eino ADK execute 工具提供流式 shell，行为与 exec 对齐：
 // 并发读取 stdout/stderr（定长块，非按行），避免官方 local.ExecuteStreaming 先排空 stdout
 // 导致 stderr 错误（如 sudo 密码提示）长时间不可见、UI 一直显示「执行中」。
@@ -55,8 +60,10 @@ func (s *EinoStreamingShell) ExecuteStreaming(ctx context.Context, input *filesy
 func runShellInBackground(ctx context.Context, command string, w *schema.StreamWriter[*filesystem.ExecuteResponse]) {
 	defer w.Close()
 
+	command = PrepareShellCommandForExecute(command)
 	cmd := exec.CommandContext(ctx, "/bin/sh", "-c", command)
-	ConfigureShellCmdForAgentExecute(cmd)
+	applyDefaultTerminalEnv(cmd)
+	attachNonInteractiveStdin(cmd)
 	stdout, err := cmd.StdoutPipe()
 	if err != nil {
 		_ = w.Send(nil, fmt.Errorf("failed to create stdout pipe: %w", err))
@@ -68,7 +75,8 @@ func runShellInBackground(ctx context.Context, command string, w *schema.StreamW
 		_ = w.Send(nil, fmt.Errorf("failed to create stderr pipe: %w", err))
 		return
 	}
-	if err := cmd.Start(); err != nil {
+	session, err := StartShellSession(cmd)
+	if err != nil {
 		_ = stdout.Close()
 		_ = stderr.Close()
 		_ = w.Send(nil, fmt.Errorf("failed to start command: %w", err))
@@ -78,14 +86,14 @@ func runShellInBackground(ctx context.Context, command string, w *schema.StreamW
 	done := make(chan struct{})
 	go func() {
 		drainShellPipes(stdout, stderr)
-		_ = cmd.Wait()
+		_ = session.Wait()
 		close(done)
 	}()
 
 	select {
 	case <-done:
 	case <-ctx.Done():
-		TerminateShellCmdTree(cmd)
+		TerminateShellCmdSession(session)
 	}
 
 	exitCode := 0
@@ -112,8 +120,10 @@ func drainShellPipes(stdout, stderr io.Reader) {
 func streamShellForeground(ctx context.Context, command string, w *schema.StreamWriter[*filesystem.ExecuteResponse]) {
 	defer w.Close()
 
+	command = PrepareShellCommandForExecute(command)
 	cmd := exec.CommandContext(ctx, "/bin/sh", "-c", command)
-	ConfigureShellCmdForAgentExecute(cmd)
+	applyDefaultTerminalEnv(cmd)
+	attachNonInteractiveStdin(cmd)
 
 	stdoutPipe, err := cmd.StdoutPipe()
 	if err != nil {
@@ -126,7 +136,8 @@ func streamShellForeground(ctx context.Context, command string, w *schema.Stream
 		_ = w.Send(nil, fmt.Errorf("failed to create stderr pipe: %w", err))
 		return
 	}
-	if err := cmd.Start(); err != nil {
+	session, err := StartShellSession(cmd)
+	if err != nil {
 		_ = stdoutPipe.Close()
 		_ = stderrPipe.Close()
 		_ = w.Send(nil, fmt.Errorf("failed to start command: %w", err))
@@ -137,7 +148,7 @@ func streamShellForeground(ctx context.Context, command string, w *schema.Stream
 	go func() {
 		select {
 		case <-ctx.Done():
-			TerminateShellCmdTree(cmd)
+			TerminateShellCmdSession(session)
 		case <-stopWatch:
 		}
 	}()
@@ -174,12 +185,12 @@ func streamShellForeground(ctx context.Context, command string, w *schema.Stream
 		}
 		hadOutput = true
 		if w.Send(&filesystem.ExecuteResponse{Output: chunk}, nil) {
-			TerminateShellCmdTree(cmd)
+			TerminateShellCmdSession(session)
 			return
 		}
 	}
 
-	waitErr := cmd.Wait()
+	waitErr := session.Wait()
 	if waitErr == nil {
 		exitCode := 0
 		_ = w.Send(&filesystem.ExecuteResponse{ExitCode: &exitCode}, nil)

internal/security/shell_execute_stream_test.go

@@ -115,3 +115,38 @@ func TestEinoStreamingShell_StderrWhileStdoutBlocks(t *testing.T) {
 		t.Fatalf("expected early stderr, got: %q", got.String())
 	}
 }
+
+// TestEinoStreamingShell_BackgroundJobDoesNotHoldPipe 模拟 cmd & 后继续前台逻辑：重定向后应快速结束。
+func TestEinoStreamingShell_BackgroundJobDoesNotHoldPipe(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping shell integration in -short")
+	}
+	shell := NewEinoStreamingShell()
+	cmd := `(sh -c 'printf x; sleep 120') & echo started; sleep 0`
+	sr, err := shell.ExecuteStreaming(context.Background(), &filesystem.ExecuteRequest{Command: cmd})
+	if err != nil {
+		t.Fatalf("ExecuteStreaming: %v", err)
+	}
+	defer sr.Close()
+
+	start := time.Now()
+	var got strings.Builder
+	for {
+		resp, rerr := sr.Recv()
+		if errors.Is(rerr, io.EOF) {
+			break
+		}
+		if rerr != nil {
+			t.Fatalf("recv: %v", rerr)
+		}
+		if resp != nil && resp.Output != "" {
+			got.WriteString(resp.Output)
+		}
+	}
+	if time.Since(start) > 3*time.Second {
+		t.Fatalf("expected fast completion, took %v output=%q", time.Since(start), got.String())
+	}
+	if !strings.Contains(got.String(), "started") {
+		t.Fatalf("expected foreground echo, got: %q", got.String())
+	}
+}

internal/security/shell_session.go

@@ -0,0 +1,47 @@
+package security
+
+import "os/exec"
+
+// ShellSession 在 Start 时记录根 shell 的进程组 ID，取消/超时时可杀整组（即使 cmd.Process 已失效）。
+type ShellSession struct {
+	Cmd     *exec.Cmd
+	rootPID int
+}
+
+// StartShellSession 配置独立进程组并启动 shell，缓存 rootPID（Unix 下即 PGID）。
+func StartShellSession(cmd *exec.Cmd) (*ShellSession, error) {
+	if err := prepareShellCmdSession(cmd); err != nil {
+		return nil, err
+	}
+	if err := cmd.Start(); err != nil {
+		return nil, err
+	}
+	pid := 0
+	if cmd.Process != nil {
+		pid = cmd.Process.Pid
+	}
+	return &ShellSession{Cmd: cmd, rootPID: pid}, nil
+}
+
+// Wait 等待 shell 退出。
+func (s *ShellSession) Wait() error {
+	if s == nil || s.Cmd == nil {
+		return nil
+	}
+	return s.Cmd.Wait()
+}
+
+// Terminate 终止 shell 及其进程组。
+func (s *ShellSession) Terminate() {
+	if s == nil {
+		return
+	}
+	terminateProcessGroup(s.rootPID, s.Cmd)
+}
+
+// TerminateShellSession 终止由 StartShellSession 启动的会话。
+func TerminateShellSession(session *ShellSession) {
+	if session != nil {
+		session.Terminate()
+	}
+}

internal/security/shell_session_test.go

@@ -0,0 +1,65 @@
+package security
+
+import (
+	"context"
+	"os/exec"
+	"runtime"
+	"testing"
+	"time"
+)
+
+func TestShellSession_TerminateUsesCachedRootPID(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("unix process group kill")
+	}
+
+	cmd := exec.Command("sh", "-c", "sleep 300")
+	ConfigureShellCmdForAgentExecute(cmd)
+
+	session, err := StartShellSession(cmd)
+	if err != nil {
+		t.Fatalf("StartShellSession: %v", err)
+	}
+
+	time.Sleep(100 * time.Millisecond)
+	session.Terminate()
+
+	done := make(chan error, 1)
+	go func() { done <- session.Wait() }()
+
+	select {
+	case <-done:
+	case <-time.After(5 * time.Second):
+		t.Fatal("session did not finish within 5s after Terminate")
+	}
+}
+
+func TestShellSession_TerminateAfterContextCancel(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("unix process group kill")
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	cmd := exec.CommandContext(ctx, "sh", "-c", "sleep 300")
+	ConfigureShellCmdForAgentExecute(cmd)
+
+	session, err := StartShellSession(cmd)
+	if err != nil {
+		t.Fatalf("StartShellSession: %v", err)
+	}
+
+	time.Sleep(100 * time.Millisecond)
+	cancel()
+	TerminateShellCmdSession(session)
+
+	done := make(chan error, 1)
+	go func() { done <- session.Wait() }()
+
+	select {
+	case <-done:
+	case <-time.After(5 * time.Second):
+		t.Fatal("session did not finish within 5s after cancel+terminate")
+	}
+}

tools/hydra.yaml

@@ -1,60 +1,95 @@
 name: "hydra"
 command: "hydra"
+args: ["-I"]
 enabled: true
 short_description: "密码暴力破解工具，支持多种协议和服务"
 description: |
-  Hydra是一个快速的网络登录破解工具，支持多种协议和服务的密码暴力破解。
+  Hydra 是网络登录口令爆破工具，支持 SSH、FTP、HTTP、SMB 等多种协议。
 
-  **主要功能：**
+  **调用约定（必读）：**
-  - 支持多种协议（SSH, FTP, HTTP, SMB等）
+  - 必须提供 **用户名**：`username`（-l）或 `username_file`（-L）至少其一
-  - 快速并行破解
+  - 必须提供 **口令**：`password`（-p）、`password_file`（-P）或 `-C`（经 `additional_args`）至少其一
-  - 支持用户名和密码字典
+  - **先用小字典试跑**（几十～几百条），确认目标可达再扩大；禁止默认使用 rockyou 等超大字典
-  - 可恢复的会话
+  - 默认已启用：找到即停（-f）、并行 4（-t）、忽略 restore（-I）；长任务请设 `output_file`
 
-  **使用场景：**
+  **CLI 顺序：** `hydra [选项] <target> <service>`（本工具已按此顺序组参，勿把 target 写在选项前）
-  - 密码强度测试
+
-  - 渗透测试
+  **使用场景：** 授权环境下的弱口令检测、密码强度评估
-  - 安全评估
+
-  - 弱密码检测
+  **注意：** 仅用于已授权目标；对无响应目标请减小 `wait_time` 或缩小字典，避免长时间挂起。
 parameters:
-  - name: "target"
-    type: "string"
-    description: "目标IP或主机名"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "service"
-    type: "string"
-    description: "服务类型（ssh, ftp, http等）"
-    required: true
-    position: 1
-    format: "positional"
   - name: "username"
     type: "string"
-    description: "单个用户名"
+    description: "单个用户名（-l）；与 username_file 二选一至少填一个"
     required: false
     flag: "-l"
     format: "flag"
   - name: "username_file"
     type: "string"
-    description: "用户名字典文件"
+    description: "用户名字典文件（-L）"
     required: false
     flag: "-L"
     format: "flag"
   - name: "password"
     type: "string"
-    description: "单个密码"
+    description: "单个密码（-p）"
     required: false
     flag: "-p"
     format: "flag"
   - name: "password_file"
     type: "string"
-    description: "密码字典文件"
+    description: "密码字典文件（-P）；优先使用小字典试跑"
     required: false
     flag: "-P"
     format: "flag"
+  - name: "stop_on_first"
+    type: "bool"
+    description: "找到一对有效账密后立即退出（-f，默认 true）"
+    required: false
+    flag: "-f"
+    format: "flag"
+    default: true
+  - name: "tasks"
+    type: "int"
+    description: "每目标并行连接数（-t）；SSH 等建议 4，默认 4"
+    required: false
+    flag: "-t"
+    format: "flag"
+    default: 4
+  - name: "wait_time"
+    type: "int"
+    description: "单次连接等待响应秒数（-w），默认 16（低于 Hydra 默认 32，减少挂起感）"
+    required: false
+    flag: "-w"
+    format: "flag"
+    default: 16
+  - name: "wait_between"
+    type: "int"
+    description: "每线程连接间隔秒数（-W），默认 1"
+    required: false
+    flag: "-W"
+    format: "flag"
+    default: 1
+  - name: "output_file"
+    type: "string"
+    description: "将结果写入文件（-o），长任务建议指定"
+    required: false
+    flag: "-o"
+    format: "flag"
+  - name: "target"
+    type: "string"
+    description: "目标 IP、主机名或 CIDR（须在选项之后）"
+    required: true
+    position: 1
+    format: "positional"
+  - name: "service"
+    type: "string"
+    description: "服务类型（ssh、ftp、http-get、http-post-form、smb 等，见 hydra -h）"
+    required: true
+    position: 2
+    format: "positional"
   - name: "additional_args"
     type: "string"
-    description: "额外的Hydra参数"
+    description: "额外参数（如 -s 端口、-S SSL、-m 模块选项、-C login:pass 文件），追加在命令末尾"
     required: false
     format: "positional"