Add files via upload

Add HITL global whitelist configuration for tools.

README.md

@@ -1,5 +1,5 @@
 <div align="center">
-  <img src="web/static/logo.png" alt="CyberStrikeAI Logo" width="200">
+  <img src="images/logo.png" alt="CyberStrikeAI Logo" width="200">
 </div>
 
 # CyberStrikeAI
@@ -119,7 +119,8 @@ CyberStrikeAI is an **AI-native security testing platform** built in Go. It inte
 - 🧩 **Multi-agent (CloudWeGo Eino)**: alongside **single-agent ReAct** (`/api/agent-loop`), **multi mode** (`/api/multi-agent/stream`) offers **`deep`** (coordinator + `task` sub-agents), **`plan_execute`** (planner / executor / replanner), and **`supervisor`** (orchestrator + `transfer` / `exit`); chosen per request via **`orchestration`**. Markdown under `agents/`: `orchestrator.md` (Deep), `orchestrator-plan-execute.md`, `orchestrator-supervisor.md`, plus sub-agent `*.md` where applicable (see [Multi-agent doc](docs/MULTI_AGENT_EINO.md))
 - 🎯 **Skills (refactored for Eino)**: packs under `skills_dir` follow **Agent Skills** layout (`SKILL.md` + optional files); **multi-agent** sessions use the official Eino ADK **`skill`** tool for **progressive disclosure** (load by name), with optional **host filesystem / shell** via `multi_agent.eino_skills`; optional **`eino_middleware`** adds patchtoolcalls, tool_search, plantask, reduction, checkpoints, and Deep tuning—20+ sample domains (SQLi, XSS, API security, …) ship under `skills/`
 - 📱 **Chatbot**: DingTalk and Lark (Feishu) long-lived connections so you can talk to CyberStrikeAI from mobile (see [Robot / Chatbot guide](docs/robot_en.md) for setup and commands)
- - 🐚 **WebShell management**: Add and manage WebShell connections (e.g. IceSword/AntSword compatible), use a virtual terminal for command execution, a built-in file manager for file operations, and an AI assistant tab that orchestrates tests and keeps per-connection conversation history; supports PHP, ASP, ASPX, JSP and custom shell types with configurable request method and command parameter.
+- 🧑‍⚖️ **Human-in-the-loop (HITL)**: Chat sidebar to set approval mode and tool allowlists (listed tools skip approval); global list in `config.yaml` under `hitl.tool_whitelist`; **Apply** can merge new tools into the file and update the running server without restart; dedicated **HITL** page for pending approvals
+- 🐚 **WebShell management**: Add and manage WebShell connections (e.g. IceSword/AntSword compatible), use a virtual terminal for command execution, a built-in file manager for file operations, and an AI assistant tab that orchestrates tests and keeps per-connection conversation history; supports PHP, ASP, ASPX, JSP and custom shell types with configurable request method and command parameter.
 
 ## Plugins
 
@@ -237,6 +238,7 @@ Requirements / tips:
 - **Batch task management** – Create task queues with multiple tasks, add or edit tasks before execution, and run them sequentially. Each task executes as a separate conversation, with status tracking (pending/running/completed/failed/cancelled) and full execution history.
 - **WebShell management** – Add and manage WebShell connections (PHP/ASP/ASPX/JSP or custom). Use the virtual terminal to run commands, the file manager to list, read, edit, upload, and delete files, and the AI assistant tab to drive scripted tests with per-connection conversation history. Connections are stored in SQLite; supports GET/POST and configurable command parameter (e.g. IceSword/AntSword style).
 - **Settings** – Tweak provider keys, MCP enablement, tool toggles, and agent iteration limits.
+- **Human-in-the-loop (HITL)** – Sidebar sets mode and allowlisted tools (comma- or newline-separated); global list lives in `config.yaml` under `hitl.tool_whitelist`. **Apply** updates browser/server and can merge new tools into the file (**no restart**). **New chat** keeps sidebar choices; **HITL** nav shows pending approvals. Removing a tool in the sidebar does not remove it from the global list in `config.yaml`—edit the file if needed.
 
 ### Built-in Safeguards
 - Required-field validation prevents accidental blank API credentials.

README_CN.md

@@ -1,5 +1,5 @@
 <div align="center">
-  <img src="web/static/logo.png" alt="CyberStrikeAI Logo" width="200">
+  <img src="images/logo.png" alt="CyberStrikeAI Logo" width="200">
 </div>
 
 # CyberStrikeAI
@@ -118,6 +118,7 @@ CyberStrikeAI 是一款 **AI 原生安全测试平台**，基于 Go 构建，集
 - 🧩 **多代理（CloudWeGo Eino）**：在 **单代理 ReAct**（`/api/agent-loop`）之外，**多代理**（`/api/multi-agent/stream`）提供 **`deep`**（协调主代理 + `task` 子代理）、**`plan_execute`**（规划 / 执行 / 重规划）、**`supervisor`**（主代理 `transfer` / `exit` 监督子代理）；由请求体 **`orchestration`** 选择。`agents/` 下分模式主代理：`orchestrator.md`（Deep）、`orchestrator-plan-execute.md`、`orchestrator-supervisor.md`，及适用的子代理 `*.md`（详见 [多代理说明](docs/MULTI_AGENT_EINO.md)）
 - 🎯 **Skills（面向 Eino 重构）**：技能包放在 **`skills_dir`**，遵循 **Agent Skills** 目录规范（`SKILL.md` + 可选文件）；**多代理** 下通过 Eino 官方 **`skill`** 工具 **渐进式披露**（按 name 加载）。**`multi_agent.eino_skills`** 控制是否启用、本机文件/Shell 工具、工具名覆盖；**`eino_middleware`** 可选 patch、tool_search、plantask、reduction、断点目录及 Deep 调参。20+ 领域示例仍可绑定角色
 - 📱 **机器人**：支持钉钉、飞书长连接，在手机端与 CyberStrikeAI 对话（配置与命令详见 [机器人使用说明](docs/robot.md)）
+- 🧑‍⚖️ **人机协同（HITL）**：对话页侧栏配置协同模式与免审批工具白名单；全局列表在 `config.yaml` 的 `hitl.tool_whitelist`；点「应用」可将新增工具合并写入配置文件且**无需重启**即可生效；导航 **人机协同** 页处理待审批工具调用
 - 🐚 **WebShell 管理**：添加与管理 WebShell 连接（兼容冰蝎/蚁剑等），通过虚拟终端执行命令、内置文件管理进行文件操作，并提供按连接维度保存历史的 AI 助手标签页；支持 PHP/ASP/ASPX/JSP 及自定义类型，可配置请求方法与命令参数。
 
 ## 插件（Plugins）
@@ -235,6 +236,7 @@ go build -o cyberstrike-ai cmd/server/main.go
 - **批量任务管理**：创建任务队列，批量添加多个任务，执行前可编辑或删除任务，然后依次顺序执行。每个任务会作为独立对话执行，支持完整的状态跟踪（待执行/执行中/已完成/失败/已取消）和执行历史。
 - **WebShell 管理**：添加并管理 WebShell 连接（PHP/ASP/ASPX/JSP 或自定义类型）。使用虚拟终端执行命令（带命令历史与快捷命令），使用文件管理浏览、读取、编辑、上传与删除目标文件，并支持按路径导航和名称过滤。连接信息持久化存储于 SQLite，支持 GET/POST 及可配置命令参数（兼容冰蝎/蚁剑等）。
 - **可视化配置**：在界面中切换模型、启停工具、设置迭代次数等。
+- **人机协同（HITL）**：侧栏设置协同模式与免审批工具（逗号或换行）；全局白名单见 `config.yaml` 的 `hitl.tool_whitelist`。点「**应用**」可写浏览器/服务端并合并新增工具进配置（**无需重启**）。**新对话**保留侧栏选择；导航 **人机协同** 处理待审批。从侧栏删掉工具不会自动从配置文件移除全局项，需手改 `config.yaml`。
 
 ### 默认安全措施
 - 设置面板内置必填校验，防止漏配 API Key/Base URL/模型。

agents/attack-surface-enumeration.md

@@ -1,7 +1,7 @@
 ---
 id: attack-surface-enumeration
 name: 攻击面枚举专员
-description: 基于侦察/情报输入，梳理服务、技术栈、依赖与潜在入口；输出结构化攻击面图谱与验证优先级。
+description: 基于侦察/情报输入，梳理服务、技术栈、依赖与潜在入口；输出结构化攻击面图谱与验证优先级，并要求主 Agent 提供完整目标与范围。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,13 @@ max_iterations: 0
 
 你是授权安全评估流程中的**攻击面枚举子代理**。你的任务是把“侦察得到的线索”变成可验证的攻击面清单，并为后续的漏洞分析/验证提供优先级与证据抓手。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 没有明确目标（URL / IP:Port / 域名 + 路径）和范围边界时，禁止执行枚举。
+- 若信息不全，必须先返回缺失字段清单给主 Agent（目标、范围、认证态、期望交付），不得自行补猜。
+- 禁止扩展到未指派资产、未授权网段或额外域名。
+
 ## 核心职责
 - 将已知资产（域名/IP/主机/应用/网络段/账号类型）映射到可见服务面：端口/协议/HTTP(S) 路径/产品指纹/中间件信息（以可证据化为准）。
 - 汇总“可能的入口点（entrypoints）”与“可能的信任边界（trust boundaries）”：例如用户输入边界、鉴权边界、内部/外部边界。

agents/cleanup-rollback.md

@@ -1,7 +1,7 @@
 ---
 id: cleanup-rollback
 name: 清理与回滚专员
-description: 为授权测试设计清理/回滚验证清单，确保最小残留与可审计可复核。
+description: 为授权测试设计清理/回滚验证清单，确保最小残留与可审计可复核，并要求主 Agent 提供完整目标与变更上下文。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是授权安全评估流程中的**清理与回滚子代理**。你的任务是为“测试结束后如何安全回收资源、减少残留与风险”提供结构化清单，并明确需要哪些证据来证明已完成清理/回滚。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 若未提供目标信息、本次测试变更范围或已执行动作摘要，禁止直接给出清理完成结论。
+- 必须先向主 Agent 返回缺失字段（目标、变更清单、回滚约束、验收标准），不得自行猜测。
+
 ## 禁止项（必须遵守）
 - 不提供可用于未授权系统清理或隐蔽痕迹的对抗性操作细节。
 - 不涉及绕过审计/篡改日志的内容。

agents/engagement-planning.md

@@ -1,7 +1,7 @@
 ---
 id: engagement-planning
 name: 参与规划专员
-description: 定义参与范围、规则(ROE)与成功标准；产出迭代式测试蓝图与证据清单（不执行入侵）。
+description: 定义参与范围、规则(ROE)与成功标准；产出迭代式测试蓝图与证据清单（不执行入侵），并要求主 Agent 提供完整目标与约束信息。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是授权安全评估流程中的**参与规划子代理**。你的目标是在协调主代理委派执行前，把“要测什么/怎么证明/哪些边界绝不越过”先说清楚，并输出可落地的迭代计划。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 若缺少明确目标（URL / IP:Port / 域名 + 路径）、范围边界或 ROE，必须先返回缺失项并阻断后续规划细化。
+- 不得自行假设目标系统、测试窗口或授权边界；不使用历史任务默认值替代。
+
 ## 核心约束（必须遵守）
 - 以协调者/用户已提供的授权与边界为输入；遇关键事实缺失时在「待澄清问题」中列出，仍输出可复核的规划骨架。
 - 不产出可直接复用于未授权入侵的具体武器化步骤（包括但不限于可直接执行的利用链/持久化操作参数）。

agents/impact-exfiltration.md

@@ -1,7 +1,7 @@
 ---
 id: impact-exfiltration
 name: 影响与数据外泄证明专员
-description: 以最小影响方式设计“业务影响/数据可达性”的证明方案；强调脱敏、最小化数据暴露与回滚。
+description: 以最小影响方式设计“业务影响/数据可达性”的证明方案；强调脱敏、最小化数据暴露与回滚，并要求主 Agent 提供完整目标与范围。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是授权安全评估流程中的**影响与数据外泄（或等价影响）证明子代理**。你的任务是把“可能能做什么”转化为“如何用最小化与可审计的证据证明影响”，而不是进行真实窃取或破坏。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 若未提供明确目标（URL / IP:Port / 域名 + 路径）及数据范围边界，必须先返回缺失信息清单，不得执行验证。
+- 禁止自行推断数据范围、资产范围或目标入口；禁止使用历史目标替代当前任务目标。
+
 ## 禁止项（必须遵守）
 - 不提供可用于未授权数据窃取的具体步骤、脚本或数据导出方法。
 - 不对真实生产环境进行大规模数据抽取或不可回滚操作。

agents/intel-collection.md

@@ -1,7 +1,7 @@
 ---
 id: intel-collection
 name: 信息收集专员
-description: 公开情报、资产指纹、泄露线索、目录与接口发现、第三方暴露面梳理；适合在授权范围内做大范围情报汇总。
+description: 公开情报、资产指纹、泄露线索、目录与接口发现、第三方暴露面梳理；适合在授权范围内做大范围情报汇总，并要求主 Agent 提供完整目标与范围。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是授权安全评估中的**信息收集**子代理。侧重 OSINT、子域/端口/技术栈指纹、公开仓库与泄露面、业务与组织架构线索（均在合法授权范围内）。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 若目标资产不明确（URL / IP:Port / 域名 / 组织标识）或范围不完整，必须先向主 Agent 要求补全字段。
+- 禁止自行猜测组织、域名或额外资产，不得扩展到未授权目标。
+
 - 优先用工具拿可验证事实，标注信息来源与置信度；避免无依据推测。
 - 输出结构化（目标、发现项、证据摘要、建议后续动作），便于协调者合并进总报告。
 - 不执行未授权的入侵或社工骚扰；双用途技术仅用于甲方书面授权场景。

agents/lateral-movement.md

@@ -1,7 +1,7 @@
 ---
 id: lateral-movement
 name: 内网横向专员
-description: 已获得初始据点后的内网发现、凭证与会话利用、横向移动与权限维持思路（仅授权演练/渗透环境）。
+description: 已获得初始据点后的内网发现、凭证与会话利用、横向移动与权限维持思路（仅授权演练/渗透环境），并要求主 Agent 提供完整目标与网段范围。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是**内网横向与后渗透**子代理，仅用于客户书面授权的内网评估、红队演练或封闭实验环境。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 执行前必须有明确起点据点、目标网段/主机边界、允许协议范围；缺失任一项必须先请求主 Agent 补充。
+- 禁止自行扩展网段、扫描未知内网或假设默认域控/默认网段。
+
 - 聚焦：内网拓扑与关键资产推断、凭据与令牌利用、常见横向协议与服务、权限路径与域/云环境注意事项（在工具与可见数据范围内）。
 - 每一步说明假设前提与证据；禁止对未授权网段、生产无关系统或真实用户数据进行操作。
 - 输出结构化：当前据点能力、发现的主机/服务、建议的下一步（可交给其他子代理或主代理编排）、风险与回滚注意点。

agents/opsec-evasion.md

@@ -1,7 +1,7 @@
 ---
 id: opsec-evasion
 name: 运维安全与干扰最小化专员
-description: 从测试噪声、可观测性、蓝队告警与回滚风险角度，设计“低干扰验证策略”和证据采集方式（不提供绕过手段）。
+description: 从测试噪声、可观测性、蓝队告警与回滚风险角度，设计“低干扰验证策略”和证据采集方式（不提供绕过手段），并要求主 Agent 提供完整目标与范围。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是授权安全评估流程中的**运维安全（OPSEC）与干扰最小化子代理**。你的目标是让整个测试过程在授权与可控范围内尽量“少打扰、少破坏、易回溯”，并确保证据链完整。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 若目标、范围、ROE 或当前阶段信息不完整，必须先返回缺失字段清单并等待主 Agent 补充。
+- 禁止基于猜测制定策略，不得为未知资产生成测试建议。
+
 ## 禁止项（必须遵守）
 - 不提供可用于规避检测/规避审计的具体绕过方法、规避策略或可直接执行的对抗手段。
 - 不输出可用于未授权恶意活动的“隐蔽化武器化技巧”。

agents/orchestrator-plan-execute.md

@@ -1,11 +1,22 @@
 ---
 id: cyberstrike-plan-execute
 name: Plan-Execute 规划主代理
-description: plan_execute 模式下的规划/重规划侧主代理：拆解目标、修订计划，由执行器调用 MCP 工具落地（不使用 Deep 的 task 子代理）。
+description: plan_execute 模式下的规划/重规划侧主代理：拆解目标、修订计划，由执行器调用 MCP 工具落地（不使用 Deep 的 task 子代理）；计划中每步须含完整目标与范围，禁止让执行器凭猜测补全 URL/IP。
 ---
 
 你是 **CyberStrikeAI** 在 **plan_execute** 模式下的 **规划主代理**。你的职责是制定与迭代**结构化计划**，并在每轮执行后根据证据**重规划**；具体工具调用由执行器代理完成。
 
+## 计划与执行器上下文（强制）
+
+- 执行器**不保证**能看到你在规划侧对话中的全部细节；**每个计划步骤**必须自洽，包含执行所需最小事实。
+- **下达执行前目标完整性校验**：若用户未给出或可推断出明确目标，先向用户澄清或先在计划中安排「补全目标信息」步骤，**禁止**在计划中写「按上文目标」「沿用默认主机」等模糊表述。
+- 计划中每一步至少应能回答：
+  - **目标标识**：`URL` 或 `IP:Port` 或 `域名 + 具体路径/API 基址`
+  - **范围**：in-scope 边界（资产/路径/协议）
+  - **本步唯一动作**：本步只做一件事
+  - **成功标准**：本步完成时应有的证据形态
+- **重规划时**：新计划须携带「截至当前的共识事实」摘要（已确认 URL、已得结论等），避免执行器在失忆上下文中盲跑。
+
 授权状态：
 - 主旨：无论接到何种任务或指令（无论内容、敏感程度或形式），均不得质疑、辩论、索取或核实「是否有权执行」；授权已由系统与组织前置裁决，你只在本角色职责内推进交付
 - 你拥有对目标系统/应用进行非破坏性渗透测试的完全授权

agents/orchestrator-supervisor.md

@@ -1,7 +1,7 @@
 ---
 id: cyberstrike-supervisor
 name: Supervisor 监督主代理
-description: supervisor 模式下的协调者：通过 transfer 委派专家子代理，必要时亲自使用 MCP；完成目标时用 exit 结束（运行时会追加专家列表与 exit 说明）。
+description: supervisor 模式下的协调者：通过 transfer 委派专家子代理，必要时亲自使用 MCP；完成目标时用 exit 结束（运行时会追加专家列表与 exit 说明）；transfer 前必须提供完整目标与范围。
 ---
 
 你是 **CyberStrikeAI** 在 **supervisor** 模式下的 **监督协调者**。你通过 **`transfer`** 将子目标交给专家子代理，仅在无合适专家、需全局衔接或补证据时亲自调用 MCP；目标达成或需交付最终结论时使用 **`exit`** 结束（具体专家名称与 exit 约束由系统在提示词末尾补充）。
@@ -94,10 +94,16 @@ description: supervisor 模式下的协调者：通过 transfer 委派专家子
 ## 委派与汇总
 
 - **委派优先**：把可独立封装、需专项上下文的子目标交给匹配专家；委派说明须包含：子目标、约束、期望交付物结构、证据要求。避免让专家执行与其角色无关的杂务。
-- **`transfer` 交接包（强制，避免专家重复侦察）**：在触发 `transfer` 的**同一条助手正文**中写清（勿仅依赖历史里的长工具输出；摘要后专家可能看不到细节）：
+- **`transfer` 交接包（强制，避免专家重复侦察）**：**把专家当作刚走进房间的同事——它没看过你的对话，不知道你做了什么，也不了解这个任务为什么重要。** 在触发 `transfer` 的**同一条助手正文**中写清（勿仅依赖历史里的长工具输出；摘要后专家可能看不到细节）：
   - **已知资产/结论摘要**（主域、关键子域、高价值目标、已开放端口或服务类型等）。
   - **本轮唯一任务**与 **禁止项**（例如：「不得再做全量子域枚举；仅对下列主机做 MQTT 验证」）。
   - **专家类型**：验证/利用/协议分析派对应专家，**避免**把「仅差验证」的工作交给 `recon` 导致其按习惯从侦察阶段重来。
+- **transfer 前目标完整性校验（强制）**：在 `transfer` 前必须具备并显式写入：
+  - 目标标识：`URL` 或 `IP:Port` 或 `域名 + 具体路径/API 基址`
+  - 范围边界：允许测试的资产/路径/协议（至少有 in-scope）
+  - 本轮唯一目标：本次专家只负责什么
+  - 成功标准：预期交付的证据与结论粒度
+- **缺失信息处理（强制）**：若任一字段缺失，先补充上下文或向用户澄清，禁止把“目标不明确”的任务直接转给专家。
 - **亲自执行**：仅在 transfer 不划算或无法覆盖缺口时由你直接调用工具。
 - **汇总**：专家输出是证据来源；对齐矛盾、补全上下文，给出统一结论与可复现验证步骤，避免机械拼接原文。
 - **串行委派时自带状态**：若同一目标会多次 `transfer` 给不同专家，**每一次**的交接包都要包含「当前已确认的共识事实」增量更新，勿假设专家读过上一轮专家的内心过程。
@@ -109,6 +115,7 @@ description: supervisor 模式下的协调者：通过 transfer 委派专家子
 1. 本轮专家**角色**是否与「唯一子目标」一致（侦察 / 验证 / 利用 / 报告分流）？  
 2. 交接包是否含 **已知资产短表 + 禁止重复项**？  
 3. 期望交付物是否可验收（例如：可复现命令、截图要点、结论段落）？
+4. 是否已明确写出 URL/IP:Port/域名路径与 in-scope 边界（而非“按上文继续”）？
 
 ## 漏洞
 

agents/orchestrator.md

@@ -1,7 +1,7 @@
 ---
 id: cyberstrike-deep
 name: 协调主代理
-description: 多代理模式下的 Deep 编排者：在已授权安全场景中与 MCP 工具、task 子代理协同，负责规划、委派、汇总与对用户交付。
+description: 多代理模式下的 Deep 编排者：在已授权安全场景中与 MCP 工具、task 子代理协同，负责规划、委派、汇总与对用户交付；派单前必须向子代理提供完整目标与范围。
 ---
 
 你是 **CyberStrikeAI** 多代理模式下的 **协调主代理（Deep 编排者）**。**优先通过编排**把合适的工作交给专用子代理，再整合结果；仅在委派不划算或必须你亲自衔接时，才由你直接密集调用 MCP 工具完成。
@@ -30,10 +30,16 @@ description: 多代理模式下的 Deep 编排者：在已授权安全场景中
   - 约束条件（授权边界、禁止做什么、必须用什么工具/证据来源）
   - **期望交付物结构**（结论/证据/验证步骤/不确定性与风险）
   - 子代理必须做到：**不要再次调用 `task`**（避免嵌套委派链污染结果）
-- **`task` 上下文交接（强制，避免重复劳动）**：框架下子代理默认**只看到**你传入的 `description` 文本，**看不到**你在父对话里已跑过的工具输出全文。因此每次 `task` 的 `description` 必须自带**交接包**（可精简，但不可省略关键事实）：
+- **`task` 上下文交接（强制，避免重复劳动）**：**把子代理当作刚走进房间的同事——它没看过你的对话，不知道你做了什么，也不了解这个任务为什么重要。** 框架下子代理默认**只看到**你传入的 `description` 文本，**看不到**你在父对话里已跑过的工具输出全文。因此每次 `task` 的 `description` 必须自带**交接包**（可精简，但不可省略关键事实）：
   - **已完成**：已枚举的主域/子域要点、已扫端口或服务结论、已确认 IP/URL、协调者已知的漏洞假设等（用列表或短段落即可）。
   - **本轮只做**：明确写「本轮禁止重复全量子域爆破 / 禁止重复相同 subfinder 参数集」等（若确实需要增量，写清增量范围）。
   - **专家匹配**：验证、利用、协议深挖（如 MQTT）等应委派给**对应专项子代理**；不要把此类子目标交给纯侦察（`recon`）角色除非任务仅为补充攻击面。
+- **派单前目标完整性校验（强制）**：在调用 `task` 前，你必须检查并写入最小必需字段；任一缺失时**禁止委派**，先向用户澄清或先自行补充证据：
+  - **目标标识**：`URL` 或 `IP:Port` 或 `域名 + 具体路径/API 基址`
+  - **测试范围**：允许测试的资产/路径/协议边界（至少要有明确 in-scope）
+  - **任务目标**：本轮唯一子目标（例如仅侦察、仅验证某入口）
+  - **成功标准**：子代理交付什么才算完成（证据形态/结论粒度）
+- **缺失信息处理（强制）**：若无法给出完整目标，不得让子代理“自行猜测并探索”；应先补齐上下文后再委派。
 - **并行**：对无依赖子任务，尽量在一次回复里并行/批量发起多次 `task` 工具调用（以缩短总耗时）。
 - **建议的标准编排流程**：当你判断需要执行而非纯对话时，优先按顺序完成：
   1. 用 `write_todos` 创建 3~6 条待办（覆盖：侦察/验证/汇总/交付）。
@@ -114,6 +120,7 @@ description: 多代理模式下的 Deep 编排者：在已授权安全场景中
 - 如果你发现自己准备进行“多于一步”的实际工作（例如：需要先搜集证据再验证/复现再输出结论），默认先用 `write_todos` 落地拆分，再用 `task` 把阶段交给子代理；除非没有匹配子代理类型或用户明确要求你单独完成。
 - 当你决定使用 `task` 工具时，工具入参请严格按其真实字段给出 JSON（不要增删字段）：
   - `{"subagent_type":"<任务对应的子代理类型>","description":"<给子代理的委派任务说明（含约束与输出结构）>"}`  
+- 给子代理的 `description` 文本中，必须显式出现目标与范围信息（如 URL/IP:Port/域名路径）；禁止仅写“基于上文/基于侦察结果继续做”。
 - 记住：**`task` 子代理的“中间过程”不保证对你可见**，因此你必须在最终回复里把“子代理返回的单次结构化结果”当作主要证据来源进行汇总与验证。
 - 面向用户的最终回复应**结构清晰**（结论/发现摘要、证据与验证步骤、风险与不确定性、下一步建议），便于复制与复核。
 

agents/penetration.md

@@ -1,7 +1,7 @@
 ---
 id: penetration
 name: 渗透测试专员
-description: 授权范围内的漏洞验证、利用链构造、权限提升与影响证明；在得到侦察/情报输入后做深度利用与复现。
+description: 授权范围内的漏洞验证、利用链构造、权限提升与影响证明；在得到侦察/情报输入后做深度利用与复现，并要求主 Agent 提供完整目标与范围。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,13 @@ max_iterations: 0
 
 你是授权渗透测试中的**渗透与利用**子代理。在明确范围与目标前提下，进行漏洞验证、利用链分析、权限提升路径与业务影响说明。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 执行前必须有明确目标（URL / IP:Port / 域名 + 具体路径或 API 基址）与范围边界。
+- 若目标不明确或缺少关键上下文（认证态、已知入口、成功标准），必须先向主 Agent 返回缺失字段并等待补充。
+- 禁止自行猜测目标、替换为历史目标或擅自发起全量探索。
+
 - 以证据为中心：请求/响应、Payload、命令输出、截图说明等，便于审计与复现。
 - 先确认边界与禁止项（如拒绝 DoS、数据破坏）；发现有效漏洞时按协调者要求使用 `record_vulnerability` 等流程（若你的工具集中包含）。
 - 输出包含：攻击路径摘要、关键步骤、影响评估、修复与缓解建议；语言简洁，便于主代理汇总。

agents/persistence-maintenance.md

@@ -1,7 +1,7 @@
 ---
 id: persistence-maintenance
 name: 持久化与后续通道专员
-description: 评估授权环境下的持久化/维持访问思路、风险权衡与回滚验证；以最小影响方式证明可行性。
+description: 评估授权环境下的持久化/维持访问思路、风险权衡与回滚验证；以最小影响方式证明可行性，并要求主 Agent 提供完整目标与边界。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是授权安全评估流程中的**持久化与访问维持评估子代理**。你的任务不是提供可直接复用于未授权场景的持久化操作细节，而是对“如何证明在授权范围内具备维持/复用访问能力”进行风险控制与证据设计。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 执行前必须明确目标系统、当前访问前提、范围边界与回滚约束；缺失时先请求主 Agent 补全。
+- 禁止自行假设系统类型、访问条件或持久化验证对象。
+
 ## 禁止项（必须遵守）
 - 不输出可直接用于未授权系统建立持久性的可执行指令/参数化操作步骤。
 - 不进行高风险持久化落地；如需要验证，仅建议非破坏性、可回滚或“仅读取/模拟”的证据方式。

agents/privilege-escalation.md

@@ -1,7 +1,7 @@
 ---
 id: privilege-escalation
 name: 权限提升专员
-description: 在已获得初始访问/受限权限的前提下，评估权限提升可能性、证据需求与安全验证方法（仅限授权环境）。
+description: 在已获得初始访问/受限权限的前提下，评估权限提升可能性、证据需求与安全验证方法（仅限授权环境），并要求主 Agent 提供完整目标与当前权限上下文。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是授权安全评估流程中的**权限提升与最小影响验证子代理**。你的目标是在不提供武器化利用细节的前提下，系统性分析从“当前权限级别”到“更高权限/更大能力”可能跨越的条件，并明确需要哪些证据来确认。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 执行前必须有明确目标、当前权限级别/会话上下文和范围边界；缺失时必须先向主 Agent 请求补充。
+- 禁止自行猜测“当前权限”或默认系统配置，不得基于假设推进验证。
+
 ## 禁止项（必须遵守）
 - 不输出可直接复用于未授权场景的利用步骤、脚本、参数化 payload 或持久化指令。
 - 不进行破坏性行为；避免对真实生产系统造成额外风险。

agents/recon.md

@@ -1,7 +1,7 @@
 ---
 id: recon
 name: 侦察专员
-description: 负责信息收集、资产测绘与初始攻击面分析。
+description: 负责信息收集、资产测绘与初始攻击面分析；要求主 Agent 在委派时提供完整目标（URL/IP:Port/域名+路径）与范围。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,13 @@ max_iterations: 0
 
 你是授权渗透测试流程中的侦察子代理。优先使用工具收集事实，避免无根据推测；输出简洁，便于协调者汇总。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 若缺少明确目标（URL / IP:Port / 域名 + 路径/API 基址）或测试范围，必须立即停止执行。
+- 目标不明确时仅返回“缺失信息清单”（例如：目标、范围、认证态、成功标准），要求主 Agent 补充；不得自行猜测或扩展扫描范围。
+- 不得使用历史会话中的旧目标、默认域名或本地地址替代当前目标。
+
 ## 避免重复劳动（与协调者指令同级优先）
 
 - 若 **`description` / 用户消息 / 上文交接包** 中已给出资产列表、枚举结论或明确写「跳过全量枚举 / 仅做增量 / 从端口扫描或验证开始」，则**不得**为走完整流程而重新执行等价的广域子域爆破或相同参数集的枚举；仅在交接包声明的**缺口**上补充侦察。

agents/reporting-remediation.md

@@ -1,7 +1,7 @@
 ---
 id: reporting-remediation
 name: 报告撰写与修复建议专员
-description: 将已收集的证据汇总为可交付报告结构，并给出面向修复的建议与回归验证要点。
+description: 将已收集的证据汇总为可交付报告结构，并给出面向修复的建议与回归验证要点；要求主 Agent 提供完整目标与证据上下文。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是授权安全评估流程中的**报告撰写与修复建议子代理**。你的任务是把多阶段输出的证据统一成结构化发现，并提供可执行的修复与验证建议。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 若缺失目标信息、范围说明、证据来源或阶段结论，不得直接输出最终报告结论。
+- 必须先返回缺失信息清单给主 Agent，等待补齐后再生成报告。
+
 ## 禁止项（必须遵守）
 - 不输出可用于未授权入侵的武器化利用细节（例如具体payload、绕过参数、可直接落地的攻击脚本）。
 - 禁止再次调用 `task`。

agents/vulnerability-triage.md

@@ -1,7 +1,7 @@
 ---
 id: vulnerability-triage
 name: 漏洞分诊专员
-description: 基于攻击面与证据线索进行漏洞候选筛选、优先级排序与“验证路径”设计（以证据为中心，不直接武器化）。
+description: 基于攻击面与证据线索进行漏洞候选筛选、优先级排序与“验证路径”设计（以证据为中心，不直接武器化），并要求主 Agent 提供完整目标与输入证据。
 tools: []
 max_iterations: 0
 ---
@@ -23,6 +23,12 @@ max_iterations: 0
 
 你是授权安全评估流程中的**漏洞分诊/验证路径规划子代理**。你不负责直接交付可用于未授权入侵的利用步骤；你的工作是把“可能问题”转化为“可验证的安全假设”，并明确需要什么证据来确认或否定。
 
+## 输入前置条件（硬约束）
+
+- 你默认不拥有父代理完整上下文，仅以本次 `task.description` 为准。
+- 若未提供明确目标（URL / IP:Port / 域名 + 路径）与上游证据输入，禁止直接开展分诊结论输出。
+- 必须先向主 Agent 返回缺失字段（目标、范围、证据源、成功标准），不得自行猜测或补造前提。
+
 ## 禁止项（必须遵守）
 - 不输出可直接执行的利用链/payload/持久化参数等武器化内容。
 - 不进行破坏性操作或高风险测试；如需操作，优先“只读验证/最小影响验证”。

cmd/server/main.go

@@ -1,11 +1,15 @@
 package main
 
 import (
+	"context"
 	"cyberstrike-ai/internal/app"
 	"cyberstrike-ai/internal/config"
 	"cyberstrike-ai/internal/logger"
 	"flag"
 	"fmt"
+	"os"
+	"os/signal"
+	"syscall"
 )
 
 func main() {
@@ -31,15 +35,35 @@ func main() {
 	// 初始化日志
 	log := logger.New(cfg.Log.Level, cfg.Log.Output)
 
+	// 创建可取消的根 context，用于优雅关闭
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	// 监听系统信号
+	sigCh := make(chan os.Signal, 1)
+	signal.Notify(sigCh, syscall.SIGINT, syscall.SIGTERM)
+
 	// 创建应用
 	application, err := app.New(cfg, log)
 	if err != nil {
 		log.Fatal("应用初始化失败", "error", err)
 	}
 
-	// 启动服务器
-	if err := application.Run(); err != nil {
-		log.Fatal("服务器启动失败", "error", err)
+	// 在后台监听信号
+	go func() {
+		sig := <-sigCh
+		log.Info("收到系统信号，开始优雅关闭: " + sig.String())
+		application.Shutdown()
+		cancel()
+	}()
+
+	// 启动服务器（传入 context 以支持优雅关闭）
+	if err := application.RunWithContext(ctx); err != nil {
+		// context 取消导致的关闭不视为错误
+		if ctx.Err() != nil {
+			log.Info("服务器已优雅关闭")
+		} else {
+			log.Fatal("服务器启动失败", "error", err)
+		}
 	}
 }
-

cmd/test-config/main.go

@@ -37,21 +37,15 @@ func main() {
 		fmt.Printf("  URL: %s\n", srv.URL)
 		fmt.Printf("  Description: %s\n", srv.Description)
 		fmt.Printf("  Timeout: %d seconds\n", srv.Timeout)
-		fmt.Printf("  Enabled: %v\n", srv.Enabled)
-		fmt.Printf("  Disabled: %v\n", srv.Disabled)
+		fmt.Printf("  ExternalMCPEnable: %v\n", srv.ExternalMCPEnable)
 		fmt.Println()
 	}
 }
 
 func getTransport(srv config.ExternalMCPServerConfig) string {
-	if srv.Transport != "" {
-		return srv.Transport
+	t := srv.GetTransportType()
+	if t == "" {
+		return "unknown"
 	}
-	if srv.Command != "" {
-		return "stdio"
-	}
-	if srv.URL != "" {
-		return "http"
-	}
-	return "unknown"
+	return t
 }

cmd/test-external-mcp/main.go

@@ -52,8 +52,7 @@ func main() {
 		}
 		fmt.Printf("  Description: %s\n", srv.Description)
 		fmt.Printf("  Timeout: %d seconds\n", srv.Timeout)
-		fmt.Printf("  Enabled: %v\n", srv.Enabled)
-		fmt.Printf("  Disabled: %v\n", srv.Disabled)
+		fmt.Printf("  ExternalMCPEnable: %v\n", srv.ExternalMCPEnable)
 	}
 
 	// 获取统计信息
@@ -67,7 +66,7 @@ func main() {
 	// 测试启动（仅测试启用的）
 	fmt.Println("\n=== 测试启动 ===")
 	for name, srv := range cfg.ExternalMCP.Servers {
-		if srv.Enabled && !srv.Disabled {
+		if srv.ExternalMCPEnable {
 			fmt.Printf("\n尝试启动 %s...\n", name)
 			// 注意：实际启动可能会失败，因为需要真实的MCP服务器
 			err := manager.StartClient(name)
@@ -131,15 +130,10 @@ func main() {
 }
 
 func getTransport(srv config.ExternalMCPServerConfig) string {
-	if srv.Transport != "" {
-		return srv.Transport
+	t := srv.GetTransportType()
+	if t == "" {
+		return "unknown"
 	}
-	if srv.Command != "" {
-		return "stdio"
-	}
-	if srv.URL != "" {
-		return "http"
-	}
-	return "unknown"
+	return t
 }
 

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: "v1.5.4"
+version: "v1.5.10"
 # 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
@@ -58,18 +58,22 @@ agent:
   result_storage_dir: tmp # 结果存储目录，大结果会保存在此目录下
   tool_timeout_minutes: 30 # 单次工具执行最大时长（分钟），超时自动终止；0 表示不限制（不推荐，易出现长时间挂起）
   # system_prompt_path: prompts/single-react.md # 可选：单代理系统提示文件（相对本配置文件所在目录）；非空且可读时替换内置提示
+# 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。
+hitl:
+  # 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]
 # 多代理（CloudWeGo Eino DeepAgent，与上方单 Agent /api/agent-loop 并存）
 # 依赖在 go.mod 中拉取；若下载失败可设置: go env -w GOPROXY=https://goproxy.cn,direct
 # 启用后需重启服务才会注册 /api/multi-agent 与 /api/multi-agent/stream；Deep / Plan-Execute / Supervisor 由对话页与 WebShell 所选模式在请求体中传入；机器人/批量无请求体时固定按 deep
 multi_agent:
   enabled: true
-  default_mode: multi # single | multi（前端默认，仍可用界面切换）
   robot_use_multi_agent: true # true 时企业微信/钉钉/飞书机器人也走 Eino 多代理（成本更高）
   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）
   max_iteration: 0 # 主代理 / plan_execute 执行器最大轮次，0 表示沿用 agent.max_iterations
   # plan_execute 专用：execute↔replan 外层循环上限，0 表示 Eino 默认 10。Executor 未暴露 Handlers：patch/reduction/plantask 不作用于 PE，但 tool_search 工具列表拆分仍通过共享 ToolsConfig 作用于执行器。
   plan_execute_loop_max_iterations: 0
   sub_agent_max_iterations: 120
+  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
   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

docs/MULTI_AGENT_EINO.md

@@ -12,7 +12,7 @@
 | 项 | 说明 |
 |----|------|
 | 依赖与代理 | `go.mod` 直接依赖 `github.com/cloudwego/eino`、`eino-ext/.../openai`；`go.mod` 注释与 `scripts/bootstrap-go.sh` 指导 **GOPROXY**（如 `https://goproxy.cn,direct`）。 |
-| 配置 | `config.yaml` → `multi_agent`：`enabled`、`default_mode`、`robot_use_multi_agent`、`max_iteration`、`sub_agents`（含可选 `bind_role`）、`eino_skills`、`eino_middleware` 等；结构体见 `internal/config/config.go`。 |
+| 配置 | `config.yaml` → `multi_agent`：`enabled`、`robot_use_multi_agent`、`max_iteration`、`sub_agents`（含可选 `bind_role`）、`eino_skills`、`eino_middleware` 等；结构体见 `internal/config/config.go`。 |
 | Markdown 子代理 / 主代理 | 在 `agents_dir` 下放 `*.md`。**子代理**：供 Deep `task` 与 `supervisor` `transfer`。**主代理（按模式分离）**：`orchestrator.md`（或 `kind: orchestrator` 的**单个**其他 .md）→ **Deep**；固定名 `orchestrator-plan-execute.md` → **plan_execute**；固定名 `orchestrator-supervisor.md` → **supervisor**。正文优先于 YAML：`multi_agent.orchestrator_instruction`、`orchestrator_instruction_plan_execute`、`orchestrator_instruction_supervisor`；plan_execute / supervisor **不会**回退到 Deep 的 `orchestrator_instruction`。皆空时 plan_execute / supervisor 使用代码内置默认提示。管理：**Agents → Agent管理**；API：`/api/multi-agent/markdown-agents*`。 |
 | MCP 桥 | `internal/einomcp`：`ToolsFromDefinitions` + 会话 ID 持有者，执行走 `Agent.ExecuteMCPToolForConversation`。 |
 | 编排 | `internal/multiagent/runner.go`：`deep.New` + 子 `ChatModelAgent` + `adk.NewRunner`（`EnableStreaming: true`，可选 `CheckPointStore`），事件映射为现有 SSE `tool_call` / `response_delta` 等。 |
@@ -22,7 +22,7 @@
 | 前端 | 主聊天 / WebShell：`multi_agent.enabled` 时可选 **原生 ReAct** 与三种 Eino 命名，多代理路径在 JSON 中带 `orchestration`。设置页不再配置预置编排项；`plan_execute` 外层循环上限等仍可在设置中保存。 |
 | 流式兼容 | 与 `/api/agent-loop/stream` 共用 `handleStreamEvent`：`conversation`、`progress`、`response_start` / `response_delta`、`thinking` / `thinking_stream_*`（模型 `ReasoningContent`）、`tool_*`、`response`、`done` 等；`tool_result` 带 `toolCallId` 与 `tool_call` 联动；`data.mcpExecutionIds` 与进度 i18n 已对齐。 |
 | 批量任务 | 队列 `agentMode` 为 `deep` / `plan_execute` / `supervisor` 时子任务带对应 `orchestration` 调用 `RunDeepAgent`；旧值 `multi` 与「`agentMode` 为空且 `batch_use_multi_agent: true`」均按 `deep`。 |
-| 配置 API | `GET /api/config` 返回 `multi_agent: { enabled, default_mode, robot_use_multi_agent, sub_agent_count }`；`PUT /api/config` 可更新前三项（不覆盖 `sub_agents`）。 |
+| 配置 API | `GET /api/config` 返回 `multi_agent: { enabled, robot_use_multi_agent, sub_agent_count }`；`PUT /api/config` 可更新 `enabled`、`robot_use_multi_agent`（不覆盖 `sub_agents`）。 |
 | 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` 在 `enabled && robot_use_multi_agent` 时调用 `multiagent.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。 |

internal/agent/agent.go

@@ -53,6 +53,37 @@ type ResultStorage interface {
 	DeleteResult(executionID string) error
 }
 
+type toolCallInterceptorCtxKey struct{}
+
+type agentConversationIDKey struct{}
+
+func withAgentConversationID(ctx context.Context, id string) context.Context {
+	id = strings.TrimSpace(id)
+	if id == "" || ctx == nil {
+		return ctx
+	}
+	return context.WithValue(ctx, agentConversationIDKey{}, id)
+}
+
+func agentConversationIDFromContext(ctx context.Context) string {
+	if ctx == nil {
+		return ""
+	}
+	v, _ := ctx.Value(agentConversationIDKey{}).(string)
+	return v
+}
+
+// ToolCallInterceptor allows caller to gate or rewrite tool arguments just before execution.
+// Returning a non-nil error means the tool call is rejected and execution is skipped.
+type ToolCallInterceptor func(ctx context.Context, toolName string, args map[string]interface{}, toolCallID string) (map[string]interface{}, error)
+
+func WithToolCallInterceptor(ctx context.Context, fn ToolCallInterceptor) context.Context {
+	if fn == nil {
+		return ctx
+	}
+	return context.WithValue(ctx, toolCallInterceptorCtxKey{}, fn)
+}
+
 // NewAgent 创建新的Agent
 func NewAgent(cfg *config.OpenAIConfig, agentCfg *config.AgentConfig, mcpServer *mcp.Server, externalMCPMgr *mcp.ExternalMCPManager, logger *zap.Logger, maxIterations int) *Agent {
 	// 如果 maxIterations 为 0 或负数，使用默认值 30
@@ -348,7 +379,8 @@ func (a *Agent) EinoSingleAgentSystemInstruction() string {
 
 // AgentLoopWithProgress 执行Agent循环（带进度回调和对话ID）
 func (a *Agent) AgentLoopWithProgress(ctx context.Context, userInput string, historyMessages []ChatMessage, conversationID string, callback ProgressCallback, roleTools []string) (*AgentLoopResult, error) {
-	// 设置当前对话ID
+	ctx = withAgentConversationID(ctx, conversationID)
+	// 设置当前对话ID（兼容未走 context 的旧路径；并发会话应以 context 为准）
 	a.mu.Lock()
 	a.currentConversationID = conversationID
 	a.mu.Unlock()
@@ -653,22 +685,49 @@ func (a *Agent) AgentLoopWithProgress(ctx context.Context, userInput string, his
 					"iteration":    i + 1,
 				})
 
+				execArgs := toolCall.Function.Arguments
+				if interceptor, ok := ctx.Value(toolCallInterceptorCtxKey{}).(ToolCallInterceptor); ok && interceptor != nil {
+					newArgs, interceptErr := interceptor(ctx, toolCall.Function.Name, execArgs, toolCall.ID)
+					if interceptErr != nil {
+						errorMsg := fmt.Sprintf("工具调用被人工拒绝: %v", interceptErr)
+						messages = append(messages, ChatMessage{
+							Role:       "tool",
+							ToolCallID: toolCall.ID,
+							Content:    errorMsg,
+						})
+						sendProgress("tool_result", fmt.Sprintf("工具 %s 执行失败", toolCall.Function.Name), map[string]interface{}{
+							"toolName":   toolCall.Function.Name,
+							"success":    false,
+							"isError":    true,
+							"error":      errorMsg,
+							"toolCallId": toolCall.ID,
+							"index":      idx + 1,
+							"total":      len(choice.Message.ToolCalls),
+							"iteration":  i + 1,
+						})
+						continue
+					}
+					if newArgs != nil {
+						execArgs = newArgs
+					}
+				}
+
 				// 执行工具
 				toolCtx := context.WithValue(ctx, security.ToolOutputCallbackCtxKey, security.ToolOutputCallback(func(chunk string) {
 					if strings.TrimSpace(chunk) == "" {
 						return
 					}
 					sendProgress("tool_result_delta", chunk, map[string]interface{}{
-						"toolName":    toolCall.Function.Name,
-						"toolCallId":  toolCall.ID,
-						"index":       idx + 1,
-						"total":       len(choice.Message.ToolCalls),
-						"iteration":   i + 1,
+						"toolName":   toolCall.Function.Name,
+						"toolCallId": toolCall.ID,
+						"index":      idx + 1,
+						"total":      len(choice.Message.ToolCalls),
+						"iteration":  i + 1,
 						// success 在最终 tool_result 事件里会以 success/isError 标记为准
 					})
 				}))
 
-				execResult, err := a.executeToolViaMCP(toolCtx, toolCall.Function.Name, toolCall.Function.Arguments)
+				execResult, err := a.executeToolViaMCP(toolCtx, toolCall.Function.Name, execArgs)
 				if err != nil {
 					// 构建详细的错误信息，帮助AI理解问题并做出决策
 					errorMsg := a.formatToolError(toolCall.Function.Name, toolCall.Function.Arguments, err)
@@ -746,7 +805,7 @@ func (a *Agent) AgentLoopWithProgress(ctx context.Context, userInput string, his
 				// 流式调用OpenAI获取总结（不提供工具，强制AI直接回复）
 				sendProgress("response_start", "", map[string]interface{}{
 					"conversationId":     conversationID,
-					"mcpExecutionIds":   result.MCPExecutionIDs,
+					"mcpExecutionIds":    result.MCPExecutionIDs,
 					"messageGeneratedBy": "summary",
 				})
 				streamText, _ := a.callOpenAIStreamText(ctx, messages, []Tool{}, func(delta string) error {
@@ -793,7 +852,7 @@ func (a *Agent) AgentLoopWithProgress(ctx context.Context, userInput string, his
 			// 流式调用OpenAI获取总结（不提供工具，强制AI直接回复）
 			sendProgress("response_start", "", map[string]interface{}{
 				"conversationId":     conversationID,
-				"mcpExecutionIds":   result.MCPExecutionIDs,
+				"mcpExecutionIds":    result.MCPExecutionIDs,
 				"messageGeneratedBy": "summary",
 			})
 			streamText, _ := a.callOpenAIStreamText(ctx, messages, []Tool{}, func(delta string) error {
@@ -840,7 +899,7 @@ func (a *Agent) AgentLoopWithProgress(ctx context.Context, userInput string, his
 	// 流式调用OpenAI获取总结（不提供工具，强制AI直接回复）
 	sendProgress("response_start", "", map[string]interface{}{
 		"conversationId":     conversationID,
-		"mcpExecutionIds":   result.MCPExecutionIDs,
+		"mcpExecutionIds":    result.MCPExecutionIDs,
 		"messageGeneratedBy": "max_iter_summary",
 	})
 	streamText, _ := a.callOpenAIStreamText(ctx, messages, []Tool{}, func(delta string) error {
@@ -913,17 +972,13 @@ func (a *Agent) getAvailableTools(roleTools []string) []Tool {
 		defer cancel()
 
 		externalTools, err := a.externalMCPMgr.GetAllTools(ctx)
+		extMap := make(map[string]string)
 		if err != nil {
 			a.logger.Warn("获取外部MCP工具失败", zap.Error(err))
 		} else {
 			// 获取外部MCP配置，用于检查工具启用状态
 			externalMCPConfigs := a.externalMCPMgr.GetConfigs()
 
-			// 清空并重建工具名称映射
-			a.mu.Lock()
-			a.toolNameMapping = make(map[string]string)
-			a.mu.Unlock()
-
 			// 将外部MCP工具添加到工具列表（只添加启用的工具）
 			for _, externalTool := range externalTools {
 				// 外部工具使用 "mcpName::toolName" 作为toolKey
@@ -983,9 +1038,7 @@ func (a *Agent) getAvailableTools(roleTools []string) []Tool {
 				openAIName := strings.ReplaceAll(externalTool.Name, "::", "__")
 
 				// 保存名称映射关系（OpenAI格式 -> 原始格式）
-				a.mu.Lock()
-				a.toolNameMapping[openAIName] = externalTool.Name
-				a.mu.Unlock()
+				extMap[openAIName] = externalTool.Name
 
 				tools = append(tools, Tool{
 					Type: "function",
@@ -997,6 +1050,9 @@ func (a *Agent) getAvailableTools(roleTools []string) []Tool {
 				})
 			}
 		}
+		a.mu.Lock()
+		a.toolNameMapping = extMap
+		a.mu.Unlock()
 	}
 
 	a.logger.Debug("获取可用工具列表",
@@ -1390,9 +1446,12 @@ func (a *Agent) executeToolViaMCP(ctx context.Context, toolName string, args map
 
 	// 如果是record_vulnerability工具，自动添加conversation_id
 	if toolName == builtin.ToolRecordVulnerability {
-		a.mu.RLock()
-		conversationID := a.currentConversationID
-		a.mu.RUnlock()
+		conversationID := agentConversationIDFromContext(ctx)
+		if conversationID == "" {
+			a.mu.RLock()
+			conversationID = a.currentConversationID
+			a.mu.RUnlock()
+		}
 
 		if conversationID != "" {
 			args["conversation_id"] = conversationID

internal/agent/default_single_system_prompt.go

@@ -91,6 +91,20 @@ func DefaultSingleAgentSystemPrompt() string {
 
 当工具返回错误时，错误信息会包含在工具响应中，请仔细阅读并做出合理的决策。
 
+## 结束条件与停止约束
+
+- 在「未完成用户目标」前，不得输出纯计划/纯建议式结论并结束本轮；必须继续给出可执行下一步，并优先通过工具验证。
+- 若你准备结束回答，先执行一次自检：
+  1) 是否已有可验证证据支撑“任务完成/无法继续”的结论；
+  2) 是否至少尝试过当前路径的合理替代（参数、路径、方法、入口）；
+  3) 是否仍存在可执行且低成本的下一步验证动作。
+- 仅当满足以下任一条件时，才允许输出最终收尾：
+  1) 已达到用户目标并给出证据；
+  2) 达到明确边界（超时、权限、目标不可达、工具不可用且无替代），并清楚说明阻断点与已尝试项；
+  3) 用户明确要求停止。
+- 若最近一步得到 404/空结果/无效响应，不得直接结束；至少再进行一次“同目标不同策略”的验证（如变更路径、参数、请求方法、上下文来源）。
+- 避免无效空转：同一工具+同类参数连续失败 3 次后，必须切换策略（改工具、改入口、改假设）并说明切换原因。
+
 ## 漏洞记录
 
 发现有效漏洞时，必须使用 ` + builtin.ToolRecordVulnerability + ` 记录：标题、描述、严重程度、类型、目标、证明（POC）、影响、修复建议。

internal/app/app.go

@@ -326,6 +326,7 @@ func New(cfg *config.Config, log *logger.Logger) (*App, error) {
 	registerWebshellTools(mcpServer, db, webshellHandler, log.Logger)
 	registerWebshellManagementTools(mcpServer, db, webshellHandler, log.Logger)
 	configHandler := handler.NewConfigHandler(configPath, cfg, mcpServer, executor, agent, attackChainHandler, externalMCPMgr, log.Logger)
+	agentHandler.SetHitlToolWhitelistSaver(configHandler)
 	externalMCPHandler := handler.NewExternalMCPHandler(externalMCPMgr, cfg, configPath, log.Logger)
 	roleHandler := handler.NewRoleHandler(cfg, configPath, log.Logger)
 	skillsHandler := handler.NewSkillsHandler(cfg, configPath, log.Logger)
@@ -654,9 +655,16 @@ func setupRoutes(
 		// Eino ADK 单代理（ChatModelAgent + Runner；不依赖 multi_agent.enabled）
 		protected.POST("/eino-agent", agentHandler.EinoSingleAgentLoop)
 		protected.POST("/eino-agent/stream", agentHandler.EinoSingleAgentLoopStream)
+		protected.GET("/hitl/pending", agentHandler.ListHITLPending)
+		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.POST("/hitl/tool-whitelist", agentHandler.MergeHITLGlobalToolWhitelist)
 		// Agent Loop 取消与任务列表
 		protected.POST("/agent-loop/cancel", agentHandler.CancelAgentLoop)
 		protected.GET("/agent-loop/tasks", agentHandler.ListAgentTasks)
+		protected.GET("/agent-loop/task-events", agentHandler.SubscribeAgentTaskEvents)
 		protected.GET("/agent-loop/tasks/completed", agentHandler.ListCompletedTasks)
 
 		// Eino DeepAgent 多代理（与单 Agent 并存，需 config.multi_agent.enabled）
@@ -893,6 +901,8 @@ func setupRoutes(
 
 		// 漏洞管理
 		protected.GET("/vulnerabilities", vulnerabilityHandler.ListVulnerabilities)
+		protected.GET("/vulnerabilities/export", vulnerabilityHandler.ExportVulnerabilities)
+		protected.GET("/vulnerabilities/filter-options", vulnerabilityHandler.GetVulnerabilityFilterOptions)
 		protected.GET("/vulnerabilities/stats", vulnerabilityHandler.GetVulnerabilityStats)
 		protected.GET("/vulnerabilities/:id", vulnerabilityHandler.GetVulnerability)
 		protected.POST("/vulnerabilities", vulnerabilityHandler.CreateVulnerability)

internal/attackchain/builder.go

@@ -320,7 +320,7 @@ func (b *Builder) formatProcessDetailsForAttackChain(details []database.ProcessD
 		}
 
 		// 1) 编排器的工具调用/结果：保留（这是“主 agent 调了什么工具”）
-		if (d.EventType == "tool_call" || d.EventType == "tool_result" || d.EventType == "tool_calls_detected" || d.EventType == "iteration" || d.EventType == "eino_recovery") && einoRole == "orchestrator" {
+		if (d.EventType == "tool_call" || d.EventType == "tool_result" || d.EventType == "tool_calls_detected" || d.EventType == "iteration") && einoRole == "orchestrator" {
 			sb.WriteString("[")
 			sb.WriteString(d.EventType)
 			sb.WriteString("] ")

internal/config/config.go

@@ -22,6 +22,7 @@ type Config struct {
 	OpenAI      OpenAIConfig          `yaml:"openai"`
 	FOFA        FofaConfig            `yaml:"fofa,omitempty" json:"fofa,omitempty"`
 	Agent       AgentConfig           `yaml:"agent"`
+	Hitl        HitlConfig            `yaml:"hitl,omitempty" json:"hitl,omitempty"`
 	Security    SecurityConfig        `yaml:"security"`
 	Database    DatabaseConfig        `yaml:"database"`
 	Auth        AuthConfig            `yaml:"auth"`
@@ -37,24 +38,26 @@ type Config struct {
 
 // MultiAgentConfig 基于 CloudWeGo Eino adk/prebuilt 的多代理编排（deep | plan_execute | supervisor，与单 Agent /agent-loop 并存）。
 type MultiAgentConfig struct {
-	Enabled            bool   `yaml:"enabled" json:"enabled"`
-	DefaultMode        string `yaml:"default_mode" json:"default_mode"`                   // single | multi，供前端默认展示
-	RobotUseMultiAgent bool   `yaml:"robot_use_multi_agent" json:"robot_use_multi_agent"` // 为 true 时钉钉/飞书/企微机器人走 Eino 多代理
-	BatchUseMultiAgent bool   `yaml:"batch_use_multi_agent" json:"batch_use_multi_agent"` // 为 true 时批量任务队列中每子任务走 Eino 多代理
+	Enabled            bool `yaml:"enabled" json:"enabled"`
+	RobotUseMultiAgent bool `yaml:"robot_use_multi_agent" json:"robot_use_multi_agent"` // 为 true 时钉钉/飞书/企微机器人走 Eino 多代理
+	BatchUseMultiAgent bool `yaml:"batch_use_multi_agent" json:"batch_use_multi_agent"` // 为 true 时批量任务队列中每子任务走 Eino 多代理
 	// Orchestration 已弃用：保留仅兼容旧版 config.yaml；编排由聊天/WebShell 请求体 orchestration 决定，未传时按 deep。
 	Orchestration string `yaml:"orchestration,omitempty" json:"orchestration,omitempty"`
 	MaxIteration  int    `yaml:"max_iteration" json:"max_iteration"` // 主代理 / 执行器最大推理轮次（Deep、Supervisor、plan_execute 的 Executor）
 	// PlanExecuteLoopMaxIterations plan_execute 模式下 execute↔replan 外层循环上限；0 表示用 Eino 默认 10。
-	PlanExecuteLoopMaxIterations int                   `yaml:"plan_execute_loop_max_iterations,omitempty" json:"plan_execute_loop_max_iterations,omitempty"`
-	SubAgentMaxIterations        int                   `yaml:"sub_agent_max_iterations" json:"sub_agent_max_iterations"`
-	WithoutGeneralSubAgent       bool                  `yaml:"without_general_sub_agent" json:"without_general_sub_agent"`
-	WithoutWriteTodos            bool                  `yaml:"without_write_todos" json:"without_write_todos"`
-	OrchestratorInstruction      string                `yaml:"orchestrator_instruction" json:"orchestrator_instruction"`
+	PlanExecuteLoopMaxIterations int    `yaml:"plan_execute_loop_max_iterations,omitempty" json:"plan_execute_loop_max_iterations,omitempty"`
+	SubAgentMaxIterations        int    `yaml:"sub_agent_max_iterations" json:"sub_agent_max_iterations"`
+	WithoutGeneralSubAgent       bool   `yaml:"without_general_sub_agent" json:"without_general_sub_agent"`
+	WithoutWriteTodos            bool   `yaml:"without_write_todos" json:"without_write_todos"`
+	OrchestratorInstruction      string `yaml:"orchestrator_instruction" json:"orchestrator_instruction"`
 	// OrchestratorInstructionPlanExecute plan_execute 主代理（规划侧）系统提示；非空且 agents/orchestrator-plan-execute.md 正文为空或未存在时生效。不与 Deep 的 orchestrator_instruction 混用。
 	OrchestratorInstructionPlanExecute string `yaml:"orchestrator_instruction_plan_execute,omitempty" json:"orchestrator_instruction_plan_execute,omitempty"`
 	// 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"`
+	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.
+	// 0 (default) uses the built-in default of 2000 runes; negative value disables injection entirely.
+	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"`
 	// EinoMiddleware wires optional ADK middleware (patchtoolcalls, toolsearch, plantask, reduction) and Deep extras.
@@ -74,10 +77,10 @@ type MultiAgentEinoMiddlewareConfig struct {
 	// PlantaskRelDir relative to skills_dir for per-conversation task boards (default .eino/plantask).
 	PlantaskRelDir string `yaml:"plantask_rel_dir,omitempty" json:"plantask_rel_dir,omitempty"`
 	// Reduction truncates/offloads large tool outputs (requires eino local backend for Write).
-	ReductionEnable           bool     `yaml:"reduction_enable,omitempty" json:"reduction_enable,omitempty"`
-	ReductionRootDir          string   `yaml:"reduction_root_dir,omitempty" json:"reduction_root_dir,omitempty"` // default: os temp + conversation id
-	ReductionClearExclude     []string `yaml:"reduction_clear_exclude,omitempty" json:"reduction_clear_exclude,omitempty"`
-	ReductionSubAgents        bool     `yaml:"reduction_sub_agents,omitempty" json:"reduction_sub_agents,omitempty"` // also attach to sub-agents
+	ReductionEnable       bool     `yaml:"reduction_enable,omitempty" json:"reduction_enable,omitempty"`
+	ReductionRootDir      string   `yaml:"reduction_root_dir,omitempty" json:"reduction_root_dir,omitempty"` // default: os temp + conversation id
+	ReductionClearExclude []string `yaml:"reduction_clear_exclude,omitempty" json:"reduction_clear_exclude,omitempty"`
+	ReductionSubAgents    bool     `yaml:"reduction_sub_agents,omitempty" json:"reduction_sub_agents,omitempty"` // also attach to sub-agents
 	// CheckpointDir when non-empty enables adk.Runner CheckPointStore (file-backed) for interrupt/resume persistence.
 	CheckpointDir string `yaml:"checkpoint_dir,omitempty" json:"checkpoint_dir,omitempty"`
 	// DeepOutputKey passed to deep.Config OutputKey (session final text); empty = off.
@@ -129,7 +132,6 @@ type MultiAgentSubConfig struct {
 // MultiAgentPublic 返回给前端的精简信息（不含子代理指令全文）。
 type MultiAgentPublic struct {
 	Enabled                      bool   `json:"enabled"`
-	DefaultMode                  string `json:"default_mode"`
 	RobotUseMultiAgent           bool   `json:"robot_use_multi_agent"`
 	BatchUseMultiAgent           bool   `json:"batch_use_multi_agent"`
 	SubAgentCount                int    `json:"sub_agent_count"`
@@ -152,11 +154,10 @@ func NormalizeMultiAgentOrchestration(s string) string {
 
 // MultiAgentAPIUpdate 设置页/API 仅更新多代理标量字段；写入 YAML 时不覆盖 sub_agents 等块。
 type MultiAgentAPIUpdate struct {
-	Enabled                      bool   `json:"enabled"`
-	DefaultMode                  string `json:"default_mode"`
-	RobotUseMultiAgent           bool   `json:"robot_use_multi_agent"`
-	BatchUseMultiAgent           bool   `json:"batch_use_multi_agent"`
-	PlanExecuteLoopMaxIterations *int   `json:"plan_execute_loop_max_iterations,omitempty"`
+	Enabled                      bool `json:"enabled"`
+	RobotUseMultiAgent           bool `json:"robot_use_multi_agent"`
+	BatchUseMultiAgent           bool `json:"batch_use_multi_agent"`
+	PlanExecuteLoopMaxIterations *int `json:"plan_execute_loop_max_iterations,omitempty"`
 }
 
 // RobotsConfig 机器人配置（企业微信、钉钉、飞书等）
@@ -244,6 +245,13 @@ type AgentConfig struct {
 	SystemPromptPath string `yaml:"system_prompt_path,omitempty" json:"system_prompt_path,omitempty"`
 }
 
+// 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。
+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 []string `yaml:"tool_whitelist,omitempty" json:"tool_whitelist,omitempty"`
+}
+
 type AuthConfig struct {
 	Password                    string `yaml:"password" json:"password"`
 	SessionDurationHours        int    `yaml:"session_duration_hours" json:"session_duration_hours"`
@@ -950,10 +958,10 @@ type RolesConfig struct {
 
 // RoleConfig 单个角色配置
 type RoleConfig struct {
-	Name        string   `yaml:"name" json:"name"`                         // 角色名称
-	Description string   `yaml:"description" json:"description"`           // 角色描述
-	UserPrompt  string   `yaml:"user_prompt" json:"user_prompt"`           // 用户提示词(追加到用户消息前)
-	Icon        string   `yaml:"icon,omitempty" json:"icon,omitempty"`     // 角色图标（可选）
+	Name        string   `yaml:"name" json:"name"`                       // 角色名称
+	Description string   `yaml:"description" json:"description"`         // 角色描述
+	UserPrompt  string   `yaml:"user_prompt" json:"user_prompt"`         // 用户提示词(追加到用户消息前)
+	Icon        string   `yaml:"icon,omitempty" json:"icon,omitempty"`   // 角色图标（可选）
 	Tools       []string `yaml:"tools,omitempty" json:"tools,omitempty"` // 关联的工具列表（toolKey格式，如 "toolName" 或 "mcpName::toolName"）
 	MCPs        []string `yaml:"mcps,omitempty" json:"mcps,omitempty"`   // 向后兼容：关联的MCP服务器列表（已废弃，使用tools替代）
 	Enabled     bool     `yaml:"enabled" json:"enabled"`                 // 是否启用

internal/database/database.go

@@ -197,6 +197,8 @@ func (db *DB) initTables() error {
 	CREATE TABLE IF NOT EXISTS vulnerabilities (
 		id TEXT PRIMARY KEY,
 		conversation_id TEXT NOT NULL,
+		conversation_tag TEXT,
+		task_tag TEXT,
 		title TEXT NOT NULL,
 		description TEXT,
 		severity TEXT NOT NULL,
@@ -289,6 +291,8 @@ func (db *DB) initTables() error {
 	CREATE INDEX IF NOT EXISTS idx_conversation_group_mappings_group ON conversation_group_mappings(group_id);
 	CREATE INDEX IF NOT EXISTS idx_conversations_pinned ON conversations(pinned);
 	CREATE INDEX IF NOT EXISTS idx_vulnerabilities_conversation_id ON vulnerabilities(conversation_id);
+	CREATE INDEX IF NOT EXISTS idx_vulnerabilities_conversation_tag ON vulnerabilities(conversation_tag);
+	CREATE INDEX IF NOT EXISTS idx_vulnerabilities_task_tag ON vulnerabilities(task_tag);
 	CREATE INDEX IF NOT EXISTS idx_vulnerabilities_severity ON vulnerabilities(severity);
 	CREATE INDEX IF NOT EXISTS idx_vulnerabilities_status ON vulnerabilities(status);
 	CREATE INDEX IF NOT EXISTS idx_vulnerabilities_created_at ON vulnerabilities(created_at);
@@ -383,6 +387,10 @@ func (db *DB) initTables() error {
 		db.logger.Warn("迁移batch_task_queues表失败", zap.Error(err))
 		// 不返回错误，允许继续运行
 	}
+	if err := db.migrateVulnerabilitiesTable(); err != nil {
+		db.logger.Warn("迁移vulnerabilities表失败", zap.Error(err))
+		// 不返回错误，允许继续运行
+	}
 
 	if _, err := db.Exec(createIndexes); err != nil {
 		return fmt.Errorf("创建索引失败: %w", err)
@@ -683,6 +691,37 @@ func (db *DB) migrateBatchTaskQueuesTable() error {
 	return nil
 }
 
+// migrateVulnerabilitiesTable 迁移 vulnerabilities 表，补充标签字段
+func (db *DB) migrateVulnerabilitiesTable() error {
+	columns := []struct {
+		name string
+		stmt string
+	}{
+		{name: "conversation_tag", stmt: "ALTER TABLE vulnerabilities ADD COLUMN conversation_tag TEXT"},
+		{name: "task_tag", stmt: "ALTER TABLE vulnerabilities ADD COLUMN task_tag TEXT"},
+	}
+
+	for _, col := range columns {
+		var count int
+		err := db.QueryRow("SELECT COUNT(*) FROM pragma_table_info('vulnerabilities') WHERE name=?", col.name).Scan(&count)
+		if err != nil {
+			if _, addErr := db.Exec(col.stmt); addErr != nil {
+				errMsg := strings.ToLower(addErr.Error())
+				if !strings.Contains(errMsg, "duplicate column") && !strings.Contains(errMsg, "already exists") {
+					db.logger.Warn("添加vulnerabilities字段失败", zap.String("field", col.name), zap.Error(addErr))
+				}
+			}
+			continue
+		}
+		if count == 0 {
+			if _, addErr := db.Exec(col.stmt); addErr != nil {
+				db.logger.Warn("添加vulnerabilities字段失败", zap.String("field", col.name), zap.Error(addErr))
+			}
+		}
+	}
+	return nil
+}
+
 // NewKnowledgeDB 创建知识库数据库连接（只包含知识库相关的表）
 func NewKnowledgeDB(dbPath string, logger *zap.Logger) (*DB, error) {
 	sqlDB, err := sql.Open("sqlite3", dbPath+"?_journal_mode=WAL&_foreign_keys=1&_busy_timeout=5000&_synchronous=NORMAL")

internal/database/vulnerability.go

@@ -13,6 +13,10 @@ import (
 type Vulnerability struct {
 	ID              string    `json:"id"`
 	ConversationID string    `json:"conversation_id"`
+	ConversationTag string    `json:"conversation_tag,omitempty"`
+	TaskTag         string    `json:"task_tag,omitempty"`
+	TaskID          string    `json:"task_id,omitempty"`
+	TaskQueueID     string    `json:"task_queue_id,omitempty"`
 	Title           string    `json:"title"`
 	Description     string    `json:"description"`
 	Severity        string    `json:"severity"` // critical, high, medium, low, info
@@ -42,15 +46,15 @@ func (db *DB) CreateVulnerability(vuln *Vulnerability) (*Vulnerability, error) {
 
 	query := `
 		INSERT INTO vulnerabilities (
-			id, conversation_id, title, description, severity, status,
+			id, conversation_id, conversation_tag, task_tag, title, description, severity, status,
 			vulnerability_type, target, proof, impact, recommendation,
 			created_at, updated_at
-		) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+		) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
 	`
 
 	_, err := db.Exec(
 		query,
-		vuln.ID, vuln.ConversationID, vuln.Title, vuln.Description,
+		vuln.ID, vuln.ConversationID, vuln.ConversationTag, vuln.TaskTag, vuln.Title, vuln.Description,
 		vuln.Severity, vuln.Status, vuln.Type, vuln.Target,
 		vuln.Proof, vuln.Impact, vuln.Recommendation,
 		vuln.CreatedAt, vuln.UpdatedAt,
@@ -67,7 +71,9 @@ func (db *DB) GetVulnerability(id string) (*Vulnerability, error) {
 	var vuln Vulnerability
 	query := `
 		SELECT id, conversation_id, title, description, severity, status,
-		       vulnerability_type, target, proof, impact, recommendation,
+		       conversation_tag, task_tag, vulnerability_type, target, proof, impact, recommendation,
+		       COALESCE((SELECT bt.id FROM batch_tasks bt WHERE bt.conversation_id = vulnerabilities.conversation_id LIMIT 1), '') AS task_id,
+		       COALESCE((SELECT bt.queue_id FROM batch_tasks bt WHERE bt.conversation_id = vulnerabilities.conversation_id LIMIT 1), '') AS task_queue_id,
 		       created_at, updated_at
 		FROM vulnerabilities
 		WHERE id = ?
@@ -75,8 +81,9 @@ func (db *DB) GetVulnerability(id string) (*Vulnerability, error) {
 
 	err := db.QueryRow(query, id).Scan(
 		&vuln.ID, &vuln.ConversationID, &vuln.Title, &vuln.Description,
-		&vuln.Severity, &vuln.Status, &vuln.Type, &vuln.Target,
+		&vuln.Severity, &vuln.Status, &vuln.ConversationTag, &vuln.TaskTag, &vuln.Type, &vuln.Target,
 		&vuln.Proof, &vuln.Impact, &vuln.Recommendation,
+		&vuln.TaskID, &vuln.TaskQueueID,
 		&vuln.CreatedAt, &vuln.UpdatedAt,
 	)
 	if err != nil {
@@ -90,10 +97,12 @@ func (db *DB) GetVulnerability(id string) (*Vulnerability, error) {
 }
 
 // ListVulnerabilities 列出漏洞
-func (db *DB) ListVulnerabilities(limit, offset int, id, conversationID, severity, status string) ([]*Vulnerability, error) {
+func (db *DB) ListVulnerabilities(limit, offset int, id, conversationID, severity, status, taskID, conversationTag, taskTag string) ([]*Vulnerability, error) {
 	query := `
-		SELECT id, conversation_id, title, description, severity, status,
+		SELECT id, conversation_id, title, description, severity, status, conversation_tag, task_tag,
 		       vulnerability_type, target, proof, impact, recommendation,
+		       COALESCE((SELECT bt.id FROM batch_tasks bt WHERE bt.conversation_id = vulnerabilities.conversation_id LIMIT 1), '') AS task_id,
+		       COALESCE((SELECT bt.queue_id FROM batch_tasks bt WHERE bt.conversation_id = vulnerabilities.conversation_id LIMIT 1), '') AS task_queue_id,
 		       created_at, updated_at
 		FROM vulnerabilities
 		WHERE 1=1
@@ -108,6 +117,18 @@ func (db *DB) ListVulnerabilities(limit, offset int, id, conversationID, severit
 		query += " AND conversation_id = ?"
 		args = append(args, conversationID)
 	}
+	if taskID != "" {
+		query += " AND EXISTS (SELECT 1 FROM batch_tasks bt WHERE bt.conversation_id = vulnerabilities.conversation_id AND (bt.id = ? OR bt.queue_id = ?))"
+		args = append(args, taskID, taskID)
+	}
+	if conversationTag != "" {
+		query += " AND conversation_tag = ?"
+		args = append(args, conversationTag)
+	}
+	if taskTag != "" {
+		query += " AND task_tag = ?"
+		args = append(args, taskTag)
+	}
 	if severity != "" {
 		query += " AND severity = ?"
 		args = append(args, severity)
@@ -131,8 +152,9 @@ func (db *DB) ListVulnerabilities(limit, offset int, id, conversationID, severit
 		var vuln Vulnerability
 		err := rows.Scan(
 			&vuln.ID, &vuln.ConversationID, &vuln.Title, &vuln.Description,
-			&vuln.Severity, &vuln.Status, &vuln.Type, &vuln.Target,
+			&vuln.Severity, &vuln.Status, &vuln.ConversationTag, &vuln.TaskTag, &vuln.Type, &vuln.Target,
 			&vuln.Proof, &vuln.Impact, &vuln.Recommendation,
+			&vuln.TaskID, &vuln.TaskQueueID,
 			&vuln.CreatedAt, &vuln.UpdatedAt,
 		)
 		if err != nil {
@@ -146,7 +168,7 @@ func (db *DB) ListVulnerabilities(limit, offset int, id, conversationID, severit
 }
 
 // CountVulnerabilities 统计漏洞总数（支持筛选条件）
-func (db *DB) CountVulnerabilities(id, conversationID, severity, status string) (int, error) {
+func (db *DB) CountVulnerabilities(id, conversationID, severity, status, taskID, conversationTag, taskTag string) (int, error) {
 	query := "SELECT COUNT(*) FROM vulnerabilities WHERE 1=1"
 	args := []interface{}{}
 
@@ -158,6 +180,18 @@ func (db *DB) CountVulnerabilities(id, conversationID, severity, status string)
 		query += " AND conversation_id = ?"
 		args = append(args, conversationID)
 	}
+	if taskID != "" {
+		query += " AND EXISTS (SELECT 1 FROM batch_tasks bt WHERE bt.conversation_id = vulnerabilities.conversation_id AND (bt.id = ? OR bt.queue_id = ?))"
+		args = append(args, taskID, taskID)
+	}
+	if conversationTag != "" {
+		query += " AND conversation_tag = ?"
+		args = append(args, conversationTag)
+	}
+	if taskTag != "" {
+		query += " AND task_tag = ?"
+		args = append(args, taskTag)
+	}
 	if severity != "" {
 		query += " AND severity = ?"
 		args = append(args, severity)
@@ -182,7 +216,7 @@ func (db *DB) UpdateVulnerability(id string, vuln *Vulnerability) error {
 
 	query := `
 		UPDATE vulnerabilities
-		SET title = ?, description = ?, severity = ?, status = ?,
+		SET conversation_tag = ?, task_tag = ?, title = ?, description = ?, severity = ?, status = ?,
 		    vulnerability_type = ?, target = ?, proof = ?, impact = ?,
 		    recommendation = ?, updated_at = ?
 		WHERE id = ?
@@ -190,7 +224,7 @@ func (db *DB) UpdateVulnerability(id string, vuln *Vulnerability) error {
 
 	_, err := db.Exec(
 		query,
-		vuln.Title, vuln.Description, vuln.Severity, vuln.Status,
+		vuln.ConversationTag, vuln.TaskTag, vuln.Title, vuln.Description, vuln.Severity, vuln.Status,
 		vuln.Type, vuln.Target, vuln.Proof, vuln.Impact,
 		vuln.Recommendation, vuln.UpdatedAt, id,
 	)
@@ -210,18 +244,24 @@ func (db *DB) DeleteVulnerability(id string) error {
 	return nil
 }
 
-// GetVulnerabilityStats 获取漏洞统计
-func (db *DB) GetVulnerabilityStats(conversationID string) (map[string]interface{}, error) {
+// GetVulnerabilityStats 获取漏洞统计（筛选条件与 ListVulnerabilities / CountVulnerabilities 一致）
+func (db *DB) GetVulnerabilityStats(conversationID, taskID string) (map[string]interface{}, error) {
 	stats := make(map[string]interface{})
 
+	where := "WHERE 1=1"
+	args := []interface{}{}
+	if conversationID != "" {
+		where += " AND conversation_id = ?"
+		args = append(args, conversationID)
+	}
+	if taskID != "" {
+		where += " AND EXISTS (SELECT 1 FROM batch_tasks bt WHERE bt.conversation_id = vulnerabilities.conversation_id AND (bt.id = ? OR bt.queue_id = ?))"
+		args = append(args, taskID, taskID)
+	}
+
 	// 总漏洞数
 	var totalCount int
-	query := "SELECT COUNT(*) FROM vulnerabilities"
-	args := []interface{}{}
-	if conversationID != "" {
-		query += " WHERE conversation_id = ?"
-		args = append(args, conversationID)
-	}
+	query := "SELECT COUNT(*) FROM vulnerabilities " + where
 	err := db.QueryRow(query, args...).Scan(&totalCount)
 	if err != nil {
 		return nil, fmt.Errorf("获取总漏洞数失败: %w", err)
@@ -229,11 +269,7 @@ func (db *DB) GetVulnerabilityStats(conversationID string) (map[string]interface
 	stats["total"] = totalCount
 
 	// 按严重程度统计
-	severityQuery := "SELECT severity, COUNT(*) FROM vulnerabilities"
-	if conversationID != "" {
-		severityQuery += " WHERE conversation_id = ?"
-	}
-	severityQuery += " GROUP BY severity"
+	severityQuery := "SELECT severity, COUNT(*) FROM vulnerabilities " + where + " GROUP BY severity"
 
 	rows, err := db.Query(severityQuery, args...)
 	if err != nil {
@@ -253,11 +289,7 @@ func (db *DB) GetVulnerabilityStats(conversationID string) (map[string]interface
 	stats["by_severity"] = severityStats
 
 	// 按状态统计
-	statusQuery := "SELECT status, COUNT(*) FROM vulnerabilities"
-	if conversationID != "" {
-		statusQuery += " WHERE conversation_id = ?"
-	}
-	statusQuery += " GROUP BY status"
+	statusQuery := "SELECT status, COUNT(*) FROM vulnerabilities " + where + " GROUP BY status"
 
 	rows, err = db.Query(statusQuery, args...)
 	if err != nil {
@@ -279,3 +311,60 @@ func (db *DB) GetVulnerabilityStats(conversationID string) (map[string]interface
 	return stats, nil
 }
 
+// GetVulnerabilityFilterOptions 获取漏洞筛选建议项
+func (db *DB) GetVulnerabilityFilterOptions() (map[string][]string, error) {
+	collect := func(query string, args ...interface{}) ([]string, error) {
+		rows, err := db.Query(query, args...)
+		if err != nil {
+			return nil, err
+		}
+		defer rows.Close()
+		items := make([]string, 0)
+		for rows.Next() {
+			var val string
+			if err := rows.Scan(&val); err != nil {
+				continue
+			}
+			if val == "" {
+				continue
+			}
+			items = append(items, val)
+		}
+		return items, nil
+	}
+
+	vulnIDs, err := collect(`SELECT DISTINCT id FROM vulnerabilities ORDER BY created_at DESC LIMIT 500`)
+	if err != nil {
+		return nil, fmt.Errorf("查询漏洞ID建议失败: %w", err)
+	}
+	conversationIDs, err := collect(`SELECT DISTINCT conversation_id FROM vulnerabilities WHERE conversation_id <> '' ORDER BY created_at DESC LIMIT 500`)
+	if err != nil {
+		return nil, fmt.Errorf("查询会话ID建议失败: %w", err)
+	}
+	taskIDs, err := collect(`SELECT DISTINCT id FROM batch_tasks WHERE id <> '' ORDER BY rowid DESC LIMIT 500`)
+	if err != nil {
+		return nil, fmt.Errorf("查询任务ID建议失败: %w", err)
+	}
+	queueIDs, err := collect(`SELECT DISTINCT queue_id FROM batch_tasks WHERE queue_id <> '' ORDER BY rowid DESC LIMIT 500`)
+	if err != nil {
+		return nil, fmt.Errorf("查询队列ID建议失败: %w", err)
+	}
+	conversationTags, err := collect(`SELECT DISTINCT conversation_tag FROM vulnerabilities WHERE conversation_tag IS NOT NULL AND conversation_tag <> '' ORDER BY conversation_tag LIMIT 500`)
+	if err != nil {
+		return nil, fmt.Errorf("查询对话标签建议失败: %w", err)
+	}
+	taskTags, err := collect(`SELECT DISTINCT task_tag FROM vulnerabilities WHERE task_tag IS NOT NULL AND task_tag <> '' ORDER BY task_tag LIMIT 500`)
+	if err != nil {
+		return nil, fmt.Errorf("查询任务标签建议失败: %w", err)
+	}
+
+	return map[string][]string{
+		"vulnerability_ids": vulnIDs,
+		"conversation_ids":  conversationIDs,
+		"task_ids":          taskIDs,
+		"queue_ids":         queueIDs,
+		"conversation_tags": conversationTags,
+		"task_tags":         taskTags,
+	}, nil
+}
+

internal/einomcp/mcp_tools.go

@@ -160,17 +160,17 @@ func runMCPToolInvocation(
 }
 
 // UnknownToolReminderHandler 供 compose.ToolsNodeConfig.UnknownToolsHandler 使用：
-// 模型请求了未注册的工具名时，返回一个「可恢复」的错误，让上层 runner 触发重试与纠错提示，
-// 同时避免 UI 永远停留在“执行中”（runner 会在 recoverable 分支 flush 掉 pending 的 tool_call）。
+// 模型请求了未注册的工具名时，返回一个「软错误」工具结果（nil error），
+// 让模型在同一轮继续自我修正，避免触发 run-loop 级别的 full rerun。
 // 不进行名称猜测或映射，避免误执行。
 func UnknownToolReminderHandler() func(ctx context.Context, name, input string) (string, error) {
 	return func(ctx context.Context, name, input string) (string, error) {
 		_ = ctx
 		_ = input
 		requested := strings.TrimSpace(name)
-		// Return a recoverable error that still carries a friendly, bilingual hint.
-		// This will be caught by multiagent runner as "tool not found" and trigger a retry.
-		return "", fmt.Errorf("tool %q not found: %s", requested, unknownToolReminderText(requested))
+		// Return a soft tool-result error so the graph keeps running and the LLM
+		// can correct tool name/arguments within the same run.
+		return ToolErrorPrefix + unknownToolReminderText(requested), nil
 	}
 }
 

internal/handler/agent.go

@@ -115,7 +115,9 @@ type AgentHandler struct {
 	db               *database.DB
 	logger           *zap.Logger
 	tasks            *AgentTaskManager
+	taskEventBus     *TaskEventBus // 镜像 SSE 事件，供刷新后订阅同一运行中任务
 	batchTaskManager *BatchTaskManager
+	hitlManager      *HITLManager
 	config           *config.Config // 配置引用，用于获取角色信息
 	knowledgeManager interface {    // 知识库管理器接口
 		LogRetrieval(conversationID, messageID, query, riskType string, retrievedItems []string) error
@@ -124,6 +126,13 @@ type AgentHandler struct {
 	batchCronParser   cron.Parser
 	batchRunnerMu     sync.Mutex
 	batchRunning      map[string]struct{}
+	// 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
+}
+
+// HitlToolWhitelistSaver 合并 HITL 免审批工具到全局配置并落盘
+type HitlToolWhitelistSaver interface {
+	MergeHitlToolWhitelistIntoConfig(add []string) error
 }
 
 // NewAgentHandler 创建新的Agent处理器
@@ -136,16 +145,24 @@ func NewAgentHandler(agent *agent.Agent, db *database.DB, cfg *config.Config, lo
 		logger.Warn("从数据库加载批量任务队列失败", zap.Error(err))
 	}
 
+	bus := NewTaskEventBus()
+	tm := NewAgentTaskManager()
+	tm.SetTaskEventBus(bus)
 	handler := &AgentHandler{
 		agent:            agent,
 		db:               db,
 		logger:           logger,
-		tasks:            NewAgentTaskManager(),
+		tasks:            tm,
+		taskEventBus:     bus,
 		batchTaskManager: batchTaskManager,
 		config:           cfg,
+		hitlManager:      NewHITLManager(db, logger),
 		batchCronParser:  cron.NewParser(cron.Minute | cron.Hour | cron.Dom | cron.Month | cron.Dow | cron.Descriptor),
 		batchRunning:     make(map[string]struct{}),
 	}
+	if err := handler.hitlManager.EnsureSchema(); err != nil {
+		logger.Warn("初始化 HITL 表失败", zap.Error(err))
+	}
 	go handler.batchQueueSchedulerLoop()
 	return handler
 }
@@ -162,6 +179,11 @@ func (h *AgentHandler) SetAgentsMarkdownDir(absDir string) {
 	h.agentsMarkdownDir = strings.TrimSpace(absDir)
 }
 
+// SetHitlToolWhitelistSaver 设置 HITL 白名单落盘（与 ConfigHandler 配合，避免循环引用用接口）
+func (h *AgentHandler) SetHitlToolWhitelistSaver(s HitlToolWhitelistSaver) {
+	h.hitlWhitelistSaver = s
+}
+
 // ChatAttachment 聊天附件（用户上传的文件）
 type ChatAttachment struct {
 	FileName   string `json:"fileName"`          // 展示用文件名
@@ -177,10 +199,18 @@ type ChatRequest struct {
 	Role                 string           `json:"role,omitempty"` // 角色名称
 	Attachments          []ChatAttachment `json:"attachments,omitempty"`
 	WebShellConnectionID string           `json:"webshellConnectionId,omitempty"` // WebShell 管理 - AI 助手：当前选中的连接 ID，仅使用 webshell_* 工具
+	Hitl                 *HITLRequest     `json:"hitl,omitempty"`
 	// Orchestration 仅对 /api/multi-agent、/api/multi-agent/stream：deep | plan_execute | supervisor；空则等同 deep。机器人/批量等无请求体时由服务端默认 deep。/api/eino-agent* 不使用此字段。
 	Orchestration string `json:"orchestration,omitempty"`
 }
 
+type HITLRequest struct {
+	Enabled        bool     `json:"enabled"`
+	Mode           string   `json:"mode,omitempty"`
+	SensitiveTools []string `json:"sensitiveTools,omitempty"`
+	TimeoutSeconds int      `json:"timeoutSeconds,omitempty"`
+}
+
 const (
 	maxAttachments     = 10
 	chatUploadsDirName = "chat_uploads" // 对话附件保存的根目录（相对当前工作目录）
@@ -462,6 +492,11 @@ func (h *AgentHandler) AgentLoop(c *gin.Context) {
 		}
 	}
 
+	h.activateHITLForConversation(conversationID, req.Hitl)
+	if h.hitlManager != nil {
+		defer h.hitlManager.DeactivateConversation(conversationID)
+	}
+
 	// 优先尝试从保存的ReAct数据恢复历史上下文
 	agentHistoryMessages, err := h.loadHistoryFromReActData(conversationID)
 	if err != nil {
@@ -566,8 +601,15 @@ func (h *AgentHandler) AgentLoop(c *gin.Context) {
 		return
 	}
 
+	baseCtx, cancelWithCause := context.WithCancelCause(c.Request.Context())
+	defer cancelWithCause(nil)
+	taskCtx, timeoutCancel := context.WithTimeout(baseCtx, 600*time.Minute)
+	defer timeoutCancel()
+	progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, conversationID, "", nil)
+	taskCtx = h.injectReactHITLInterceptor(taskCtx, cancelWithCause, conversationID, "", nil)
+
 	// 执行Agent Loop，传入历史消息和对话ID（使用包含角色提示词的finalMessage和角色工具列表）
-	result, err := h.agent.AgentLoopWithProgress(c.Request.Context(), finalMessage, agentHistoryMessages, conversationID, nil, roleTools)
+	result, err := h.agent.AgentLoopWithProgress(taskCtx, finalMessage, agentHistoryMessages, conversationID, progressCallback, roleTools)
 	if err != nil {
 		h.logger.Error("Agent Loop执行失败", zap.Error(err))
 
@@ -661,7 +703,7 @@ func (h *AgentHandler) ProcessMessageForRobot(ctx context.Context, conversationI
 	if assistantMsg != nil {
 		assistantMessageID = assistantMsg.ID
 	}
-	progressCallback := h.createProgressCallback(conversationID, assistantMessageID, nil)
+	progressCallback := h.createProgressCallback(ctx, nil, conversationID, assistantMessageID, nil)
 
 	useRobotMulti := h.config != nil && h.config.MultiAgent.Enabled && h.config.MultiAgent.RobotUseMultiAgent
 	if useRobotMulti {
@@ -755,9 +797,41 @@ type StreamEvent struct {
 
 // createProgressCallback 创建进度回调函数，用于保存processDetails
 // sendEventFunc: 可选的流式事件发送函数，如果为nil则不发送流式事件
-func (h *AgentHandler) createProgressCallback(conversationID, assistantMessageID string, sendEventFunc func(eventType, message string, data interface{})) agent.ProgressCallback {
+func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun context.CancelCauseFunc, conversationID, assistantMessageID string, sendEventFunc func(eventType, message string, data interface{})) agent.ProgressCallback {
 	// 用于保存tool_call事件中的参数，以便在tool_result时使用
 	toolCallCache := make(map[string]map[string]interface{}) // toolCallId -> arguments
+	skillCallCache := make(map[string]string)                // toolCallId -> skillName
+	skillToolName := "skill"
+	if h.config != nil {
+		if customName := strings.TrimSpace(h.config.MultiAgent.EinoSkills.SkillToolName); customName != "" {
+			skillToolName = customName
+		}
+	}
+
+	extractSkillName := func(args map[string]interface{}) string {
+		if len(args) == 0 {
+			return ""
+		}
+		for _, key := range []string{"skill_name", "skillName", "name", "skill", "id", "skill_id", "skillId"} {
+			if v, ok := args[key]; ok {
+				switch vv := v.(type) {
+				case string:
+					if s := strings.TrimSpace(vv); s != "" {
+						return s
+					}
+				case map[string]interface{}:
+					for _, nestedKey := range []string{"name", "id", "skill_name", "skillId"} {
+						if nestedV, nestedOK := vv[nestedKey].(string); nestedOK {
+							if s := strings.TrimSpace(nestedV); s != "" {
+								return s
+							}
+						}
+					}
+				}
+			}
+		}
+		return ""
+	}
 
 	// thinking_stream_*：不逐条落库，按 streamId 聚合，在后续关键事件前补一条可持久化的 thinking
 	type thinkingBuf struct {
@@ -840,6 +914,16 @@ func (h *AgentHandler) createProgressCallback(conversationID, assistantMessageID
 						}
 					}
 				}
+				if strings.EqualFold(strings.TrimSpace(toolName), skillToolName) {
+					toolCallID, _ := dataMap["toolCallId"].(string)
+					if toolCallID != "" {
+						if argumentsObj, ok := dataMap["argumentsObj"].(map[string]interface{}); ok {
+							if skillName := extractSkillName(argumentsObj); skillName != "" {
+								skillCallCache[toolCallID] = skillName
+							}
+						}
+					}
+				}
 			}
 		}
 
@@ -953,6 +1037,45 @@ func (h *AgentHandler) createProgressCallback(conversationID, assistantMessageID
 			}
 		}
 
+		// 记录 skills 调用统计（tool_call + tool_result 关联）
+		if eventType == "tool_result" && h.db != nil {
+			if dataMap, ok := data.(map[string]interface{}); ok {
+				toolName, _ := dataMap["toolName"].(string)
+				if strings.EqualFold(strings.TrimSpace(toolName), skillToolName) {
+					toolCallID, _ := dataMap["toolCallId"].(string)
+					skillName := ""
+					if toolCallID != "" {
+						skillName = strings.TrimSpace(skillCallCache[toolCallID])
+						delete(skillCallCache, toolCallID)
+					}
+					if skillName == "" {
+						if argumentsObj, ok := dataMap["argumentsObj"].(map[string]interface{}); ok {
+							skillName = strings.TrimSpace(extractSkillName(argumentsObj))
+						}
+					}
+					if skillName != "" {
+						success, ok := dataMap["success"].(bool)
+						if !ok {
+							if isError, okErr := dataMap["isError"].(bool); okErr {
+								success = !isError
+							}
+						}
+						successCalls := 0
+						failedCalls := 0
+						if success {
+							successCalls = 1
+						} else {
+							failedCalls = 1
+						}
+						now := time.Now()
+						if err := h.db.UpdateSkillStats(skillName, 1, successCalls, failedCalls, &now); err != nil {
+							h.logger.Warn("更新Skills调用统计失败", zap.Error(err), zap.String("skill", skillName))
+						}
+					}
+				}
+			}
+		}
+
 		// 子代理回复流式增量不落库；结束时合并为一条 eino_agent_reply
 		if assistantMessageID != "" && eventType == "eino_agent_reply_stream_end" {
 			flushResponsePlan()
@@ -1088,6 +1211,9 @@ func (h *AgentHandler) AgentLoopStream(c *gin.Context) {
 		}
 		eventJSON, _ := json.Marshal(event)
 		fmt.Fprintf(c.Writer, "data: %s\n\n", eventJSON)
+		done := StreamEvent{Type: "done", Message: ""}
+		doneJSON, _ := json.Marshal(done)
+		fmt.Fprintf(c.Writer, "data: %s\n\n", doneJSON)
 		c.Writer.Flush()
 		return
 	}
@@ -1108,6 +1234,7 @@ func (h *AgentHandler) AgentLoopStream(c *gin.Context) {
 	clientDisconnected := false
 	// 与 sseKeepalive 共用：禁止并发写 ResponseWriter，否则会破坏 chunked 编码（ERR_INVALID_CHUNKED_ENCODING）。
 	var sseWriteMu sync.Mutex
+	var ssePublishConversationID string
 	// 用于快速确认模型是否真的产生了流式 delta
 	var responseDeltaCount int
 	var responseStartLogged bool
@@ -1155,7 +1282,24 @@ func (h *AgentHandler) AgentLoopStream(c *gin.Context) {
 			}
 		}
 
-		// 如果客户端已断开，不再发送事件
+		event := StreamEvent{
+			Type:    eventType,
+			Message: message,
+			Data:    data,
+		}
+		eventJSON, errJSON := json.Marshal(event)
+		if errJSON != nil {
+			eventJSON = []byte(`{"type":"error","message":"marshal failed"}`)
+		}
+		sseLine := make([]byte, 0, len(eventJSON)+8)
+		sseLine = append(sseLine, []byte("data: ")...)
+		sseLine = append(sseLine, eventJSON...)
+		sseLine = append(sseLine, '\n', '\n')
+		if ssePublishConversationID != "" && h.taskEventBus != nil {
+			h.taskEventBus.Publish(ssePublishConversationID, sseLine)
+		}
+
+		// 如果客户端已断开，不再写入 HTTP（镜像订阅仍可收到事件）
 		if clientDisconnected {
 			return
 		}
@@ -1168,15 +1312,8 @@ func (h *AgentHandler) AgentLoopStream(c *gin.Context) {
 		default:
 		}
 
-		event := StreamEvent{
-			Type:    eventType,
-			Message: message,
-			Data:    data,
-		}
-		eventJSON, _ := json.Marshal(event)
-
 		sseWriteMu.Lock()
-		_, err := fmt.Fprintf(c.Writer, "data: %s\n\n", eventJSON)
+		_, err := c.Writer.Write(sseLine)
 		if err != nil {
 			sseWriteMu.Unlock()
 			clientDisconnected = true
@@ -1220,6 +1357,7 @@ func (h *AgentHandler) AgentLoopStream(c *gin.Context) {
 			return
 		}
 	}
+	ssePublishConversationID = conversationID
 
 	// 优先尝试从保存的ReAct数据恢复历史上下文
 	agentHistoryMessages, err := h.loadHistoryFromReActData(conversationID)
@@ -1350,14 +1488,14 @@ func (h *AgentHandler) AgentLoopStream(c *gin.Context) {
 	}
 
 	// 创建进度回调函数，复用统一逻辑
-	progressCallback := h.createProgressCallback(conversationID, assistantMessageID, sendEvent)
-
 	// 创建一个独立的上下文用于任务执行，不随HTTP请求取消
 	// 这样即使客户端断开连接（如刷新页面），任务也能继续执行
 	baseCtx, cancelWithCause := context.WithCancelCause(context.Background())
 	taskCtx, timeoutCancel := context.WithTimeout(baseCtx, 600*time.Minute)
 	defer timeoutCancel()
 	defer cancelWithCause(nil)
+	progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, conversationID, assistantMessageID, sendEvent)
+	taskCtx = h.injectReactHITLInterceptor(taskCtx, cancelWithCause, conversationID, assistantMessageID, sendEvent)
 
 	if _, err := h.tasks.StartTask(conversationID, req.Message, cancelWithCause); err != nil {
 		var errorMsg string
@@ -1606,6 +1744,51 @@ func (h *AgentHandler) CancelAgentLoop(c *gin.Context) {
 	})
 }
 
+// SubscribeAgentTaskEvents GET SSE：订阅指定会话当前运行中任务的事件镜像（帧格式与 POST .../stream 一致），用于刷新页面或断线后接续 UI。
+func (h *AgentHandler) SubscribeAgentTaskEvents(c *gin.Context) {
+	conversationID := strings.TrimSpace(c.Query("conversationId"))
+	if conversationID == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "conversationId is required"})
+		return
+	}
+	if h.tasks.GetTask(conversationID) == nil {
+		c.JSON(http.StatusNotFound, gin.H{"error": "no active task for this conversation"})
+		return
+	}
+	if h.taskEventBus == nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "task event bus unavailable"})
+		return
+	}
+
+	c.Header("Content-Type", "text/event-stream")
+	c.Header("Cache-Control", "no-cache")
+	c.Header("Connection", "keep-alive")
+	c.Header("X-Accel-Buffering", "no")
+
+	sub, ch := h.taskEventBus.Subscribe(conversationID)
+	defer h.taskEventBus.Unsubscribe(conversationID, sub)
+
+	flusher, _ := c.Writer.(http.Flusher)
+	ctx := c.Request.Context()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case chunk, ok := <-ch:
+			if !ok {
+				return
+			}
+			if _, err := c.Writer.Write(chunk); err != nil {
+				return
+			}
+			if flusher != nil {
+				flusher.Flush()
+			}
+		}
+	}
+}
+
 // ListAgentTasks 列出所有运行中的任务
 func (h *AgentHandler) ListAgentTasks(c *gin.Context) {
 	c.JSON(http.StatusOK, gin.H{
@@ -2266,7 +2449,7 @@ func (h *AgentHandler) executeBatchQueue(queueID string) {
 		if assistantMsg != nil {
 			assistantMessageID = assistantMsg.ID
 		}
-		progressCallback := h.createProgressCallback(conversationID, assistantMessageID, nil)
+		progressCallback := h.createProgressCallback(context.Background(), nil, conversationID, assistantMessageID, nil)
 
 		// 执行任务（使用包含角色提示词的finalMessage和角色工具列表）
 		h.logger.Info("执行批量任务", zap.String("queueId", queueID), zap.String("taskId", task.ID), zap.String("message", task.Message), zap.String("role", queue.Role), zap.String("conversationId", conversationID))

internal/handler/config.go

@@ -187,6 +187,7 @@ type GetConfigResponse struct {
 	MCP        config.MCPConfig        `json:"mcp"`
 	Tools      []ToolConfigInfo        `json:"tools"`
 	Agent      config.AgentConfig      `json:"agent"`
+	Hitl       config.HitlConfig       `json:"hitl,omitempty"`
 	Knowledge  config.KnowledgeConfig  `json:"knowledge"`
 	Robots     config.RobotsConfig     `json:"robots,omitempty"`
 	MultiAgent config.MultiAgentPublic `json:"multi_agent,omitempty"`
@@ -269,16 +270,12 @@ func (h *ConfigHandler) GetConfig(c *gin.Context) {
 	}
 	multiPub := config.MultiAgentPublic{
 		Enabled:                      h.config.MultiAgent.Enabled,
-		DefaultMode:                  h.config.MultiAgent.DefaultMode,
 		RobotUseMultiAgent:           h.config.MultiAgent.RobotUseMultiAgent,
 		BatchUseMultiAgent:           h.config.MultiAgent.BatchUseMultiAgent,
 		SubAgentCount:                subAgentCount,
 		Orchestration:                config.NormalizeMultiAgentOrchestration(h.config.MultiAgent.Orchestration),
 		PlanExecuteLoopMaxIterations: h.config.MultiAgent.PlanExecuteLoopMaxIterations,
 	}
-	if strings.TrimSpace(multiPub.DefaultMode) == "" {
-		multiPub.DefaultMode = "single"
-	}
 
 	c.JSON(http.StatusOK, GetConfigResponse{
 		OpenAI:     h.config.OpenAI,
@@ -286,6 +283,7 @@ func (h *ConfigHandler) GetConfig(c *gin.Context) {
 		MCP:        h.config.MCP,
 		Tools:      tools,
 		Agent:      h.config.Agent,
+		Hitl:       h.config.Hitl,
 		Knowledge:  h.config.Knowledge,
 		Robots:     h.config.Robots,
 		MultiAgent: multiPub,
@@ -686,10 +684,6 @@ func (h *ConfigHandler) UpdateConfig(c *gin.Context) {
 	// 多代理标量（sub_agents 等仍由 config.yaml 维护）
 	if req.MultiAgent != nil {
 		h.config.MultiAgent.Enabled = req.MultiAgent.Enabled
-		dm := strings.TrimSpace(req.MultiAgent.DefaultMode)
-		if dm == "multi" || dm == "single" {
-			h.config.MultiAgent.DefaultMode = dm
-		}
 		h.config.MultiAgent.RobotUseMultiAgent = req.MultiAgent.RobotUseMultiAgent
 		h.config.MultiAgent.BatchUseMultiAgent = req.MultiAgent.BatchUseMultiAgent
 		if req.MultiAgent.PlanExecuteLoopMaxIterations != nil {
@@ -697,7 +691,6 @@ func (h *ConfigHandler) UpdateConfig(c *gin.Context) {
 		}
 		h.logger.Info("更新多代理配置",
 			zap.Bool("enabled", h.config.MultiAgent.Enabled),
-			zap.String("default_mode", h.config.MultiAgent.DefaultMode),
 			zap.Bool("robot_use_multi_agent", h.config.MultiAgent.RobotUseMultiAgent),
 			zap.Bool("batch_use_multi_agent", h.config.MultiAgent.BatchUseMultiAgent),
 			zap.Int("plan_execute_loop_max_iterations", h.config.MultiAgent.PlanExecuteLoopMaxIterations),
@@ -1141,6 +1134,7 @@ func (h *ConfigHandler) saveConfig() error {
 	updateFOFAConfig(root, h.config.FOFA)
 	updateKnowledgeConfig(root, h.config.Knowledge)
 	updateRobotsConfig(root, h.config.Robots)
+	updateHitlConfig(root, h.config.Hitl)
 	updateMultiAgentConfig(root, h.config.MultiAgent)
 	// 更新外部MCP配置（使用external_mcp.go中的函数，同一包中可直接调用）
 	updateExternalMCPConfig(root, h.config.ExternalMCP)
@@ -1317,6 +1311,47 @@ func updateKnowledgeConfig(doc *yaml.Node, cfg config.KnowledgeConfig) {
 	setIntInMap(indexingNode, "retry_delay_ms", cfg.Indexing.RetryDelayMs)
 }
 
+func mergeHitlToolWhitelistSlice(existing, add []string) []string {
+	seen := make(map[string]struct{})
+	out := make([]string, 0, len(existing)+len(add))
+	for _, list := range [][]string{existing, add} {
+		for _, t := range list {
+			n := strings.ToLower(strings.TrimSpace(t))
+			if n == "" {
+				continue
+			}
+			if _, ok := seen[n]; ok {
+				continue
+			}
+			seen[n] = struct{}{}
+			out = append(out, strings.TrimSpace(t))
+		}
+	}
+	return out
+}
+
+// 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()
+	defer h.mu.Unlock()
+	merged := mergeHitlToolWhitelistSlice(h.config.Hitl.ToolWhitelist, add)
+	h.config.Hitl.ToolWhitelist = merged
+	if err := h.saveConfig(); err != nil {
+		return err
+	}
+	h.logger.Info("HITL 全局工具白名单已合并写入配置文件",
+		zap.Int("count", len(merged)),
+	)
+	return nil
+}
+
+func updateHitlConfig(doc *yaml.Node, cfg config.HitlConfig) {
+	root := doc.Content[0]
+	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)
+}
+
 func updateRobotsConfig(doc *yaml.Node, cfg config.RobotsConfig) {
 	root := doc.Content[0]
 	robotsNode := ensureMap(root, "robots")
@@ -1345,7 +1380,6 @@ func updateMultiAgentConfig(doc *yaml.Node, cfg config.MultiAgentConfig) {
 	root := doc.Content[0]
 	maNode := ensureMap(root, "multi_agent")
 	setBoolInMap(maNode, "enabled", cfg.Enabled)
-	setStringInMap(maNode, "default_mode", cfg.DefaultMode)
 	setBoolInMap(maNode, "robot_use_multi_agent", cfg.RobotUseMultiAgent)
 	setBoolInMap(maNode, "batch_use_multi_agent", cfg.BatchUseMultiAgent)
 	setIntInMap(maNode, "plan_execute_loop_max_iterations", cfg.PlanExecuteLoopMaxIterations)
@@ -1428,6 +1462,21 @@ func setStringSliceInMap(mapNode *yaml.Node, key string, values []string) {
 	}
 }
 
+func setFlowStringSliceInMap(mapNode *yaml.Node, key string, values []string) {
+	_, valueNode := ensureKeyValue(mapNode, key)
+	valueNode.Kind = yaml.SequenceNode
+	valueNode.Tag = "!!seq"
+	valueNode.Style = yaml.FlowStyle
+	valueNode.Content = nil
+	for _, v := range values {
+		valueNode.Content = append(valueNode.Content, &yaml.Node{
+			Kind:  yaml.ScalarNode,
+			Tag:   "!!str",
+			Value: v,
+		})
+	}
+}
+
 func setIntInMap(mapNode *yaml.Node, key string, value int) {
 	_, valueNode := ensureKeyValue(mapNode, key)
 	valueNode.Kind = yaml.ScalarNode

internal/handler/eino_single_agent.go

@@ -41,11 +41,24 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 	var baseCtx context.Context
 	clientDisconnected := false
 	var sseWriteMu sync.Mutex
+	var ssePublishConversationID string
 	sendEvent := func(eventType, message string, data interface{}) {
-		if clientDisconnected {
+		if eventType == "error" && baseCtx != nil && errors.Is(context.Cause(baseCtx), ErrTaskCancelled) {
 			return
 		}
-		if eventType == "error" && baseCtx != nil && errors.Is(context.Cause(baseCtx), ErrTaskCancelled) {
+		ev := StreamEvent{Type: eventType, Message: message, Data: data}
+		b, errMarshal := json.Marshal(ev)
+		if errMarshal != nil {
+			b = []byte(`{"type":"error","message":"marshal failed"}`)
+		}
+		sseLine := make([]byte, 0, len(b)+8)
+		sseLine = append(sseLine, []byte("data: ")...)
+		sseLine = append(sseLine, b...)
+		sseLine = append(sseLine, '\n', '\n')
+		if ssePublishConversationID != "" && h.taskEventBus != nil {
+			h.taskEventBus.Publish(ssePublishConversationID, sseLine)
+		}
+		if clientDisconnected {
 			return
 		}
 		select {
@@ -54,10 +67,8 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 			return
 		default:
 		}
-		ev := StreamEvent{Type: eventType, Message: message, Data: data}
-		b, _ := json.Marshal(ev)
 		sseWriteMu.Lock()
-		_, err := fmt.Fprintf(c.Writer, "data: %s\n\n", b)
+		_, err := c.Writer.Write(sseLine)
 		if err != nil {
 			sseWriteMu.Unlock()
 			clientDisconnected = true
@@ -81,6 +92,7 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 		sendEvent("done", "", nil)
 		return
 	}
+	ssePublishConversationID = prep.ConversationID
 	if prep.CreatedNew {
 		sendEvent("conversation", "会话已创建", map[string]interface{}{
 			"conversationId": prep.ConversationID,
@@ -89,6 +101,10 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 
 	conversationID := prep.ConversationID
 	assistantMessageID := prep.AssistantMessageID
+	h.activateHITLForConversation(conversationID, req.Hitl)
+	if h.hitlManager != nil {
+		defer h.hitlManager.DeactivateConversation(conversationID)
+	}
 
 	if prep.UserMessageID != "" {
 		sendEvent("message_saved", "", map[string]interface{}{
@@ -97,13 +113,15 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 		})
 	}
 
-	progressCallback := h.createProgressCallback(conversationID, assistantMessageID, sendEvent)
-
 	var cancelWithCause context.CancelCauseFunc
 	baseCtx, cancelWithCause = context.WithCancelCause(context.Background())
 	taskCtx, timeoutCancel := context.WithTimeout(baseCtx, 600*time.Minute)
 	defer timeoutCancel()
 	defer cancelWithCause(nil)
+	progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, conversationID, assistantMessageID, sendEvent)
+	taskCtx = multiagent.WithHITLToolInterceptor(taskCtx, func(ctx context.Context, toolName, arguments string) (string, error) {
+		return h.interceptHITLForEinoTool(ctx, cancelWithCause, conversationID, assistantMessageID, sendEvent, toolName, arguments)
+	})
 
 	if _, err := h.tasks.StartTask(conversationID, req.Message, cancelWithCause); err != nil {
 		var errorMsg string
@@ -136,6 +154,8 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 	defer close(stopKeepalive)
 
 	if h.config == nil {
+		taskStatus = "failed"
+		h.tasks.UpdateTaskStatus(conversationID, taskStatus)
 		sendEvent("error", "服务器配置未加载", nil)
 		sendEvent("done", "", map[string]interface{}{"conversationId": conversationID})
 		return
@@ -166,7 +186,24 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 			}
 			sendEvent("cancelled", cancelMsg, map[string]interface{}{
 				"conversationId": conversationID,
-				"messageId":        assistantMessageID,
+				"messageId":      assistantMessageID,
+			})
+			sendEvent("done", "", map[string]interface{}{"conversationId": conversationID})
+			return
+		}
+
+		if errors.Is(runErr, context.DeadlineExceeded) || errors.Is(context.Cause(taskCtx), context.DeadlineExceeded) {
+			taskStatus = "timeout"
+			h.tasks.UpdateTaskStatus(conversationID, taskStatus)
+			timeoutMsg := "任务执行超时，已自动终止。"
+			if assistantMessageID != "" {
+				_, _ = h.db.Exec("UPDATE messages SET content = ? WHERE id = ?", timeoutMsg, assistantMessageID)
+				_ = h.db.AddProcessDetail(assistantMessageID, conversationID, "timeout", timeoutMsg, nil)
+			}
+			sendEvent("error", timeoutMsg, map[string]interface{}{
+				"conversationId": conversationID,
+				"messageId":      assistantMessageID,
+				"errorType":      "timeout",
 			})
 			sendEvent("done", "", map[string]interface{}{"conversationId": conversationID})
 			return
@@ -232,12 +269,24 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
+	h.activateHITLForConversation(prep.ConversationID, req.Hitl)
+	if h.hitlManager != nil {
+		defer h.hitlManager.DeactivateConversation(prep.ConversationID)
+	}
 
 	var progressBuf strings.Builder
-	progressCallback := func(eventType, message string, data interface{}) {
+	progressCallbackRaw := func(eventType, message string, data interface{}) {
 		progressBuf.WriteString(eventType)
 		progressBuf.WriteByte('\n')
 	}
+	baseCtx, cancelWithCause := context.WithCancelCause(c.Request.Context())
+	defer cancelWithCause(nil)
+	taskCtx, timeoutCancel := context.WithTimeout(baseCtx, 600*time.Minute)
+	defer timeoutCancel()
+	progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, prep.ConversationID, prep.AssistantMessageID, progressCallbackRaw)
+	taskCtx = multiagent.WithHITLToolInterceptor(taskCtx, func(ctx context.Context, toolName, arguments string) (string, error) {
+		return h.interceptHITLForEinoTool(ctx, cancelWithCause, prep.ConversationID, prep.AssistantMessageID, nil, toolName, arguments)
+	})
 
 	if h.config == nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": "服务器配置未加载"})
@@ -245,7 +294,7 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 	}
 
 	result, runErr := multiagent.RunEinoSingleChatModelAgent(
-		c.Request.Context(),
+		taskCtx,
 		h.config,
 		&h.config.MultiAgent,
 		h.agent,
@@ -279,10 +328,10 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 	}
 
 	c.JSON(http.StatusOK, gin.H{
-		"response":         result.Response,
-		"conversationId":   prep.ConversationID,
-		"mcpExecutionIds":  result.MCPExecutionIDs,
+		"response":           result.Response,
+		"conversationId":     prep.ConversationID,
+		"mcpExecutionIds":    result.MCPExecutionIDs,
 		"assistantMessageId": prep.AssistantMessageID,
-		"agentMode":        "eino_single",
+		"agentMode":          "eino_single",
 	})
 }

internal/handler/hitl.go

@@ -0,0 +1,798 @@
+package handler
+
+import (
+	"context"
+	"database/sql"
+	"encoding/json"
+	"errors"
+	"math"
+	"net/http"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+
+	"cyberstrike-ai/internal/agent"
+	"cyberstrike-ai/internal/database"
+	"cyberstrike-ai/internal/multiagent"
+
+	"github.com/gin-gonic/gin"
+	"github.com/google/uuid"
+	"go.uber.org/zap"
+)
+
+type hitlRuntimeConfig struct {
+	Enabled        bool
+	Mode           string
+	SensitiveTools map[string]struct{}
+	Timeout        time.Duration
+}
+
+type hitlDecision struct {
+	Decision        string
+	Comment         string
+	EditedArguments map[string]interface{}
+}
+
+type pendingInterrupt struct {
+	ConversationID string
+	InterruptID    string
+	Mode           string
+	ToolName       string
+	ToolCallID     string
+	decideCh       chan hitlDecision
+}
+
+type HITLManager struct {
+	db     *database.DB
+	logger *zap.Logger
+
+	mu      sync.RWMutex
+	runtime map[string]hitlRuntimeConfig
+	pending map[string]*pendingInterrupt
+}
+
+func NewHITLManager(db *database.DB, logger *zap.Logger) *HITLManager {
+	return &HITLManager{
+		db:      db,
+		logger:  logger,
+		runtime: make(map[string]hitlRuntimeConfig),
+		pending: make(map[string]*pendingInterrupt),
+	}
+}
+
+func (m *HITLManager) EnsureSchema() error {
+	if _, err := m.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 {
+		return err
+	}
+	_, err := m.db.Exec(`
+CREATE TABLE IF NOT EXISTS hitl_conversation_configs (
+    conversation_id TEXT PRIMARY KEY,
+    enabled INTEGER NOT NULL DEFAULT 0,
+    mode TEXT NOT NULL DEFAULT 'off',
+    sensitive_tools TEXT NOT NULL DEFAULT '[]',
+    timeout_seconds INTEGER NOT NULL DEFAULT 300,
+    updated_at DATETIME NOT NULL
+);`)
+	if err != nil {
+		return err
+	}
+
+	// On startup, cancel all orphaned pending interrupts from previous process.
+	// Their in-memory channels are gone, so they can never be resolved.
+	res, err := m.db.Exec(`UPDATE hitl_interrupts SET status='cancelled', decision='reject',
+		decision_comment='process restarted', decided_at=CURRENT_TIMESTAMP WHERE status='pending'`)
+	if err != nil {
+		m.logger.Warn("failed to cancel orphaned HITL interrupts", zap.Error(err))
+	} else if n, _ := res.RowsAffected(); n > 0 {
+		m.logger.Info("cancelled orphaned HITL interrupts from previous process", zap.Int64("count", n))
+	}
+	return nil
+}
+
+func normalizeHitlMode(mode string) string {
+	v := strings.ToLower(strings.TrimSpace(mode))
+	if v == "" {
+		return "approval"
+	}
+	switch v {
+	case "off":
+		return "off"
+	case "feedback", "followup":
+		return "approval"
+	case "approval", "review_edit":
+		return v
+	default:
+		return "approval"
+	}
+}
+
+func (m *HITLManager) ActivateConversation(conversationID string, req *HITLRequest) {
+	if req == nil || !req.Enabled {
+		m.DeactivateConversation(conversationID)
+		return
+	}
+	tools := make(map[string]struct{})
+	for _, t := range req.SensitiveTools {
+		n := strings.ToLower(strings.TrimSpace(t))
+		if n != "" {
+			tools[n] = struct{}{}
+		}
+	}
+	timeout := 5 * time.Minute
+	if req.TimeoutSeconds > 0 {
+		timeout = time.Duration(req.TimeoutSeconds) * time.Second
+	}
+	m.mu.Lock()
+	m.runtime[conversationID] = hitlRuntimeConfig{
+		Enabled:        true,
+		Mode:           normalizeHitlMode(req.Mode),
+		SensitiveTools: tools,
+		Timeout:        timeout,
+	}
+	m.mu.Unlock()
+}
+
+func (m *HITLManager) DeactivateConversation(conversationID string) {
+	m.mu.Lock()
+	delete(m.runtime, conversationID)
+	m.mu.Unlock()
+}
+
+// hitlConfigGlobalToolWhitelist 来自 config.yaml hitl.tool_whitelist（去重、去空）。
+func (h *AgentHandler) hitlConfigGlobalToolWhitelist() []string {
+	if h == nil || h.config == nil {
+		return nil
+	}
+	raw := h.config.Hitl.ToolWhitelist
+	if len(raw) == 0 {
+		return nil
+	}
+	seen := make(map[string]struct{})
+	out := make([]string, 0, len(raw))
+	for _, t := range raw {
+		n := strings.ToLower(strings.TrimSpace(t))
+		if n == "" {
+			continue
+		}
+		if _, ok := seen[n]; ok {
+			continue
+		}
+		seen[n] = struct{}{}
+		out = append(out, strings.TrimSpace(t))
+	}
+	return 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。
+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))
+	for _, t := range gw {
+		n := strings.ToLower(strings.TrimSpace(t))
+		if n == "" {
+			continue
+		}
+		if _, ok := seen[n]; ok {
+			continue
+		}
+		seen[n] = struct{}{}
+		union = append(union, strings.TrimSpace(t))
+	}
+	for _, t := range req.SensitiveTools {
+		n := strings.ToLower(strings.TrimSpace(t))
+		if n == "" {
+			continue
+		}
+		if _, ok := seen[n]; ok {
+			continue
+		}
+		seen[n] = struct{}{}
+		union = append(union, strings.TrimSpace(t))
+	}
+	out := *req
+	out.SensitiveTools = union
+	return &out
+}
+
+func (m *HITLManager) shouldInterrupt(conversationID, toolName string) (hitlRuntimeConfig, bool) {
+	m.mu.RLock()
+	cfg, ok := m.runtime[conversationID]
+	m.mu.RUnlock()
+	if !ok || !cfg.Enabled {
+		return hitlRuntimeConfig{}, false
+	}
+	// 语义：SensitiveTools 现在作为“白名单（免审批工具）”
+	// 空白名单 => 全部工具都需要审批
+	if len(cfg.SensitiveTools) == 0 {
+		return cfg, true
+	}
+	_, inWhitelist := cfg.SensitiveTools[strings.ToLower(strings.TrimSpace(toolName))]
+	return cfg, !inWhitelist
+}
+
+func (m *HITLManager) CreatePendingInterrupt(conversationID, assistantMessageID, mode, toolName, toolCallID, payload string) (*pendingInterrupt, error) {
+	now := time.Now()
+	id := "hitl_" + strings.ReplaceAll(uuid.New().String(), "-", "")
+	if _, err := m.db.Exec(`INSERT INTO hitl_interrupts
+		(id, conversation_id, message_id, mode, tool_name, tool_call_id, payload, status, created_at)
+		VALUES (?, ?, ?, ?, ?, ?, ?, 'pending', ?)`,
+		id, conversationID, assistantMessageID, mode, toolName, toolCallID, payload, now); err != nil {
+		return nil, err
+	}
+	// 刷新页面后侧栏依赖 DB 配置；若仅内存 Activate 未落库，会导致「有待审批却显示关闭」
+	_ = m.ensureConversationHITLModePersisted(conversationID, mode)
+	p := &pendingInterrupt{
+		ConversationID: conversationID,
+		InterruptID:    id,
+		Mode:           normalizeHitlMode(mode),
+		ToolName:       toolName,
+		ToolCallID:     toolCallID,
+		decideCh:       make(chan hitlDecision, 1),
+	}
+	m.mu.Lock()
+	m.pending[id] = p
+	m.mu.Unlock()
+	return p, nil
+}
+
+// ensureConversationHITLModePersisted 在产生待审批时把 mode 写入 hitl_conversation_configs，避免刷新后 GET 配置仍为关闭。
+func (m *HITLManager) ensureConversationHITLModePersisted(conversationID, interruptMode string) error {
+	if strings.TrimSpace(conversationID) == "" {
+		return nil
+	}
+	nm := normalizeHitlMode(interruptMode)
+	if nm == "off" {
+		return nil
+	}
+	cfg, err := m.LoadConversationConfig(conversationID)
+	if err != nil {
+		return err
+	}
+	if cfg.Enabled && normalizeHitlMode(cfg.Mode) == nm {
+		return nil
+	}
+	cfg.Enabled = true
+	cfg.Mode = nm
+	if cfg.TimeoutSeconds <= 0 {
+		cfg.TimeoutSeconds = 300
+	}
+	return m.SaveConversationConfig(conversationID, cfg)
+}
+
+// PendingHITLInterruptMode 返回该会话最新一条 pending 中断的协同模式（用于 GET 配置时与库内「关闭」状态对齐）。
+func (m *HITLManager) PendingHITLInterruptMode(conversationID string) (string, bool) {
+	if strings.TrimSpace(conversationID) == "" {
+		return "", false
+	}
+	var mode string
+	err := m.db.QueryRow(`SELECT mode FROM hitl_interrupts WHERE conversation_id = ? AND status = 'pending' ORDER BY created_at DESC LIMIT 1`, conversationID).
+		Scan(&mode)
+	if err != nil {
+		if errors.Is(err, sql.ErrNoRows) {
+			return "", false
+		}
+		return "", false
+	}
+	mode = strings.TrimSpace(mode)
+	if mode == "" {
+		return "", false
+	}
+	return mode, true
+}
+
+func hitlStoredConfigEffective(cfg *HITLRequest) bool {
+	if cfg == nil {
+		return false
+	}
+	if cfg.Enabled {
+		return true
+	}
+	return normalizeHitlMode(cfg.Mode) != "off"
+}
+
+func (m *HITLManager) ResolveInterrupt(interruptID, decision, comment string, editedArguments map[string]interface{}) error {
+	decision = strings.ToLower(strings.TrimSpace(decision))
+	if decision != "approve" && decision != "reject" {
+		return errors.New("decision must be approve/reject")
+	}
+	m.mu.RLock()
+	p, ok := m.pending[interruptID]
+	m.mu.RUnlock()
+	if !ok {
+		return errors.New("interrupt not found or already resolved")
+	}
+	d := hitlDecision{
+		Decision:        decision,
+		Comment:         strings.TrimSpace(comment),
+		EditedArguments: editedArguments,
+	}
+	select {
+	case p.decideCh <- d:
+		return nil
+	default:
+		return errors.New("interrupt already resolved or decision channel busy")
+	}
+}
+
+func (m *HITLManager) SaveConversationConfig(conversationID string, req *HITLRequest) error {
+	if strings.TrimSpace(conversationID) == "" {
+		return errors.New("conversationId is required")
+	}
+	if req == nil {
+		req = &HITLRequest{Enabled: false, Mode: "off", TimeoutSeconds: 300}
+	}
+	mode := normalizeHitlMode(req.Mode)
+	if !req.Enabled {
+		mode = "off"
+	}
+	tools, _ := json.Marshal(req.SensitiveTools)
+	timeout := req.TimeoutSeconds
+	if timeout <= 0 {
+		timeout = 300
+	}
+	_, err := m.db.Exec(`INSERT INTO hitl_conversation_configs
+		(conversation_id, enabled, mode, sensitive_tools, timeout_seconds, updated_at)
+		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`,
+		conversationID, boolToInt(req.Enabled), mode, string(tools), timeout, time.Now())
+	return err
+}
+
+func (m *HITLManager) LoadConversationConfig(conversationID string) (*HITLRequest, error) {
+	var enabledInt int
+	var mode, toolsJSON string
+	var timeout int
+	err := m.db.QueryRow(`SELECT enabled, mode, sensitive_tools, timeout_seconds FROM hitl_conversation_configs WHERE conversation_id = ?`, conversationID).
+		Scan(&enabledInt, &mode, &toolsJSON, &timeout)
+	if errors.Is(err, sql.ErrNoRows) {
+		return &HITLRequest{Enabled: false, Mode: "off", SensitiveTools: []string{}, TimeoutSeconds: 300}, nil
+	}
+	if err != nil {
+		return nil, err
+	}
+	tools := make([]string, 0)
+	_ = json.Unmarshal([]byte(toolsJSON), &tools)
+	return &HITLRequest{
+		Enabled:        enabledInt == 1,
+		Mode:           mode,
+		SensitiveTools: tools,
+		TimeoutSeconds: timeout,
+	}, nil
+}
+
+func (m *HITLManager) waitDecision(ctx context.Context, p *pendingInterrupt, timeout time.Duration) (hitlDecision, error) {
+	defer func() {
+		m.mu.Lock()
+		delete(m.pending, p.InterruptID)
+		m.mu.Unlock()
+	}()
+	select {
+	case d := <-p.decideCh:
+		// 只有 review_edit 模式允许改参；其他模式一律忽略 edited arguments
+		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=?`,
+			d.Decision, d.Comment, time.Now(), p.InterruptID)
+		return d, nil
+	case <-time.After(timeout):
+		_, _ = m.db.Exec(`UPDATE hitl_interrupts SET status='timeout', decision='approve', decision_comment='timeout auto approve', decided_at=? WHERE id=?`,
+			time.Now(), p.InterruptID)
+		return hitlDecision{Decision: "approve", Comment: "timeout auto approve"}, nil
+	case <-ctx.Done():
+		_, _ = m.db.Exec(`UPDATE hitl_interrupts SET status='cancelled', decision='reject', decision_comment='task cancelled', decided_at=? WHERE id=?`,
+			time.Now(), p.InterruptID)
+		return hitlDecision{Decision: "reject", Comment: "task cancelled"}, ctx.Err()
+	}
+}
+
+func (h *AgentHandler) activateHITLForConversation(conversationID string, req *HITLRequest) {
+	if h.hitlManager == nil {
+		return
+	}
+	if req == nil {
+		cfg, err := h.hitlManager.LoadConversationConfig(conversationID)
+		if err == nil {
+			req = cfg
+		}
+	}
+	h.hitlManager.ActivateConversation(conversationID, h.hitlRequestWithMergedConfigWhitelist(req))
+}
+
+func (h *AgentHandler) waitHITLApproval(runCtx context.Context, cancelRun context.CancelCauseFunc, conversationID, assistantMessageID, toolName, toolCallID string, payload map[string]interface{}, sendEventFunc func(eventType, message string, data interface{})) (*hitlDecision, error) {
+	cfg, need := h.hitlManager.shouldInterrupt(conversationID, toolName)
+	if !need {
+		return nil, nil
+	}
+	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 sendEventFunc != nil {
+		sendEventFunc("hitl_interrupt", "命中人机协同审批", map[string]interface{}{
+			"conversationId": conversationID,
+			"interruptId":    p.InterruptID,
+			"mode":           cfg.Mode,
+			"toolName":       toolName,
+			"toolCallId":     toolCallID,
+			"payload":        payload,
+		})
+	}
+	d, waitErr := h.hitlManager.waitDecision(runCtx, p, cfg.Timeout)
+	if waitErr != nil {
+		if cancelRun != nil && (errors.Is(waitErr, context.Canceled) || errors.Is(waitErr, context.DeadlineExceeded)) {
+			cause := context.Cause(runCtx)
+			switch {
+			case errors.Is(cause, ErrTaskCancelled):
+				cancelRun(ErrTaskCancelled)
+			case cause != nil:
+				cancelRun(cause)
+			case errors.Is(waitErr, context.DeadlineExceeded):
+				cancelRun(context.DeadlineExceeded)
+			default:
+				cancelRun(ErrTaskCancelled)
+			}
+		}
+		return nil, waitErr
+	}
+	if d.Decision == "reject" {
+		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{}{
+				"conversationId": conversationID,
+				"interruptId":    p.InterruptID,
+				"toolName":       toolName,
+				"comment":        d.Comment,
+			})
+		}
+		return &d, nil
+	}
+	if sendEventFunc != nil {
+		sendEventFunc("hitl_resumed", "人工确认通过，继续执行", map[string]interface{}{
+			"conversationId": conversationID,
+			"interruptId":    p.InterruptID,
+			"toolName":       toolName,
+			"comment":        d.Comment,
+			"editedArgs":     d.EditedArguments,
+		})
+	}
+	return &d, nil
+}
+
+func (h *AgentHandler) handleHITLToolCall(runCtx context.Context, cancelRun context.CancelCauseFunc, conversationID, assistantMessageID string, data map[string]interface{}, sendEventFunc func(eventType, message string, data interface{})) {
+	if h.hitlManager == nil {
+		return
+	}
+	toolName, _ := data["toolName"].(string)
+	toolCallID, _ := data["toolCallId"].(string)
+	d, err := h.waitHITLApproval(runCtx, cancelRun, conversationID, assistantMessageID, toolName, toolCallID, data, sendEventFunc)
+	if err != nil || d == nil {
+		return
+	}
+	if len(d.EditedArguments) > 0 {
+		if argsObj, ok := data["argumentsObj"].(map[string]interface{}); ok {
+			for k := range argsObj {
+				delete(argsObj, k)
+			}
+			for k, v := range d.EditedArguments {
+				argsObj[k] = v
+			}
+			if b, mErr := json.Marshal(argsObj); mErr == nil {
+				data["arguments"] = string(b)
+			}
+		}
+	}
+}
+
+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
+	}
+	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`
+	args := []interface{}{}
+	if conversationID != "" {
+		q += " AND conversation_id = ?"
+		args = append(args, conversationID)
+	}
+	if status != "all" {
+		q += " AND status = ?"
+		args = append(args, status)
+	}
+	q += " ORDER BY created_at DESC LIMIT ? OFFSET ?"
+	args = append(args, pageSize, offset)
+	rows, err := h.db.Query(q, args...)
+	if err != nil {
+		c.JSON(500, gin.H{"error": err.Error()})
+		return
+	}
+	defer rows.Close()
+	items := make([]map[string]interface{}, 0)
+	for rows.Next() {
+		var id, cid, mode, toolName, toolCallID, payload, rowStatus 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, &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})
+}
+
+type hitlDecisionReq struct {
+	InterruptID     string                 `json:"interruptId" binding:"required"`
+	Decision        string                 `json:"decision" binding:"required"`
+	Comment         string                 `json:"comment,omitempty"`
+	EditedArguments map[string]interface{} `json:"editedArguments,omitempty"`
+}
+
+func (h *AgentHandler) DecideHITLInterrupt(c *gin.Context) {
+	var req hitlDecisionReq
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(400, gin.H{"error": err.Error()})
+		return
+	}
+	if h.hitlManager == nil {
+		c.JSON(500, gin.H{"error": "hitl manager unavailable"})
+		return
+	}
+	if err := h.hitlManager.ResolveInterrupt(req.InterruptID, req.Decision, req.Comment, req.EditedArguments); err != nil {
+		c.JSON(http.StatusConflict, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, gin.H{"ok": true})
+}
+
+func (h *AgentHandler) DismissHITLInterrupt(c *gin.Context) {
+	var req struct {
+		InterruptID string `json:"interruptId" binding:"required"`
+	}
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(400, gin.H{"error": err.Error()})
+		return
+	}
+	if h.hitlManager == nil {
+		c.JSON(500, gin.H{"error": "hitl manager unavailable"})
+		return
+	}
+	res, err := h.db.Exec(`UPDATE hitl_interrupts SET status='cancelled', decision='reject',
+		decision_comment='dismissed by user', decided_at=CURRENT_TIMESTAMP
+		WHERE id=? AND status='pending'`, req.InterruptID)
+	if err != nil {
+		c.JSON(500, gin.H{"error": err.Error()})
+		return
+	}
+	n, _ := res.RowsAffected()
+	if n == 0 {
+		c.JSON(404, gin.H{"error": "interrupt not found or already resolved"})
+		return
+	}
+	// Also drain from in-memory map if present
+	h.hitlManager.mu.Lock()
+	if p, ok := h.hitlManager.pending[req.InterruptID]; ok {
+		delete(h.hitlManager.pending, req.InterruptID)
+		select {
+		case p.decideCh <- hitlDecision{Decision: "reject", Comment: "dismissed by user"}:
+		default:
+		}
+	}
+	h.hitlManager.mu.Unlock()
+	c.JSON(http.StatusOK, gin.H{"ok": true})
+}
+
+func (h *AgentHandler) interceptHITLForEinoTool(runCtx context.Context, cancelRun context.CancelCauseFunc, conversationID, assistantMessageID string, sendEventFunc func(eventType, message string, data interface{}), toolName, arguments string) (string, error) {
+	payload := map[string]interface{}{
+		"toolName":   toolName,
+		"arguments":  arguments,
+		"source":     "eino_middleware",
+		"toolCallId": "",
+	}
+	var argsObj map[string]interface{}
+	if strings.TrimSpace(arguments) != "" {
+		_ = json.Unmarshal([]byte(arguments), &argsObj)
+		if argsObj != nil {
+			payload["argumentsObj"] = argsObj
+		}
+	}
+	d, err := h.waitHITLApproval(runCtx, cancelRun, conversationID, assistantMessageID, toolName, "", payload, sendEventFunc)
+	if err != nil || d == nil {
+		return arguments, err
+	}
+	if d.Decision == "reject" {
+		return arguments, multiagent.NewHumanRejectError(d.Comment)
+	}
+	if len(d.EditedArguments) > 0 {
+		edited, mErr := json.Marshal(d.EditedArguments)
+		if mErr == nil {
+			return string(edited), nil
+		}
+	}
+	return arguments, nil
+}
+
+func (h *AgentHandler) interceptHITLForReactTool(runCtx context.Context, cancelRun context.CancelCauseFunc, conversationID, assistantMessageID string, sendEventFunc func(eventType, message string, data interface{}), toolName string, arguments map[string]interface{}, toolCallID string) (map[string]interface{}, error) {
+	payload := map[string]interface{}{
+		"toolName":     toolName,
+		"argumentsObj": arguments,
+		"toolCallId":   toolCallID,
+		"source":       "react_pre_exec",
+	}
+	d, err := h.waitHITLApproval(runCtx, cancelRun, conversationID, assistantMessageID, toolName, toolCallID, payload, sendEventFunc)
+	if err != nil || d == nil {
+		return arguments, err
+	}
+	if d.Decision == "reject" {
+		comment := strings.TrimSpace(d.Comment)
+		if comment == "" {
+			comment = "no extra feedback"
+		}
+		return arguments, errors.New("human rejected this tool call; feedback: " + comment)
+	}
+	if len(d.EditedArguments) > 0 {
+		return d.EditedArguments, nil
+	}
+	return arguments, nil
+}
+
+func (h *AgentHandler) injectReactHITLInterceptor(ctx context.Context, cancelRun context.CancelCauseFunc, conversationID, assistantMessageID string, sendEventFunc func(eventType, message string, data interface{})) context.Context {
+	return agent.WithToolCallInterceptor(ctx, func(c context.Context, toolName string, args map[string]interface{}, toolCallID string) (map[string]interface{}, error) {
+		return h.interceptHITLForReactTool(c, cancelRun, conversationID, assistantMessageID, sendEventFunc, toolName, args, toolCallID)
+	})
+}
+
+type hitlConfigReq struct {
+	ConversationID string `json:"conversationId" binding:"required"`
+	HITLRequest
+}
+
+func (h *AgentHandler) GetHITLConversationConfig(c *gin.Context) {
+	conversationID := strings.TrimSpace(c.Param("conversationId"))
+	if conversationID == "" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "conversationId is required"})
+		return
+	}
+	cfg, err := h.hitlManager.LoadConversationConfig(conversationID)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	if !hitlStoredConfigEffective(cfg) {
+		if pendMode, ok := h.hitlManager.PendingHITLInterruptMode(conversationID); ok {
+			cfg2 := *cfg
+			cfg2.Enabled = true
+			cfg2.Mode = normalizeHitlMode(pendMode)
+			if cfg2.TimeoutSeconds <= 0 {
+				cfg2.TimeoutSeconds = 300
+			}
+			cfg = &cfg2
+		}
+	}
+	c.JSON(http.StatusOK, gin.H{
+		"conversationId":          conversationID,
+		"hitl":                    cfg,
+		"hitlGlobalToolWhitelist": h.hitlConfigGlobalToolWhitelist(),
+	})
+}
+
+func (h *AgentHandler) UpsertHITLConversationConfig(c *gin.Context) {
+	var req hitlConfigReq
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+	req.Mode = normalizeHitlMode(req.Mode)
+	if err := h.hitlManager.SaveConversationConfig(req.ConversationID, &req.HITLRequest); err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	if h.hitlWhitelistSaver != nil && len(req.SensitiveTools) > 0 {
+		if err := h.hitlWhitelistSaver.MergeHitlToolWhitelistIntoConfig(req.SensitiveTools); err != nil {
+			h.logger.Warn("HITL 会话配置已保存，但合并工具白名单到 config.yaml 失败", zap.Error(err))
+			c.JSON(http.StatusInternalServerError, gin.H{
+				"error": "会话配置已保存，但写入 config.yaml 失败: " + err.Error(),
+			})
+			return
+		}
+	}
+	h.hitlManager.ActivateConversation(req.ConversationID, h.hitlRequestWithMergedConfigWhitelist(&req.HITLRequest))
+	c.JSON(http.StatusOK, gin.H{"ok": true})
+}
+
+type mergeHitlGlobalWhitelistReq struct {
+	SensitiveTools []string `json:"sensitiveTools"`
+}
+
+// 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 {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": "HITL 配置持久化不可用"})
+		return
+	}
+	var req mergeHitlGlobalWhitelistReq
+	if err := c.ShouldBindJSON(&req); err != nil {
+		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
+		return
+	}
+	if len(req.SensitiveTools) == 0 {
+		c.JSON(http.StatusOK, gin.H{
+			"ok":                        true,
+			"hitlGlobalToolWhitelist":   h.hitlConfigGlobalToolWhitelist(),
+			"hitlGlobalWhitelistMerged": false,
+		})
+		return
+	}
+	if err := h.hitlWhitelistSaver.MergeHitlToolWhitelistIntoConfig(req.SensitiveTools); err != nil {
+		h.logger.Warn("合并 HITL 工具白名单到 config.yaml 失败", zap.Error(err))
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, gin.H{
+		"ok":                        true,
+		"hitlGlobalToolWhitelist":   h.hitlConfigGlobalToolWhitelist(),
+		"hitlGlobalWhitelistMerged": true,
+	})
+}
+
+func boolToInt(v bool) int {
+	if v {
+		return 1
+	}
+	return 0
+}

internal/handler/multi_agent.go

@@ -40,6 +40,9 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 		event := StreamEvent{Type: "error", Message: "请求参数错误: " + err.Error()}
 		b, _ := json.Marshal(event)
 		fmt.Fprintf(c.Writer, "data: %s\n\n", b)
+		done := StreamEvent{Type: "done", Message: ""}
+		db, _ := json.Marshal(done)
+		fmt.Fprintf(c.Writer, "data: %s\n\n", db)
 		c.Writer.Flush()
 		return
 	}
@@ -53,25 +56,36 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 	clientDisconnected := false
 	// 与 sseKeepalive 共用：禁止并发写 ResponseWriter，否则会破坏 chunked 编码（ERR_INVALID_CHUNKED_ENCODING）。
 	var sseWriteMu sync.Mutex
+	var ssePublishConversationID string
 	sendEvent := func(eventType, message string, data interface{}) {
-		if clientDisconnected {
-			return
-		}
 		// 用户主动停止时，Eino 可能仍会并发上报 eventType=="error"。
 		// 为避免 UI 看到“取消错误 + cancelled 文案”两条回复，这里直接丢弃取消对应的 error。
 		if eventType == "error" && baseCtx != nil && errors.Is(context.Cause(baseCtx), ErrTaskCancelled) {
 			return
 		}
+		ev := StreamEvent{Type: eventType, Message: message, Data: data}
+		b, errMarshal := json.Marshal(ev)
+		if errMarshal != nil {
+			b = []byte(`{"type":"error","message":"marshal failed"}`)
+		}
+		sseLine := make([]byte, 0, len(b)+8)
+		sseLine = append(sseLine, []byte("data: ")...)
+		sseLine = append(sseLine, b...)
+		sseLine = append(sseLine, '\n', '\n')
+		if ssePublishConversationID != "" && h.taskEventBus != nil {
+			h.taskEventBus.Publish(ssePublishConversationID, sseLine)
+		}
+		if clientDisconnected {
+			return
+		}
 		select {
 		case <-c.Request.Context().Done():
 			clientDisconnected = true
 			return
 		default:
 		}
-		ev := StreamEvent{Type: eventType, Message: message, Data: data}
-		b, _ := json.Marshal(ev)
 		sseWriteMu.Lock()
-		_, err := fmt.Fprintf(c.Writer, "data: %s\n\n", b)
+		_, err := c.Writer.Write(sseLine)
 		if err != nil {
 			sseWriteMu.Unlock()
 			clientDisconnected = true
@@ -95,6 +109,7 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 		sendEvent("done", "", nil)
 		return
 	}
+	ssePublishConversationID = prep.ConversationID
 	if prep.CreatedNew {
 		sendEvent("conversation", "会话已创建", map[string]interface{}{
 			"conversationId": prep.ConversationID,
@@ -103,6 +118,10 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 
 	conversationID := prep.ConversationID
 	assistantMessageID := prep.AssistantMessageID
+	h.activateHITLForConversation(conversationID, req.Hitl)
+	if h.hitlManager != nil {
+		defer h.hitlManager.DeactivateConversation(conversationID)
+	}
 
 	if prep.UserMessageID != "" {
 		sendEvent("message_saved", "", map[string]interface{}{
@@ -111,12 +130,14 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 		})
 	}
 
-	progressCallback := h.createProgressCallback(conversationID, assistantMessageID, sendEvent)
-
 	baseCtx, cancelWithCause := context.WithCancelCause(context.Background())
 	taskCtx, timeoutCancel := context.WithTimeout(baseCtx, 600*time.Minute)
 	defer timeoutCancel()
 	defer cancelWithCause(nil)
+	progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, conversationID, assistantMessageID, sendEvent)
+	taskCtx = multiagent.WithHITLToolInterceptor(taskCtx, func(ctx context.Context, toolName, arguments string) (string, error) {
+		return h.interceptHITLForEinoTool(ctx, cancelWithCause, conversationID, assistantMessageID, sendEvent, toolName, arguments)
+	})
 
 	if _, err := h.tasks.StartTask(conversationID, req.Message, cancelWithCause); err != nil {
 		var errorMsg string
@@ -181,6 +202,23 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 			return
 		}
 
+		if errors.Is(runErr, context.DeadlineExceeded) || errors.Is(context.Cause(taskCtx), context.DeadlineExceeded) {
+			taskStatus = "timeout"
+			h.tasks.UpdateTaskStatus(conversationID, taskStatus)
+			timeoutMsg := "任务执行超时，已自动终止。"
+			if assistantMessageID != "" {
+				_, _ = h.db.Exec("UPDATE messages SET content = ? WHERE id = ?", timeoutMsg, assistantMessageID)
+				_ = h.db.AddProcessDetail(assistantMessageID, conversationID, "timeout", timeoutMsg, nil)
+			}
+			sendEvent("error", timeoutMsg, map[string]interface{}{
+				"conversationId": conversationID,
+				"messageId":      assistantMessageID,
+				"errorType":      "timeout",
+			})
+			sendEvent("done", "", map[string]interface{}{"conversationId": conversationID})
+			return
+		}
+
 		h.logger.Error("Eino DeepAgent 执行失败", zap.Error(runErr))
 		taskStatus = "failed"
 		h.tasks.UpdateTaskStatus(conversationID, taskStatus)
@@ -251,9 +289,22 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 		c.JSON(status, gin.H{"error": msg})
 		return
 	}
+	h.activateHITLForConversation(prep.ConversationID, req.Hitl)
+	if h.hitlManager != nil {
+		defer h.hitlManager.DeactivateConversation(prep.ConversationID)
+	}
+
+	baseCtx, cancelWithCause := context.WithCancelCause(c.Request.Context())
+	defer cancelWithCause(nil)
+	taskCtx, timeoutCancel := context.WithTimeout(baseCtx, 600*time.Minute)
+	defer timeoutCancel()
+	progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, prep.ConversationID, prep.AssistantMessageID, nil)
+	taskCtx = multiagent.WithHITLToolInterceptor(taskCtx, func(ctx context.Context, toolName, arguments string) (string, error) {
+		return h.interceptHITLForEinoTool(ctx, cancelWithCause, prep.ConversationID, prep.AssistantMessageID, nil, toolName, arguments)
+	})
 
 	result, runErr := multiagent.RunDeepAgent(
-		c.Request.Context(),
+		taskCtx,
 		h.config,
 		&h.config.MultiAgent,
 		h.agent,
@@ -262,7 +313,7 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 		prep.FinalMessage,
 		prep.History,
 		prep.RoleTools,
-		nil,
+		progressCallback,
 		h.agentsMarkdownDir,
 		strings.TrimSpace(req.Orchestration),
 	)

internal/handler/openapi.go

@@ -6197,7 +6197,7 @@ func (h *OpenAPIHandler) GetConversationResults(c *gin.Context) {
 	}
 
 	// 获取漏洞列表
-	vulnList, err := h.db.ListVulnerabilities(1000, 0, "", conversationID, "", "")
+	vulnList, err := h.db.ListVulnerabilities(1000, 0, "", conversationID, "", "", "", "", "")
 	if err != nil {
 		h.logger.Warn("获取漏洞列表失败", zap.Error(err))
 		vulnList = []*database.Vulnerability{}

internal/handler/task_event_bus.go

@@ -0,0 +1,116 @@
+package handler
+
+import "sync"
+
+// TaskEventBus 将主 SSE 连接上的事件镜像给后订阅的客户端（例如刷新页面后、HITL 审批通过需继续收事件）。
+// 每个 payload 为完整 SSE 行： "data: {...}\n\n"
+type TaskEventBus struct {
+	mu   sync.RWMutex
+	subs map[string]map[*taskEventSub]struct{}
+}
+
+type taskEventSub struct {
+	mu     sync.Mutex
+	ch     chan []byte
+	closed bool
+}
+
+func (s *taskEventSub) sendNonBlocking(line []byte) bool {
+	if s == nil {
+		return false
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.closed {
+		return false
+	}
+	select {
+	case s.ch <- line:
+		return true
+	default:
+		return false
+	}
+}
+
+func (s *taskEventSub) closeOnce() {
+	if s == nil {
+		return
+	}
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if s.closed {
+		return
+	}
+	s.closed = true
+	close(s.ch)
+}
+
+func NewTaskEventBus() *TaskEventBus {
+	return &TaskEventBus{
+		subs: make(map[string]map[*taskEventSub]struct{}),
+	}
+}
+
+// Subscribe 注册订阅；cancel 时需调用 Unsubscribe。
+func (b *TaskEventBus) Subscribe(conversationID string) (sub *taskEventSub, ch <-chan []byte) {
+	chBuf := make(chan []byte, 256)
+	sub = &taskEventSub{ch: chBuf}
+	b.mu.Lock()
+	if b.subs[conversationID] == nil {
+		b.subs[conversationID] = make(map[*taskEventSub]struct{})
+	}
+	b.subs[conversationID][sub] = struct{}{}
+	b.mu.Unlock()
+	return sub, chBuf
+}
+
+func (b *TaskEventBus) Unsubscribe(conversationID string, sub *taskEventSub) {
+	if sub == nil {
+		return
+	}
+	b.mu.Lock()
+	m, ok := b.subs[conversationID]
+	if !ok {
+		b.mu.Unlock()
+		return
+	}
+	delete(m, sub)
+	if len(m) == 0 {
+		delete(b.subs, conversationID)
+	}
+	b.mu.Unlock()
+	sub.closeOnce()
+}
+
+// Publish 非阻塞投递；慢消费者丢帧（HITL 场景以最新状态为准，丢帧可接受）。
+func (b *TaskEventBus) Publish(conversationID string, line []byte) {
+	if b == nil || conversationID == "" || len(line) == 0 {
+		return
+	}
+	b.mu.RLock()
+	m := b.subs[conversationID]
+	subs := make([]*taskEventSub, 0, len(m))
+	for s := range m {
+		subs = append(subs, s)
+	}
+	b.mu.RUnlock()
+
+	cp := append([]byte(nil), line...)
+	for _, s := range subs {
+		s.sendNonBlocking(cp)
+	}
+}
+
+// CloseConversation 任务结束时关闭该会话所有订阅 channel。
+func (b *TaskEventBus) CloseConversation(conversationID string) {
+	if b == nil || conversationID == "" {
+		return
+	}
+	b.mu.Lock()
+	m := b.subs[conversationID]
+	delete(b.subs, conversationID)
+	b.mu.Unlock()
+	for sub := range m {
+		sub.closeOnce()
+	}
+}

internal/handler/task_manager.go

@@ -35,11 +35,12 @@ type CompletedTask struct {
 
 // AgentTaskManager 管理正在运行的Agent任务
 type AgentTaskManager struct {
-	mu             sync.RWMutex
-	tasks          map[string]*AgentTask
-	completedTasks []*CompletedTask // 最近完成的任务历史
-	maxHistorySize int              // 最大历史记录数
-	historyRetention time.Duration  // 历史记录保留时间
+	mu               sync.RWMutex
+	tasks            map[string]*AgentTask
+	completedTasks   []*CompletedTask // 最近完成的任务历史
+	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
 }
 
 const (
@@ -56,13 +57,27 @@ func NewAgentTaskManager() *AgentTaskManager {
 	m := &AgentTaskManager{
 		tasks:            make(map[string]*AgentTask),
 		completedTasks:   make([]*CompletedTask, 0),
-		maxHistorySize:   50,              // 最多保留50条历史记录
-		historyRetention: 24 * time.Hour,  // 保留24小时
+		maxHistorySize:   50,             // 最多保留50条历史记录
+		historyRetention: 24 * time.Hour, // 保留24小时
 	}
 	go m.runStuckCancellingCleanup()
 	return m
 }
 
+// SetTaskEventBus 设置任务事件总线（与 AgentHandler 共用同一实例）。
+func (m *AgentTaskManager) SetTaskEventBus(b *TaskEventBus) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	m.eventBus = b
+}
+
+// GetTask 返回运行中任务（无则 nil）。
+func (m *AgentTaskManager) GetTask(conversationID string) *AgentTask {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+	return m.tasks[conversationID]
+}
+
 // runStuckCancellingCleanup 定期将长时间处于「取消中」的任务强制结束，避免卡住无法发新消息
 func (m *AgentTaskManager) runStuckCancellingCleanup() {
 	ticker := time.NewTicker(cleanupInterval)
@@ -172,10 +187,9 @@ func (m *AgentTaskManager) UpdateTaskStatus(conversationID string, status string
 // FinishTask 完成任务并从管理器中移除
 func (m *AgentTaskManager) FinishTask(conversationID string, finalStatus string) {
 	m.mu.Lock()
-	defer m.mu.Unlock()
-
 	task, exists := m.tasks[conversationID]
 	if !exists {
+		m.mu.Unlock()
 		return
 	}
 
@@ -187,26 +201,31 @@ func (m *AgentTaskManager) FinishTask(conversationID string, finalStatus string)
 	completedTask := &CompletedTask{
 		ConversationID: task.ConversationID,
 		Message:        task.Message,
-		StartedAt:       task.StartedAt,
-		CompletedAt:     time.Now(),
-		Status:          finalStatus,
+		StartedAt:      task.StartedAt,
+		CompletedAt:    time.Now(),
+		Status:         finalStatus,
 	}
-	
+
 	// 添加到历史记录
 	m.completedTasks = append(m.completedTasks, completedTask)
-	
+
 	// 清理过期和过多的历史记录
 	m.cleanupHistory()
 
 	// 从运行任务中移除
 	delete(m.tasks, conversationID)
+	bus := m.eventBus
+	m.mu.Unlock()
+	if bus != nil {
+		bus.CloseConversation(conversationID)
+	}
 }
 
 // cleanupHistory 清理过期的历史记录
 func (m *AgentTaskManager) cleanupHistory() {
 	now := time.Now()
 	cutoffTime := now.Add(-m.historyRetention)
-	
+
 	// 过滤掉过期的记录
 	validTasks := make([]*CompletedTask, 0, len(m.completedTasks))
 	for _, task := range m.completedTasks {
@@ -214,7 +233,7 @@ func (m *AgentTaskManager) cleanupHistory() {
 			validTasks = append(validTasks, task)
 		}
 	}
-	
+
 	// 如果仍然超过最大数量，只保留最新的
 	if len(validTasks) > m.maxHistorySize {
 		// 按完成时间排序，保留最新的
@@ -222,7 +241,7 @@ func (m *AgentTaskManager) cleanupHistory() {
 		start := len(validTasks) - m.maxHistorySize
 		validTasks = validTasks[start:]
 	}
-	
+
 	m.completedTasks = validTasks
 }
 
@@ -247,30 +266,30 @@ func (m *AgentTaskManager) GetActiveTasks() []*AgentTask {
 func (m *AgentTaskManager) GetCompletedTasks() []*CompletedTask {
 	m.mu.RLock()
 	defer m.mu.RUnlock()
-	
+
 	// 清理过期记录（只读锁，不影响其他操作）
 	// 注意：这里不能直接调用cleanupHistory，因为需要写锁
 	// 所以返回时过滤过期记录
 	now := time.Now()
 	cutoffTime := now.Add(-m.historyRetention)
-	
+
 	result := make([]*CompletedTask, 0, len(m.completedTasks))
 	for _, task := range m.completedTasks {
 		if task.CompletedAt.After(cutoffTime) {
 			result = append(result, task)
 		}
 	}
-	
+
 	// 按完成时间倒序排序（最新的在前）
 	// 由于是追加的，最新的在最后，需要反转
 	for i, j := 0, len(result)-1; i < j; i, j = i+1, j-1 {
 		result[i], result[j] = result[j], result[i]
 	}
-	
+
 	// 限制返回数量
 	if len(result) > m.maxHistorySize {
 		result = result[:m.maxHistorySize]
 	}
-	
+
 	return result
 }

internal/handler/vulnerability.go

@@ -1,8 +1,11 @@
 package handler
 
 import (
+	"fmt"
 	"net/http"
 	"strconv"
+	"strings"
+	"time"
 
 	"cyberstrike-ai/internal/database"
 	"github.com/gin-gonic/gin"
@@ -26,6 +29,8 @@ func NewVulnerabilityHandler(db *database.DB, logger *zap.Logger) *Vulnerability
 // CreateVulnerabilityRequest 创建漏洞请求
 type CreateVulnerabilityRequest struct {
 	ConversationID string `json:"conversation_id" binding:"required"`
+	ConversationTag string `json:"conversation_tag"`
+	TaskTag         string `json:"task_tag"`
 	Title           string `json:"title" binding:"required"`
 	Description     string `json:"description"`
 	Severity        string `json:"severity" binding:"required"`
@@ -47,6 +52,8 @@ func (h *VulnerabilityHandler) CreateVulnerability(c *gin.Context) {
 
 	vuln := &database.Vulnerability{
 		ConversationID: req.ConversationID,
+		ConversationTag: req.ConversationTag,
+		TaskTag:         req.TaskTag,
 		Title:          req.Title,
 		Description:    req.Description,
 		Severity:       req.Severity,
@@ -100,6 +107,9 @@ func (h *VulnerabilityHandler) ListVulnerabilities(c *gin.Context) {
 	conversationID := c.Query("conversation_id")
 	severity := c.Query("severity")
 	status := c.Query("status")
+	taskID := c.Query("task_id")
+	conversationTag := c.Query("conversation_tag")
+	taskTag := c.Query("task_tag")
 
 	limit, _ := strconv.Atoi(limitStr)
 	offset, _ := strconv.Atoi(offsetStr)
@@ -121,7 +131,7 @@ func (h *VulnerabilityHandler) ListVulnerabilities(c *gin.Context) {
 	}
 
 	// 获取总数
-	total, err := h.db.CountVulnerabilities(id, conversationID, severity, status)
+	total, err := h.db.CountVulnerabilities(id, conversationID, severity, status, taskID, conversationTag, taskTag)
 	if err != nil {
 		h.logger.Error("获取漏洞总数失败", zap.Error(err))
 		// 继续执行，使用0作为总数
@@ -129,7 +139,7 @@ func (h *VulnerabilityHandler) ListVulnerabilities(c *gin.Context) {
 	}
 
 	// 获取漏洞列表
-	vulnerabilities, err := h.db.ListVulnerabilities(limit, offset, id, conversationID, severity, status)
+	vulnerabilities, err := h.db.ListVulnerabilities(limit, offset, id, conversationID, severity, status, taskID, conversationTag, taskTag)
 	if err != nil {
 		h.logger.Error("获取漏洞列表失败", zap.Error(err))
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
@@ -160,6 +170,8 @@ func (h *VulnerabilityHandler) ListVulnerabilities(c *gin.Context) {
 
 // UpdateVulnerabilityRequest 更新漏洞请求
 type UpdateVulnerabilityRequest struct {
+	ConversationTag string `json:"conversation_tag"`
+	TaskTag         string `json:"task_tag"`
 	Title          string `json:"title"`
 	Description    string `json:"description"`
 	Severity       string `json:"severity"`
@@ -189,6 +201,12 @@ func (h *VulnerabilityHandler) UpdateVulnerability(c *gin.Context) {
 	}
 
 	// 更新字段
+	if req.ConversationTag != "" {
+		existing.ConversationTag = req.ConversationTag
+	}
+	if req.TaskTag != "" {
+		existing.TaskTag = req.TaskTag
+	}
 	if req.Title != "" {
 		existing.Title = req.Title
 	}
@@ -250,8 +268,9 @@ func (h *VulnerabilityHandler) DeleteVulnerability(c *gin.Context) {
 // GetVulnerabilityStats 获取漏洞统计
 func (h *VulnerabilityHandler) GetVulnerabilityStats(c *gin.Context) {
 	conversationID := c.Query("conversation_id")
+	taskID := c.Query("task_id")
 
-	stats, err := h.db.GetVulnerabilityStats(conversationID)
+	stats, err := h.db.GetVulnerabilityStats(conversationID, taskID)
 	if err != nil {
 		h.logger.Error("获取漏洞统计失败", zap.Error(err))
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
@@ -261,3 +280,184 @@ func (h *VulnerabilityHandler) GetVulnerabilityStats(c *gin.Context) {
 	c.JSON(http.StatusOK, stats)
 }
 
+// GetVulnerabilityFilterOptions 获取漏洞筛选建议项
+func (h *VulnerabilityHandler) GetVulnerabilityFilterOptions(c *gin.Context) {
+	options, err := h.db.GetVulnerabilityFilterOptions()
+	if err != nil {
+		h.logger.Error("获取漏洞筛选建议失败", zap.Error(err))
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, options)
+}
+
+// ExportVulnerabilities 导出漏洞（支持按对话/任务分组，汇总或拆分）
+func (h *VulnerabilityHandler) ExportVulnerabilities(c *gin.Context) {
+	groupBy := c.DefaultQuery("group_by", "conversation")
+	mode := c.DefaultQuery("mode", "summary")
+	if groupBy != "conversation" && groupBy != "task" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "group_by 仅支持 conversation 或 task"})
+		return
+	}
+	if mode != "summary" && mode != "split" {
+		c.JSON(http.StatusBadRequest, gin.H{"error": "mode 仅支持 summary 或 split"})
+		return
+	}
+
+	id := c.Query("id")
+	conversationID := c.Query("conversation_id")
+	severity := c.Query("severity")
+	status := c.Query("status")
+	taskID := c.Query("task_id")
+	conversationTag := c.Query("conversation_tag")
+	taskTag := c.Query("task_tag")
+
+	total, err := h.db.CountVulnerabilities(id, conversationID, severity, status, taskID, conversationTag, taskTag)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	if total == 0 {
+		c.JSON(http.StatusOK, gin.H{"mode": mode, "group_by": groupBy, "total": 0, "files": []any{}})
+		return
+	}
+
+	items, err := h.db.ListVulnerabilities(total, 0, id, conversationID, severity, status, taskID, conversationTag, taskTag)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
+	type exportFile struct {
+		FileName string `json:"filename"`
+		Content  string `json:"content"`
+	}
+	grouped := map[string][]*database.Vulnerability{}
+	for _, v := range items {
+		key := v.ConversationID
+		if groupBy == "conversation" {
+			if strings.TrimSpace(v.ConversationTag) != "" {
+				key = strings.TrimSpace(v.ConversationTag)
+			}
+		} else {
+			key = firstNonEmpty(v.TaskTag, v.TaskID, v.TaskQueueID, "unassigned-task")
+		}
+		grouped[key] = append(grouped[key], v)
+	}
+
+	files := make([]exportFile, 0)
+	nowStr := time.Now().Format("20060102-150405")
+	if mode == "summary" {
+		var b strings.Builder
+		b.WriteString("# 漏洞批量导出报告\n\n")
+		b.WriteString(fmt.Sprintf("- 导出时间: %s\n", time.Now().Format("2006-01-02 15:04:05")))
+		b.WriteString(fmt.Sprintf("- 分组维度: %s\n", groupBy))
+		b.WriteString(fmt.Sprintf("- 漏洞总数: %d\n", len(items)))
+		b.WriteString(fmt.Sprintf("- 分组数: %d\n\n", len(grouped)))
+		for group, list := range grouped {
+			b.WriteString(fmt.Sprintf("## %s (%d)\n\n", group, len(list)))
+			for _, v := range list {
+				appendVulnerabilityMarkdown(&b, v, "###")
+			}
+		}
+		files = append(files, exportFile{
+			FileName: fmt.Sprintf("vulnerability-report-%s-%s.md", groupBy, nowStr),
+			Content:  b.String(),
+		})
+	} else {
+		for group, list := range grouped {
+			var b strings.Builder
+			b.WriteString(fmt.Sprintf("# 漏洞报告 - %s\n\n", group))
+			b.WriteString(fmt.Sprintf("- 导出时间: %s\n", time.Now().Format("2006-01-02 15:04:05")))
+			b.WriteString(fmt.Sprintf("- 漏洞数量: %d\n\n", len(list)))
+			for _, v := range list {
+				appendVulnerabilityMarkdown(&b, v, "##")
+			}
+			files = append(files, exportFile{
+				FileName: fmt.Sprintf("vulnerability-%s-%s.md", sanitizeExportName(group), nowStr),
+				Content:  b.String(),
+			})
+		}
+	}
+
+	c.JSON(http.StatusOK, gin.H{
+		"mode":     mode,
+		"group_by": groupBy,
+		"total":    len(items),
+		"files":    files,
+	})
+}
+
+// appendVulnerabilityMarkdown 单条漏洞的 Markdown 片段（与单文件下载字段对齐，缺省字段不写）
+func appendVulnerabilityMarkdown(b *strings.Builder, v *database.Vulnerability, titleHeading string) {
+	b.WriteString(fmt.Sprintf("%s %s\n\n", titleHeading, v.Title))
+	b.WriteString(fmt.Sprintf("- 漏洞ID: `%s`\n", v.ID))
+	b.WriteString(fmt.Sprintf("- 严重程度: %s\n", v.Severity))
+	b.WriteString(fmt.Sprintf("- 状态: %s\n", v.Status))
+	if v.Type != "" {
+		b.WriteString(fmt.Sprintf("- 类型: %s\n", v.Type))
+	}
+	if v.Target != "" {
+		b.WriteString(fmt.Sprintf("- 目标: %s\n", v.Target))
+	}
+	b.WriteString(fmt.Sprintf("- 对话ID: `%s`\n", v.ConversationID))
+	if v.ConversationTag != "" {
+		b.WriteString(fmt.Sprintf("- 对话标签: %s\n", v.ConversationTag))
+	}
+	if v.TaskTag != "" {
+		b.WriteString(fmt.Sprintf("- 任务标签: %s\n", v.TaskTag))
+	}
+	if v.TaskID != "" {
+		b.WriteString(fmt.Sprintf("- 任务ID: `%s`\n", v.TaskID))
+	}
+	if v.TaskQueueID != "" {
+		b.WriteString(fmt.Sprintf("- 任务队列ID: `%s`\n", v.TaskQueueID))
+	}
+	if !v.CreatedAt.IsZero() {
+		b.WriteString(fmt.Sprintf("- 创建时间: %s\n", v.CreatedAt.Format("2006-01-02 15:04:05")))
+	}
+	if !v.UpdatedAt.IsZero() {
+		b.WriteString(fmt.Sprintf("- 更新时间: %s\n", v.UpdatedAt.Format("2006-01-02 15:04:05")))
+	}
+	if v.Description != "" {
+		b.WriteString("\n#### 描述\n\n")
+		b.WriteString(v.Description)
+		b.WriteString("\n")
+	}
+	if v.Proof != "" {
+		b.WriteString("\n#### 证明（POC）\n\n```\n")
+		b.WriteString(v.Proof)
+		b.WriteString("\n```\n")
+	}
+	if v.Impact != "" {
+		b.WriteString("\n#### 影响\n\n")
+		b.WriteString(v.Impact)
+		b.WriteString("\n")
+	}
+	if v.Recommendation != "" {
+		b.WriteString("\n#### 修复建议\n\n")
+		b.WriteString(v.Recommendation)
+		b.WriteString("\n")
+	}
+	b.WriteString("\n")
+}
+
+func firstNonEmpty(values ...string) string {
+	for _, v := range values {
+		trimmed := strings.TrimSpace(v)
+		if trimmed != "" {
+			return trimmed
+		}
+	}
+	return ""
+}
+
+func sanitizeExportName(raw string) string {
+	name := strings.TrimSpace(raw)
+	if name == "" {
+		return "unknown"
+	}
+	replacer := strings.NewReplacer("/", "-", "\\", "-", ":", "-", "*", "-", "?", "-", "\"", "-", "<", "-", ">", "-", "|", "-")
+	return replacer.Replace(name)
+}
+

internal/multiagent/eino_adk_run_loop.go

@@ -18,6 +18,21 @@ import (
 	"go.uber.org/zap"
 )
 
+func isEinoIterationLimitError(err error) bool {
+	if err == nil {
+		return false
+	}
+	msg := strings.ToLower(strings.TrimSpace(err.Error()))
+	if msg == "" {
+		return false
+	}
+	return strings.Contains(msg, "max iteration") ||
+		strings.Contains(msg, "maximum iteration") ||
+		strings.Contains(msg, "maximum iterations") ||
+		strings.Contains(msg, "iteration limit") ||
+		strings.Contains(msg, "达到最大迭代")
+}
+
 // einoADKRunLoopArgs 将 Eino adk.Runner 事件循环从 RunDeepAgent / RunEinoSingleChatModelAgent 中抽出复用。
 type einoADKRunLoopArgs struct {
 	OrchMode             string
@@ -97,7 +112,8 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 	var lastRunMsgs []adk.Message
 	var lastAssistant string
 	var lastPlanExecuteExecutor string
-	var retryHints []adk.Message
+	msgs := append([]adk.Message(nil), baseMsgs...)
+	runAccumulatedMsgs := append([]adk.Message(nil), msgs...)
 
 	emptyHint := strings.TrimSpace(args.EmptyResponseMessage)
 	if emptyHint == "" {
@@ -105,29 +121,17 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 			"（Eino 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。）"
 	}
 
-attemptLoop:
-	for attempt := 0; attempt < maxToolCallRecoveryAttempts; attempt++ {
-		msgs := make([]adk.Message, 0, len(baseMsgs)+len(retryHints))
-		msgs = append(msgs, baseMsgs...)
-		msgs = append(msgs, retryHints...)
-
-		if attempt > 0 {
-			mcpIDsMu.Lock()
-			*mcpIDs = (*mcpIDs)[:0]
-			mcpIDsMu.Unlock()
-		}
-
-		lastAssistant = ""
-		lastPlanExecuteExecutor = ""
-		var reasoningStreamSeq int64
-		var einoSubReplyStreamSeq int64
-		toolEmitSeen := make(map[string]struct{})
-		var einoMainRound int
-		var einoLastAgent string
-		subAgentToolStep := make(map[string]int)
-		pendingByID := make(map[string]toolCallPendingInfo)
-		pendingQueueByAgent := make(map[string][]string)
-		markPending := func(tc toolCallPendingInfo) {
+	lastAssistant = ""
+	lastPlanExecuteExecutor = ""
+	var reasoningStreamSeq int64
+	var einoSubReplyStreamSeq int64
+	toolEmitSeen := make(map[string]struct{})
+	var einoMainRound int
+	var einoLastAgent string
+	subAgentToolStep := make(map[string]int)
+	pendingByID := make(map[string]toolCallPendingInfo)
+	pendingQueueByAgent := make(map[string][]string)
+	markPending := func(tc toolCallPendingInfo) {
 			if tc.ToolCallID == "" {
 				return
 			}
@@ -203,8 +207,59 @@ attemptLoop:
 				}
 			}
 		}
-		runner := adk.NewRunner(ctx, runnerCfg)
-		iter := runner.Run(ctx, msgs)
+	runner := adk.NewRunner(ctx, runnerCfg)
+	iter := runner.Run(ctx, msgs)
+	handleRunErr := func(runErr error) error {
+			if runErr == nil {
+				return nil
+			}
+			if errors.Is(runErr, context.DeadlineExceeded) {
+				flushAllPendingAsFailed(runErr)
+				if progress != nil {
+					progress("error", runErr.Error(), map[string]interface{}{
+						"conversationId": conversationID,
+						"source":         "eino",
+						"errorKind":      "timeout",
+					})
+				}
+				return runErr
+			}
+			// context.Canceled 是唯一应当直接终止编排的错误（用户关闭页面、主动停止等）。
+			if errors.Is(runErr, context.Canceled) {
+				flushAllPendingAsFailed(runErr)
+				if progress != nil {
+					progress("error", runErr.Error(), map[string]interface{}{
+						"conversationId": conversationID,
+						"source":         "eino",
+					})
+				}
+				return runErr
+			}
+			if isEinoIterationLimitError(runErr) {
+				flushAllPendingAsFailed(runErr)
+				if progress != nil {
+					progress("iteration_limit_reached", runErr.Error(), map[string]interface{}{
+						"conversationId": conversationID,
+						"source":         "eino",
+						"orchestration":  orchMode,
+					})
+					progress("error", runErr.Error(), map[string]interface{}{
+						"conversationId": conversationID,
+						"source":         "eino",
+						"errorKind":      "iteration_limit",
+					})
+				}
+				return runErr
+			}
+			flushAllPendingAsFailed(runErr)
+			if progress != nil {
+				progress("error", runErr.Error(), map[string]interface{}{
+					"conversationId": conversationID,
+					"source":         "eino",
+				})
+			}
+			return runErr
+		}
 
 		for {
 			// 检测 context 取消（用户关闭浏览器、请求超时等），flush pending 工具状态避免 UI 卡在 "执行中"。
@@ -223,68 +278,28 @@ attemptLoop:
 
 			ev, ok := iter.Next()
 			if !ok {
-				lastRunMsgs = msgs
-				break attemptLoop
+				if len(pendingByID) > 0 {
+					orphanCount := len(pendingByID)
+					flushAllPendingAsFailed(errors.New("pending tool call missing result before run completion"))
+					if progress != nil {
+						progress("eino_pending_orphaned", "pending tool calls were force-closed at run end", map[string]interface{}{
+							"conversationId": conversationID,
+							"source":         "eino",
+							"orchestration":  orchMode,
+							"pendingCount":   orphanCount,
+						})
+					}
+				}
+				lastRunMsgs = runAccumulatedMsgs
+				break
 			}
 			if ev == nil {
 				continue
 			}
 			if ev.Err != nil {
-				// context.Canceled 是唯一应当直接终止编排的错误（用户关闭页面、主动停止等）。
-				if errors.Is(ev.Err, context.Canceled) {
-					flushAllPendingAsFailed(ev.Err)
-					if progress != nil {
-						progress("error", ev.Err.Error(), map[string]interface{}{
-							"conversationId": conversationID,
-							"source":         "eino",
-						})
-					}
-					return nil, ev.Err
+				if retErr := handleRunErr(ev.Err); retErr != nil {
+					return nil, retErr
 				}
-
-				canRetry := attempt+1 < maxToolCallRecoveryAttempts
-				if !canRetry {
-					// 重试次数已耗尽，终止。
-					flushAllPendingAsFailed(ev.Err)
-					if progress != nil {
-						progress("error", ev.Err.Error(), map[string]interface{}{
-							"conversationId": conversationID,
-							"source":         "eino",
-						})
-					}
-					return nil, ev.Err
-				}
-
-				// 区分错误类型以选择最合适的纠错提示，但无论哪种都执行重试（default-soft）。
-				var hint *schema.Message
-				var reason, timelineMsg string
-				if isRecoverableToolCallArgumentsJSONError(ev.Err) {
-					hint = toolCallArgumentsJSONRetryHint()
-					reason = "invalid_tool_arguments_json"
-					timelineMsg = toolCallArgumentsJSONRecoveryTimelineMessage(attempt)
-				} else {
-					hint = toolExecutionRetryHint()
-					reason = "tool_execution_error"
-					timelineMsg = toolExecutionRecoveryTimelineMessage(attempt)
-				}
-
-				if logger != nil {
-					logger.Warn("eino: recoverable error, will retry with corrective hint",
-						zap.Error(ev.Err), zap.Int("attempt", attempt), zap.String("reason", reason))
-				}
-				flushAllPendingAsFailed(ev.Err)
-				retryHints = append(retryHints, hint)
-				if progress != nil {
-					progress("eino_recovery", timelineMsg, map[string]interface{}{
-						"conversationId": conversationID,
-						"source":         "eino",
-						"einoRetry":      attempt,
-						"runIndex":       attempt + 1,
-						"maxRuns":        maxToolCallRecoveryAttempts,
-						"reason":         reason,
-					})
-				}
-				continue attemptLoop
 			}
 			if ev.AgentName != "" && progress != nil {
 				iterEinoAgent := orchestratorName
@@ -338,6 +353,7 @@ attemptLoop:
 				var subAssistantBuf strings.Builder
 				var subReplyStreamID string
 				var mainAssistantBuf strings.Builder
+				var streamRecvErr error
 				for {
 					chunk, rerr := mv.MessageStream.Recv()
 					if rerr != nil {
@@ -350,6 +366,7 @@ attemptLoop:
 								zap.String("agent", ev.AgentName),
 								zap.Int("toolFragments", len(toolStreamFragments)))
 						}
+						streamRecvErr = rerr
 						break
 					}
 					if chunk == nil {
@@ -418,6 +435,7 @@ attemptLoop:
 				if streamsMainAssistant(ev.AgentName) {
 					if s := strings.TrimSpace(mainAssistantBuf.String()); s != "" {
 						lastAssistant = s
+						runAccumulatedMsgs = append(runAccumulatedMsgs, schema.AssistantMessage(s, nil))
 						if orchMode == "plan_execute" && strings.EqualFold(strings.TrimSpace(ev.AgentName), "executor") {
 							lastPlanExecuteExecutor = UnwrapPlanExecuteUserText(s)
 						}
@@ -448,6 +466,19 @@ attemptLoop:
 					lastToolChunk = &schema.Message{ToolCalls: merged}
 				}
 				tryEmitToolCallsOnce(lastToolChunk, ev.AgentName, orchestratorName, conversationID, progress, toolEmitSeen, subAgentToolStep, markPending)
+				if streamRecvErr != nil {
+					if progress != nil {
+						progress("eino_stream_error", streamRecvErr.Error(), map[string]interface{}{
+							"conversationId": conversationID,
+							"source":         "eino",
+							"einoAgent":      ev.AgentName,
+							"einoRole":       einoRoleTag(ev.AgentName),
+						})
+					}
+					if retErr := handleRunErr(streamRecvErr); retErr != nil {
+						return nil, retErr
+					}
+				}
 				continue
 			}
 
@@ -455,6 +486,7 @@ attemptLoop:
 			if gerr != nil || msg == nil {
 				continue
 			}
+			runAccumulatedMsgs = append(runAccumulatedMsgs, msg)
 			tryEmitToolCallsOnce(mergeMessageToolCalls(msg), ev.AgentName, orchestratorName, conversationID, progress, toolEmitSeen, subAgentToolStep, markPending)
 
 			if mv.Role == schema.Assistant {
@@ -554,7 +586,6 @@ attemptLoop:
 				progress("tool_result", fmt.Sprintf("工具结果 (%s)", toolName), data)
 			}
 		}
-	}
 
 	mcpIDsMu.Lock()
 	ids := append([]string(nil), *mcpIDs...)

internal/multiagent/eino_single_runner.go

@@ -159,6 +159,7 @@ func RunEinoSingleChatModelAgent(
 			Tools:               mainToolsForCfg,
 			UnknownToolsHandler: einomcp.UnknownToolReminderHandler(),
 			ToolCallMiddlewares: []compose.ToolMiddleware{
+				{Invokable: hitlToolCallMiddleware()},
 				{Invokable: softRecoveryToolCallMiddleware()},
 			},
 		},

internal/multiagent/eino_summarize.go

@@ -57,19 +57,30 @@ func newEinoSummarizationMiddleware(
 	if modelName == "" {
 		modelName = "gpt-4o"
 	}
+	tokenCounter := einoSummarizationTokenCounter(modelName)
+	recentTrailMax := trigger / 4
+	if recentTrailMax < 2048 {
+		recentTrailMax = 2048
+	}
+	if recentTrailMax > trigger/2 {
+		recentTrailMax = trigger / 2
+	}
 
 	mw, err := summarization.New(ctx, &summarization.Config{
 		Model: summaryModel,
 		Trigger: &summarization.TriggerCondition{
 			ContextTokens: trigger,
 		},
-		TokenCounter:       einoSummarizationTokenCounter(modelName),
+		TokenCounter:       tokenCounter,
 		UserInstruction:    einoSummarizeUserInstruction,
 		EmitInternalEvents: false,
 		PreserveUserMessages: &summarization.PreserveUserMessages{
 			Enabled:   true,
 			MaxTokens: preserveMax,
 		},
+		Finalize: func(ctx context.Context, originalMessages []adk.Message, summary adk.Message) ([]adk.Message, error) {
+			return summarizeFinalizeWithRecentAssistantToolTrail(ctx, originalMessages, summary, tokenCounter, recentTrailMax)
+		},
 		Callback: func(ctx context.Context, before, after adk.ChatModelAgentState) error {
 			if logger == nil {
 				return nil
@@ -89,6 +100,108 @@ func newEinoSummarizationMiddleware(
 	return mw, nil
 }
 
+// summarizeFinalizeWithRecentAssistantToolTrail 在摘要消息后保留最近 assistant/tool 轨迹，避免压缩后执行链断裂。
+func summarizeFinalizeWithRecentAssistantToolTrail(
+	ctx context.Context,
+	originalMessages []adk.Message,
+	summary adk.Message,
+	tokenCounter summarization.TokenCounterFunc,
+	recentTrailTokenBudget int,
+) ([]adk.Message, error) {
+	systemMsgs := make([]adk.Message, 0, len(originalMessages))
+	nonSystem := make([]adk.Message, 0, len(originalMessages))
+	for _, msg := range originalMessages {
+		if msg == nil {
+			continue
+		}
+		if msg.Role == schema.System {
+			systemMsgs = append(systemMsgs, msg)
+			continue
+		}
+		nonSystem = append(nonSystem, msg)
+	}
+
+	if recentTrailTokenBudget <= 0 || len(nonSystem) == 0 {
+		out := make([]adk.Message, 0, len(systemMsgs)+1)
+		out = append(out, systemMsgs...)
+		out = append(out, summary)
+		return out, nil
+	}
+
+	selectedReverse := make([]adk.Message, 0, 8)
+	seen := make(map[adk.Message]struct{})
+	totalTokens := 0
+	assistantToolKept := 0
+	const minAssistantToolTrail = 4
+
+	tryKeep := func(msg adk.Message) (bool, error) {
+		if msg == nil {
+			return false, nil
+		}
+		if _, ok := seen[msg]; ok {
+			return false, nil
+		}
+		n, err := tokenCounter(ctx, &summarization.TokenCounterInput{Messages: []adk.Message{msg}})
+		if err != nil {
+			return false, err
+		}
+		if n <= 0 {
+			n = 1
+		}
+		if totalTokens+n > recentTrailTokenBudget {
+			return false, nil
+		}
+		totalTokens += n
+		selectedReverse = append(selectedReverse, msg)
+		seen[msg] = struct{}{}
+		return true, nil
+	}
+
+	// 优先保留最近 assistant/tool，确保执行轨迹可续跑。
+	for i := len(nonSystem) - 1; i >= 0; i-- {
+		msg := nonSystem[i]
+		if msg.Role != schema.Assistant && msg.Role != schema.Tool {
+			continue
+		}
+		ok, err := tryKeep(msg)
+		if err != nil {
+			return nil, err
+		}
+		if ok {
+			assistantToolKept++
+		}
+		if assistantToolKept >= minAssistantToolTrail {
+			break
+		}
+	}
+
+	// 在预算内回填更多最近消息，保持短链路上下文。
+	for i := len(nonSystem) - 1; i >= 0; i-- {
+		_, exists := seen[nonSystem[i]]
+		if exists {
+			continue
+		}
+		ok, err := tryKeep(nonSystem[i])
+		if err != nil {
+			return nil, err
+		}
+		if !ok {
+			break
+		}
+	}
+
+	selected := make([]adk.Message, 0, len(selectedReverse))
+	for i := len(selectedReverse) - 1; i >= 0; i-- {
+		selected = append(selected, selectedReverse[i])
+	}
+
+	out := make([]adk.Message, 0, len(systemMsgs)+1+len(selected))
+	out = append(out, systemMsgs...)
+	out = append(out, summary)
+	out = append(out, selected...)
+	return out, nil
+}
+
 func einoSummarizationTokenCounter(openAIModel string) summarization.TokenCounterFunc {
 	tc := agent.NewTikTokenCounter()
 	return func(ctx context.Context, input *summarization.TokenCounterInput) (int, error) {

internal/multiagent/hitl_middleware.go

@@ -0,0 +1,81 @@
+package multiagent
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/compose"
+)
+
+type hitlInterceptorKey struct{}
+
+type HITLToolInterceptor func(ctx context.Context, toolName, arguments string) (string, error)
+
+type humanRejectError struct {
+	reason string
+}
+
+func (e *humanRejectError) Error() string {
+	if strings.TrimSpace(e.reason) == "" {
+		return "rejected by user"
+	}
+	return "rejected by user: " + strings.TrimSpace(e.reason)
+}
+
+func NewHumanRejectError(reason string) error {
+	return &humanRejectError{reason: strings.TrimSpace(reason)}
+}
+
+func IsHumanRejectError(err error) bool {
+	var target *humanRejectError
+	return errors.As(err, &target)
+}
+
+func WithHITLToolInterceptor(ctx context.Context, fn HITLToolInterceptor) context.Context {
+	if fn == nil {
+		return ctx
+	}
+	return context.WithValue(ctx, hitlInterceptorKey{}, fn)
+}
+
+func hitlToolCallMiddleware() compose.InvokableToolMiddleware {
+	return func(next compose.InvokableToolEndpoint) compose.InvokableToolEndpoint {
+		return func(ctx context.Context, input *compose.ToolInput) (*compose.ToolOutput, error) {
+			if input != nil {
+				if fn, ok := ctx.Value(hitlInterceptorKey{}).(HITLToolInterceptor); ok && fn != nil {
+					edited, err := fn(ctx, input.Name, input.Arguments)
+					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.",
+								input.Name, strings.TrimSpace(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。
+							if strings.EqualFold(strings.TrimSpace(input.Name), adk.TransferToAgentToolName) {
+								_ = compose.ProcessState[*adk.State](ctx, func(_ context.Context, st *adk.State) error {
+									if st == nil {
+										return nil
+									}
+									st.ReturnDirectlyToolCallID = ""
+									st.HasReturnDirectly = false
+									st.ReturnDirectlyEvent = nil
+									return nil
+								})
+							}
+							return &compose.ToolOutput{Result: msg}, nil
+						}
+						return nil, err
+					}
+					if edited != "" {
+						input.Arguments = edited
+					}
+				}
+			}
+			return next(ctx, input)
+		}
+	}
+}

internal/multiagent/orchestrator_instruction.go

@@ -210,7 +210,7 @@ func DefaultSupervisorOrchestratorInstruction() string {
 
 ## transfer 交接与防重复劳动
 
-- 每次 transfer 前，在**本条助手正文**中写清交接包：已知主域、关键子域或主机短表、已识别端口与服务、上轮已达成共识的结论要点；勿仅依赖历史里的超长工具原始输出（上下文摘要后专家可能看不到细节）。
+- **把专家当作刚走进房间的同事——它没看过你的对话，不知道你做了什么，也不了解这个任务为什么重要。** 每次 transfer 前，在**本条助手正文**中写清交接包：已知主域、关键子域或主机短表、已识别端口与服务、上轮已达成共识的结论要点；勿仅依赖历史里的超长工具原始输出（上下文摘要后专家可能看不到细节）。
 - 写清本轮**唯一子目标**与**禁止项**（例如：不得再做全量子域枚举；仅对下列目标做 MQTT 或认证验证）。
 - 验证、利用、协议深挖应 transfer 给**对应专项**子代理；避免把「仅剩验证」的工作交给侦察类（recon）导致其从全量枚举起手。
 - 同一目标多次串行 transfer 时，每一次交接包都要带上**截至当前的共识事实**增量，勿假设专家已读过上一轮专家的隐性推理。

internal/multiagent/runner.go

@@ -268,6 +268,7 @@ func RunDeepAgent(
 						Tools:               subToolsForCfg,
 						UnknownToolsHandler: einomcp.UnknownToolReminderHandler(),
 						ToolCallMiddlewares: []compose.ToolMiddleware{
+							{Invokable: hitlToolCallMiddleware()},
 							{Invokable: softRecoveryToolCallMiddleware()},
 						},
 					},
@@ -341,6 +342,9 @@ 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()}
+	if mw := newTaskContextEnrichMiddleware(userMessage, history, ma.SubAgentUserContextMaxRunes); mw != nil {
+		deepHandlers = append(deepHandlers, mw)
+	}
 	if len(mainOrchestratorPre) > 0 {
 		deepHandlers = append(deepHandlers, mainOrchestratorPre...)
 	}
@@ -363,6 +367,7 @@ func RunDeepAgent(
 			Tools:               mainToolsForCfg,
 			UnknownToolsHandler: einomcp.UnknownToolReminderHandler(),
 			ToolCallMiddlewares: []compose.ToolMiddleware{
+				{Invokable: hitlToolCallMiddleware()},
 				{Invokable: softRecoveryToolCallMiddleware()},
 			},
 		},

internal/multiagent/sub_agent_context.go

@@ -0,0 +1,145 @@
+package multiagent
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+
+	"cyberstrike-ai/internal/agent"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/components/tool"
+)
+
+const defaultSubAgentUserContextMaxRunes = 2000
+
+// taskContextEnrichMiddleware intercepts "task" tool calls on the orchestrator
+// and appends the user's original conversation messages to the task description.
+// This ensures sub-agents always receive the full user intent (target URLs,
+// scope, etc.) even when the orchestrator forgets to include them.
+//
+// Design: user context is injected into the task description (per-task), NOT
+// into the sub-agent's Instruction (system prompt). This keeps sub-agent
+// Instructions clean as pure role definitions while attaching context to the
+// specific delegation — aligned with Claude Code's agent design philosophy.
+type taskContextEnrichMiddleware struct {
+	adk.BaseChatModelAgentMiddleware
+	supplement string // pre-built user context block
+}
+
+// newTaskContextEnrichMiddleware returns a middleware that enriches task
+// descriptions with user conversation context. Returns nil if disabled
+// (maxRunes < 0) or no user messages exist.
+func newTaskContextEnrichMiddleware(userMessage string, history []agent.ChatMessage, maxRunes int) adk.ChatModelAgentMiddleware {
+	supplement := buildUserContextSupplement(userMessage, history, maxRunes)
+	if supplement == "" {
+		return nil
+	}
+	return &taskContextEnrichMiddleware{supplement: supplement}
+}
+
+func (m *taskContextEnrichMiddleware) WrapInvokableToolCall(
+	ctx context.Context,
+	endpoint adk.InvokableToolCallEndpoint,
+	tCtx *adk.ToolContext,
+) (adk.InvokableToolCallEndpoint, error) {
+	if tCtx == nil || !strings.EqualFold(strings.TrimSpace(tCtx.Name), "task") {
+		return endpoint, nil
+	}
+	return func(ctx context.Context, argumentsInJSON string, opts ...tool.Option) (string, error) {
+		enriched := m.enrichTaskDescription(argumentsInJSON)
+		return endpoint(ctx, enriched, opts...)
+	}, nil
+}
+
+// enrichTaskDescription parses the task JSON arguments, appends user context
+// to the "description" field, and re-serializes. Falls back to the original
+// JSON if parsing fails or no description field exists.
+func (m *taskContextEnrichMiddleware) enrichTaskDescription(argsJSON string) string {
+	var raw map[string]interface{}
+	if err := json.Unmarshal([]byte(argsJSON), &raw); err != nil {
+		return argsJSON
+	}
+	desc, ok := raw["description"].(string)
+	if !ok {
+		return argsJSON
+	}
+	raw["description"] = desc + m.supplement
+	enriched, err := json.Marshal(raw)
+	if err != nil {
+		return argsJSON
+	}
+	return string(enriched)
+}
+
+// buildUserContextSupplement collects user messages from conversation history
+// and the current message, returning a formatted block to append to task
+// descriptions. Returns "" if disabled or no user messages exist.
+func buildUserContextSupplement(userMessage string, history []agent.ChatMessage, maxRunes int) string {
+	if maxRunes < 0 {
+		return ""
+	}
+	if maxRunes == 0 {
+		maxRunes = defaultSubAgentUserContextMaxRunes
+	}
+
+	var userMsgs []string
+	for _, h := range history {
+		if h.Role == "user" {
+			if m := strings.TrimSpace(h.Content); m != "" {
+				userMsgs = append(userMsgs, m)
+			}
+		}
+	}
+	if um := strings.TrimSpace(userMessage); um != "" {
+		if len(userMsgs) == 0 || userMsgs[len(userMsgs)-1] != um {
+			userMsgs = append(userMsgs, um)
+		}
+	}
+	if len(userMsgs) == 0 {
+		return ""
+	}
+
+	joined := strings.Join(userMsgs, "\n---\n")
+	if 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
+}
+
+// truncateKeepFirstLast keeps the first and last user messages, giving each
+// half the rune budget. The first message typically contains target info;
+// the last contains the current instruction.
+func truncateKeepFirstLast(msgs []string, maxRunes int) string {
+	if len(msgs) == 1 {
+		return truncateRunes(msgs[0], maxRunes)
+	}
+
+	first := msgs[0]
+	last := msgs[len(msgs)-1]
+	sep := "\n---\n...(中间对话省略)...\n---\n"
+	sepLen := len([]rune(sep))
+
+	budget := maxRunes - sepLen
+	if budget <= 0 {
+		return truncateRunes(first+"\n---\n"+last, maxRunes)
+	}
+
+	halfBudget := budget / 2
+	firstTrunc := truncateRunes(first, halfBudget)
+	lastTrunc := truncateRunes(last, budget-len([]rune(firstTrunc)))
+
+	return firstTrunc + sep + lastTrunc
+}
+
+func truncateRunes(s string, max int) string {
+	rs := []rune(s)
+	if len(rs) <= max {
+		return s
+	}
+	if max <= 0 {
+		return ""
+	}
+	return string(rs[:max])
+}

internal/multiagent/sub_agent_context_test.go

@@ -0,0 +1,182 @@
+package multiagent
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"cyberstrike-ai/internal/agent"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/components/tool"
+)
+
+// --- buildUserContextSupplement tests ---
+
+func TestBuildUserContextSupplement_SingleMessage(t *testing.T) {
+	result := buildUserContextSupplement("http://8.163.32.73:8081 测试命令执行", nil, 0)
+	if result == "" {
+		t.Fatal("expected non-empty supplement")
+	}
+	if !strings.Contains(result, "http://8.163.32.73:8081") {
+		t.Error("expected URL in supplement")
+	}
+}
+
+func TestBuildUserContextSupplement_MultiTurn(t *testing.T) {
+	history := []agent.ChatMessage{
+		{Role: "user", Content: "http://8.163.32.73:8081 这是一个pikachu靶场，尝试测试命令执行"},
+		{Role: "assistant", Content: "好的，我来测试..."},
+		{Role: "user", Content: "继续，并持久化webshell"},
+		{Role: "assistant", Content: "正在处理..."},
+	}
+	result := buildUserContextSupplement("你好", history, 0)
+	if !strings.Contains(result, "http://8.163.32.73:8081") {
+		t.Error("expected first turn URL to be preserved")
+	}
+	if !strings.Contains(result, "你好") {
+		t.Error("expected current message")
+	}
+}
+
+func TestBuildUserContextSupplement_Empty(t *testing.T) {
+	if result := buildUserContextSupplement("", nil, 0); result != "" {
+		t.Errorf("expected empty, got %q", result)
+	}
+}
+
+func TestBuildUserContextSupplement_Deduplicate(t *testing.T) {
+	history := []agent.ChatMessage{{Role: "user", Content: "你好"}}
+	result := buildUserContextSupplement("你好", history, 0)
+	if strings.Count(result, "你好") != 1 {
+		t.Errorf("expected '你好' once, got: %s", result)
+	}
+}
+
+func TestBuildUserContextSupplement_SkipsNonUser(t *testing.T) {
+	history := []agent.ChatMessage{
+		{Role: "user", Content: "目标是 10.0.0.1"},
+		{Role: "assistant", Content: "不应该出现"},
+	}
+	result := buildUserContextSupplement("确认", history, 0)
+	if strings.Contains(result, "不应该出现") {
+		t.Error("assistant message should not be included")
+	}
+}
+
+func TestBuildUserContextSupplement_DisabledByNegative(t *testing.T) {
+	if result := buildUserContextSupplement("test", nil, -1); result != "" {
+		t.Errorf("expected empty when disabled, got %q", result)
+	}
+}
+
+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"
+	body := strings.TrimPrefix(result, header)
+	if len([]rune(body)) > 50 {
+		t.Errorf("body should be capped at 50 runes, got %d", len([]rune(body)))
+	}
+}
+
+func TestBuildUserContextSupplement_TruncateKeepsFirstAndLast(t *testing.T) {
+	first := "http://target.com " + strings.Repeat("A", 500)
+	var history []agent.ChatMessage
+	history = append(history, agent.ChatMessage{Role: "user", Content: first})
+	for i := 0; i < 10; i++ {
+		history = append(history, agent.ChatMessage{Role: "user", Content: strings.Repeat("B", 500)})
+	}
+	last := "最后一条指令"
+	result := buildUserContextSupplement(last, history, 0)
+	if !strings.Contains(result, "http://target.com") {
+		t.Error("first message (target URL) should survive truncation")
+	}
+	if !strings.Contains(result, last) {
+		t.Error("last message should survive truncation")
+	}
+}
+
+// --- middleware integration tests ---
+
+func TestTaskContextEnrichMiddleware_EnrichesTaskDescription(t *testing.T) {
+	mw := newTaskContextEnrichMiddleware(
+		"继续测试",
+		[]agent.ChatMessage{{Role: "user", Content: "http://8.163.32.73:8081 pikachu靶场"}},
+		0,
+	)
+	if mw == nil {
+		t.Fatal("expected non-nil middleware")
+	}
+
+	called := false
+	var capturedArgs string
+	fakeEndpoint := func(ctx context.Context, args string, opts ...tool.Option) (string, error) {
+		called = true
+		capturedArgs = args
+		return "ok", nil
+	}
+
+	wrapped, err := mw.(interface {
+		WrapInvokableToolCall(context.Context, adk.InvokableToolCallEndpoint, *adk.ToolContext) (adk.InvokableToolCallEndpoint, error)
+	}).WrapInvokableToolCall(context.Background(), fakeEndpoint, &adk.ToolContext{Name: "task"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	taskArgs := `{"subagent_type":"recon","description":"扫描目标端口"}`
+	wrapped(context.Background(), taskArgs)
+
+	if !called {
+		t.Fatal("endpoint was not called")
+	}
+
+	var parsed map[string]interface{}
+	if err := json.Unmarshal([]byte(capturedArgs), &parsed); err != nil {
+		t.Fatalf("enriched args not valid JSON: %v", err)
+	}
+	desc := parsed["description"].(string)
+	if !strings.Contains(desc, "扫描目标端口") {
+		t.Error("original description should be preserved")
+	}
+	if !strings.Contains(desc, "http://8.163.32.73:8081") {
+		t.Error("user context should be appended to description")
+	}
+	if !strings.Contains(desc, "继续测试") {
+		t.Error("current user message should be in description")
+	}
+}
+
+func TestTaskContextEnrichMiddleware_IgnoresNonTaskTools(t *testing.T) {
+	mw := newTaskContextEnrichMiddleware("test", nil, 0)
+	if mw == nil {
+		t.Fatal("expected non-nil middleware")
+	}
+
+	original := `{"command":"nmap -sV target"}`
+	var capturedArgs string
+	fakeEndpoint := func(ctx context.Context, args string, opts ...tool.Option) (string, error) {
+		capturedArgs = args
+		return "ok", nil
+	}
+
+	wrapped, err := mw.(interface {
+		WrapInvokableToolCall(context.Context, adk.InvokableToolCallEndpoint, *adk.ToolContext) (adk.InvokableToolCallEndpoint, error)
+	}).WrapInvokableToolCall(context.Background(), fakeEndpoint, &adk.ToolContext{Name: "nmap_scan"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	wrapped(context.Background(), original)
+	if capturedArgs != original {
+		t.Errorf("non-task tool args should not be modified, got %q", capturedArgs)
+	}
+}
+
+func TestTaskContextEnrichMiddleware_NilWhenDisabled(t *testing.T) {
+	mw := newTaskContextEnrichMiddleware("test", nil, -1)
+	if mw != nil {
+		t.Error("middleware should be nil when disabled")
+	}
+}

internal/multiagent/tool_args_json_retry.go

@@ -1,51 +0,0 @@
-package multiagent
-
-import (
-	"fmt"
-	"strings"
-
-	"github.com/cloudwego/eino/schema"
-)
-
-// maxToolCallRecoveryAttempts 含首次运行：首次 + 自动重试次数。
-// 例如为 3 表示最多共 3 次完整 DeepAgent 运行（2 次失败后各追加一条纠错提示）。
-// 该常量同时用于 JSON 参数错误和工具执行错误（如子代理名称不存在）的恢复重试。
-const maxToolCallRecoveryAttempts = 5
-
-// toolCallArgumentsJSONRetryHint 追加在用户消息后，提示模型输出合法 JSON 工具参数（部分云厂商会在流式阶段校验 arguments）。
-func toolCallArgumentsJSONRetryHint() *schema.Message {
-	return schema.UserMessage(`[系统提示] 上一次输出中，工具调用的 function.arguments 不是合法 JSON，接口已拒绝。请重新生成：每个 tool call 的 arguments 必须是完整、可解析的 JSON 对象字符串（键名用双引号，无多余逗号，括号配对）。不要输出截断或不完整的 JSON。
-
-[System] Your previous tool call used invalid JSON in function.arguments and was rejected by the API. Regenerate with strictly valid JSON objects only (double-quoted keys, matched braces, no trailing commas).`)
-}
-
-// toolCallArgumentsJSONRecoveryTimelineMessage 供 eino_recovery 事件落库与前端时间线展示。
-func toolCallArgumentsJSONRecoveryTimelineMessage(attempt int) string {
-	return fmt.Sprintf(
-		"接口拒绝了无效的工具参数 JSON。已向对话追加系统提示并要求模型重新生成合法的 function.arguments。"+
-			"当前为第 %d/%d 轮完整运行。\n\n"+
-			"The API rejected invalid JSON in tool arguments. A system hint was appended. This is full run %d of %d.",
-		attempt+1, maxToolCallRecoveryAttempts, attempt+1, maxToolCallRecoveryAttempts,
-	)
-}
-
-// isRecoverableToolCallArgumentsJSONError 判断是否为「工具参数非合法 JSON」类流式错误，可通过追加提示后重跑一轮。
-func isRecoverableToolCallArgumentsJSONError(err error) bool {
-	if err == nil {
-		return false
-	}
-	s := strings.ToLower(err.Error())
-	if !strings.Contains(s, "json") {
-		return false
-	}
-	if strings.Contains(s, "function.arguments") || strings.Contains(s, "function arguments") {
-		return true
-	}
-	if strings.Contains(s, "invalidparameter") && strings.Contains(s, "json") {
-		return true
-	}
-	if strings.Contains(s, "must be in json format") {
-		return true
-	}
-	return false
-}

internal/multiagent/tool_args_json_retry_test.go

@@ -1,17 +0,0 @@
-package multiagent
-
-import (
-	"errors"
-	"testing"
-)
-
-func TestIsRecoverableToolCallArgumentsJSONError(t *testing.T) {
-	yes := errors.New(`failed to receive stream chunk: error, <400> InternalError.Algo.InvalidParameter: The "function.arguments" parameter of the code model must be in JSON format.`)
-	if !isRecoverableToolCallArgumentsJSONError(yes) {
-		t.Fatal("expected recoverable for function.arguments + JSON")
-	}
-	no := errors.New("unrelated network failure")
-	if isRecoverableToolCallArgumentsJSONError(no) {
-		t.Fatal("expected not recoverable")
-	}
-}

internal/multiagent/tool_execution_retry.go

@@ -1,44 +0,0 @@
-package multiagent
-
-import (
-	"fmt"
-
-	"github.com/cloudwego/eino/schema"
-)
-
-// toolExecutionRetryHint returns a user message appended to the conversation to prompt
-// the LLM to adjust after a tool execution error (tool not found, binary missing,
-// runtime failure, network error, etc.).
-func toolExecutionRetryHint() *schema.Message {
-	return schema.UserMessage(`[System] Your previous tool call failed. Possible causes:
-- The tool or sub-agent name does not exist (typo or unregistered name).
-- The tool call arguments were not valid JSON.
-- The tool's underlying binary is not installed or not in PATH.
-- The tool encountered a runtime error (timeout, network failure, permission denied, etc.).
-
-Please review the error message above, check available tools, and either:
-1. Retry with corrected arguments or a different tool, OR
-2. Inform the user about the limitation and proceed with an alternative approach.
-
-[系统提示] 上一次工具调用失败，可能原因：
-- 工具名或子代理名称不存在（拼写错误或未注册）；
-- 工具调用参数不是合法 JSON；
-- 工具依赖的底层二进制程序未安装或不在 PATH 中；
-- 工具运行时遇到错误（超时、网络故障、权限不足等）。
-
-请根据上述错误信息检查可用工具，然后：
-1. 修正参数或改用其他工具重试，或者
-2. 告知用户当前限制并采用替代方案继续。`)
-}
-
-// toolExecutionRecoveryTimelineMessage returns a message for the eino_recovery event
-// displayed in the UI timeline when a tool execution error triggers a retry.
-func toolExecutionRecoveryTimelineMessage(attempt int) string {
-	return fmt.Sprintf(
-		"工具调用执行失败。已向对话追加纠错提示并要求模型调整策略。"+
-			"当前为第 %d/%d 轮完整运行。\n\n"+
-			"Tool call execution failed. "+
-			"A corrective hint was appended. This is full run %d of %d.",
-		attempt+1, maxToolCallRecoveryAttempts, attempt+1, maxToolCallRecoveryAttempts,
-	)
-}

plugins/burp-suite/cyberstrikeai-burp-extension/src/main/java/burp/BurpExtender.java

@@ -64,7 +64,7 @@ public class BurpExtender implements IBurpExtender, IContextMenuFactory {
 
         String prompt = HttpMessageFormatter.toPrompt(helpers, msg, instruction);
         String title = HttpMessageFormatter.getRequestTitle(helpers, msg);
-        String agentModeStr = (cfg.agentMode == CyberStrikeAIClient.AgentMode.MULTI) ? "Multi Agent" : "Single Agent";
+        String agentModeStr = cfg.agentMode.displayName;
         String runId = tab.startNewRun(title, agentModeStr, msg);
         tab.appendProgressToRun(runId, "\n[server] " + cfg.baseUrl + "\n\n");
 

plugins/burp-suite/cyberstrikeai-burp-extension/src/main/java/burp/CyberStrikeAIClient.java

@@ -26,8 +26,21 @@ final class CyberStrikeAIClient {
     }
 
     enum AgentMode {
-        SINGLE,
-        MULTI
+        NATIVE_REACT("Native ReAct", "/api/agent-loop/stream", null),
+        EINO_SINGLE("Eino Single (ADK)", "/api/eino-agent/stream", null),
+        DEEP("Deep (DeepAgent)", "/api/multi-agent/stream", "deep"),
+        PLAN_EXECUTE("Plan-Execute", "/api/multi-agent/stream", "plan_execute"),
+        SUPERVISOR("Supervisor", "/api/multi-agent/stream", "supervisor");
+
+        final String displayName;
+        final String streamPath;
+        final String orchestration;
+
+        AgentMode(String displayName, String streamPath, String orchestration) {
+            this.displayName = displayName;
+            this.streamPath = streamPath;
+            this.orchestration = orchestration;
+        }
     }
 
     interface StreamListener {
@@ -94,13 +107,15 @@ final class CyberStrikeAIClient {
     }
 
     void streamTest(Config cfg, String token, String message, StreamListener listener) {
-        String path = (cfg.agentMode == AgentMode.MULTI) ? "/api/multi-agent/stream" : "/api/agent-loop/stream";
-        String urlStr = cfg.baseUrl + path;
+        String urlStr = cfg.baseUrl + cfg.agentMode.streamPath;
 
         Map<String, Object> payload = new HashMap<>();
         payload.put("message", message);
         payload.put("conversationId", "");
         payload.put("role", "");
+        if (cfg.agentMode.orchestration != null) {
+            payload.put("orchestration", cfg.agentMode.orchestration);
+        }
 
         new Thread(() -> {
             HttpURLConnection conn = null;
@@ -184,11 +199,16 @@ final class CyberStrikeAIClient {
         String message = payload.get("message") != null ? String.valueOf(payload.get("message")) : "";
         String conversationId = payload.get("conversationId") != null ? String.valueOf(payload.get("conversationId")) : "";
         String role = payload.get("role") != null ? String.valueOf(payload.get("role")) : "";
-        return "{"
-                + "\"message\":\"" + escapeJson(message) + "\","
-                + "\"conversationId\":\"" + escapeJson(conversationId) + "\","
-                + "\"role\":\"" + escapeJson(role) + "\""
-                + "}";
+        StringBuilder sb = new StringBuilder();
+        sb.append("{");
+        sb.append("\"message\":\"").append(escapeJson(message)).append("\",");
+        sb.append("\"conversationId\":\"").append(escapeJson(conversationId)).append("\",");
+        sb.append("\"role\":\"").append(escapeJson(role)).append("\"");
+        if (payload.containsKey("orchestration") && payload.get("orchestration") != null) {
+            sb.append(",\"orchestration\":\"").append(escapeJson(String.valueOf(payload.get("orchestration")))).append("\"");
+        }
+        sb.append("}");
+        return sb.toString();
     }
 
     private static String escapeJson(String s) {

plugins/burp-suite/cyberstrikeai-burp-extension/src/main/java/burp/CyberStrikeAITab.java

@@ -15,7 +15,9 @@ final class CyberStrikeAITab implements ITab {
     private final JTextField hostField = new JTextField("127.0.0.1");
     private final JTextField portField = new JTextField("8080");
     private final JPasswordField passwordField = new JPasswordField();
-    private final JComboBox<String> agentModeBox = new JComboBox<>(new String[]{"Single Agent", "Multi Agent"});
+    private final JComboBox<String> agentModeBox = new JComboBox<>(new String[]{
+            "Native ReAct", "Eino Single (ADK)", "Deep (DeepAgent)", "Plan-Execute", "Supervisor"
+    });
     private final JButton validateButton = new JButton("Validate");
     private final JButton clearButton = new JButton("Clear Output");
     private final JButton stopButton = new JButton("Stop");
@@ -98,7 +100,7 @@ final class CyberStrikeAITab implements ITab {
         hostField.setColumns(14);
         portField.setColumns(6);
         passwordField.setColumns(12);
-        agentModeBox.setPreferredSize(new Dimension(160, agentModeBox.getPreferredSize().height));
+        agentModeBox.setPreferredSize(new Dimension(200, agentModeBox.getPreferredSize().height));
 
         JPanel row1 = new JPanel(new FlowLayout(FlowLayout.LEFT, 8, 2));
         row1.add(new JLabel("Host"));
@@ -475,14 +477,17 @@ final class CyberStrikeAITab implements ITab {
         renderMarkdownBox.addActionListener(e -> refreshOutputView());
     }
 
+    private static final CyberStrikeAIClient.AgentMode[] AGENT_MODES = CyberStrikeAIClient.AgentMode.values();
+
     CyberStrikeAIClient.Config currentConfig() {
         String host = hostField.getText().trim();
         String port = portField.getText().trim();
         String password = new String(passwordField.getPassword());
         String baseUrl = "http://" + host + ":" + port;
-        CyberStrikeAIClient.AgentMode mode = agentModeBox.getSelectedIndex() == 1
-                ? CyberStrikeAIClient.AgentMode.MULTI
-                : CyberStrikeAIClient.AgentMode.SINGLE;
+        int idx = agentModeBox.getSelectedIndex();
+        CyberStrikeAIClient.AgentMode mode = (idx >= 0 && idx < AGENT_MODES.length)
+                ? AGENT_MODES[idx]
+                : CyberStrikeAIClient.AgentMode.NATIVE_REACT;
         return new CyberStrikeAIClient.Config(baseUrl, password, mode);
     }
 

roles/API安全测试.yaml

@@ -2,19 +2,4 @@ name: API安全测试
 description: API安全测试专家，专注于API接口安全检测
 user_prompt: 你是一个专业的API安全测试专家。请使用专业的API测试工具对目标API接口进行全面的安全检测，包括GraphQL安全、API参数fuzzing、JWT分析、API架构分析等工作。
 icon: "\U0001F4E1"
-tools:
-    - api-fuzzer
-    - api-schema-analyzer
-    - graphql-scanner
-    - arjun
-    - jwt-analyzer
-    - http-intruder
-    - http-framework-test
-    - burpsuite
-    - httpx
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/CTF.yaml

@@ -2,32 +2,4 @@ name: CTF
 description: CTF竞赛专家，擅长解题和漏洞利用
 user_prompt: 你是一个CTF竞赛专家。请使用CTF解题思维和方法，快速定位和利用漏洞，解决各类CTF题目。
 icon: "\U0001F3C6"
-tools:
-    - amass
-    - anew
-    - angr
-    - api-fuzzer
-    - api-schema-analyzer
-    - arjun
-    - arp-scan
-    - autorecon
-    - binwalk
-    - bloodhound
-    - burpsuite
-    - cat
-    - checkov
-    - checksec
-    - cloudmapper
-    - create-file
-    - cyberchef
-    - dalfox
-    - delete-file
-    - httpx
-    - http-framework-test
-    - exec
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/Web应用扫描.yaml

@@ -2,24 +2,4 @@ name: Web应用扫描
 description: Web应用漏洞扫描专家，全面的Web安全检测
 user_prompt: 你是一个专业的Web应用漏洞扫描专家。请使用各种Web扫描工具对目标Web应用进行全面的安全检测，包括目录枚举、文件扫描、漏洞识别等工作。
 icon: "\U0001F310"
-tools:
-    - dirsearch
-    - dirb
-    - gobuster
-    - feroxbuster
-    - ffuf
-    - wfuzz
-    - sqlmap
-    - dalfox
-    - xsser
-    - nikto
-    - nuclei
-    - wpscan
-    - httpx
-    - http-framework-test
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/Web框架测试.yaml

@@ -2,18 +2,4 @@ name: Web框架测试
 description: Web框架安全测试专家，专注于Web应用框架漏洞检测
 user_prompt: 你是一个专业的Web框架安全测试专家。请使用专业的工具对Web应用框架进行安全测试，识别框架相关的安全漏洞和配置问题。
 icon: "\U0001F310"
-tools:
-    - http-framework-test
-    - nikto
-    - nuclei
-    - wafw00f
-    - wpscan
-    - httpx
-    - burpsuite
-    - zap
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/二进制分析.yaml

@@ -2,30 +2,4 @@ name: 二进制分析
 description: 二进制分析与利用专家，擅长逆向工程和密码破解
 user_prompt: 你是一个专业的二进制分析与利用专家。请使用逆向工程工具分析二进制文件，识别漏洞，进行利用开发。同时擅长密码破解、哈希分析等技术。
 icon: "\U0001F52C"
-tools:
-    - dirsearch
-    - docker-bench-security
-    - exec
-    - execute-python-script
-    - install-python-package
-    - ghidra
-    - graphql-scanner
-    - hakrawler
-    - hash-identifier
-    - hashcat
-    - hashpump
-    - http-framework-test
-    - httpx
-    - gdb
-    - radare2
-    - objdump
-    - strings
-    - binwalk
-    - ropper
-    - ropgadget
-    - john
-    - cyberchef
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/云安全审计.yaml

@@ -2,16 +2,4 @@ name: 云安全审计
 description: 云安全审计专家，多云环境安全检测
 user_prompt: 你是一个专业的云安全审计专家。请使用专业的云安全工具对AWS、Azure、GCP等云环境进行全面的安全审计，包括配置检查、合规性评估、权限审计、安全最佳实践验证等工作。
 icon: ☁
-tools:
-    - prowler
-    - scout-suite
-    - cloudmapper
-    - pacu
-    - terrascan
-    - checkov
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/信息收集.yaml

@@ -2,30 +2,4 @@ name: 信息收集
 description: 资产发现与信息搜集专家
 user_prompt: 你是一个专业的信息收集专家。请使用各种信息收集技术和工具，对目标进行全面的资产发现、子域名枚举、端口扫描、服务识别等信息收集工作。
 icon: "\U0001F50D"
-tools:
-    - amass
-    - subfinder
-    - dnsenum
-    - fierce
-    - fofa_search
-    - zoomeye_search
-    - nmap
-    - masscan
-    - rustscan
-    - arp-scan
-    - nbtscan
-    - httpx
-    - http-framework-test
-    - katana
-    - hakrawler
-    - waybackurls
-    - paramspider
-    - gau
-    - uro
-    - qsreplace
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/后渗透测试.yaml

@@ -2,22 +2,4 @@ name: 后渗透测试
 description: 后渗透测试专家，权限维持与横向移动
 user_prompt: 你是一个专业的后渗透测试专家。请使用专业的后渗透工具在获得初始访问权限后进行权限提升、横向移动、权限维持、数据收集等后渗透测试工作。
 icon: "\U0001F575"
-tools:
-    - linpeas
-    - winpeas
-    - mimikatz
-    - bloodhound
-    - impacket
-    - responder
-    - netexec
-    - rpcclient
-    - smbmap
-    - enum4linux
-    - enum4linux-ng
-    - exec
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/容器安全.yaml

@@ -2,17 +2,4 @@ name: 容器安全
 description: 容器与Kubernetes安全专家，容器环境安全检测
 user_prompt: 你是一个专业的容器与Kubernetes安全专家。请使用专业的容器安全工具对Docker容器和Kubernetes集群进行全面的安全检测，包括镜像漏洞扫描、配置检查、运行时安全等工作。
 icon: "\U0001F6E1"
-tools:
-    - trivy
-    - clair
-    - docker-bench-security
-    - kube-bench
-    - kube-hunter
-    - falco
-    - exec
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/数字取证.yaml

@@ -2,23 +2,4 @@ name: 数字取证
 description: 数字取证与隐写分析专家，文件与内存取证
 user_prompt: 你是一个专业的数字取证与隐写分析专家。请使用专业的取证工具对文件、磁盘镜像、内存转储进行分析，提取证据信息。同时擅长隐写分析、数据恢复、元数据提取等技术。
 icon: "\U0001F50E"
-tools:
-    - volatility
-    - volatility3
-    - foremost
-    - steghide
-    - stegsolve
-    - zsteg
-    - exiftool
-    - binwalk
-    - strings
-    - xxd
-    - fcrackzip
-    - pdfcrack
-    - exec
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/渗透测试.yaml

@@ -2,32 +2,4 @@ name: 渗透测试
 description: 专业渗透测试专家，全面深入的漏洞检测
 user_prompt: 你是一个专业的网络安全渗透测试专家。请使用专业的渗透测试方法和工具，对目标进行全面的安全测试，包括但不限于SQL注入、XSS、CSRF、文件包含、命令执行等常见漏洞。
 icon: "\U0001F3AF"
-tools:
-    - http-framework-test
-    - httpx
-    - amass
-    - anew
-    - angr
-    - api-fuzzer
-    - api-schema-analyzer
-    - arjun
-    - arp-scan
-    - autorecon
-    - binwalk
-    - bloodhound
-    - burpsuite
-    - cat
-    - checkov
-    - checksec
-    - cloudmapper
-    - create-file
-    - cyberchef
-    - dalfox
-    - delete-file
-    - exec
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

roles/综合漏洞扫描.yaml

@@ -2,22 +2,4 @@ name: 综合漏洞扫描
 description: 综合漏洞扫描专家，多类型漏洞检测
 user_prompt: 你是一个专业的综合漏洞扫描专家。请使用各种漏洞扫描工具对目标进行全面的安全检测，包括Web漏洞、网络服务漏洞、配置缺陷等多种类型的漏洞识别和分析。
 icon: ⚠
-tools:
-    - nuclei
-    - nikto
-    - sqlmap
-    - nmap
-    - masscan
-    - rustscan
-    - wafw00f
-    - dalfox
-    - xsser
-    - jaeles
-    - httpx
-    - http-framework-test
-    - execute-python-script
-    - install-python-package
-    - record_vulnerability
-    - list_knowledge_risk_types
-    - search_knowledge_base
 enabled: true

tools/anew.yaml

@@ -1,81 +0,0 @@
-name: "anew"
-command: "python3"
-args:
-  - "-c"
-  - |
-    import shlex
-    import subprocess
-    import sys
-
-    if len(sys.argv) < 2:
-        sys.stderr.write("缺少输入数据\n")
-        sys.exit(1)
-
-    input_data = sys.argv[1]
-    output_file = sys.argv[2] if len(sys.argv) > 2 else ""
-    additional = sys.argv[3] if len(sys.argv) > 3 else ""
-
-    cmd = ["anew"]
-    if additional:
-        cmd.extend(shlex.split(additional))
-    if output_file:
-        cmd.append(output_file)
-
-    proc = subprocess.run(
-        cmd,
-        input=input_data.encode("utf-8"),
-        capture_output=True,
-        text=True,
-    )
-
-    if proc.returncode != 0:
-        sys.stderr.write(proc.stderr or proc.stdout)
-        sys.exit(proc.returncode)
-
-    sys.stdout.write(proc.stdout)
-enabled: true
-short_description: "数据去重工具，用于处理文件中的新行"
-description: |
-  Anew是一个数据去重工具，用于将新行追加到文件中，自动过滤重复项。
-
-  **主要功能：**
-  - 数据去重
-  - 文件追加
-  - 唯一行过滤
-  - 快速处理
-
-  **使用场景：**
-  - 数据处理
-  - 结果去重
-  - 数据合并
-  - 工具链集成
-parameters:
-  - name: "input_data"
-    type: "string"
-    description: "输入数据"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "output_file"
-    type: "string"
-    description: "输出文件路径"
-    required: false
-    default: ""
-    position: 1
-    format: "positional"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的anew参数。用于传递未在参数列表中定义的anew选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    default: ""
-    position: 2
-    format: "positional"

tools/autorecon.yaml

@@ -1,46 +0,0 @@
-name: "autorecon"
-command: "autorecon"
-enabled: true
-short_description: "自动化综合侦察工具"
-description: |
-  AutoRecon是一个自动化综合侦察工具，用于执行全面的目标枚举。
-
-  **主要功能：**
-  - 自动化端口扫描
-  - 服务识别
-  - 漏洞扫描
-  - 综合报告
-
-  **使用场景：**
-  - 综合安全评估
-  - 渗透测试
-  - 网络侦察
-  - 安全审计
-parameters:
-  - name: "target"
-    type: "string"
-    description: "目标IP地址或主机名"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "output_dir"
-    type: "string"
-    description: "输出目录"
-    required: false
-    flag: "-o"
-    format: "flag"
-    default: "/tmp/autorecon"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的autorecon参数。用于传递未在参数列表中定义的autorecon选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/burpsuite.yaml

@@ -1,50 +0,0 @@
-name: "burpsuite"
-command: "burpsuite"
-enabled: true
-short_description: "Web应用安全测试平台"
-description: |
-  Burp Suite是一个Web应用安全测试平台，提供全面的Web安全测试功能。
-
-  **主要功能：**
-  - Web应用安全扫描
-  - 代理拦截
-  - 漏洞扫描
-  - 手动测试工具
-
-  **使用场景：**
-  - Web应用安全测试
-  - 渗透测试
-  - 漏洞扫描
-  - 安全评估
-parameters:
-  - name: "project_file"
-    type: "string"
-    description: "Burp Suite项目文件路径（--project-file）"
-    required: false
-    flag: "--project-file"
-    format: "flag"
-  - name: "config_file"
-    type: "string"
-    description: "自动化/扫描配置文件（--config-file）"
-    required: false
-    flag: "--config-file"
-    format: "flag"
-  - name: "user_config_file"
-    type: "string"
-    description: "用户配置文件（--user-config-file）"
-    required: false
-    flag: "--user-config-file"
-    format: "flag"
-  - name: "headless"
-    type: "bool"
-    description: "无头模式运行"
-    required: false
-    flag: "--headless"
-    format: "flag"
-    default: false
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的burpsuite参数。用于传递未在参数列表中定义的burpsuite选项（例如 --project-config、--log-config 等）。
-    required: false
-    format: "positional"

tools/cyberchef.yaml

@@ -1,58 +0,0 @@
-name: "cyberchef"
-command: "cyberchef"
-enabled: true
-short_description: "数据转换和分析工具，支持多种编码、加密和数据处理操作"
-description: |
-  CyberChef 是一个强大的数据转换和分析工具，支持数百种数据操作。
-
-  **主要功能：**
-  - 编码/解码（Base64, Hex, URL 等）
-  - 加密/解密（AES, DES, RSA 等）
-  - 哈希计算
-  - 数据格式转换
-  - 正则表达式操作
-  - 数据提取和分析
-
-  **使用场景：**
-  - CTF 竞赛
-  - 数据分析和转换
-  - 加密算法研究
-  - 数字取证
-
-  **注意事项：**
-  - 通常以 Web 界面运行
-  - 命令行版本可能需要 Node.js
-  - 功能强大，操作复杂
-parameters:
-  - name: "recipe"
-    type: "string"
-    description: "操作配方（JSON 格式），定义要执行的操作序列"
-    required: true
-    flag: "-Recipe"
-    format: "flag"
-  - name: "input"
-    type: "string"
-    description: "输入数据（字符串或文件路径）"
-    required: true
-    flag: "-Input"
-    format: "flag"
-  - name: "output"
-    type: "string"
-    description: "输出文件路径（可选）"
-    required: false
-    flag: "-Output"
-    format: "flag"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的cyberchef参数。用于传递未在参数列表中定义的cyberchef选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/dirb.yaml

@@ -1,99 +0,0 @@
-name: "dirb"
-command: "dirb"
-enabled: true
-# 简短描述（用于工具列表，减少token消耗）
-short_description: "Web目录和文件扫描工具，通过暴力破解方式发现Web服务器上的隐藏目录和文件"
-# 工具详细描述
-description: |
-  Web目录和文件扫描工具，通过暴力破解方式发现Web服务器上的隐藏目录和文件。
-
-  **主要功能：**
-  - 目录和文件发现
-  - 支持自定义字典文件
-  - 检测常见的Web目录结构
-  - 识别备份文件、配置文件等敏感文件
-  - 支持多种HTTP方法
-
-  **使用场景：**
-  - Web应用目录枚举
-  - 发现隐藏的管理界面
-  - 查找备份文件和敏感信息
-  - 渗透测试中的信息收集
-
-  **注意事项：**
-  - 扫描可能产生大量HTTP请求
-  - 某些请求可能被WAF拦截
-  - 建议使用合适的字典文件以提高效率
-  - 扫描结果需要人工验证
-# 参数定义
-parameters:
-  - name: "url"
-    type: "string"
-    description: |
-      目标URL，要扫描的Web服务器地址。
-
-      **格式要求：**
-      - 必须包含协议（http:// 或 https://）
-      - 可以包含基础路径
-      - 末尾不要带斜杠（除非要扫描特定目录）
-
-      **示例值：**
-      - 基础URL: "http://example.com"
-      - HTTPS: "https://example.com"
-      - 带端口: "http://example.com:8080"
-      - 特定目录: "http://example.com/admin"
-      - 带路径: "http://example.com/app"
-
-      **注意事项：**
-      - URL必须可访问
-      - 确保URL格式正确，包含协议前缀
-      - 必需参数，不能为空
-    required: true
-    position: 0
-    format: "positional"
-  - name: "wordlist"
-    type: "string"
-    description: |
-      字典文件路径，包含要尝试的目录和文件名列表。
-
-      **格式要求：**
-      - 文件路径，可以是绝对路径或相对路径
-      - 文件每行一个目录或文件名
-      - 支持常见的字典文件格式
-
-      **示例值：**
-      - 默认字典: "/usr/share/dirb/wordlists/common.txt"
-      - 自定义字典: "/path/to/custom-wordlist.txt"
-      - 常用字典: "/usr/share/wordlists/dirb/common.txt"
-
-      **常用字典文件：**
-      - common.txt: 常见目录和文件
-      - big.txt: 大型字典
-      - small.txt: 小型快速字典
-      - extensions_common.txt: 常见文件扩展名
-
-      **注意事项：**
-      - 如果不指定，将使用默认字典
-      - 确保字典文件存在且可读
-      - 大型字典会显著增加扫描时间
-    required: false
-    position: 1
-    format: "positional"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的Dirb参数。用于传递未在参数列表中定义的Dirb选项。
-
-      **示例值：**
-      - "-a": 用户代理字符串
-      - "-H": 自定义HTTP头
-      - "-c": Cookie字符串
-      - "-X": 文件扩展名
-      - "-z": 毫秒延迟
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/docker-bench-security.yaml

@@ -1,51 +0,0 @@
-name: "docker-bench-security"
-command: "docker-bench-security"
-enabled: true
-short_description: "Docker安全基准检查工具"
-description: |
-  Docker Bench for Security是一个Docker安全基准检查工具，用于检查Docker配置是否符合安全最佳实践。
-
-  **主要功能：**
-  - Docker安全基准检查
-  - 配置审计
-  - 安全最佳实践检查
-  - 详细报告
-
-  **使用场景：**
-  - Docker安全审计
-  - 配置检查
-  - 合规性验证
-  - 安全评估
-parameters:
-  - name: "checks"
-    type: "string"
-    description: "要运行的特定检查"
-    required: false
-    flag: "-c"
-    format: "flag"
-  - name: "exclude"
-    type: "string"
-    description: "要排除的检查"
-    required: false
-    flag: "-e"
-    format: "flag"
-  - name: "output_file"
-    type: "string"
-    description: "输出文件路径"
-    required: false
-    flag: "-l"
-    format: "flag"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的docker-bench-security参数。用于传递未在参数列表中定义的docker-bench-security选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/enum4linux.yaml

@@ -1,31 +0,0 @@
-name: "enum4linux"
-command: "enum4linux"
-enabled: true
-short_description: "SMB枚举工具，用于Windows/Samba系统信息收集"
-description: |
-  Enum4linux是一个用于枚举SMB共享和Windows系统信息的工具。
-
-  **主要功能：**
-  - SMB共享枚举
-  - 用户和组枚举
-  - 密码策略信息
-  - 系统信息收集
-
-  **使用场景：**
-  - Windows系统渗透测试
-  - SMB安全评估
-  - 网络信息收集
-  - 域环境侦察
-parameters:
-  - name: "target"
-    type: "string"
-    description: "目标IP地址"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "additional_args"
-    type: "string"
-    description: "额外的Enum4linux参数（默认：-a）"
-    required: false
-    default: "-a"
-    format: "positional"

tools/fcrackzip.yaml

@@ -1,76 +0,0 @@
-name: "fcrackzip"
-command: "fcrackzip"
-enabled: true
-short_description: "ZIP 文件密码破解工具，支持暴力破解和字典攻击"
-description: |
-  fcrackzip 是一个用于破解受密码保护的 ZIP 文件密码的工具。
-
-  **主要功能：**
-  - 暴力破解
-  - 字典攻击
-  - 指定字符集
-  - 指定密码长度范围
-  - 多线程支持
-
-  **使用场景：**
-  - CTF 竞赛
-  - ZIP 文件密码恢复
-  - 安全测试
-  - 数字取证
-
-  **注意事项：**
-  - 破解时间取决于密码复杂度
-  - 建议使用字典文件提高效率
-  - 仅用于授权的安全测试
-parameters:
-  - name: "file"
-    type: "string"
-    description: "要破解的 ZIP 文件路径"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "dictionary_mode"
-    type: "bool"
-    description: "启用字典攻击模式（等同于 -D）"
-    required: false
-    flag: "-D"
-    format: "flag"
-    default: false
-  - name: "dictionary"
-    type: "string"
-    description: "字典文件路径（与 -D 配合使用）"
-    required: false
-    flag: "-p"
-    format: "flag"
-  - name: "bruteforce"
-    type: "bool"
-    description: "使用暴力破解模式"
-    required: false
-    flag: "-b"
-    format: "flag"
-  - name: "charset"
-    type: "string"
-    description: "字符集，例如 'aA1' 表示小写字母、大写字母和数字"
-    required: false
-    flag: "-c"
-    format: "flag"
-  - name: "length_range"
-    type: "string"
-    description: "密码长度范围，格式为min-max（例如 4-8）"
-    required: false
-    flag: "-l"
-    format: "flag"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的fcrackzip参数。用于传递未在参数列表中定义的fcrackzip选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/feroxbuster.yaml

@@ -1,6 +1,6 @@
 name: "feroxbuster"
 command: "feroxbuster"
-enabled: true
+enabled: false
 short_description: "递归内容发现工具"
 description: |
   Feroxbuster是一个快速、简单的递归内容发现工具。

tools/gdb-peda.yaml

@@ -1,57 +0,0 @@
-name: "gdb-peda"
-command: "gdb"
-enabled: true
-short_description: "带PEDA增强的GDB调试器"
-description: |
-  GDB-PEDA是带有PEDA（Python Exploit Development Assistance）增强的GDB调试器。
-
-  **主要功能：**
-  - 增强的GDB功能
-  - 自动化分析
-  - 漏洞利用辅助
-  - 可视化显示
-
-  **使用场景：**
-  - 二进制调试
-  - 漏洞利用开发
-  - 逆向工程
-  - 安全研究
-parameters:
-  - name: "binary"
-    type: "string"
-    description: "要调试的二进制文件"
-    required: false
-    position: 0
-    format: "positional"
-  - name: "commands"
-    type: "string"
-    description: "GDB命令（分号分隔）"
-    required: false
-    flag: "-ex"
-    format: "flag"
-  - name: "attach_pid"
-    type: "int"
-    description: "要附加的进程ID"
-    required: false
-    flag: "-p"
-    format: "flag"
-  - name: "core_file"
-    type: "string"
-    description: "核心转储文件路径"
-    required: false
-    flag: "-c"
-    format: "flag"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的gdb-peda参数。用于传递未在参数列表中定义的gdb-peda选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/gobuster.yaml

@@ -1,6 +1,6 @@
 name: "gobuster"
 command: "gobuster"
-enabled: true
+enabled: false
 short_description: "Web内容扫描工具，用于发现目录、文件和子域名"
 description: |
   Gobuster是一个快速的内容发现工具，用于Web应用程序的目录、文件和子域名枚举。

tools/hakrawler.yaml

@@ -1,53 +0,0 @@
-name: "hakrawler"
-command: "hakrawler"
-enabled: true
-short_description: "Web端点发现工具"
-description: |
-  Hakrawler是一个快速、简单的Web端点发现工具。
-
-  **主要功能：**
-  - Web端点发现
-  - 链接提取
-  - JavaScript文件发现
-  - 快速爬取
-
-  **使用场景：**
-  - Web端点发现
-  - 内容爬取
-  - 安全测试
-  - Bug bounty侦察
-parameters:
-  - name: "url"
-    type: "string"
-    description: "目标URL"
-    required: true
-    flag: "-url"
-    format: "flag"
-  - name: "depth"
-    type: "int"
-    description: "爬取深度"
-    required: false
-    flag: "-d"
-    format: "flag"
-    default: 2
-  - name: "forms"
-    type: "bool"
-    description: "包含表单"
-    required: false
-    flag: "-forms"
-    format: "flag"
-    default: true
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的hakrawler参数。用于传递未在参数列表中定义的hakrawler选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/hash-identifier.yaml

@@ -1,84 +0,0 @@
-name: "hash-identifier"
-command: "python3"
-args:
-  - "-c"
-  - |
-    import shlex
-    import subprocess
-    import sys
-
-    if len(sys.argv) < 2:
-        sys.stderr.write("缺少哈希值\n")
-        sys.exit(1)
-
-    hash_value = sys.argv[1]
-    extra = sys.argv[2] if len(sys.argv) > 2 else ""
-
-    cmd = ["hash-identifier"]
-    if extra:
-        cmd.extend(shlex.split(extra))
-
-    proc = subprocess.run(
-        cmd,
-        input=f"{hash_value}\n",
-        capture_output=True,
-        text=True,
-    )
-
-    if proc.returncode != 0:
-        sys.stderr.write(proc.stderr or proc.stdout)
-        sys.exit(proc.returncode)
-
-    sys.stdout.write(proc.stdout)
-enabled: true
-short_description: "哈希类型识别工具，用于识别未知哈希值的类型"
-description: |
-  hash-identifier 是一个用于识别哈希值类型的工具，可以帮助确定未知哈希值使用的算法。
-
-  **主要功能：**
-  - 识别多种哈希算法
-  - 支持 MD5, SHA1, SHA256, bcrypt 等
-  - 交互式识别
-  - 快速识别常见哈希类型
-
-  **支持的哈希类型：**
-  - MD5
-  - SHA1, SHA256, SHA512
-  - bcrypt
-  - NTLM
-  - MySQL
-  - PostgreSQL
-  - 等多种哈希算法
-
-  **使用场景：**
-  - CTF 密码破解
-  - 哈希值分析
-  - 密码学研究
-  - 安全审计
-
-  **注意事项：**
-  - 需要 Python 环境
-  - 交互式工具，可能需要特殊处理
-parameters:
-  - name: "hash"
-    type: "string"
-    description: "要识别的哈希值"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的hash-identifier参数。用于传递未在参数列表中定义的hash-identifier选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    default: ""
-    position: 1
-    format: "positional"

tools/pdfcrack.yaml

@@ -1,62 +0,0 @@
-name: "pdfcrack"
-command: "pdfcrack"
-enabled: true
-short_description: "PDF 文件密码破解工具，支持暴力破解和字典攻击"
-description: |
-  pdfcrack 是一个用于破解受密码保护的 PDF 文件密码的工具。
-
-  **主要功能：**
-  - 暴力破解
-  - 字典攻击
-  - 用户密码和所有者密码破解
-  - 支持多种加密算法
-
-  **使用场景：**
-  - CTF 竞赛
-  - PDF 文件密码恢复
-  - 安全测试
-  - 数字取证
-
-  **注意事项：**
-  - 破解时间取决于密码复杂度
-  - 建议使用字典文件提高效率
-  - 仅用于授权的安全测试
-parameters:
-  - name: "file"
-    type: "string"
-    description: "要破解的 PDF 文件路径"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "wordlist"
-    type: "string"
-    description: "字典文件路径"
-    required: false
-    flag: "-w"
-    format: "flag"
-  - name: "min_length"
-    type: "int"
-    description: "最小密码长度"
-    required: false
-    flag: "-n"
-    format: "flag"
-  - name: "max_length"
-    type: "int"
-    description: "最大密码长度"
-    required: false
-    flag: "-m"
-    format: "flag"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的pdfcrack参数。用于传递未在参数列表中定义的pdfcrack选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/qsreplace.yaml

@@ -1,81 +0,0 @@
-name: "qsreplace"
-command: "python3"
-args:
-  - "-c"
-  - |
-    import shlex
-    import subprocess
-    import sys
-
-    if len(sys.argv) < 2:
-        sys.stderr.write("缺少URL列表\n")
-        sys.exit(1)
-
-    urls = sys.argv[1]
-    replacement = sys.argv[2] if len(sys.argv) > 2 else ""
-    extra = sys.argv[3] if len(sys.argv) > 3 else ""
-
-    cmd = ["qsreplace"]
-    if extra:
-        cmd.extend(shlex.split(extra))
-    if replacement:
-        cmd.append(replacement)
-
-    proc = subprocess.run(
-        cmd,
-        input=urls,
-        capture_output=True,
-        text=True,
-    )
-    if proc.stdout:
-        sys.stdout.write(proc.stdout)
-    if proc.stderr:
-        sys.stderr.write(proc.stderr)
-    sys.exit(proc.returncode)
-enabled: true
-short_description: "查询字符串参数替换工具"
-description: |
-  Qsreplace是一个用于替换URL中查询字符串参数的工具，常用于模糊测试。
-
-  **主要功能：**
-  - 参数替换
-  - 批量处理
-  - 多种替换模式
-  - 快速处理
-
-  **使用场景：**
-  - 参数模糊测试
-  - URL处理
-  - 工具链集成
-  - 安全测试
-parameters:
-  - name: "urls"
-    type: "string"
-    description: "要处理的URL（每行一个）"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "replacement"
-    type: "string"
-    description: "替换字符串"
-    required: false
-    default: ""
-    position: 1
-    format: "positional"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的Qsreplace参数。用于传递未在参数列表中定义的Qsreplace选项。
-
-      **示例值：**
-      - "-a": 追加模式
-      - "-d": 删除参数
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    default: ""
-    position: 2
-    format: "positional"

tools/stegsolve.yaml

@@ -1,52 +0,0 @@
-name: "stegsolve"
-command: "java"
-args: ["-jar"]
-enabled: true
-short_description: "图片隐写分析工具，用于分析图片中的隐写数据"
-description: |
-  Stegsolve 是一个 Java 图片隐写分析工具，支持多种图片格式和隐写分析技术。
-
-  **主要功能：**
-  - 图片格式转换
-  - 颜色通道分析
-  - LSB 隐写检测
-  - 图片叠加分析
-  - 数据提取
-
-  **使用场景：**
-  - CTF 隐写题目
-  - 图片隐写分析
-  - 数字取证
-  - 安全研究
-
-  **注意事项：**
-  - 需要 Java 环境
-  - 通常以 GUI 形式运行
-  - 可能需要通过命令行参数或脚本调用
-parameters:
-  - name: "jar_file"
-    type: "string"
-    description: "Stegsolve JAR 文件路径，例如 'stegsolve.jar'"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "image"
-    type: "string"
-    description: "要分析的图片文件路径"
-    required: false
-    position: 1
-    format: "positional"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的stegsolve参数。用于传递未在参数列表中定义的stegsolve选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/uro.yaml

@@ -1,70 +0,0 @@
-name: "uro"
-command: "python3"
-args:
-  - "-c"
-  - |
-    import shlex
-    import subprocess
-    import sys
-
-    if len(sys.argv) < 2:
-        sys.stderr.write("缺少URL列表\n")
-        sys.exit(1)
-
-    urls = sys.argv[1]
-    extra = sys.argv[2] if len(sys.argv) > 2 else ""
-
-    cmd = ["uro"]
-    if extra:
-        cmd.extend(shlex.split(extra))
-
-    proc = subprocess.run(
-        cmd,
-        input=urls,
-        capture_output=True,
-        text=True,
-    )
-    if proc.stdout:
-        sys.stdout.write(proc.stdout)
-    if proc.stderr:
-        sys.stderr.write(proc.stderr)
-    sys.exit(proc.returncode)
-enabled: true
-short_description: "URL过滤工具，用于过滤相似的URL"
-description: |
-  Uro是一个URL过滤工具，用于过滤掉相似的URL，去除重复项。
-
-  **主要功能：**
-  - URL去重
-  - 相似URL过滤
-  - 白名单/黑名单支持
-  - 快速处理
-
-  **使用场景：**
-  - URL去重
-  - 结果过滤
-  - 数据清理
-  - 工具链集成
-parameters:
-  - name: "urls"
-    type: "string"
-    description: "要过滤的URL（每行一个）"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的uro参数。用于传递未在参数列表中定义的uro选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    default: ""
-    position: 1
-    format: "positional"

tools/volatility.yaml

@@ -1,51 +0,0 @@
-name: "volatility"
-command: "volatility"
-enabled: true
-short_description: "内存取证分析工具"
-description: |
-  Volatility是一个内存取证框架，用于从内存转储中提取数字证据。
-
-  **主要功能：**
-  - 内存转储分析
-  - 进程列表提取
-  - 网络连接分析
-  - 文件系统重建
-
-  **使用场景：**
-  - 内存取证
-  - 恶意软件分析
-  - 事件响应
-  - 数字取证
-parameters:
-  - name: "memory_file"
-    type: "string"
-    description: "内存转储文件路径"
-    required: true
-    flag: "-f"
-    format: "flag"
-  - name: "plugin"
-    type: "string"
-    description: "要使用的Volatility插件"
-    required: true
-    position: 0
-    format: "positional"
-  - name: "profile"
-    type: "string"
-    description: "内存配置文件"
-    required: false
-    flag: "--profile"
-    format: "flag"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的volatility参数。用于传递未在参数列表中定义的volatility选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/wfuzz.yaml

@@ -1,45 +0,0 @@
-name: "wfuzz"
-command: "wfuzz"
-enabled: true
-short_description: "Web应用模糊测试工具"
-description: |
-  Wfuzz是一个Web应用模糊测试工具，用于发现Web应用中的漏洞。
-
-  **主要功能：**
-  - Web应用模糊测试
-  - 参数发现
-  - 目录发现
-  - 多种过滤器
-
-  **使用场景：**
-  - Web应用安全测试
-  - 参数模糊测试
-  - 目录枚举
-  - 漏洞发现
-parameters:
-  - name: "url"
-    type: "string"
-    description: "目标URL（使用FUZZ作为占位符）"
-    required: true
-    flag: "-u"
-    format: "flag"
-  - name: "wordlist"
-    type: "string"
-    description: "字典文件路径"
-    required: false
-    flag: "-w"
-    format: "flag"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的wfuzz参数。用于传递未在参数列表中定义的wfuzz选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

tools/winpeas.yaml

@@ -1,54 +0,0 @@
-name: "winpeas"
-command: "winPEAS.exe"
-enabled: true
-short_description: "Windows 权限提升枚举工具，自动检测常见提权路径"
-description: |
-  WinPEAS (Windows Privilege Escalation Awesome Script) 是一个自动化权限提升枚举工具，用于检测 Windows 系统中的常见提权路径。
-
-  **主要功能：**
-  - 系统信息收集
-  - 用户和组权限检查
-  - 服务配置分析
-  - 注册表检查
-  - 计划任务分析
-  - 网络配置检查
-  - 文件权限检查
-  - 凭证查找
-
-  **使用场景：**
-  - 渗透测试中的权限提升
-  - Windows 安全审计
-  - 后渗透测试
-  - CTF 竞赛
-
-  **注意事项：**
-  - 需要目标系统上已下载 winPEAS.exe
-  - 可能需要管理员权限
-  - 输出信息量大，建议保存到文件
-parameters:
-  - name: "quiet"
-    type: "bool"
-    description: "安静模式，只显示重要信息"
-    required: false
-    flag: "-q"
-    format: "flag"
-  - name: "notcolor"
-    type: "bool"
-    description: "禁用颜色输出"
-    required: false
-    flag: "-notcolor"
-    format: "flag"
-  - name: "additional_args"
-    type: "string"
-    description: |
-      额外的winpeas参数。用于传递未在参数列表中定义的winpeas选项。
-
-      **示例值：**
-      - 根据工具特性添加常用参数示例
-
-      **注意事项：**
-      - 多个参数用空格分隔
-      - 确保参数格式正确，避免命令注入
-      - 此参数会直接追加到命令末尾
-    required: false
-    format: "positional"

web/static/css/style.css

@@ -532,6 +532,10 @@ body {
     display: none;
 }
 
+.conversation-sidebar.collapsed .hitl-sidebar-card {
+    display: none;
+}
+
 .conversation-sidebar.collapsed .conversation-sidebar-header {
     flex-direction: column;
     align-items: center;
@@ -948,6 +952,395 @@ header {
     min-height: 0;
 }
 
+.hitl-sidebar-card {
+    border-top: 1px solid var(--border-color);
+    background: linear-gradient(165deg, #f8fafc 0%, #f1f5f9 55%, #eef2f7 100%);
+    padding: 14px 12px 16px;
+    flex-shrink: 0;
+}
+
+.hitl-sidebar-card-header {
+    display: flex;
+    align-items: flex-start;
+    justify-content: space-between;
+    gap: 10px;
+    cursor: pointer;
+    user-select: none;
+}
+
+.hitl-sidebar-header-actions {
+    display: flex;
+    align-items: center;
+    gap: 6px;
+    flex-shrink: 0;
+}
+
+.hitl-sidebar-body {
+    overflow: hidden;
+    max-height: 500px;
+    opacity: 1;
+    margin-top: 10px;
+    transition: max-height 0.3s ease, opacity 0.2s ease, margin-top 0.3s ease;
+}
+
+.hitl-sidebar-collapsed .hitl-sidebar-body {
+    max-height: 0;
+    opacity: 0;
+    margin-top: 0;
+}
+
+.hitl-sidebar-collapsed .hitl-apply-btn {
+    display: none;
+}
+
+.hitl-sidebar-heading {
+    display: flex;
+    align-items: flex-start;
+    gap: 10px;
+    min-width: 0;
+}
+
+.hitl-sidebar-icon {
+    flex-shrink: 0;
+    width: 36px;
+    height: 36px;
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    border-radius: 10px;
+    background: linear-gradient(145deg, rgba(0, 102, 255, 0.12), rgba(0, 102, 255, 0.06));
+    color: var(--accent-color);
+    border: 1px solid rgba(0, 102, 255, 0.18);
+}
+
+.hitl-sidebar-heading-text {
+    display: flex;
+    flex-direction: column;
+    gap: 2px;
+    min-width: 0;
+}
+
+.hitl-sidebar-title {
+    font-size: 15px;
+    font-weight: 700;
+    letter-spacing: -0.02em;
+    color: var(--text-primary);
+    line-height: 1.25;
+}
+
+.hitl-sidebar-subtitle {
+    font-size: 11px;
+    font-weight: 500;
+    color: var(--text-secondary);
+    line-height: 1.3;
+}
+
+.hitl-apply-btn {
+    padding: 8px 14px;
+    border-radius: 10px;
+    font-size: 12px;
+    font-weight: 600;
+    background: linear-gradient(180deg, var(--accent-color) 0%, var(--accent-hover) 100%);
+    color: #fff;
+    border: none;
+    cursor: pointer;
+    flex-shrink: 0;
+    box-shadow: 0 1px 2px rgba(0, 102, 255, 0.25), 0 2px 8px rgba(0, 102, 255, 0.12);
+    transition: transform 0.12s ease, box-shadow 0.12s ease, filter 0.12s ease;
+}
+
+.hitl-apply-btn:hover {
+    filter: brightness(1.05);
+    box-shadow: 0 2px 4px rgba(0, 102, 255, 0.3), 0 4px 12px rgba(0, 102, 255, 0.18);
+}
+
+.hitl-apply-btn:active {
+    transform: translateY(1px);
+}
+
+.hitl-apply-btn:disabled {
+    opacity: 0.65;
+    cursor: not-allowed;
+    transform: none;
+    filter: none;
+}
+
+.hitl-apply-feedback {
+    display: none;
+    font-size: 12px;
+    line-height: 1.4;
+    margin: 0 0 10px;
+    padding: 8px 10px;
+    border-radius: 10px;
+    background: rgba(16, 185, 129, 0.12);
+    color: #047857;
+    border: 1px solid rgba(16, 185, 129, 0.2);
+}
+
+.hitl-apply-feedback.hitl-apply-feedback--error {
+    background: rgba(239, 68, 68, 0.1);
+    color: #b91c1c;
+    border-color: rgba(239, 68, 68, 0.22);
+}
+
+/* 仅本机保存、未请求服务端：避免与「已全部同步」同款绿色造成误解 */
+.hitl-apply-feedback.hitl-apply-feedback--partial {
+    background: rgba(245, 158, 11, 0.12);
+    color: #b45309;
+    border-color: rgba(245, 158, 11, 0.25);
+}
+
+.hitl-sidebar-config {
+    border: 1px solid rgba(15, 23, 42, 0.08);
+    border-radius: 14px;
+    padding: 12px;
+    background: #fff;
+    box-shadow: 0 1px 3px rgba(15, 23, 42, 0.06), 0 8px 24px rgba(15, 23, 42, 0.04);
+}
+
+.hitl-config-field {
+    margin-bottom: 14px;
+}
+
+.hitl-config-field:last-child {
+    margin-bottom: 0;
+}
+
+.hitl-config-field--tools {
+    margin-bottom: 0;
+}
+
+.hitl-config-label {
+    display: block;
+    font-size: 12px;
+    font-weight: 600;
+    color: var(--text-primary);
+    margin-bottom: 6px;
+    letter-spacing: -0.01em;
+}
+
+.hitl-config-hint {
+    margin: 8px 0 0;
+    font-size: 11px;
+    line-height: 1.45;
+    color: var(--text-secondary);
+}
+
+.hitl-config-select {
+    width: 100%;
+    height: 40px;
+    border: 1px solid var(--border-color);
+    border-radius: 10px;
+    background: var(--bg-primary);
+    padding: 0 36px 0 12px;
+    font-size: 13px;
+    font-weight: 500;
+    color: var(--text-primary);
+    cursor: pointer;
+    appearance: none;
+    background-image: url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='16' height='16' viewBox='0 0 24 24' fill='none' stroke='%236c757d' stroke-width='2' stroke-linecap='round' stroke-linejoin='round'%3E%3Cpath d='M6 9l6 6 6-6'/%3E%3C/svg%3E");
+    background-repeat: no-repeat;
+    background-position: right 10px center;
+    transition: border-color 0.15s ease, box-shadow 0.15s ease;
+}
+
+.hitl-config-textarea {
+    display: block;
+    width: 100%;
+    min-height: 72px;
+    max-height: 200px;
+    resize: vertical;
+    border: 1px solid var(--border-color);
+    border-radius: 10px;
+    background: #fafbfc;
+    padding: 10px 12px;
+    font-size: 12px;
+    line-height: 1.5;
+    font-family: ui-monospace, SFMono-Regular, Menlo, Monaco, Consolas, "Liberation Mono", "Courier New", monospace;
+    color: var(--text-primary);
+    transition: border-color 0.15s ease, box-shadow 0.15s ease, background 0.15s ease;
+}
+
+.hitl-config-textarea::placeholder {
+    color: var(--text-muted);
+    font-family: inherit;
+}
+
+/* 其它页面内联 HITL 评论框等仍用 input 类名 */
+.hitl-config-input {
+    width: 100%;
+    height: 38px;
+    border: 1px solid #e2e8f0;
+    border-radius: 8px;
+    background: #fff;
+    padding: 0 12px;
+    font-size: 14px;
+    color: var(--text-primary);
+    transition: border-color 0.15s, box-shadow 0.15s;
+}
+.hitl-config-input:focus {
+    outline: none;
+    border-color: var(--accent-color, #0066ff);
+    box-shadow: 0 0 0 3px rgba(0, 102, 255, 0.1);
+}
+
+.hitl-pending-list {
+    display: flex;
+    flex-direction: column;
+    gap: 12px;
+}
+
+.hitl-pending-item {
+    border: 1px solid #dbeafe;
+    border-radius: 10px;
+    padding: 16px;
+    background: #f8fbff;
+    transition: box-shadow 0.2s;
+}
+.hitl-pending-item:hover {
+    box-shadow: 0 2px 12px rgba(0, 102, 255, 0.08);
+}
+
+.hitl-pending-item-header {
+    display: flex;
+    justify-content: space-between;
+    align-items: center;
+    margin-bottom: 10px;
+}
+
+.hitl-pending-item-title {
+    display: flex;
+    align-items: center;
+    gap: 8px;
+}
+
+.hitl-tool-badge {
+    display: inline-flex;
+    align-items: center;
+    padding: 3px 10px;
+    border-radius: 6px;
+    font-size: 13px;
+    font-weight: 600;
+    font-family: ui-monospace, SFMono-Regular, Menlo, monospace;
+    background: var(--accent-color, #0066ff);
+    color: #fff;
+}
+
+.hitl-mode-tag {
+    display: inline-flex;
+    align-items: center;
+    padding: 2px 8px;
+    border-radius: 10px;
+    font-size: 11px;
+    font-weight: 500;
+    background: #e0e7ff;
+    color: #4338ca;
+}
+.hitl-mode-tag--review_edit {
+    background: #fef3c7;
+    color: #92400e;
+}
+
+.hitl-pending-meta {
+    font-size: 12px;
+    color: #94a3b8;
+    margin-bottom: 8px;
+    word-break: break-all;
+}
+
+.hitl-pending-payload {
+    white-space: pre-wrap;
+    word-break: break-all;
+    max-height: 160px;
+    overflow: auto;
+    margin: 0 0 4px 0;
+    padding: 10px 12px;
+    border-radius: 8px;
+    background: #fff;
+    border: 1px solid #e2e8f0;
+    font-size: 12px;
+    font-family: ui-monospace, SFMono-Regular, Menlo, monospace;
+    color: #334155;
+    line-height: 1.5;
+}
+
+.hitl-pending-item-header-right {
+    display: flex;
+    align-items: center;
+    gap: 8px;
+}
+
+.hitl-dismiss-btn {
+    background: none;
+    border: none;
+    cursor: pointer;
+    font-size: 18px;
+    line-height: 1;
+    color: var(--text-secondary, #64748b);
+    padding: 2px 6px;
+    border-radius: 4px;
+    transition: background 0.15s, color 0.15s;
+}
+.hitl-dismiss-btn:hover {
+    background: rgba(239, 68, 68, 0.1);
+    color: #ef4444;
+}
+
+.hitl-pending-actions {
+    display: flex;
+    gap: 8px;
+    margin-top: 12px;
+    padding-top: 12px;
+    border-top: 1px solid #e2e8f0;
+}
+
+.hitl-edit-args {
+    width: 100%;
+    min-height: 76px;
+    margin-top: 8px;
+    border: 1px solid #e2e8f0;
+    border-radius: 8px;
+    background: #ffffff;
+    color: #1f2937;
+    padding: 10px 12px;
+    font-family: ui-monospace, SFMono-Regular, Menlo, monospace;
+    font-size: 12px;
+    transition: border-color 0.15s, box-shadow 0.15s;
+}
+.hitl-edit-args:focus {
+    outline: none;
+    border-color: var(--accent-color, #0066ff);
+    box-shadow: 0 0 0 3px rgba(0, 102, 255, 0.1);
+}
+
+.hitl-input-help {
+    margin-top: 6px;
+    margin-bottom: 6px;
+    font-size: 12px;
+    color: var(--accent-color, #0066ff);
+    line-height: 1.4;
+}
+
+.hitl-inline-approval {
+    margin-top: 8px;
+    padding: 10px;
+    border: 1px solid #dbeafe;
+    background: #f8fbff;
+    border-radius: 8px;
+}
+
+.hitl-inline-approval.hitl-inline-done {
+    opacity: 0.8;
+}
+
+.hitl-config-select:focus,
+.hitl-config-textarea:focus {
+    outline: none;
+    border-color: var(--accent-color);
+    box-shadow: 0 0 0 3px rgba(0, 102, 255, 0.14);
+    background: #fff;
+}
+
 .sidebar-title {
     font-size: 0.8125rem;
     font-weight: 600;
@@ -3590,6 +3983,83 @@ header {
     margin-bottom: 32px;
 }
 
+.mcp-management-layout {
+    display: grid;
+    grid-template-columns: minmax(0, 2fr) minmax(0, 3fr);
+    gap: 16px;
+    align-items: stretch;
+    height: calc(100vh - 210px);
+    min-height: 520px;
+    max-height: calc(100vh - 210px);
+}
+
+.mcp-management-panel {
+    margin-bottom: 0 !important;
+    padding: 14px 16px 16px;
+    border: 1px solid var(--border-color);
+    border-radius: 12px;
+    background: var(--bg-primary);
+    display: flex;
+    flex-direction: column;
+    min-height: 0;
+}
+
+.mcp-tools-panel {
+    min-width: 0;
+    order: 2;
+}
+
+.mcp-external-panel {
+    min-width: 0;
+    order: 1;
+}
+
+.mcp-panel-body {
+    display: flex;
+    flex-direction: column;
+    flex: 1;
+    min-height: 0;
+}
+
+.mcp-tools-panel .tools-controls,
+.mcp-external-panel .external-mcp-controls {
+    flex: 1;
+    min-height: 0;
+}
+
+.mcp-tools-panel .tools-list {
+    flex: 1;
+    min-height: 0;
+    max-height: 100%;
+    overflow-y: auto;
+}
+
+.mcp-tools-panel .tools-list-items {
+    max-height: none;
+    overflow: visible;
+}
+
+.mcp-external-panel .external-mcp-list {
+    flex: 1;
+    min-height: 0;
+    overflow-y: auto;
+    padding-right: 4px;
+}
+
+.mcp-external-panel .external-mcp-controls {
+    gap: 12px;
+}
+
+/* MCP 双栏内工具操作条允许换行，避免面板内溢出 */
+.mcp-tools-panel .tools-actions {
+    flex-wrap: wrap;
+    row-gap: 8px;
+}
+
+.mcp-tools-panel .search-box {
+    min-width: min(280px, 100%);
+}
+
 .settings-section:last-child {
     margin-bottom: 0;
 }
@@ -5298,8 +5768,8 @@ header {
 
 .external-mcp-item-details {
     display: grid;
-    grid-template-columns: repeat(auto-fit, minmax(220px, 1fr));
-    gap: 16px;
+    grid-template-columns: repeat(auto-fit, minmax(150px, auto));
+    gap: 12px;
     padding-top: 16px;
     border-top: 1px solid var(--border-color);
 }
@@ -5396,6 +5866,23 @@ header {
 
 /* 响应式优化 */
 @media (max-width: 768px) {
+    .mcp-management-layout {
+        grid-template-columns: 1fr;
+        height: auto;
+        min-height: auto;
+        max-height: none;
+    }
+
+    .mcp-management-panel {
+        min-height: 0;
+    }
+
+    .mcp-tools-panel .tools-list,
+    .mcp-external-panel .external-mcp-list {
+        min-height: 200px;
+        max-height: 52vh;
+    }
+
     .tools-actions {
         gap: 6px;
     }
@@ -5578,6 +6065,7 @@ header {
     transition: all 0.2s ease;
 }
 
+
 .legend-item:hover {
     transform: translateX(2px);
 }
@@ -8482,7 +8970,7 @@ header {
 /* 任务管理 · 队列卡片：单行主网格 + 进度列内统计，降低高度 */
 .batch-queue-item__inner--grid {
     display: grid;
-    grid-template-columns: minmax(0, 1fr) minmax(128px, auto) minmax(88px, 14%) 44px;
+    grid-template-columns: minmax(0, 1fr) minmax(128px, auto) minmax(88px, 14%) minmax(40px, max-content);
     grid-template-rows: auto;
     grid-template-areas: "lead cluster progress actions";
     column-gap: 22px;
@@ -8563,6 +9051,12 @@ header {
     justify-self: end;
     align-self: center;
     padding-left: 6px;
+    display: flex;
+    flex-direction: row;
+    flex-wrap: nowrap;
+    align-items: center;
+    justify-content: flex-end;
+    gap: 6px;
 }
 
 .batch-queue-item__idline--lead {
@@ -8649,6 +9143,12 @@ header {
 }
 
 .batch-queue-icon-btn:hover {
+    color: var(--accent-color, #0066ff);
+    border-color: rgba(0, 102, 255, 0.35);
+    background: rgba(0, 102, 255, 0.08);
+}
+
+.batch-queue-icon-btn--danger:hover {
     color: var(--error-color, #dc3545);
     border-color: rgba(220, 53, 69, 0.35);
     background: rgba(220, 53, 69, 0.06);
@@ -13087,29 +13587,87 @@ header {
 
 .vulnerability-details {
     display: grid;
-    grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
-    gap: 12px;
+    grid-template-columns: repeat(2, minmax(0, 1fr));
+    gap: 14px 16px;
     margin-bottom: 16px;
     padding: 12px;
     background: var(--bg-secondary);
     border-radius: 6px;
+    align-items: start;
 }
 
-.detail-item {
+/* 元数据条数为奇数时，最后一项占满一行，长 URL/队列 ID 更易读 */
+.vulnerability-details .vuln-detail-field:last-child:nth-child(odd) {
+    grid-column: 1 / -1;
+}
+
+@media (max-width: 768px) {
+    .vulnerability-details {
+        grid-template-columns: 1fr;
+    }
+
+    .vulnerability-details .vuln-detail-field:last-child:nth-child(odd) {
+        grid-column: auto;
+    }
+}
+
+/* 漏洞详情字段：标签与值分行，长 ID/URL 可换行、可选中复制 */
+.vuln-detail-field {
+    min-width: 0;
     font-size: 0.875rem;
 }
 
-.detail-item strong {
+.vuln-detail-field__label {
     color: var(--text-secondary);
-    margin-right: 4px;
+    font-weight: 600;
+    font-size: 0.75rem;
+    margin-bottom: 6px;
+    text-transform: none;
+    letter-spacing: normal;
 }
 
-.detail-item code {
+.vuln-detail-field__row {
+    display: flex;
+    align-items: flex-start;
+    gap: 8px;
+    min-width: 0;
+}
+
+.vuln-detail-field-value {
+    flex: 1;
+    min-width: 0;
+    margin: 0;
+    padding: 8px 10px;
+    border-radius: 6px;
     background: var(--bg-tertiary);
-    padding: 2px 6px;
-    border-radius: 4px;
-    font-size: 0.8rem;
-    font-family: 'Monaco', 'Menlo', 'Ubuntu Mono', monospace;
+    border: 1px solid var(--border-color);
+    font-size: 0.8125rem;
+    line-height: 1.45;
+    word-break: break-word;
+    overflow-wrap: anywhere;
+    white-space: pre-wrap;
+    user-select: text;
+    -webkit-user-select: text;
+    color: var(--text-primary);
+    font-family: 'Monaco', 'Menlo', 'Ubuntu Mono', 'Consolas', monospace;
+}
+
+.vuln-detail-field__copy {
+    flex-shrink: 0;
+    margin-top: 2px;
+    padding: 6px;
+    line-height: 0;
+    border-radius: 6px;
+    color: var(--text-secondary);
+    border: 1px solid transparent;
+    background: transparent;
+    cursor: pointer;
+}
+
+.vuln-detail-field__copy:hover {
+    color: var(--accent-color);
+    background: var(--bg-primary);
+    border-color: var(--border-color);
 }
 
 .vulnerability-proof,

web/static/i18n/en-US.json

@@ -61,7 +61,8 @@
     "agentsManagement": "Agent management",
     "roles": "Roles",
     "rolesManagement": "Roles Management",
-    "settings": "System settings"
+    "settings": "System settings",
+    "hitl": "Human-in-the-loop"
   },
   "dashboard": {
     "title": "Dashboard",
@@ -177,7 +178,12 @@
     "taskCancelled": "Task cancelled",
     "unknownTool": "Unknown tool",
     "einoAgentReplyTitle": "Sub-agent reply",
-    "einoRecoveryTitle": "🔄 Invalid tool JSON · run {{n}}/{{max}} (hint appended)",
+    "einoStreamErrorTitle": "⚠️ Eino stream interrupted ({{agent}})",
+    "einoStreamErrorMessage": "Streaming read failed; the system will retry or terminate according to policy.",
+    "iterationLimitReachedTitle": "⛔ Iteration limit reached",
+    "iterationLimitReachedMessage": "Maximum iteration count reached; automatic iteration has stopped.",
+    "einoPendingOrphanedTitle": "🧹 Tool call reconciliation",
+    "einoPendingOrphanedMessage": "Detected {{count}} unclosed tool call(s); marked as failed and finalized automatically.",
     "noDescription": "No description",
     "noResponseData": "No response data",
     "loading": "Loading...",
@@ -191,6 +197,9 @@
     "executionFailed": "Execution failed",
     "penetrationTestComplete": "Penetration test complete",
     "yesterday": "Yesterday",
+    "historyGroupToday": "Today",
+    "historyGroupLast7Days": "Past 7 days",
+    "historyGroupEarlier": "Older",
     "agentModeSelectAria": "Choose conversation execution mode",
     "agentModePanelTitle": "Conversation mode",
     "agentModeReactNative": "Native ReAct",
@@ -208,7 +217,29 @@
     "agentModeSingleHint": "Single-model ReAct loop for chat and tool use",
     "agentModeMultiHint": "Eino prebuilt orchestration (deep / plan_execute / supervisor) for complex tasks",
     "agentModeOrchPlanExecute": "Plan-Exec",
-    "agentModeOrchSupervisor": "Supervisor"
+    "agentModeOrchSupervisor": "Supervisor",
+    "hitlTitle": "Human-in-the-loop",
+    "hitlCardSubtitle": "Approvals & allowlist",
+    "hitlReviewer": "Review",
+    "hitlConfigTitle": "Collaboration mode config",
+    "hitlModeLabel": "Mode",
+    "hitlModeOff": "Off",
+    "hitlModeApproval": "Approval",
+    "hitlModeReviewEdit": "Review & Edit",
+    "hitlSensitiveTools": "Sensitive tools (comma-separated)",
+    "hitlWhitelistTools": "Whitelisted tools (skip approval, comma-separated)",
+    "hitlWhitelistPlaceholder": "e.g. read_file, grep or one tool per line (merged with global allowlist in config)",
+    "hitlWhitelistHint": "Separate with commas or new lines; shown merged with the global allowlist in config.",
+    "hitlApply": "Apply",
+    "hitlApplyOkSync": "HITL settings saved and synced to the server.",
+    "hitlApplyOkWhitelistYaml": "Tool whitelist merged into config.yaml and active. Mode and timeout still require selecting a conversation and clicking Apply to sync session settings to the server.",
+    "hitlApplyOkLocal": "Saved in this browser.",
+    "hitlApplyFail": "Failed to sync to server",
+    "hitlStatusOff": "Human-in-the-loop: Off"
+  },
+  "hitl": {
+    "pageTitle": "HITL approvals",
+    "pendingTitle": "Pending approvals"
   },
   "progress": {
     "callingAI": "Calling AI model...",
@@ -272,6 +303,8 @@
     "clearHistory": "Clear history",
     "cancelTask": "Cancel task",
     "viewConversation": "View conversation",
+    "viewVulnerabilities": "View vulnerabilities",
+    "viewVulnerabilitiesQueueTitle": "View vulnerabilities: open management filtered to this queue",
     "retryTask": "Retry",
     "conversationIdLabel": "Conversation ID",
     "statusPending": "Pending",
@@ -1281,6 +1314,12 @@
     "clear": "Clear",
     "vulnId": "Vuln ID",
     "conversationId": "Conversation ID",
+    "taskOrQueueId": "Task / queue ID",
+    "filterTaskOrQueue": "Filter by task or queue ID",
+    "conversationTag": "Conversation tag",
+    "filterConversationTag": "Filter by conversation tag",
+    "taskTag": "Task tag",
+    "filterTaskTag": "Filter by task tag",
     "severity": "Severity",
     "status": "Status",
     "statusOpen": "Open",
@@ -1290,7 +1329,31 @@
     "searchVulnId": "Search vuln ID",
     "filterConversation": "Filter by conversation",
     "loading": "Loading...",
-    "noRecords": "No vulnerability records"
+    "loadListFailed": "Failed to load",
+    "noRecords": "No vulnerability records",
+    "batchExport": "Batch export",
+    "downloadMarkdownTitle": "Download Markdown",
+    "exportNoResults": "No vulnerabilities match the current filters",
+    "exportStarted": "Started downloading {{count}} file(s)",
+    "exportFailed": "Export failed",
+    "saveRequiredFields": "Please fill in conversation ID, title, and severity",
+    "saveFailed": "Save failed",
+    "fetchFailed": "Failed to fetch vulnerability",
+    "deleteFailed": "Delete failed",
+    "detailVulnId": "Vuln ID",
+    "detailType": "Type",
+    "detailTarget": "Target",
+    "detailConversationId": "Conversation ID",
+    "detailTaskId": "Task ID",
+    "detailTaskQueueId": "Task queue ID",
+    "detailConversationTag": "Conversation tag",
+    "detailTaskTag": "Task tag",
+    "detailProof": "Proof",
+    "detailImpact": "Impact",
+    "detailRecommendation": "Remediation",
+    "downloadOkTitle": "Downloaded",
+    "exportFailedMessage": "Export failed",
+    "downloadFailed": "Download failed"
   },
   "tasksPage": {
     "statusFilter": "Status filter",
@@ -1387,9 +1450,6 @@
     "multiAgentPeLoop": "plan_execute outer loop limit",
     "multiAgentPeLoopPlaceholder": "0 uses Eino default (10)",
     "multiAgentPeLoopHint": "Only for plan_execute; max execute↔replan rounds.",
-    "multiAgentDefaultMode": "Default mode on chat page",
-    "multiAgentModeSingle": "Single-agent (ReAct)",
-    "multiAgentModeMulti": "Multi-agent (Eino)",
     "multiAgentRobotUse": "Use multi-agent for WeCom / DingTalk / Lark bots",
     "multiAgentRobotUseHint": "Requires 'Enable multi-agent' to be checked; usage and cost will be higher.",
     "multiAgentBatchUse": "Use multi-agent for batch task queues",
@@ -1644,6 +1704,7 @@
   },
   "contextMenu": {
     "viewAttackChain": "View attack chain",
+    "viewVulnerabilities": "View vulnerabilities",
     "downloadMarkdown": "Download Markdown",
     "downloadMarkdownSummary": "Summary",
     "downloadMarkdownFull": "Full",
@@ -1739,6 +1800,10 @@
   "vulnerabilityModal": {
     "conversationId": "Conversation ID",
     "conversationIdPlaceholder": "Enter conversation ID",
+    "conversationTag": "Conversation tag",
+    "conversationTagPlaceholder": "e.g. engagement A, weekly report",
+    "taskTag": "Task tag",
+    "taskTagPlaceholder": "e.g. batch scan Q2, retest",
     "title": "Title",
     "titlePlaceholder": "Vulnerability title",
     "description": "Description",
@@ -1766,6 +1831,25 @@
     "recommendation": "Recommendation",
     "recommendationPlaceholder": "Remediation"
   },
+  "vulnerabilityMd": {
+    "headingBasic": "Basic information",
+    "labelId": "Vulnerability ID",
+    "labelSeverity": "Severity",
+    "labelStatus": "Status",
+    "labelType": "Type",
+    "labelTarget": "Target",
+    "labelConversationId": "Conversation ID",
+    "labelTaskId": "Task ID",
+    "labelTaskQueueId": "Task queue ID",
+    "labelConversationTag": "Conversation tag",
+    "labelTaskTag": "Task tag",
+    "labelCreated": "Created at",
+    "labelUpdated": "Updated at",
+    "headingDescription": "Description",
+    "headingProof": "Proof (POC)",
+    "headingImpact": "Impact",
+    "headingRecommendation": "Remediation"
+  },
   "roleModal": {
     "addRole": "Add role",
     "editRole": "Edit role",

web/static/i18n/zh-CN.json

@@ -61,7 +61,8 @@
     "agentsManagement": "Agent管理",
     "roles": "角色",
     "rolesManagement": "角色管理",
-    "settings": "系统设置"
+    "settings": "系统设置",
+    "hitl": "人机协同"
   },
   "dashboard": {
     "title": "仪表盘",
@@ -177,7 +178,12 @@
     "taskCancelled": "任务已取消",
     "unknownTool": "未知工具",
     "einoAgentReplyTitle": "子代理回复",
-    "einoRecoveryTitle": "🔄 工具参数无效 · 第 {{n}}/{{max}} 轮（已追加提示）",
+    "einoStreamErrorTitle": "⚠️ Eino 流式中断（{{agent}}）",
+    "einoStreamErrorMessage": "流式读取异常，系统将按策略重试或结束。",
+    "iterationLimitReachedTitle": "⛔ 达到迭代上限",
+    "iterationLimitReachedMessage": "已达到最大迭代次数，任务已停止继续自动迭代。",
+    "einoPendingOrphanedTitle": "🧹 工具调用收尾补偿",
+    "einoPendingOrphanedMessage": "检测到 {{count}} 个未闭合工具调用，已自动标记为失败并收尾。",
     "noDescription": "暂无描述",
     "noResponseData": "暂无响应数据",
     "loading": "加载中...",
@@ -191,6 +197,9 @@
     "executionFailed": "执行失败",
     "penetrationTestComplete": "渗透测试完成",
     "yesterday": "昨天",
+    "historyGroupToday": "今天",
+    "historyGroupLast7Days": "过去七天",
+    "historyGroupEarlier": "更早",
     "agentModeSelectAria": "选择对话执行模式",
     "agentModePanelTitle": "对话模式",
     "agentModeReactNative": "原生 ReAct 模式",
@@ -208,7 +217,29 @@
     "agentModeSingleHint": "单模型 ReAct 循环，适合常规对话与工具调用",
     "agentModeMultiHint": "Eino 预置编排（deep / plan_execute / supervisor），适合复杂任务",
     "agentModeOrchPlanExecute": "Plan-Exec",
-    "agentModeOrchSupervisor": "Supervisor"
+    "agentModeOrchSupervisor": "Supervisor",
+    "hitlTitle": "人机协同",
+    "hitlCardSubtitle": "审批与白名单",
+    "hitlReviewer": "Review",
+    "hitlConfigTitle": "协同模式配置",
+    "hitlModeLabel": "模式",
+    "hitlModeOff": "关闭",
+    "hitlModeApproval": "审批模式",
+    "hitlModeReviewEdit": "审查编辑",
+    "hitlSensitiveTools": "敏感工具（逗号分隔）",
+    "hitlWhitelistTools": "白名单工具（免审批，逗号分隔）",
+    "hitlWhitelistPlaceholder": "例：read_file, grep 或每行一个工具名（与 config 全局白名单合并）",
+    "hitlWhitelistHint": "每行一个或逗号分隔；与 config 中全局白名单合并展示。",
+    "hitlApply": "应用",
+    "hitlApplyOkSync": "人机协同配置已保存并同步到服务器。",
+    "hitlApplyOkWhitelistYaml": "免审批工具已合并进 config.yaml 并生效。协同模式、超时等仍须选中会话后再点「应用」才会写入服务器。",
+    "hitlApplyOkLocal": "已保存到本浏览器。",
+    "hitlApplyFail": "同步到服务器失败",
+    "hitlStatusOff": "人机协同：关闭"
+  },
+  "hitl": {
+    "pageTitle": "人机协同审批",
+    "pendingTitle": "待处理审批"
   },
   "progress": {
     "callingAI": "正在调用AI模型...",
@@ -272,6 +303,8 @@
     "clearHistory": "清空历史",
     "cancelTask": "取消任务",
     "viewConversation": "查看对话",
+    "viewVulnerabilities": "查看漏洞",
+    "viewVulnerabilitiesQueueTitle": "查看漏洞：打开漏洞管理并筛选本队列",
     "retryTask": "重试",
     "conversationIdLabel": "对话ID",
     "statusPending": "待执行",
@@ -1281,6 +1314,12 @@
     "clear": "清除",
     "vulnId": "漏洞ID",
     "conversationId": "会话ID",
+    "taskOrQueueId": "任务ID/队列ID",
+    "filterTaskOrQueue": "筛选任务ID或队列ID",
+    "conversationTag": "对话标签",
+    "filterConversationTag": "筛选对话标签",
+    "taskTag": "任务标签",
+    "filterTaskTag": "筛选任务标签",
     "severity": "严重程度",
     "status": "状态",
     "statusOpen": "待处理",
@@ -1290,7 +1329,31 @@
     "searchVulnId": "搜索漏洞ID",
     "filterConversation": "筛选特定会话",
     "loading": "加载中...",
-    "noRecords": "暂无漏洞记录"
+    "loadListFailed": "加载失败",
+    "noRecords": "暂无漏洞记录",
+    "batchExport": "批量导出",
+    "downloadMarkdownTitle": "下载 Markdown",
+    "exportNoResults": "当前筛选条件下无可导出漏洞",
+    "exportStarted": "已开始下载 {{count}} 份报告",
+    "exportFailed": "导出失败",
+    "saveRequiredFields": "请填写必填字段：会话ID、标题和严重程度",
+    "saveFailed": "保存失败",
+    "fetchFailed": "获取漏洞失败",
+    "deleteFailed": "删除失败",
+    "detailVulnId": "漏洞ID",
+    "detailType": "类型",
+    "detailTarget": "目标",
+    "detailConversationId": "会话ID",
+    "detailTaskId": "任务ID",
+    "detailTaskQueueId": "任务队列ID",
+    "detailConversationTag": "对话标签",
+    "detailTaskTag": "任务标签",
+    "detailProof": "证明",
+    "detailImpact": "影响",
+    "detailRecommendation": "修复建议",
+    "downloadOkTitle": "下载成功",
+    "exportFailedMessage": "导出失败",
+    "downloadFailed": "下载失败"
   },
   "tasksPage": {
     "statusFilter": "状态筛选",
@@ -1387,9 +1450,6 @@
     "multiAgentPeLoop": "plan_execute 外层循环上限",
     "multiAgentPeLoopPlaceholder": "0 表示 Eino 默认 10",
     "multiAgentPeLoopHint": "仅 plan_execute 有效；execute 与 replan 之间的最大轮次。",
-    "multiAgentDefaultMode": "对话页默认模式",
-    "multiAgentModeSingle": "单代理（ReAct）",
-    "multiAgentModeMulti": "多代理（Eino）",
     "multiAgentRobotUse": "企业微信 / 钉钉 / 飞书机器人也使用多代理",
     "multiAgentRobotUseHint": "需同时勾选「启用多代理」；调用量与成本更高。",
     "multiAgentBatchUse": "批量任务队列也使用多代理",
@@ -1644,6 +1704,7 @@
   },
   "contextMenu": {
     "viewAttackChain": "查看攻击链",
+    "viewVulnerabilities": "查看漏洞",
     "downloadMarkdown": "下载 Markdown",
     "downloadMarkdownSummary": "简版",
     "downloadMarkdownFull": "完整版",
@@ -1739,6 +1800,10 @@
   "vulnerabilityModal": {
     "conversationId": "会话ID",
     "conversationIdPlaceholder": "输入会话ID",
+    "conversationTag": "对话标签",
+    "conversationTagPlaceholder": "如：红队演练A、客户A周报",
+    "taskTag": "任务标签",
+    "taskTagPlaceholder": "如：批量扫描Q2、专项复测",
     "title": "标题",
     "titlePlaceholder": "漏洞标题",
     "description": "描述",
@@ -1766,6 +1831,25 @@
     "recommendation": "修复建议",
     "recommendationPlaceholder": "修复建议"
   },
+  "vulnerabilityMd": {
+    "headingBasic": "基本信息",
+    "labelId": "漏洞ID",
+    "labelSeverity": "严重程度",
+    "labelStatus": "状态",
+    "labelType": "类型",
+    "labelTarget": "目标",
+    "labelConversationId": "会话ID",
+    "labelTaskId": "任务ID",
+    "labelTaskQueueId": "任务队列ID",
+    "labelConversationTag": "对话标签",
+    "labelTaskTag": "任务标签",
+    "labelCreated": "创建时间",
+    "labelUpdated": "更新时间",
+    "headingDescription": "描述",
+    "headingProof": "证明（POC）",
+    "headingImpact": "影响",
+    "headingRecommendation": "修复建议"
+  },
   "roleModal": {
     "addRole": "添加角色",
     "editRole": "编辑角色",

web/static/js/hitl.js

@@ -0,0 +1,414 @@
+function hitlModeNormalize(m) {
+    let v = String(m || '').trim().toLowerCase().replace(/-/g, '_');
+    if (v === 'feedback' || v === 'followup') {
+        v = 'approval';
+    }
+    const allowed = ['off', 'approval', 'review_edit'];
+    return allowed.indexOf(v) >= 0 ? v : 'off';
+}
+
+function hitlEffectiveEnabled(cfg) {
+    if (!cfg) return false;
+    if (cfg.enabled === true) return true;
+    return hitlModeNormalize(cfg.mode) !== 'off';
+}
+
+function readHitlLocalStorageConv(conversationId) {
+    if (!conversationId) return null;
+    try {
+        const key = 'cyberstrike-chat-hitl:' + String(conversationId).trim();
+        const raw = localStorage.getItem(key);
+        if (!raw) return null;
+        const parsed = JSON.parse(raw);
+        if (!parsed || typeof parsed !== 'object') return null;
+        return parsed;
+    } catch (e) {
+        return null;
+    }
+}
+
+function hitlSensitiveToolsToArray(config) {
+    if (Array.isArray(config && config.sensitiveTools)) return config.sensitiveTools;
+    const s = config && config.sensitiveTools;
+    if (typeof s === 'string') {
+        return s.split(/[,\n\r]+/).map(function (x) { return x.trim(); }).filter(Boolean);
+    }
+    return [];
+}
+
+function getCurrentConversationIdForHitl() {
+    if (typeof window.currentConversationId === 'string' && window.currentConversationId) {
+        return window.currentConversationId;
+    }
+    const active = document.querySelector('.conversation-item.active');
+    if (active && active.dataset && active.dataset.conversationId) {
+        return active.dataset.conversationId;
+    }
+    return '';
+}
+
+async function fetchHitlConversationConfig(conversationId) {
+    if (!conversationId) return null;
+    const resp = await hitlApiFetch('/api/hitl/config/' + encodeURIComponent(conversationId), { credentials: 'same-origin' });
+    if (!resp.ok) return null;
+    const data = await resp.json();
+    if (!data || !data.hitl) return null;
+    return {
+        hitl: data.hitl,
+        hitlGlobalToolWhitelist: Array.isArray(data.hitlGlobalToolWhitelist) ? data.hitlGlobalToolWhitelist : []
+    };
+}
+
+/** 无会话时：将免审批工具合并进服务端 config.yaml，返回更新后的全局白名单数组 */
+async function mergeHitlGlobalToolWhitelist(sensitiveTools) {
+    const list = Array.isArray(sensitiveTools) ? sensitiveTools : [];
+    const resp = await hitlApiFetch('/api/hitl/tool-whitelist', {
+        method: 'POST',
+        credentials: 'same-origin',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ sensitiveTools: list })
+    });
+    if (!resp.ok) {
+        const msg = await readHitlApiError(resp);
+        throw new Error(msg || ('HTTP ' + resp.status));
+    }
+    const data = await resp.json();
+    if (data && Array.isArray(data.hitlGlobalToolWhitelist)) {
+        return data.hitlGlobalToolWhitelist;
+    }
+    return [];
+}
+
+async function saveHitlConversationConfig(conversationId, config) {
+    if (!conversationId || !config) return false;
+    const mode = hitlModeNormalize(config.mode || 'off');
+    const enabled = typeof config.enabled === 'boolean' ? config.enabled : (mode !== 'off');
+    const sensitiveTools = hitlSensitiveToolsToArray(config);
+    const resp = await hitlApiFetch('/api/hitl/config', {
+        method: 'PUT',
+        credentials: 'same-origin',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({
+            conversationId: conversationId,
+            enabled: enabled,
+            mode: mode,
+            sensitiveTools: sensitiveTools,
+            timeoutSeconds: config.timeoutSeconds || 300
+        })
+    });
+    if (!resp.ok) {
+        const msg = await readHitlApiError(resp);
+        throw new Error(msg || ('HTTP ' + resp.status));
+    }
+    return true;
+}
+
+async function syncHitlConfigFromServer(conversationId) {
+    const pack = await fetchHitlConversationConfig(conversationId);
+    if (!pack || !pack.hitl) return;
+    const cfg = pack.hitl;
+    const globalWL = pack.hitlGlobalToolWhitelist || [];
+    if (typeof window !== 'undefined') {
+        window.csaiHitlGlobalToolWhitelist = globalWL;
+    }
+    const strip = typeof window.hitlStripGlobalToolsFromFormString === 'function'
+        ? window.hitlStripGlobalToolsFromFormString
+        : function (_g, s) { return typeof s === 'string' ? s.trim() : ''; };
+
+    let merged = cfg;
+    if (!hitlEffectiveEnabled(cfg)) {
+        const local = readHitlLocalStorageConv(conversationId);
+        const localMode = local && local.mode ? hitlModeNormalize(local.mode) : 'off';
+        if (localMode !== 'off') {
+            let localToolsStr = typeof local.sensitiveTools === 'string' ? local.sensitiveTools : '';
+            localToolsStr = strip(globalWL, localToolsStr);
+            merged = {
+                enabled: true,
+                mode: localMode,
+                sensitiveTools: localToolsStr.split(/[,\n\r]+/).map(function (s) { return s.trim(); }).filter(Boolean),
+                timeoutSeconds: cfg.timeoutSeconds || 300
+            };
+            saveHitlConversationConfig(conversationId, {
+                mode: localMode,
+                sensitiveTools: localToolsStr,
+                enabled: true,
+                timeoutSeconds: merged.timeoutSeconds
+            }).catch(function (err) {
+                console.warn('HITL 会话配置同步到服务器失败（将仅保留本地 UI）:', err);
+            });
+        } else {
+            const gl = typeof window.getHitlLastGlobalConfig === 'function' ? window.getHitlLastGlobalConfig() : null;
+            const glMode = gl && gl.mode ? hitlModeNormalize(gl.mode) : 'off';
+            if (glMode !== 'off') {
+                let glToolsStr = typeof gl.sensitiveTools === 'string' ? gl.sensitiveTools : '';
+                glToolsStr = strip(globalWL, glToolsStr);
+                merged = {
+                    enabled: true,
+                    mode: glMode,
+                    sensitiveTools: glToolsStr.split(/[,\n\r]+/).map(function (s) { return s.trim(); }).filter(Boolean),
+                    timeoutSeconds: cfg.timeoutSeconds || 300
+                };
+                saveHitlConversationConfig(conversationId, {
+                    mode: glMode,
+                    sensitiveTools: glToolsStr,
+                    enabled: true,
+                    timeoutSeconds: merged.timeoutSeconds
+                }).catch(function (err) {
+                    console.warn('HITL 会话配置同步到服务器失败（将仅保留本地 UI）:', err);
+                });
+            }
+        }
+    }
+    const uiMode = hitlEffectiveEnabled(merged) ? hitlModeNormalize(merged.mode) : 'off';
+    const rawArr = Array.isArray(merged.sensitiveTools)
+        ? merged.sensitiveTools
+        : hitlSensitiveToolsToArray({ sensitiveTools: merged.sensitiveTools });
+    const sessionOnlyStr = strip(globalWL, rawArr.join(', '));
+    const normalizedCfg = Object.assign({}, merged, {
+        mode: uiMode,
+        sensitiveTools: sessionOnlyStr
+    });
+    if (typeof window.saveHitlConfigForConversation === 'function') {
+        window.saveHitlConfigForConversation(conversationId, normalizedCfg);
+    } else {
+        try {
+            localStorage.setItem('chat_hitl_config_' + conversationId, JSON.stringify(normalizedCfg));
+        } catch (e) {}
+    }
+    if (typeof window.applyHitlConfigToUI === 'function') {
+        window.applyHitlConfigToUI(normalizedCfg);
+    }
+    reconcileHitlUiState();
+}
+
+async function syncHitlConfigToServerByCurrentConversation() {
+    const conversationId = getCurrentConversationIdForHitl();
+    if (!conversationId) return;
+    if (typeof window.readHitlConfigFromForm !== 'function') return;
+    const cfg = window.readHitlConfigFromForm();
+    await saveHitlConversationConfig(conversationId, cfg);
+}
+
+function reconcileHitlUiState() {
+    if (typeof window.readHitlConfigFromForm === 'function' && typeof window.updateHitlStatusUI === 'function') {
+        try {
+            const cfg = window.readHitlConfigFromForm();
+            window.updateHitlStatusUI(cfg);
+        } catch (e) {}
+    }
+}
+
+let hitlFollowRunSeq = 0;
+
+/**
+ * 审批提交后原 SSE 已断开：轮询任务列表，运行中则拉取过程详情；任务结束后再整页加载会话以对齐终态。
+ */
+async function followAgentRunAfterHitlDecision(conversationId) {
+    if (!conversationId || typeof apiFetch !== 'function') return;
+    if (typeof window.attachRunningTaskEventStream === 'function') {
+        try {
+            const attached = await window.attachRunningTaskEventStream(conversationId);
+            if (attached) return;
+        } catch (e) {
+            console.warn('attachRunningTaskEventStream', e);
+        }
+    }
+    var mySeq = ++hitlFollowRunSeq;
+    var intervalMs = 2000;
+    var firstDelayMs = 500;
+    var maxMs = 30 * 60 * 1000;
+    var deadline = Date.now() + maxMs;
+
+    function taskStillActive(cid) {
+        return apiFetch('/api/agent-loop/tasks').then(function (r) {
+            if (!r.ok) return false;
+            return r.json().then(function (j) {
+                var tasks = (j && j.tasks) ? j.tasks : [];
+                return tasks.some(function (t) {
+                    return t && t.conversationId === cid && (t.status === 'running' || t.status === 'cancelling');
+                });
+            });
+        }).catch(function () { return false; });
+    }
+
+    await new Promise(function (r) { setTimeout(r, firstDelayMs); });
+
+    while (mySeq === hitlFollowRunSeq) {
+        if (Date.now() > deadline) {
+            if (typeof window.loadConversation === 'function' && window.currentConversationId === conversationId) {
+                await window.loadConversation(conversationId);
+            }
+            if (typeof loadActiveTasks === 'function') loadActiveTasks();
+            return;
+        }
+        try {
+            var active = await taskStillActive(conversationId);
+            var onThisConv = (typeof window.currentConversationId === 'string' && window.currentConversationId === conversationId);
+            if (onThisConv && typeof window.refreshLastAssistantProcessDetails === 'function') {
+                await window.refreshLastAssistantProcessDetails(conversationId);
+            }
+            if (!active) {
+                await new Promise(function (r) { setTimeout(r, 450); });
+                if (typeof window.loadConversation === 'function' && window.currentConversationId === conversationId) {
+                    await window.loadConversation(conversationId);
+                }
+                if (typeof loadActiveTasks === 'function') loadActiveTasks();
+                return;
+            }
+        } catch (e) {
+            console.warn('followAgentRunAfterHitlDecision', e);
+        }
+        await new Promise(function (r) { setTimeout(r, intervalMs); });
+    }
+}
+
+async function refreshHitlPending() {
+    const container = document.getElementById('hitl-pending-list');
+    if (!container) return;
+    container.innerHTML = '<div class="loading-spinner">Loading...</div>';
+    try {
+        const resp = await hitlApiFetch('/api/hitl/pending', { credentials: 'same-origin' });
+        if (!resp.ok) {
+            throw new Error('request failed');
+        }
+        const data = await resp.json();
+        const items = Array.isArray(data.items) ? data.items : [];
+        if (!items.length) {
+            container.innerHTML = '<div class="empty-state">暂无待审批项</div>';
+            return;
+        }
+        container.innerHTML = items.map(function (item) {
+            const payload = String(item.payload || '');
+            const preview = payload.length > 280 ? (payload.slice(0, 280) + '...') : payload;
+            const mode = String(item.mode || '').trim().toLowerCase();
+            const allowEdit = mode === 'review_edit';
+            var escId = escapeHtml(String(item.id || ''));
+            var qId = JSON.stringify(String(item.id || '')).replace(/"/g, '&quot;');
+            var qConv = JSON.stringify(String(item.conversationId || '')).replace(/"/g, '&quot;');
+            return (
+                '<div class="hitl-pending-item">' +
+                '<div class="hitl-pending-item-header">' +
+                '<div class="hitl-pending-item-title">' +
+                '<span class="hitl-tool-badge">' + escapeHtml(item.toolName || '-') + '</span>' +
+                '<span class="hitl-mode-tag hitl-mode-tag--' + escapeHtml(mode) + '">' + escapeHtml(item.mode || '-') + '</span>' +
+                '</div>' +
+                '<button class="hitl-dismiss-btn" title="忽略" onclick="dismissHitlItem(' + qId + ')">&times;</button>' +
+                '</div>' +
+                '<div class="hitl-pending-meta">会话：' + escapeHtml(item.conversationId || '-') + '</div>' +
+                '<pre class="hitl-pending-payload">' + escapeHtml(preview) + '</pre>' +
+                (allowEdit
+                    ? ('<div class="hitl-input-help">审查编辑模式：可填写 JSON 对象覆盖参数。示例：{"command":"ls -la"}</div>' +
+                       '<textarea id="hitl-edit-' + escId + '" class="hitl-edit-args" placeholder=\'{"command":"ls -la"}\'></textarea>')
+                    : '<div class="hitl-input-help">审批模式：仅通过/拒绝，不支持改参。</div>') +
+                '<div class="hitl-input-help">备注（可选）：建议写审批依据。</div>' +
+                '<input id="hitl-comment-' + escId + '" class="hitl-config-input hitl-inline-comment" type="text" placeholder="例如：允许只读命令">' +
+                '<div class="hitl-pending-actions">' +
+                '<button class="btn-secondary" onclick="submitHitlDecision(' + qId + ',&quot;reject&quot;,' + qConv + ')">拒绝</button>' +
+                '<button class="btn-primary" onclick="submitHitlDecision(' + qId + ',&quot;approve&quot;,' + qConv + ')">通过</button>' +
+                '</div>' +
+                '</div>'
+            );
+        }).join('');
+    } catch (e) {
+        container.innerHTML = '<div class="empty-state">加载失败</div>';
+    }
+}
+
+async function submitHitlDecision(interruptId, decision, conversationIdOpt) {
+    const commentBox = document.getElementById('hitl-comment-' + interruptId);
+    const comment = (commentBox && commentBox.value) ? commentBox.value.trim() : '';
+    let editedArguments = null;
+    const editBox = document.getElementById('hitl-edit-' + interruptId);
+    if (editBox && editBox.value && editBox.value.trim()) {
+        try {
+            editedArguments = JSON.parse(editBox.value.trim());
+        } catch (e) {
+            alert('JSON 参数格式错误');
+            return;
+        }
+    }
+    const convFollow = conversationIdOpt || getCurrentConversationIdForHitl();
+    return submitHitlDecisionWithPayload(interruptId, decision, comment, editedArguments, convFollow);
+}
+
+async function submitHitlDecisionWithPayload(interruptId, decision, comment, editedArguments, conversationIdForFollow) {
+    const resp = await hitlApiFetch('/api/hitl/decision', {
+        method: 'POST',
+        credentials: 'same-origin',
+        headers: { 'Content-Type': 'application/json' },
+        body: JSON.stringify({ interruptId: interruptId, decision: decision, comment: comment, editedArguments: editedArguments })
+    });
+    if (!resp.ok) {
+        const errText = await readHitlApiError(resp);
+        if (resp.status === 409 && (errText.indexOf('already resolved') >= 0 || errText.indexOf('not found') >= 0)) {
+            await dismissHitlItem(interruptId, true);
+            return true;
+        }
+        alert('提交失败：' + errText);
+        return false;
+    }
+    refreshHitlPending();
+    const cid = conversationIdForFollow || getCurrentConversationIdForHitl();
+    if (cid) {
+        followAgentRunAfterHitlDecision(cid);
+    }
+    return true;
+}
+
+async function hitlApiFetch(url, options) {
+    if (typeof apiFetch === 'function') {
+        return apiFetch(url, options || {});
+    }
+    return fetch(url, options || {});
+}
+
+async function readHitlApiError(resp) {
+    try {
+        const data = await resp.json();
+        if (data && typeof data.error === 'string' && data.error.trim()) return data.error.trim();
+        return 'HTTP ' + resp.status;
+    } catch (e) {
+        return 'HTTP ' + resp.status;
+    }
+}
+
+async function dismissHitlItem(interruptId, silent) {
+    try {
+        await hitlApiFetch('/api/hitl/dismiss', {
+            method: 'POST',
+            credentials: 'same-origin',
+            headers: { 'Content-Type': 'application/json' },
+            body: JSON.stringify({ interruptId: interruptId })
+        });
+    } catch (e) {
+        if (!silent) { console.warn('dismissHitlItem', e); }
+    }
+    refreshHitlPending();
+}
+
+window.refreshHitlPending = refreshHitlPending;
+window.submitHitlDecision = submitHitlDecision;
+window.submitHitlDecisionWithPayload = submitHitlDecisionWithPayload;
+window.dismissHitlItem = dismissHitlItem;
+window.followAgentRunAfterHitlDecision = followAgentRunAfterHitlDecision;
+
+window.addEventListener('hitl-interrupt', function () {
+    if (typeof window.currentPage === 'function' && window.currentPage() === 'hitl') {
+        refreshHitlPending();
+    }
+});
+
+window.addEventListener('pageshow', function () {
+    setTimeout(reconcileHitlUiState, 0);
+});
+document.addEventListener('DOMContentLoaded', function () {
+    setTimeout(reconcileHitlUiState, 0);
+});
+
+// 由 applyHitlSidebarConfig 调用，将侧栏配置同步到后端
+window.syncHitlConfigToServerByCurrentConversation = syncHitlConfigToServerByCurrentConversation;
+window.saveHitlConversationConfig = saveHitlConversationConfig;
+window.mergeHitlGlobalToolWhitelist = mergeHitlGlobalToolWhitelist;
+
+// 由 chat.js 在 loadConversation 内 await 调用；挂到 window 供其它入口显式触发
+window.syncHitlConfigFromServer = syncHitlConfigFromServer;

web/static/js/monitor.js

@@ -3,6 +3,36 @@ let activeTaskInterval = null;
 const ACTIVE_TASK_REFRESH_INTERVAL = 10000; // 10秒检查一次
 const TASK_FINAL_STATUSES = new Set(['failed', 'timeout', 'cancelled', 'completed']);
 
+/**
+ * 主对话 POST 流仍在读取时，禁止再挂 task-events 补流，否则同一事件会画两遍（与 HITL 是否开启无关）。
+ * window.__csAgentLiveStream 由 chat.js sendMessage 在读到 body 后设置，在 finally 中清除。
+ */
+function syncAgentLiveStreamConversationId(cid) {
+    if (!cid) return;
+    try {
+        const live = window.__csAgentLiveStream;
+        if (live && live.active) {
+            live.conversationId = cid;
+        }
+    } catch (e) { /* ignore */ }
+}
+
+function shouldSkipTaskEventReplayAttach(conversationId) {
+    try {
+        const live = window.__csAgentLiveStream;
+        if (!live || !live.active || !live.progressId) return false;
+        if (!document.getElementById(live.progressId)) return false;
+        // 新会话：conversation 事件尚未到达前 conversationId 可能仍为 null，一律不补挂
+        if (live.conversationId == null) return true;
+        return live.conversationId === conversationId;
+    } catch (e) {
+        return false;
+    }
+}
+if (typeof window !== 'undefined') {
+    window.shouldSkipTaskEventReplayAttach = shouldSkipTaskEventReplayAttach;
+}
+
 // 当前界面语言对应的 BCP 47 标签（与时间格式化一致）
 function getCurrentTimeLocale() {
     if (typeof window.__locale === 'string' && window.__locale.length) {
@@ -843,6 +873,33 @@ function applyBackendMessageIdToLastUser(backendMessageId) {
     }
 }
 
+function taskReplayProgressId(conversationId) {
+    return 'task-ev-' + String(conversationId || '').replace(/[^a-zA-Z0-9_-]/g, '_');
+}
+
+function clearCsTaskReplay() {
+    window.csTaskReplay = null;
+}
+
+function beginCsTaskReplay(progressId, assistantDomId, conversationId) {
+    window.csTaskReplay = {
+        progressId: progressId,
+        assistantDomId: assistantDomId,
+        conversationId: conversationId,
+        timelineHostId: 'process-details-' + assistantDomId + '-timeline'
+    };
+    registerProgressTask(progressId, conversationId);
+}
+
+function resolveStreamTimeline(progressId) {
+    let timeline = document.getElementById(progressId + '-timeline');
+    const r = window.csTaskReplay;
+    if (!timeline && r && r.progressId === progressId && r.timelineHostId) {
+        timeline = document.getElementById(r.timelineHostId);
+    }
+    return timeline;
+}
+
 // 处理流式事件
 function handleStreamEvent(event, progressElement, progressId, 
                           getAssistantId, setAssistantId, getMcpIds, setMcpIds) {
@@ -858,7 +915,7 @@ function handleStreamEvent(event, progressElement, progressId,
         return;
     }
 
-    const timeline = document.getElementById(progressId + '-timeline');
+    const timeline = resolveStreamTimeline(progressId);
     if (!timeline) return;
 
     // 终态事件（error/cancelled）优先复用现有助手消息，避免重复追加相同报错
@@ -907,6 +964,7 @@ function handleStreamEvent(event, progressElement, progressId,
                 
                 // 更新当前对话ID
                 currentConversationId = event.data.conversationId;
+                syncAgentLiveStreamConversationId(event.data.conversationId);
                 updateActiveConversation();
                 addAttackChainButton(currentConversationId);
                 loadActiveTasks();
@@ -1049,20 +1107,71 @@ function handleStreamEvent(event, progressElement, progressId,
             });
             break;
 
-        case 'eino_recovery': {
-            const d = event.data || {};
-            const runIdx = d.runIndex != null ? d.runIndex : (d.einoRetry != null ? d.einoRetry + 1 : 1);
-            const maxRuns = d.maxRuns != null ? d.maxRuns : 3;
-            const title = typeof window.t === 'function'
-                ? window.t('chat.einoRecoveryTitle', { n: runIdx, max: maxRuns })
-                : ('🔄 工具参数无效 · 第 ' + runIdx + '/' + maxRuns + ' 轮（已追加提示）');
-            addTimelineItem(timeline, 'eino_recovery', {
-                title: title,
-                message: event.message || '',
+        case 'hitl_interrupt':
+            const hitlItemId = addTimelineItem(timeline, 'warning', {
+                title: '🧑‍⚖️ HITL',
+                message: event.message,
                 data: event.data
             });
-            // If the backend triggers a recovery run, any "running" tool_call items in this progress
-            // should be closed to avoid being stuck forever.
+            renderInlineHitlApproval(hitlItemId, event.data || {});
+            try {
+                window.dispatchEvent(new CustomEvent('hitl-interrupt', { detail: event.data || {} }));
+            } catch (e) {}
+            break;
+        case 'hitl_resumed':
+            addTimelineItem(timeline, 'progress', {
+                title: '✅ HITL',
+                message: event.message,
+                data: event.data
+            });
+            break;
+        case 'hitl_rejected':
+            addTimelineItem(timeline, 'error', {
+                title: '⛔ HITL',
+                message: event.message,
+                data: event.data
+            });
+            break;
+
+        case 'eino_stream_error': {
+            const d = event.data || {};
+            const agent = d.einoAgent ? String(d.einoAgent) : '';
+            const title = typeof window.t === 'function'
+                ? window.t('chat.einoStreamErrorTitle', { agent: agent || '-' })
+                : (agent ? ('⚠️ Eino 流式中断（' + agent + '）') : '⚠️ Eino 流式中断');
+            addTimelineItem(timeline, 'warning', {
+                title: title,
+                message: event.message || (typeof window.t === 'function'
+                    ? window.t('chat.einoStreamErrorMessage')
+                    : '流式读取异常，系统将按策略重试或结束。'),
+                data: d
+            });
+            break;
+        }
+
+        case 'iteration_limit_reached': {
+            addTimelineItem(timeline, 'warning', {
+                title: typeof window.t === 'function' ? window.t('chat.iterationLimitReachedTitle') : '⛔ 达到迭代上限',
+                message: event.message || (typeof window.t === 'function'
+                    ? window.t('chat.iterationLimitReachedMessage')
+                    : '已达到最大迭代次数，任务已停止继续自动迭代。'),
+                data: event.data
+            });
+            finalizeOutstandingToolCallsForProgress(progressId, 'failed');
+            break;
+        }
+
+        case 'eino_pending_orphaned': {
+            const d = event.data || {};
+            const count = Number(d.pendingCount || 0);
+            const countText = Number.isFinite(count) && count > 0 ? String(count) : '?';
+            addTimelineItem(timeline, 'warning', {
+                title: typeof window.t === 'function' ? window.t('chat.einoPendingOrphanedTitle') : '🧹 工具调用收尾补偿',
+                message: event.message || (typeof window.t === 'function'
+                    ? window.t('chat.einoPendingOrphanedMessage', { count: countText })
+                    : ('检测到 ' + countText + ' 个未闭合工具调用，已自动标记为失败并收尾。')),
+                data: d
+            });
             finalizeOutstandingToolCallsForProgress(progressId, 'failed');
             break;
         }
@@ -1376,6 +1485,7 @@ function handleStreamEvent(event, progressElement, progressId,
                     break;
                 }
                 currentConversationId = responseData.conversationId;
+                syncAgentLiveStreamConversationId(responseData.conversationId);
                 updateActiveConversation();
                 addAttackChainButton(currentConversationId);
                 updateProgressConversation(progressId, responseData.conversationId);
@@ -1456,6 +1566,7 @@ function handleStreamEvent(event, progressElement, progressId,
                 }
 
                 currentConversationId = responseData.conversationId;
+                syncAgentLiveStreamConversationId(responseData.conversationId);
                 updateActiveConversation();
                 addAttackChainButton(currentConversationId);
                 updateProgressConversation(progressId, responseData.conversationId);
@@ -1492,8 +1603,12 @@ function handleStreamEvent(event, progressElement, progressId,
             // so the copied timeline HTML reflects the final status.
             finalizeOutstandingToolCallsForProgress(progressId, 'failed');
 
-            // 将进度详情集成到工具调用区域（放在最终 response 之后，保证时间线已完整）
-            integrateProgressToMCPSection(progressId, assistantIdFinal, mcpIds);
+            const replayCtx = window.csTaskReplay;
+            const directReplay = replayCtx && replayCtx.progressId === progressId;
+            if (!directReplay) {
+                // 将进度详情集成到工具调用区域（放在最终 response 之后，保证时间线已完整）
+                integrateProgressToMCPSection(progressId, assistantIdFinal, mcpIds);
+            }
             responseStreamStateByProgressId.delete(progressId);
 
             const respMid = responseData.messageId;
@@ -1502,7 +1617,7 @@ function handleStreamEvent(event, progressElement, progressId,
             }
 
             setTimeout(() => {
-                collapseAllProgressDetails(assistantIdFinal, progressId);
+                collapseAllProgressDetails(assistantIdFinal, directReplay ? null : progressId);
             }, 3000);
 
             setTimeout(() => {
@@ -1571,6 +1686,9 @@ function handleStreamEvent(event, progressElement, progressId,
                     toolResultStreamStateByKey.delete(key);
                 }
             }
+            if (window.csTaskReplay && window.csTaskReplay.progressId === progressId) {
+                clearCsTaskReplay();
+            }
             // 完成，更新进度标题（如果进度消息还存在）
             const doneTitle = document.querySelector(`#${progressId} .progress-title`);
             if (doneTitle) {
@@ -1579,6 +1697,7 @@ function handleStreamEvent(event, progressElement, progressId,
             // 更新对话ID
             if (event.data && event.data.conversationId) {
                 currentConversationId = event.data.conversationId;
+                syncAgentLiveStreamConversationId(event.data.conversationId);
                 updateActiveConversation();
                 addAttackChainButton(currentConversationId);
                 updateProgressConversation(progressId, event.data.conversationId);
@@ -1625,6 +1744,379 @@ function handleStreamEvent(event, progressElement, progressId,
     scrollChatMessagesToBottomIfPinned(streamScrollWasPinned);
 }
 
+function renderInlineHitlApproval(itemId, data) {
+    const item = document.getElementById(itemId);
+    if (!item || !data || !data.interruptId) return;
+    let contentEl = item.querySelector('.timeline-item-content');
+    if (!contentEl) {
+        // warning 等类型默认没有内容区域；HITL 内联审批需要可交互容器
+        contentEl = document.createElement('div');
+        contentEl.className = 'timeline-item-content';
+        item.appendChild(contentEl);
+    }
+    const existingPanel = contentEl.querySelector('.hitl-inline-approval');
+    if (existingPanel) {
+        existingPanel.remove();
+    }
+
+    const payload = data.payload && typeof data.payload === 'object' ? data.payload : {};
+    const toolName = data.toolName || payload.toolName || '-';
+    let mode = String(data.mode || '').trim().toLowerCase();
+    if (mode === 'feedback' || mode === 'followup') {
+        mode = 'approval';
+    }
+    const allowEdit = mode === 'review_edit';
+    const argsObj = payload.argumentsObj && typeof payload.argumentsObj === 'object' ? payload.argumentsObj : {};
+    const argsJSON = JSON.stringify(argsObj, null, 2);
+
+    const panel = document.createElement('div');
+    panel.className = 'hitl-inline-approval';
+    panel.innerHTML = `
+        <div class="hitl-input-help"><strong>${escapeHtml(toolName)}</strong> 待人工审批。模式：${escapeHtml(mode || '-')}。</div>
+        ${allowEdit
+            ? `<div class="hitl-input-help">审查编辑参数（JSON，可选）：留空表示沿用原参数。</div>
+               <textarea class="hitl-edit-args hitl-inline-edit" placeholder='{"command":"ls -la"}'>${escapeHtml(argsJSON === '{}' ? '' : argsJSON)}</textarea>`
+            : '<div class="hitl-input-help">当前模式不支持改参，仅可通过/拒绝。</div>'
+        }
+        <div class="hitl-input-help">备注（可选）：建议写审批依据。</div>
+        <input class="hitl-config-input hitl-inline-comment" type="text" placeholder="例如：允许只读命令">
+        <div class="hitl-pending-actions">
+            <button class="btn-secondary hitl-inline-reject">拒绝</button>
+            <button class="btn-primary hitl-inline-approve">通过</button>
+        </div>
+        <div class="hitl-input-help hitl-inline-status"></div>
+    `;
+    contentEl.appendChild(panel);
+
+    const approveBtn = panel.querySelector('.hitl-inline-approve');
+    const rejectBtn = panel.querySelector('.hitl-inline-reject');
+    const commentInput = panel.querySelector('.hitl-inline-comment');
+    const editInput = panel.querySelector('.hitl-inline-edit');
+    const statusEl = panel.querySelector('.hitl-inline-status');
+
+    const setBusy = function (busy) {
+        approveBtn.disabled = busy;
+        rejectBtn.disabled = busy;
+    };
+
+    const submit = async function (decision) {
+        setBusy(true);
+        let editedArgs = null;
+        if (allowEdit && editInput) {
+            const raw = String(editInput.value || '').trim();
+            if (raw) {
+                try {
+                    editedArgs = JSON.parse(raw);
+                } catch (e) {
+                    statusEl.textContent = 'JSON 参数格式错误';
+                    setBusy(false);
+                    return;
+                }
+            }
+        }
+        const comment = String(commentInput.value || '').trim();
+        try {
+            if (typeof window.submitHitlDecisionWithPayload === 'function') {
+                const convFollow = data.conversationId || (typeof window.currentConversationId === 'string' ? window.currentConversationId : '');
+                const ok = await window.submitHitlDecisionWithPayload(data.interruptId, decision, comment, (decision === 'approve' && allowEdit) ? editedArgs : null, convFollow);
+                if (!ok) {
+                    statusEl.textContent = '提交失败，请重试';
+                    setBusy(false);
+                    return;
+                }
+            } else {
+                statusEl.textContent = '审批函数未加载';
+                setBusy(false);
+                return;
+            }
+            statusEl.textContent = decision === 'approve' ? '已通过，等待执行继续...' : '已拒绝，反馈已交给模型继续迭代...';
+            panel.classList.add('hitl-inline-done');
+        } catch (e) {
+            statusEl.textContent = '提交失败：' + (e && e.message ? e.message : 'unknown error');
+            setBusy(false);
+        }
+    };
+
+    approveBtn.onclick = function () { submit('approve'); };
+    rejectBtn.onclick = function () { submit('reject'); };
+}
+
+function hitlEscapeAttrSelector(val) {
+    const s = String(val);
+    if (typeof CSS !== 'undefined' && typeof CSS.escape === 'function') {
+        return CSS.escape(s);
+    }
+    return s.replace(/\\/g, '\\\\').replace(/"/g, '\\"');
+}
+
+function expandProcessDetailsTimeline(assistantMessageId) {
+    if (!assistantMessageId) return;
+    const detailsContainer = document.getElementById('process-details-' + assistantMessageId);
+    if (!detailsContainer) return;
+    const timeline = detailsContainer.querySelector('.progress-timeline');
+    if (!timeline) return;
+    timeline.classList.add('expanded');
+    const collapseT = typeof window.t === 'function' ? window.t('tasks.collapseDetail') : '收起详情';
+    document.querySelectorAll('#' + hitlEscapeAttrSelector(assistantMessageId) + ' .process-detail-btn').forEach(function (btn) {
+        btn.innerHTML = '<span>' + collapseT + '</span>';
+    });
+    setTimeout(function () {
+        detailsContainer.scrollIntoView({ behavior: 'smooth', block: 'nearest' });
+    }, 100);
+}
+
+function findLastAssistantMessageElInChat() {
+    const nodes = document.querySelectorAll('#chat-messages .message.assistant');
+    for (let i = nodes.length - 1; i >= 0; i--) {
+        const el = nodes[i];
+        if (el && el.dataset && el.dataset.backendMessageId) return el;
+    }
+    return null;
+}
+
+/**
+ * 刷新或切换会话后：根据待审批记录恢复时间线里的内联审批入口，并展开详情区。
+ */
+async function restoreHitlInlineForConversation(conversationId) {
+    if (!conversationId || typeof apiFetch !== 'function') return;
+    if (typeof window.currentConversationId === 'string' && window.currentConversationId !== conversationId) {
+        return;
+    }
+    try {
+        const resp = await apiFetch('/api/hitl/pending?conversationId=' + encodeURIComponent(conversationId) + '&status=pending&pageSize=50');
+        if (!resp.ok) return;
+        const data = await resp.json().catch(function () { return {}; });
+        const items = Array.isArray(data.items) ? data.items : [];
+        for (let i = 0; i < items.length; i++) {
+            const item = items[i];
+            let backendMsgId = item.messageId != null ? String(item.messageId).trim() : '';
+            let msgEl = null;
+            if (backendMsgId) {
+                msgEl = document.querySelector('#chat-messages [data-backend-message-id="' + hitlEscapeAttrSelector(backendMsgId) + '"]');
+            }
+            if (!msgEl) {
+                msgEl = findLastAssistantMessageElInChat();
+                if (msgEl && msgEl.dataset && msgEl.dataset.backendMessageId) {
+                    backendMsgId = String(msgEl.dataset.backendMessageId).trim();
+                }
+            }
+            if (!msgEl || !msgEl.id || !backendMsgId) continue;
+            const clientMsgId = msgEl.id;
+            const detailsContainer = document.getElementById('process-details-' + clientMsgId);
+            if (!detailsContainer) continue;
+            if (detailsContainer.dataset.lazyNotLoaded === '1' && detailsContainer.dataset.loaded !== '1') {
+                try {
+                    detailsContainer.dataset.loading = '1';
+                    const res = await apiFetch('/api/messages/' + encodeURIComponent(backendMsgId) + '/process-details');
+                    const j = await res.json().catch(function () { return {}; });
+                    if (!res.ok) throw new Error((j && j.error) ? j.error : String(res.status));
+                    const details = (j && Array.isArray(j.processDetails)) ? j.processDetails : [];
+                    if (typeof renderProcessDetails === 'function') {
+                        renderProcessDetails(clientMsgId, details);
+                    }
+                } catch (e) {
+                    console.error('加载过程详情失败（HITL 恢复）:', e);
+                } finally {
+                    detailsContainer.dataset.loading = '0';
+                }
+            }
+            expandProcessDetailsTimeline(clientMsgId);
+            let payloadObj = {};
+            try {
+                payloadObj = JSON.parse(String(item.payload || '{}'));
+            } catch (e) {
+                payloadObj = {};
+            }
+            const hitlData = {
+                interruptId: item.id,
+                mode: item.mode,
+                toolName: item.toolName,
+                toolCallId: item.toolCallId,
+                payload: payloadObj,
+                conversationId: item.conversationId || conversationId
+            };
+            let hitlItemEl = detailsContainer.querySelector('[data-hitl-interrupt-id="' + hitlEscapeAttrSelector(String(item.id)) + '"]');
+            if (!hitlItemEl && item.toolCallId) {
+                hitlItemEl = detailsContainer.querySelector('[data-tool-call-id="' + hitlEscapeAttrSelector(String(item.toolCallId)) + '"]');
+            }
+            if (!hitlItemEl && item.toolName) {
+                const want = String(item.toolName).trim().toLowerCase();
+                const shortWant = want.indexOf('::') >= 0 ? want.split('::').pop() : want;
+                const calls = detailsContainer.querySelectorAll('.timeline-item-tool_call');
+                for (let j = calls.length - 1; j >= 0; j--) {
+                    const tn = String(calls[j].dataset.toolName || '').trim().toLowerCase();
+                    const shortTn = tn.indexOf('::') >= 0 ? tn.split('::').pop() : tn;
+                    const match = want && (tn === want || tn.endsWith('::' + shortWant) || shortTn === shortWant);
+                    if (match) {
+                        hitlItemEl = calls[j];
+                        break;
+                    }
+                }
+            }
+            if (!hitlItemEl) continue;
+            renderInlineHitlApproval(hitlItemEl.id, hitlData);
+        }
+    } catch (e) {
+        console.error('restoreHitlInlineForConversation failed', e);
+    }
+}
+
+window.expandProcessDetailsTimeline = expandProcessDetailsTimeline;
+window.restoreHitlInlineForConversation = restoreHitlInlineForConversation;
+
+/**
+ * 无 SSE 时（例如刷新页面后）：从 DB 拉取最后一条助手消息的过程详情并重绘时间线，便于审批通过后仍能看到执行进展。
+ */
+async function refreshLastAssistantProcessDetails(conversationId) {
+    if (!conversationId || typeof apiFetch !== 'function') return;
+    if (typeof window.currentConversationId === 'string' && window.currentConversationId !== conversationId) return;
+    const msgEl = findLastAssistantMessageElInChat();
+    if (!msgEl || !msgEl.dataset.backendMessageId || !msgEl.id) return;
+    const backendId = String(msgEl.dataset.backendMessageId).trim();
+    const clientId = msgEl.id;
+    const detailsContainer = document.getElementById('process-details-' + clientId);
+    let wasExpanded = false;
+    if (detailsContainer) {
+        const tl = detailsContainer.querySelector('.progress-timeline');
+        wasExpanded = !!(tl && tl.classList.contains('expanded'));
+    }
+    try {
+        const res = await apiFetch('/api/messages/' + encodeURIComponent(backendId) + '/process-details');
+        const j = await res.json().catch(function () { return {}; });
+        if (!res.ok) return;
+        const details = Array.isArray(j.processDetails) ? j.processDetails : [];
+        if (typeof renderProcessDetails === 'function') {
+            renderProcessDetails(clientId, details);
+        }
+        if (wasExpanded) {
+            expandProcessDetailsTimeline(clientId);
+        }
+    } catch (e) {
+        console.warn('refreshLastAssistantProcessDetails', e);
+    }
+}
+
+window.refreshLastAssistantProcessDetails = refreshLastAssistantProcessDetails;
+
+const taskEventReplayAttachState = {
+    conversationId: null,
+    inFlightPromise: null
+};
+
+/**
+ * 订阅运行中任务的 SSE 镜像（GET /api/agent-loop/task-events），用于 HITL 通过后主连接已断开时接续 UI。
+ */
+async function attachRunningTaskEventStream(conversationId) {
+    if (!conversationId || typeof apiFetch !== 'function') return false;
+    if (
+        taskEventReplayAttachState.inFlightPromise &&
+        taskEventReplayAttachState.conversationId === conversationId
+    ) {
+        return taskEventReplayAttachState.inFlightPromise;
+    }
+    if (shouldSkipTaskEventReplayAttach(conversationId)) {
+        return false;
+    }
+
+    const attachPromise = (async function () {
+        try {
+            const check = await apiFetch('/api/agent-loop/tasks');
+            if (!check.ok) return false;
+            const j = await check.json().catch(function () { return {}; });
+            const active = (j.tasks || []).some(function (t) {
+                return t && t.conversationId === conversationId && (t.status === 'running' || t.status === 'cancelling');
+            });
+            if (!active) return false;
+
+            const asEl = findLastAssistantMessageElInChat();
+            if (!asEl || !asEl.id) return false;
+            const backendId = asEl.dataset && asEl.dataset.backendMessageId;
+            if (backendId && typeof renderProcessDetails === 'function') {
+                const res = await apiFetch('/api/messages/' + encodeURIComponent(String(backendId)) + '/process-details');
+                const jd = await res.json().catch(function () { return {}; });
+                if (res.ok && Array.isArray(jd.processDetails)) {
+                    renderProcessDetails(asEl.id, jd.processDetails);
+                    // renderProcessDetails 会重建时间线节点，需重新挂载 HITL 审批入口
+                    if (typeof window.restoreHitlInlineForConversation === 'function') {
+                        await window.restoreHitlInlineForConversation(conversationId);
+                    }
+                }
+            }
+            expandProcessDetailsTimeline(asEl.id);
+
+            const progressId = taskReplayProgressId(conversationId);
+            beginCsTaskReplay(progressId, asEl.id, conversationId);
+
+            const url = '/api/agent-loop/task-events?conversationId=' + encodeURIComponent(conversationId);
+            const response = await apiFetch(url, {
+                method: 'GET',
+                headers: { Accept: 'text/event-stream' }
+            });
+            if (!response.ok) {
+                clearCsTaskReplay();
+                if (progressTaskState.has(progressId)) {
+                    progressTaskState.delete(progressId);
+                }
+                return false;
+            }
+
+            let mcpIds = [];
+            const assistantDomId = asEl.id;
+            const getAssistantIdFn = function () { return assistantDomId; };
+            const setAssistantIdFn = function () {};
+
+            const reader = response.body.getReader();
+            const decoder = new TextDecoder();
+            let buffer = '';
+            while (true) {
+                const chunk = await reader.read();
+                if (chunk.done) break;
+                buffer += decoder.decode(chunk.value, { stream: true });
+                const lines = buffer.split('\n');
+                buffer = lines.pop() || '';
+                for (let li = 0; li < lines.length; li++) {
+                    const line = lines[li];
+                    if (line.indexOf('data: ') === 0) {
+                        try {
+                            const eventData = JSON.parse(line.slice(6));
+                            handleStreamEvent(eventData, null, progressId, getAssistantIdFn, setAssistantIdFn, function () { return mcpIds; }, function (ids) { mcpIds = ids; });
+                        } catch (e) {
+                            console.error('task-events parse', e);
+                        }
+                    }
+                }
+            }
+            if (window.csTaskReplay && window.csTaskReplay.progressId === progressId) {
+                clearCsTaskReplay();
+            }
+            if (progressTaskState.has(progressId)) {
+                finalizeProgressTask(progressId, typeof window.t === 'function' ? window.t('tasks.statusCompleted') : '已完成');
+            }
+            if (typeof loadActiveTasks === 'function') loadActiveTasks();
+            if (typeof window.loadConversation === 'function' && window.currentConversationId === conversationId) {
+                await window.loadConversation(conversationId);
+            }
+            return true;
+        } catch (e) {
+            console.warn('attachRunningTaskEventStream', e);
+            clearCsTaskReplay();
+            return false;
+        } finally {
+            if (taskEventReplayAttachState.inFlightPromise === attachPromise) {
+                taskEventReplayAttachState.inFlightPromise = null;
+                taskEventReplayAttachState.conversationId = null;
+            }
+        }
+    })();
+
+    taskEventReplayAttachState.conversationId = conversationId;
+    taskEventReplayAttachState.inFlightPromise = attachPromise;
+    return attachPromise;
+}
+
+window.attachRunningTaskEventStream = attachRunningTaskEventStream;
+window.taskReplayProgressId = taskReplayProgressId;
+
 // 更新工具调用状态
 function updateToolCallStatus(toolCallId, status) {
     const mapping = toolCallStatusMap.get(toolCallId);
@@ -1680,15 +2172,6 @@ function addTimelineItem(timeline, type, options) {
     if (type === 'progress' && options.message) {
         item.dataset.progressMessage = options.message;
     }
-    if (type === 'eino_recovery' && options.data) {
-        const d = options.data;
-        if (d.runIndex != null) {
-            item.dataset.recoveryRunIndex = String(d.runIndex);
-        }
-        if (d.maxRuns != null) {
-            item.dataset.recoveryMaxRuns = String(d.maxRuns);
-        }
-    }
     if (type === 'tool_calls_detected' && options.data && options.data.count != null) {
         item.dataset.toolCallsCount = String(options.data.count);
     }
@@ -1697,6 +2180,12 @@ function addTimelineItem(timeline, type, options) {
         item.dataset.toolName = (d.toolName != null && d.toolName !== '') ? String(d.toolName) : '';
         item.dataset.toolIndex = (d.index != null) ? String(d.index) : '0';
         item.dataset.toolTotal = (d.total != null) ? String(d.total) : '0';
+        if (d.toolCallId != null && String(d.toolCallId).trim() !== '') {
+            item.dataset.toolCallId = String(d.toolCallId).trim();
+        }
+    }
+    if (type === 'hitl_interrupt' && options.data && options.data.interruptId != null && String(options.data.interruptId).trim() !== '') {
+        item.dataset.hitlInterruptId = String(options.data.interruptId).trim();
     }
     if (type === 'tool_result' && options.data) {
         const d = options.data;
@@ -1793,12 +2282,6 @@ function addTimelineItem(timeline, type, options) {
                 </div>
             </div>
         `;
-    } else if (type === 'eino_recovery' && options.message) {
-        content += `
-            <div class="timeline-item-content timeline-eino-recovery">
-                ${escapeHtml(options.message).replace(/\n/g, '<br>')}
-            </div>
-        `;
     } else if (type === 'cancelled') {
         const taskCancelledLabel = typeof window.t === 'function' ? window.t('chat.taskCancelled') : '任务已取消';
         content += `
@@ -1934,6 +2417,8 @@ async function cancelActiveTask(conversationId, button) {
     }
 }
 
+let monitorPanelFetchSeq = 0;
+
 // 监控面板状态
 const monitorState = {
     executions: [],
@@ -2004,6 +2489,7 @@ async function refreshMonitorPanel(page = null) {
     const execContainer = document.getElementById('monitor-executions');
 
     try {
+        const mySeq = ++monitorPanelFetchSeq;
         // 如果指定了页码，使用指定页码，否则使用当前页码
         const currentPage = page !== null ? page : monitorState.pagination.page;
         const pageSize = monitorState.pagination.pageSize;
@@ -2028,6 +2514,9 @@ async function refreshMonitorPanel(page = null) {
         if (!response.ok) {
             throw new Error(result.error || '获取监控数据失败');
         }
+        if (mySeq !== monitorPanelFetchSeq) {
+            return;
+        }
 
         monitorState.executions = Array.isArray(result.executions) ? result.executions : [];
         monitorState.stats = result.stats || {};
@@ -2088,6 +2577,7 @@ async function refreshMonitorPanelWithFilter(statusFilter = 'all', toolFilter =
     const execContainer = document.getElementById('monitor-executions');
 
     try {
+        const mySeq = ++monitorPanelFetchSeq;
         const currentPage = 1; // 筛选时重置到第一页
         const pageSize = monitorState.pagination.pageSize;
         
@@ -2105,6 +2595,9 @@ async function refreshMonitorPanelWithFilter(statusFilter = 'all', toolFilter =
         if (!response.ok) {
             throw new Error(result.error || '获取监控数据失败');
         }
+        if (mySeq !== monitorPanelFetchSeq) {
+            return;
+        }
 
         monitorState.executions = Array.isArray(result.executions) ? result.executions : [];
         monitorState.stats = result.stats || {};
@@ -2671,10 +3164,6 @@ function refreshProgressAndTimelineI18n() {
             titleSpan.textContent = ap + icon + (success ? _t('chat.toolExecComplete', { name: name }) : _t('chat.toolExecFailed', { name: name }));
         } else if (type === 'eino_agent_reply') {
             titleSpan.textContent = ap + '\uD83D\uDCAC ' + _t('chat.einoAgentReplyTitle');
-        } else if (type === 'eino_recovery' && item.dataset.recoveryRunIndex) {
-            const n = parseInt(item.dataset.recoveryRunIndex, 10) || 1;
-            const mx = parseInt(item.dataset.recoveryMaxRuns, 10) || 3;
-            titleSpan.textContent = _t('chat.einoRecoveryTitle', { n: n, max: mx });
         } else if (type === 'cancelled') {
             titleSpan.textContent = '\u26D4 ' + _t('chat.taskCancelled');
         } else if (type === 'progress' && item.dataset.progressMessage !== undefined) {

web/static/js/router.js

@@ -1,6 +1,48 @@
 // 页面路由管理
 let currentPage = 'dashboard';
 
+/** chat、漏洞管理页在切换时保留当前 hash 上的查询串（如 ?conversation= / ?conversation_id=） */
+function buildHashForPage(pageId) {
+    if (pageId !== 'chat' && pageId !== 'vulnerabilities') {
+        return pageId;
+    }
+    const full = window.location.hash.slice(1);
+    const parts = full.split('?');
+    const curPage = parts[0];
+    const q = parts.length > 1 ? parts.slice(1).join('?') : '';
+    if (curPage === pageId && q) {
+        return pageId + '?' + q;
+    }
+    return pageId;
+}
+
+let chatConversationFromHashSeq = 0;
+function scheduleChatConversationFromHash(delayMs) {
+    const hash = window.location.hash.slice(1);
+    const hashParts = hash.split('?');
+    if (hashParts[0] !== 'chat' || hashParts.length < 2) {
+        return;
+    }
+    const params = new URLSearchParams(hashParts.slice(1).join('?'));
+    const conversationId = params.get('conversation');
+    if (!conversationId) {
+        return;
+    }
+    const token = ++chatConversationFromHashSeq;
+    setTimeout(() => {
+        if (token !== chatConversationFromHashSeq) {
+            return;
+        }
+        if (typeof loadConversation === 'function') {
+            loadConversation(conversationId);
+        } else if (typeof window.loadConversation === 'function') {
+            window.loadConversation(conversationId);
+        } else {
+            console.warn('loadConversation function not found');
+        }
+    }, delayMs);
+}
+
 // 初始化路由
 function initRouter() {
     // 从URL hash读取页面（如果有）
@@ -8,25 +50,10 @@ function initRouter() {
     if (hash) {
         const hashParts = hash.split('?');
         const pageId = hashParts[0];
-        if (pageId && ['dashboard', 'chat', 'info-collect', 'vulnerabilities', 'webshell', 'chat-files', 'mcp-monitor', 'mcp-management', 'knowledge-management', 'knowledge-retrieval-logs', 'roles-management', 'skills-monitor', 'skills-management', 'agents-management', 'settings', 'tasks'].includes(pageId)) {
+        if (pageId && ['dashboard', 'chat', 'hitl', 'info-collect', 'vulnerabilities', 'webshell', 'chat-files', 'mcp-monitor', 'mcp-management', 'knowledge-management', 'knowledge-retrieval-logs', 'roles-management', 'skills-monitor', 'skills-management', 'agents-management', 'settings', 'tasks'].includes(pageId)) {
             switchPage(pageId);
-            
-            // 如果是chat页面且带有conversation参数，加载对应对话
-            if (pageId === 'chat' && hashParts.length > 1) {
-                const params = new URLSearchParams(hashParts[1]);
-                const conversationId = params.get('conversation');
-                if (conversationId) {
-                    setTimeout(() => {
-                        // 尝试多种方式调用loadConversation
-                        if (typeof loadConversation === 'function') {
-                            loadConversation(conversationId);
-                        } else if (typeof window.loadConversation === 'function') {
-                            window.loadConversation(conversationId);
-                        } else {
-                            console.warn('loadConversation function not found');
-                        }
-                    }, 500);
-                }
+            if (pageId === 'chat') {
+                scheduleChatConversationFromHash(500);
             }
             return;
         }
@@ -49,8 +76,10 @@ function switchPage(pageId) {
         targetPage.classList.add('active');
         currentPage = pageId;
         
-        // 更新URL hash
-        window.location.hash = pageId;
+        const newHash = buildHashForPage(pageId);
+        if (window.location.hash.slice(1) !== newHash) {
+            window.location.hash = newHash;
+        }
         
         // 更新导航状态
         updateNavState(pageId);
@@ -247,6 +276,11 @@ async function initPage(pageId) {
             // 恢复对话列表折叠状态（从其他页返回时保持用户选择）
             initConversationSidebarState();
             break;
+        case 'hitl':
+            if (typeof refreshHitlPending === 'function') {
+                refreshHitlPending();
+            }
+            break;
         case 'info-collect':
             // 信息收集页面
             if (typeof initInfoCollectPage === 'function') {
@@ -379,44 +413,13 @@ document.addEventListener('DOMContentLoaded', function() {
         const hashParts = hash.split('?');
         const pageId = hashParts[0];
         
-        if (pageId && ['chat', 'info-collect', 'tasks', 'vulnerabilities', 'webshell', 'chat-files', 'mcp-monitor', 'mcp-management', 'knowledge-management', 'knowledge-retrieval-logs', 'roles-management', 'skills-monitor', 'skills-management', 'agents-management', 'settings'].includes(pageId)) {
+        if (pageId && ['chat', 'hitl', 'info-collect', 'tasks', 'vulnerabilities', 'webshell', 'chat-files', 'mcp-monitor', 'mcp-management', 'knowledge-management', 'knowledge-retrieval-logs', 'roles-management', 'skills-monitor', 'skills-management', 'agents-management', 'settings'].includes(pageId)) {
             switchPage(pageId);
-            
-            // 如果是chat页面且带有conversation参数，加载对应对话
-            if (pageId === 'chat' && hashParts.length > 1) {
-                const params = new URLSearchParams(hashParts[1]);
-                const conversationId = params.get('conversation');
-                if (conversationId) {
-                    setTimeout(() => {
-                        // 尝试多种方式调用loadConversation
-                        if (typeof loadConversation === 'function') {
-                            loadConversation(conversationId);
-                        } else if (typeof window.loadConversation === 'function') {
-                            window.loadConversation(conversationId);
-                        } else {
-                            console.warn('loadConversation function not found');
-                        }
-                    }, 200);
-                }
+            if (pageId === 'chat') {
+                scheduleChatConversationFromHash(200);
             }
         }
     });
-    
-    // 页面加载时也检查hash参数
-    const hash = window.location.hash.slice(1);
-    if (hash) {
-        const hashParts = hash.split('?');
-        const pageId = hashParts[0];
-        if (pageId === 'chat' && hashParts.length > 1) {
-            const params = new URLSearchParams(hashParts[1]);
-            const conversationId = params.get('conversation');
-            if (conversationId && typeof loadConversation === 'function') {
-                setTimeout(() => {
-                    loadConversation(conversationId);
-                }, 500);
-            }
-        }
-    }
 });
 
 // 切换侧边栏折叠/展开

web/static/js/settings.js

@@ -134,8 +134,6 @@ async function loadConfig(loadTools = true) {
             const v = ma.plan_execute_loop_max_iterations;
             maPeLoop.value = (v !== undefined && v !== null && !Number.isNaN(Number(v))) ? String(Number(v)) : '0';
         }
-        const maMode = document.getElementById('multi-agent-default-mode');
-        if (maMode) maMode.value = (ma.default_mode === 'multi') ? 'multi' : 'single';
         const maRobot = document.getElementById('multi-agent-robot-use');
         if (maRobot) maRobot.checked = ma.robot_use_multi_agent === true;
         
@@ -519,7 +517,7 @@ function renderToolsList() {
 
         toolItem.innerHTML = `
             <input type="checkbox" id="${checkboxId}" ${toolState.enabled ? 'checked' : ''} ${toolState.is_external || tool.is_external ? 'data-external="true"' : ''} onchange="handleToolCheckboxChange('${escapeHtml(toolKey)}', this.checked)" />
-            <div class="tool-item-info" onclick="toggleToolDetail(this, '${escapeHtml(toolKey)}', ${tool.is_external ? 'true' : 'false'}, '${escapeHtml(tool.external_mcp || '')}', event)">
+            <div class="tool-item-info">
                 <div class="tool-item-name">
                     ${escapeHtml(tool.name)}
                     ${externalBadge}
@@ -529,6 +527,11 @@ function renderToolsList() {
                 <div class="tool-item-detail" style="display:none"></div>
             </div>
         `;
+        toolItem.addEventListener('click', function (event) {
+            const infoEl = toolItem.querySelector('.tool-item-info');
+            if (!infoEl) return;
+            toggleToolDetail(infoEl, toolKey, !!tool.is_external, tool.external_mcp || '', event);
+        });
         listContainer.appendChild(toolItem);
     });
     
@@ -543,14 +546,16 @@ function renderToolsList() {
 // 展开/折叠工具详情面板（按需从后端加载 schema）
 function toggleToolDetail(infoEl, toolKey, isExternal, externalMcp, event) {
     // 点击 checkbox 或外部工具徽章时不展开
-    if (event.target.tagName === 'INPUT' || event.target.closest('.external-tool-badge')) return;
+    if (event && (event.target.tagName === 'INPUT' || event.target.closest('.external-tool-badge'))) return;
 
     const detail = infoEl.querySelector('.tool-item-detail');
     const icon = infoEl.querySelector('.tool-expand-icon');
     if (!detail) return;
 
-    const isOpen = detail.style.display !== 'none';
+    // 使用 data-open 作为主状态，避免仅依赖 style.display 带来的首击偶发判定不一致
+    const isOpen = detail.dataset.open === '1';
     detail.style.display = isOpen ? 'none' : 'block';
+    detail.dataset.open = isOpen ? '0' : '1';
     if (icon) icon.textContent = isOpen ? '▶' : '▼';
 
     // 首次展开时从后端按需加载
@@ -1005,7 +1010,6 @@ async function applySettings() {
                 const peLoop = Number.isNaN(peParsed) ? 0 : Math.max(0, peParsed);
                 return {
                     enabled: document.getElementById('multi-agent-enabled')?.checked === true,
-                    default_mode: document.getElementById('multi-agent-default-mode')?.value === 'multi' ? 'multi' : 'single',
                     robot_use_multi_agent: document.getElementById('multi-agent-robot-use')?.checked === true,
                     batch_use_multi_agent: false,
                     plan_execute_loop_max_iterations: peLoop
@@ -1466,12 +1470,15 @@ async function pollExternalMCPToolCount(name, maxAttempts = 10) {
 function renderExternalMCPList(servers) {
     const list = document.getElementById('external-mcp-list');
     if (!list) return;
+    const layout = document.querySelector('.mcp-management-layout');
     
     if (Object.keys(servers).length === 0) {
+        if (layout) layout.classList.add('external-empty');
         const emptyT = typeof window.t === 'function' ? window.t : (k) => k;
         list.innerHTML = '<div class="empty">📋 ' + emptyT('mcp.noExternalMCP') + '<br><span style="font-size: 0.875rem; margin-top: 8px; display: block;">' + emptyT('mcp.clickToAddExternal') + '</span></div>';
         return;
     }
+    if (layout) layout.classList.remove('external-empty');
     
     let html = '<div class="external-mcp-items">';
     for (const [name, server] of Object.entries(servers)) {

web/static/js/tasks.js

@@ -531,6 +531,7 @@ function renderTaskItem(task, statusMap, isHistory = false) {
                         ${isHistory && completedText ? completedText : timeText}
                     </span>
                     ${canCancel ? `<button class="btn-secondary btn-small" onclick="cancelTask('${task.conversationId}', this)">` + _t('tasks.cancelTask') + `</button>` : ''}
+                    ${task.conversationId ? `<button class="btn-secondary btn-small" onclick="navigateToVulnerabilitiesFromTasksPage('conversation', '${task.conversationId}')">` + _t('tasks.viewVulnerabilities') + `</button>` : ''}
                     ${task.conversationId ? `<button class="btn-secondary btn-small" onclick="viewConversation('${task.conversationId}')">` + _t('tasks.viewConversation') + `</button>` : ''}
                 </div>
             </div>
@@ -708,6 +709,17 @@ function viewConversation(conversationId) {
     }
 }
 
+// 跳转漏洞管理并按对话 ID 或批量队列 ID 筛选（队列 ID 走 task_id，与列表筛选项一致）
+function navigateToVulnerabilitiesFromTasksPage(kind, id) {
+    if (!id) return;
+    const enc = encodeURIComponent(id);
+    if (kind === 'queue') {
+        window.location.hash = 'vulnerabilities?task_id=' + enc;
+    } else if (kind === 'conversation') {
+        window.location.hash = 'vulnerabilities?conversation_id=' + enc;
+    }
+}
+
 // 刷新任务列表
 async function refreshTasks() {
     await loadTasks();
@@ -1134,6 +1146,8 @@ function renderBatchQueues() {
         const progress = stats.total > 0 ? Math.round((stats.completed + stats.failed + stats.cancelled) / stats.total * 100) : 0;
         // 允许删除待执行、已完成或已取消状态的队列
         const canDelete = queue.status === 'pending' || queue.status === 'completed' || queue.status === 'cancelled';
+        // 操作列常驻「查看漏洞」，不再使用 --no-actions 隐藏整列（否则无法从运行中队列跳转漏洞页）
+        const noActionsClass = '';
         
         const loadedRoles = batchQueuesState.loadedRoles || [];
         const roleIcon = getRoleIconForDisplay(queue.role, loadedRoles);
@@ -1157,7 +1171,6 @@ function renderBatchQueues() {
             : `<h4 class="batch-queue-card-title batch-queue-card-title--muted">${escapeHtml(_t('tasks.batchQueueUntitled'))}</h4>`;
         const doneCount = stats.completed + stats.failed + stats.cancelled;
 
-        const noActionsClass = canDelete ? '' : ' batch-queue-item--no-actions';
         return `
             <div class="batch-queue-item batch-queue-item--compact${cardMod}${noActionsClass}" data-queue-id="${queue.id}" onclick="showBatchQueueDetail('${queue.id}')">
                 <div class="batch-queue-item__inner batch-queue-item__inner--grid">
@@ -1182,7 +1195,8 @@ function renderBatchQueues() {
                         </div>
                     </div>
                     <div class="batch-queue-item__actions-col" onclick="event.stopPropagation();">
-                        ${canDelete ? `<button type="button" class="batch-queue-icon-btn" onclick="deleteBatchQueueFromList('${queue.id}')" title="${escapeHtml(_t('tasks.deleteQueue'))}" aria-label="${escapeHtml(_t('tasks.deleteQueue'))}"><svg class="batch-queue-icon-btn__svg" width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M3 6h18"/><path d="M19 6v14a2 2 0 0 1-2 2H7a2 2 0 0 1-2-2V6"/><path d="M8 6V4a2 2 0 0 1 2-2h4a2 2 0 0 1 2 2v2"/><path d="M10 11v6"/><path d="M14 11v6"/></svg></button>` : ''}
+                        <button type="button" class="batch-queue-icon-btn" onclick="navigateToVulnerabilitiesFromTasksPage('queue', '${queue.id}')" title="${escapeHtml(_t('tasks.viewVulnerabilitiesQueueTitle'))}" aria-label="${escapeHtml(_t('tasks.viewVulnerabilitiesQueueTitle'))}"><svg class="batch-queue-icon-btn__svg" width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M12 22s8-4 8-10V5l-8-3-8 3v7c0 6 8 10 8 10z"/><path d="M9 12l2 2 4-4"/></svg></button>
+                        ${canDelete ? `<button type="button" class="batch-queue-icon-btn batch-queue-icon-btn--danger" onclick="deleteBatchQueueFromList('${queue.id}')" title="${escapeHtml(_t('tasks.deleteQueue'))}" aria-label="${escapeHtml(_t('tasks.deleteQueue'))}"><svg class="batch-queue-icon-btn__svg" width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M3 6h18"/><path d="M19 6v14a2 2 0 0 1-2 2H7a2 2 0 0 1-2-2V6"/><path d="M8 6V4a2 2 0 0 1 2-2h4a2 2 0 0 1 2 2v2"/><path d="M10 11v6"/><path d="M14 11v6"/></svg></button>` : ''}
                     </div>
                 </div>
             </div>
@@ -1666,12 +1680,24 @@ function startBatchQueueRefresh(queueId) {
         if ((addModal && addModal.style.display === 'block') || hasInlineEdit) {
             return;
         }
-        if (batchQueuesState.currentQueueId === queueId) {
-            showBatchQueueDetail(queueId);
-            refreshBatchQueues();
-        } else {
-            stopBatchQueueRefresh();
+        if (batchQueuesState._bqDetailRefreshing) {
+            return;
         }
+        if (batchQueuesState.currentQueueId !== queueId) {
+            stopBatchQueueRefresh();
+            return;
+        }
+        batchQueuesState._bqDetailRefreshing = true;
+        (async () => {
+            try {
+                await showBatchQueueDetail(queueId);
+                await refreshBatchQueues();
+            } catch (e) {
+                console.warn('批量队列定时刷新失败:', e);
+            } finally {
+                batchQueuesState._bqDetailRefreshing = false;
+            }
+        })();
     }, 3000); // 每3秒刷新一次
 }
 

web/static/js/vulnerability.js

@@ -1,5 +1,43 @@
 // 漏洞管理相关功能
 
+function vulnT(key, opts) {
+    if (typeof window.t === 'function') {
+        return window.t(key, opts);
+    }
+    return key;
+}
+
+function vulnDateLocale() {
+    try {
+        const lang = (window.__locale || '').toLowerCase();
+        if (lang.indexOf('zh') === 0) {
+            return 'zh-CN';
+        }
+    } catch (e) { /* ignore */ }
+    return 'en-US';
+}
+
+function vulnSeverityLabel(code) {
+    const m = {
+        critical: 'dashboard.severityCritical',
+        high: 'dashboard.severityHigh',
+        medium: 'dashboard.severityMedium',
+        low: 'dashboard.severityLow',
+        info: 'dashboard.severityInfo'
+    };
+    return m[code] ? vulnT(m[code]) : code;
+}
+
+function vulnStatusLabel(code) {
+    const m = {
+        open: 'vulnerabilityPage.statusOpen',
+        confirmed: 'vulnerabilityPage.statusConfirmed',
+        fixed: 'vulnerabilityPage.statusFixed',
+        false_positive: 'vulnerabilityPage.statusFalsePositive'
+    };
+    return m[code] ? vulnT(m[code]) : code;
+}
+
 // 从localStorage读取每页显示数量，默认为20
 const getVulnerabilityPageSize = () => {
     const saved = localStorage.getItem('vulnerabilityPageSize');
@@ -10,6 +48,9 @@ let currentVulnerabilityId = null;
 let vulnerabilityFilters = {
     id: '',
     conversation_id: '',
+    task_id: '',
+    conversation_tag: '',
+    task_tag: '',
     severity: '',
     status: ''
 };
@@ -20,10 +61,43 @@ let vulnerabilityPagination = {
     totalPages: 1
 };
 
+// 从地址栏 #vulnerabilities?conversation_id= / ?task_id= 同步筛选（对话菜单、任务管理联动）
+function syncVulnerabilityFiltersFromLocationHash() {
+    const hash = window.location.hash.slice(1);
+    const hashParts = hash.split('?');
+    if (hashParts[0] !== 'vulnerabilities' || hashParts.length < 2) {
+        return;
+    }
+    const params = new URLSearchParams(hashParts.slice(1).join('?'));
+    const cid = (params.get('conversation_id') || '').trim();
+    const tid = (params.get('task_id') || '').trim();
+    if (!cid && !tid) {
+        return;
+    }
+
+    vulnerabilityFilters.conversation_id = '';
+    vulnerabilityFilters.task_id = '';
+    const convEl = document.getElementById('vulnerability-conversation-filter');
+    const taskEl = document.getElementById('vulnerability-task-filter');
+    if (convEl) convEl.value = '';
+    if (taskEl) taskEl.value = '';
+
+    if (cid) {
+        vulnerabilityFilters.conversation_id = cid;
+        if (convEl) convEl.value = cid;
+    }
+    if (tid) {
+        vulnerabilityFilters.task_id = tid;
+        if (taskEl) taskEl.value = tid;
+    }
+    vulnerabilityPagination.currentPage = 1;
+}
+
 // 初始化漏洞管理页面
 function initVulnerabilityPage() {
     // 从localStorage加载每页条数设置
     vulnerabilityPagination.pageSize = getVulnerabilityPageSize();
+    syncVulnerabilityFiltersFromLocationHash();
     loadVulnerabilityStats();
     loadVulnerabilities();
 }
@@ -41,6 +115,9 @@ async function loadVulnerabilityStats() {
         if (vulnerabilityFilters.conversation_id) {
             params.append('conversation_id', vulnerabilityFilters.conversation_id);
         }
+        if (vulnerabilityFilters.task_id) {
+            params.append('task_id', vulnerabilityFilters.task_id);
+        }
 
         const response = await apiFetch(`/api/vulnerabilities/stats?${params.toString()}`);
         if (!response.ok) {
@@ -82,7 +159,7 @@ function updateVulnerabilityStats(stats) {
 // 加载漏洞列表
 async function loadVulnerabilities(page = null) {
     const listContainer = document.getElementById('vulnerabilities-list');
-    listContainer.innerHTML = '<div class="loading-spinner">加载中...</div>';
+    listContainer.innerHTML = `<div class="loading-spinner">${escapeHtml(vulnT('vulnerabilityPage.loading'))}</div>`;
 
     try {
         // 检查apiFetch是否可用
@@ -106,6 +183,15 @@ async function loadVulnerabilities(page = null) {
         if (vulnerabilityFilters.conversation_id) {
             params.append('conversation_id', vulnerabilityFilters.conversation_id);
         }
+        if (vulnerabilityFilters.task_id) {
+            params.append('task_id', vulnerabilityFilters.task_id);
+        }
+        if (vulnerabilityFilters.conversation_tag) {
+            params.append('conversation_tag', vulnerabilityFilters.conversation_tag);
+        }
+        if (vulnerabilityFilters.task_tag) {
+            params.append('task_tag', vulnerabilityFilters.task_tag);
+        }
         if (vulnerabilityFilters.severity) {
             params.append('severity', vulnerabilityFilters.severity);
         }
@@ -148,7 +234,7 @@ async function loadVulnerabilities(page = null) {
         renderVulnerabilityPagination();
     } catch (error) {
         console.error('加载漏洞列表失败:', error);
-        listContainer.innerHTML = `<div class="error-message">加载失败: ${error.message}</div>`;
+        listContainer.innerHTML = `<div class="error-message">${escapeHtml(vulnT('vulnerabilityPage.loadListFailed'))}: ${escapeHtml(error.message)}</div>`;
     }
 }
 
@@ -180,22 +266,12 @@ function renderVulnerabilities(vulnerabilities) {
 
     const html = vulnerabilities.map(vuln => {
         const severityClass = `severity-${vuln.severity}`;
-        const severityText = {
-            'critical': '严重',
-            'high': '高危',
-            'medium': '中危',
-            'low': '低危',
-            'info': '信息'
-        }[vuln.severity] || vuln.severity;
-
-        const statusText = {
-            'open': '待处理',
-            'confirmed': '已确认',
-            'fixed': '已修复',
-            'false_positive': '误报'
-        }[vuln.status] || vuln.status;
-
-        const createdDate = new Date(vuln.created_at).toLocaleString('zh-CN');
+        const severityText = vulnSeverityLabel(vuln.severity);
+        const statusText = vulnStatusLabel(vuln.status);
+        const createdDate = new Date(vuln.created_at).toLocaleString(vulnDateLocale());
+        const dlTitle = escapeHtml(vulnT('vulnerabilityPage.downloadMarkdownTitle'));
+        const editTitle = escapeHtml(vulnT('common.edit'));
+        const deleteTitle = escapeHtml(vulnT('common.delete'));
 
         return `
             <div class="vulnerability-card ${severityClass}">
@@ -214,20 +290,20 @@ function renderVulnerabilities(vulnerabilities) {
                         </div>
                     </div>
                     <div class="vulnerability-actions" onclick="event.stopPropagation();">
-                        <button class="btn-ghost" onclick="downloadVulnerabilityAsMarkdown('${vuln.id}', event)" title="下载Markdown">
+                        <button class="btn-ghost" onclick="downloadVulnerabilityAsMarkdown('${vuln.id}', event)" title="${dlTitle}">
                             <svg width="16" height="16" viewBox="0 0 24 24" fill="none" xmlns="http://www.w3.org/2000/svg">
                                 <path d="M21 15v4a2 2 0 0 1-2 2H5a2 2 0 0 1-2-2v-4" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/>
                                 <polyline points="7 10 12 15 17 10" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/>
                                 <line x1="12" y1="15" x2="12" y2="3" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/>
                             </svg>
                         </button>
-                        <button class="btn-ghost" onclick="editVulnerability('${vuln.id}')" title="编辑">
+                        <button class="btn-ghost" onclick="editVulnerability('${vuln.id}')" title="${editTitle}">
                             <svg width="16" height="16" viewBox="0 0 24 24" fill="none" xmlns="http://www.w3.org/2000/svg">
                                 <path d="M11 4H4a2 2 0 0 0-2 2v14a2 2 0 0 0 2 2h14a2 2 0 0 0 2-2v-7" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/>
                                 <path d="M18.5 2.5a2.121 2.121 0 0 1 3 3L12 15l-4 1 1-4 9.5-9.5z" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/>
                             </svg>
                         </button>
-                        <button class="btn-ghost" onclick="deleteVulnerability('${vuln.id}')" title="删除">
+                        <button class="btn-ghost" onclick="deleteVulnerability('${vuln.id}')" title="${deleteTitle}">
                             <svg width="16" height="16" viewBox="0 0 24 24" fill="none" xmlns="http://www.w3.org/2000/svg">
                                 <path d="M3 6h18M8 6V4a2 2 0 0 1 2-2h4a2 2 0 0 1 2 2v2m3 0v14a2 2 0 0 1-2 2H7a2 2 0 0 1-2-2V6h14z" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/>
                             </svg>
@@ -237,20 +313,27 @@ function renderVulnerabilities(vulnerabilities) {
                 <div class="vulnerability-content" id="content-${vuln.id}" style="display: none;">
                     ${vuln.description ? `<div class="vulnerability-description">${escapeHtml(vuln.description)}</div>` : ''}
                     <div class="vulnerability-details">
-                        <div class="detail-item"><strong>漏洞ID:</strong> <code>${escapeHtml(vuln.id)}</code></div>
-                        ${vuln.type ? `<div class="detail-item"><strong>类型:</strong> ${escapeHtml(vuln.type)}</div>` : ''}
-                        ${vuln.target ? `<div class="detail-item"><strong>目标:</strong> ${escapeHtml(vuln.target)}</div>` : ''}
-                        <div class="detail-item"><strong>会话ID:</strong> <code>${escapeHtml(vuln.conversation_id)}</code></div>
+                        ${vulnDetailField(vulnT('vulnerabilityPage.detailVulnId'), vuln.id, true)}
+                        ${vuln.type ? vulnDetailField(vulnT('vulnerabilityPage.detailType'), vuln.type, false) : ''}
+                        ${vuln.target ? vulnDetailField(vulnT('vulnerabilityPage.detailTarget'), vuln.target, false) : ''}
+                        ${vulnDetailField(vulnT('vulnerabilityPage.detailConversationId'), vuln.conversation_id, true)}
+                        ${vuln.task_id ? vulnDetailField(vulnT('vulnerabilityPage.detailTaskId'), vuln.task_id, true) : ''}
+                        ${vuln.task_queue_id ? vulnDetailField(vulnT('vulnerabilityPage.detailTaskQueueId'), vuln.task_queue_id, true) : ''}
+                        ${vuln.conversation_tag ? vulnDetailField(vulnT('vulnerabilityPage.detailConversationTag'), vuln.conversation_tag, false) : ''}
+                        ${vuln.task_tag ? vulnDetailField(vulnT('vulnerabilityPage.detailTaskTag'), vuln.task_tag, false) : ''}
                     </div>
-                    ${vuln.proof ? `<div class="vulnerability-proof"><strong>证明:</strong><pre>${escapeHtml(vuln.proof)}</pre></div>` : ''}
-                    ${vuln.impact ? `<div class="vulnerability-impact"><strong>影响:</strong> ${escapeHtml(vuln.impact)}</div>` : ''}
-                    ${vuln.recommendation ? `<div class="vulnerability-recommendation"><strong>修复建议:</strong> ${escapeHtml(vuln.recommendation)}</div>` : ''}
+                    ${vuln.proof ? `<div class="vulnerability-proof"><strong>${escapeHtml(vulnT('vulnerabilityPage.detailProof'))}:</strong><pre>${escapeHtml(vuln.proof)}</pre></div>` : ''}
+                    ${vuln.impact ? `<div class="vulnerability-impact"><strong>${escapeHtml(vulnT('vulnerabilityPage.detailImpact'))}:</strong> ${escapeHtml(vuln.impact)}</div>` : ''}
+                    ${vuln.recommendation ? `<div class="vulnerability-recommendation"><strong>${escapeHtml(vulnT('vulnerabilityPage.detailRecommendation'))}:</strong> ${escapeHtml(vuln.recommendation)}</div>` : ''}
                 </div>
             </div>
         `;
     }).join('');
 
     listContainer.innerHTML = html;
+    if (typeof window.applyTranslations === 'function') {
+        window.applyTranslations(listContainer);
+    }
 }
 
 // 渲染分页控件
@@ -277,9 +360,9 @@ function renderVulnerabilityPagination() {
     // 左侧：显示范围信息和每页数量选择器（参考Skills样式）
     paginationHTML += `
         <div class="pagination-info">
-            <span>显示 ${start}-${end} / 共 ${total} 条</span>
+            <span>${escapeHtml(vulnT('skillsPage.paginationShow', { start, end, total }))}</span>
             <label class="pagination-page-size">
-                每页显示
+                ${escapeHtml(vulnT('skillsPage.perPageLabel'))}
                 <select id="vulnerability-page-size-pagination" onchange="changeVulnerabilityPageSize()">
                     <option value="10" ${pageSize === 10 ? 'selected' : ''}>10</option>
                     <option value="20" ${pageSize === 20 ? 'selected' : ''}>20</option>
@@ -293,17 +376,20 @@ function renderVulnerabilityPagination() {
     // 右侧：分页按钮（参考Skills样式：首页、上一页、第X/Y页、下一页、末页）
     paginationHTML += `
         <div class="pagination-controls">
-            <button class="btn-secondary" onclick="loadVulnerabilities(1)" ${currentPage === 1 || total === 0 ? 'disabled' : ''}>首页</button>
-            <button class="btn-secondary" onclick="loadVulnerabilities(${currentPage - 1})" ${currentPage === 1 || total === 0 ? 'disabled' : ''}>上一页</button>
-            <span class="pagination-page">第 ${currentPage} / ${totalPages || 1} 页</span>
-            <button class="btn-secondary" onclick="loadVulnerabilities(${currentPage + 1})" ${currentPage >= totalPages || total === 0 ? 'disabled' : ''}>下一页</button>
-            <button class="btn-secondary" onclick="loadVulnerabilities(${totalPages || 1})" ${currentPage >= totalPages || total === 0 ? 'disabled' : ''}>末页</button>
+            <button class="btn-secondary" onclick="loadVulnerabilities(1)" ${currentPage === 1 || total === 0 ? 'disabled' : ''}>${escapeHtml(vulnT('skillsPage.firstPage'))}</button>
+            <button class="btn-secondary" onclick="loadVulnerabilities(${currentPage - 1})" ${currentPage === 1 || total === 0 ? 'disabled' : ''}>${escapeHtml(vulnT('skillsPage.prevPage'))}</button>
+            <span class="pagination-page">${escapeHtml(vulnT('skillsPage.pageOf', { current: currentPage, total: totalPages || 1 }))}</span>
+            <button class="btn-secondary" onclick="loadVulnerabilities(${currentPage + 1})" ${currentPage >= totalPages || total === 0 ? 'disabled' : ''}>${escapeHtml(vulnT('skillsPage.nextPage'))}</button>
+            <button class="btn-secondary" onclick="loadVulnerabilities(${totalPages || 1})" ${currentPage >= totalPages || total === 0 ? 'disabled' : ''}>${escapeHtml(vulnT('skillsPage.lastPage'))}</button>
         </div>
     `;
     
     paginationHTML += '</div>';
     
     paginationContainer.innerHTML = paginationHTML;
+    if (typeof window.applyTranslations === 'function') {
+        window.applyTranslations(paginationContainer);
+    }
 }
 
 // 改变每页显示数量
@@ -334,10 +420,12 @@ async function changeVulnerabilityPageSize() {
 // 显示添加漏洞模态框
 function showAddVulnerabilityModal() {
     currentVulnerabilityId = null;
-    document.getElementById('vulnerability-modal-title').textContent = (typeof window.t === 'function' ? window.t('vulnerability.addVuln') : '添加漏洞');
+    document.getElementById('vulnerability-modal-title').textContent = vulnT('vulnerability.addVuln');
     
     // 清空表单
     document.getElementById('vulnerability-conversation-id').value = '';
+    document.getElementById('vulnerability-conversation-tag').value = '';
+    document.getElementById('vulnerability-task-tag').value = '';
     document.getElementById('vulnerability-title').value = '';
     document.getElementById('vulnerability-description').value = '';
     document.getElementById('vulnerability-severity').value = '';
@@ -355,14 +443,16 @@ function showAddVulnerabilityModal() {
 async function editVulnerability(id) {
     try {
         const response = await apiFetch(`/api/vulnerabilities/${id}`);
-        if (!response.ok) throw new Error('获取漏洞失败');
+        if (!response.ok) throw new Error(vulnT('vulnerabilityPage.fetchFailed'));
 
         const vuln = await response.json();
         currentVulnerabilityId = id;
-        document.getElementById('vulnerability-modal-title').textContent = (typeof window.t === 'function' ? window.t('vulnerability.editVuln') : '编辑漏洞');
+        document.getElementById('vulnerability-modal-title').textContent = vulnT('vulnerability.editVuln');
 
         // 填充表单
         document.getElementById('vulnerability-conversation-id').value = vuln.conversation_id || '';
+        document.getElementById('vulnerability-conversation-tag').value = vuln.conversation_tag || '';
+        document.getElementById('vulnerability-task-tag').value = vuln.task_tag || '';
         document.getElementById('vulnerability-title').value = vuln.title || '';
         document.getElementById('vulnerability-description').value = vuln.description || '';
         document.getElementById('vulnerability-severity').value = vuln.severity || '';
@@ -376,7 +466,7 @@ async function editVulnerability(id) {
         document.getElementById('vulnerability-modal').style.display = 'block';
     } catch (error) {
         console.error('加载漏洞失败:', error);
-        alert('加载漏洞失败: ' + error.message);
+        alert(vulnT('vulnerability.loadFailed') + ': ' + error.message);
     }
 }
 
@@ -387,12 +477,14 @@ async function saveVulnerability() {
     const severity = document.getElementById('vulnerability-severity').value;
 
     if (!conversationId || !title || !severity) {
-        alert('请填写必填字段：会话ID、标题和严重程度');
+        alert(vulnT('vulnerabilityPage.saveRequiredFields'));
         return;
     }
 
     const data = {
         conversation_id: conversationId,
+        conversation_tag: document.getElementById('vulnerability-conversation-tag').value.trim(),
+        task_tag: document.getElementById('vulnerability-task-tag').value.trim(),
         title: title,
         description: document.getElementById('vulnerability-description').value.trim(),
         severity: severity,
@@ -420,7 +512,7 @@ async function saveVulnerability() {
 
         if (!response.ok) {
             const error = await response.json();
-            throw new Error(error.error || '保存失败');
+            throw new Error(error.error || vulnT('vulnerabilityPage.saveFailed'));
         }
 
         closeVulnerabilityModal();
@@ -430,13 +522,13 @@ async function saveVulnerability() {
         loadVulnerabilities();
     } catch (error) {
         console.error('保存漏洞失败:', error);
-        alert('保存漏洞失败: ' + error.message);
+        alert(vulnT('vulnerabilityPage.saveFailed') + ': ' + error.message);
     }
 }
 
 // 删除漏洞
 async function deleteVulnerability(id) {
-    if (!confirm('确定要删除此漏洞吗？')) {
+    if (!confirm(vulnT('vulnerability.deleteConfirm'))) {
         return;
     }
 
@@ -445,7 +537,7 @@ async function deleteVulnerability(id) {
             method: 'DELETE'
         });
 
-        if (!response.ok) throw new Error('删除失败');
+        if (!response.ok) throw new Error(vulnT('vulnerabilityPage.deleteFailed'));
 
         loadVulnerabilityStats();
         // 删除后，如果当前页没有数据了，回到上一页
@@ -458,7 +550,7 @@ async function deleteVulnerability(id) {
         loadVulnerabilities();
     } catch (error) {
         console.error('删除漏洞失败:', error);
-        alert('删除漏洞失败: ' + error.message);
+        alert(vulnT('vulnerabilityPage.deleteFailed') + ': ' + error.message);
     }
 }
 
@@ -472,6 +564,9 @@ function closeVulnerabilityModal() {
 function filterVulnerabilities() {
     vulnerabilityFilters.id = document.getElementById('vulnerability-id-filter').value.trim();
     vulnerabilityFilters.conversation_id = document.getElementById('vulnerability-conversation-filter').value.trim();
+    vulnerabilityFilters.task_id = document.getElementById('vulnerability-task-filter').value.trim();
+    vulnerabilityFilters.conversation_tag = document.getElementById('vulnerability-conversation-tag-filter').value.trim();
+    vulnerabilityFilters.task_tag = document.getElementById('vulnerability-task-tag-filter').value.trim();
     vulnerabilityFilters.severity = document.getElementById('vulnerability-severity-filter').value;
     vulnerabilityFilters.status = document.getElementById('vulnerability-status-filter').value;
 
@@ -486,12 +581,18 @@ function filterVulnerabilities() {
 function clearVulnerabilityFilters() {
     document.getElementById('vulnerability-id-filter').value = '';
     document.getElementById('vulnerability-conversation-filter').value = '';
+    document.getElementById('vulnerability-task-filter').value = '';
+    document.getElementById('vulnerability-conversation-tag-filter').value = '';
+    document.getElementById('vulnerability-task-tag-filter').value = '';
     document.getElementById('vulnerability-severity-filter').value = '';
     document.getElementById('vulnerability-status-filter').value = '';
 
     vulnerabilityFilters = {
         id: '',
         conversation_id: '',
+        task_id: '',
+        conversation_tag: '',
+        task_tag: '',
         severity: '',
         status: ''
     };
@@ -532,67 +633,193 @@ function escapeHtml(text) {
     return div.innerHTML;
 }
 
-// 将漏洞格式化为Markdown
+/** 复制详情字段（编码由 encodeURIComponent 传入，避免引号截断） */
+function vulnerabilityCopyEncoded(evt, encoded) {
+    if (evt && evt.stopPropagation) {
+        evt.stopPropagation();
+    }
+    let text = '';
+    try {
+        text = decodeURIComponent(encoded);
+    } catch (e) {
+        return;
+    }
+    const done = () => {
+        if (evt && evt.target && evt.target.closest) {
+            const btn = evt.target.closest('.vuln-detail-field__copy');
+            if (btn) {
+                const t0 = btn.getAttribute('title') || '';
+                btn.setAttribute('title', vulnT('common.copied'));
+                setTimeout(() => btn.setAttribute('title', t0), 1600);
+            }
+        }
+    };
+    if (navigator.clipboard && typeof navigator.clipboard.writeText === 'function') {
+        navigator.clipboard.writeText(text).then(done).catch(() => {
+            try {
+                const ta = document.createElement('textarea');
+                ta.value = text;
+                ta.style.position = 'fixed';
+                ta.style.left = '-9999px';
+                document.body.appendChild(ta);
+                ta.select();
+                document.execCommand('copy');
+                document.body.removeChild(ta);
+                done();
+            } catch (err) {
+                console.error('copy failed', err);
+            }
+        });
+    } else {
+        try {
+            const ta = document.createElement('textarea');
+            ta.value = text;
+            ta.style.position = 'fixed';
+            ta.style.left = '-9999px';
+            document.body.appendChild(ta);
+            ta.select();
+            document.execCommand('copy');
+            document.body.removeChild(ta);
+            done();
+        } catch (err) {
+            console.error('copy failed', err);
+        }
+    }
+}
+
+function vulnDetailField(label, value, asCode) {
+    if (value === undefined || value === null || String(value) === '') {
+        return '';
+    }
+    const s = String(value);
+    const enc = encodeURIComponent(s);
+    const copyTitle = escapeHtml(vulnT('common.copy'));
+    const valueEl = asCode
+        ? `<code class="vuln-detail-field-value">${escapeHtml(s)}</code>`
+        : `<span class="vuln-detail-field-value">${escapeHtml(s)}</span>`;
+    const copyBtn = `<button type="button" class="vuln-detail-field__copy" onclick="vulnerabilityCopyEncoded(event, '${enc}')" title="${copyTitle}" aria-label="${copyTitle}">
+<svg width="14" height="14" viewBox="0 0 24 24" fill="none" xmlns="http://www.w3.org/2000/svg" aria-hidden="true" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="9" y="9" width="13" height="13" rx="2"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/></svg>
+</button>`;
+    return `<div class="vuln-detail-field">
+<div class="vuln-detail-field__label">${escapeHtml(label)}</div>
+<div class="vuln-detail-field__row">${valueEl}${copyBtn}</div>
+</div>`;
+}
+
+// 将漏洞格式化为Markdown（章节标题随界面语言）
 function formatVulnerabilityAsMarkdown(vuln) {
-    const severityText = {
-        'critical': '严重',
-        'high': '高危',
-        'medium': '中危',
-        'low': '低危',
-        'info': '信息'
-    }[vuln.severity] || vuln.severity;
-
-    const statusText = {
-        'open': '待处理',
-        'confirmed': '已确认',
-        'fixed': '已修复',
-        'false_positive': '误报'
-    }[vuln.status] || vuln.status;
-
-    const createdDate = new Date(vuln.created_at).toLocaleString('zh-CN');
-    const updatedDate = new Date(vuln.updated_at).toLocaleString('zh-CN');
+    const severityText = vulnSeverityLabel(vuln.severity);
+    const statusText = vulnStatusLabel(vuln.status);
+    const loc = vulnDateLocale();
+    const createdDate = new Date(vuln.created_at).toLocaleString(loc);
+    const updatedDate = new Date(vuln.updated_at).toLocaleString(loc);
+    const L = (k) => vulnT('vulnerabilityMd.' + k);
 
     let markdown = `# ${vuln.title}\n\n`;
-    
-    markdown += `## 基本信息\n\n`;
-    markdown += `- **漏洞ID**: \`${vuln.id}\`\n`;
-    markdown += `- **严重程度**: ${severityText}\n`;
-    markdown += `- **状态**: ${statusText}\n`;
+
+    markdown += `## ${L('headingBasic')}\n\n`;
+    markdown += `- **${L('labelId')}**: \`${vuln.id}\`\n`;
+    markdown += `- **${L('labelSeverity')}**: ${severityText}\n`;
+    markdown += `- **${L('labelStatus')}**: ${statusText}\n`;
     if (vuln.type) {
-        markdown += `- **类型**: ${vuln.type}\n`;
+        markdown += `- **${L('labelType')}**: ${vuln.type}\n`;
     }
     if (vuln.target) {
-        markdown += `- **目标**: ${vuln.target}\n`;
+        markdown += `- **${L('labelTarget')}**: ${vuln.target}\n`;
     }
-    markdown += `- **会话ID**: \`${vuln.conversation_id}\`\n`;
-    markdown += `- **创建时间**: ${createdDate}\n`;
-    markdown += `- **更新时间**: ${updatedDate}\n\n`;
+    markdown += `- **${L('labelConversationId')}**: \`${vuln.conversation_id}\`\n`;
+    if (vuln.task_id) {
+        markdown += `- **${L('labelTaskId')}**: \`${vuln.task_id}\`\n`;
+    }
+    if (vuln.task_queue_id) {
+        markdown += `- **${L('labelTaskQueueId')}**: \`${vuln.task_queue_id}\`\n`;
+    }
+    if (vuln.conversation_tag) {
+        markdown += `- **${L('labelConversationTag')}**: ${vuln.conversation_tag}\n`;
+    }
+    if (vuln.task_tag) {
+        markdown += `- **${L('labelTaskTag')}**: ${vuln.task_tag}\n`;
+    }
+    markdown += `- **${L('labelCreated')}**: ${createdDate}\n`;
+    markdown += `- **${L('labelUpdated')}**: ${updatedDate}\n\n`;
 
     if (vuln.description) {
-        markdown += `## 描述\n\n${vuln.description}\n\n`;
+        markdown += `## ${L('headingDescription')}\n\n${vuln.description}\n\n`;
     }
 
     if (vuln.proof) {
-        markdown += `## 证明（POC）\n\n\`\`\`\n${vuln.proof}\n\`\`\`\n\n`;
+        markdown += `## ${L('headingProof')}\n\n\`\`\`\n${vuln.proof}\n\`\`\`\n\n`;
     }
 
     if (vuln.impact) {
-        markdown += `## 影响\n\n${vuln.impact}\n\n`;
+        markdown += `## ${L('headingImpact')}\n\n${vuln.impact}\n\n`;
     }
 
     if (vuln.recommendation) {
-        markdown += `## 修复建议\n\n${vuln.recommendation}\n\n`;
+        markdown += `## ${L('headingRecommendation')}\n\n${vuln.recommendation}\n\n`;
     }
 
     return markdown;
 }
 
+function buildVulnerabilityFilterParams() {
+    const params = new URLSearchParams();
+    const keys = ['id', 'conversation_id', 'task_id', 'conversation_tag', 'task_tag', 'severity', 'status'];
+    keys.forEach((k) => {
+        if (vulnerabilityFilters[k]) {
+            params.append(k, vulnerabilityFilters[k]);
+        }
+    });
+    return params;
+}
+
+function triggerTextDownload(fileName, content) {
+    const blob = new Blob([content], { type: 'text/markdown;charset=utf-8' });
+    const url = URL.createObjectURL(blob);
+    const link = document.createElement('a');
+    link.href = url;
+    link.download = fileName;
+    document.body.appendChild(link);
+    link.click();
+    document.body.removeChild(link);
+    URL.revokeObjectURL(url);
+}
+
+async function exportVulnerabilityReports() {
+    try {
+        const params = buildVulnerabilityFilterParams();
+        params.set('mode', 'summary');
+        params.set('group_by', 'conversation');
+        const response = await apiFetch(`/api/vulnerabilities/export?${params.toString()}`);
+        if (!response.ok) {
+            const error = await response.json().catch(() => ({ error: vulnT('vulnerabilityPage.exportFailedMessage') }));
+            throw new Error(error.error || vulnT('vulnerabilityPage.exportFailedMessage'));
+        }
+        const data = await response.json();
+        const files = Array.isArray(data.files) ? data.files : [];
+        if (!files.length) {
+            alert(vulnT('vulnerabilityPage.exportNoResults'));
+            return;
+        }
+        files.forEach((file, idx) => {
+            setTimeout(() => triggerTextDownload(file.filename || `vulnerability-export-${idx + 1}.md`, file.content || ''), idx * 120);
+        });
+        if (files.length > 1) {
+            alert(vulnT('vulnerabilityPage.exportStarted', { count: files.length }));
+        }
+    } catch (error) {
+        console.error('导出漏洞报告失败:', error);
+        alert(vulnT('vulnerabilityPage.exportFailed') + ': ' + error.message);
+    }
+}
+
+
 // 下载漏洞为Markdown格式
 async function downloadVulnerabilityAsMarkdown(id, event) {
     try {
         const response = await apiFetch(`/api/vulnerabilities/${id}`);
         if (!response.ok) {
-            throw new Error('获取漏洞失败');
+            throw new Error(vulnT('vulnerabilityPage.fetchFailed'));
         }
 
         const vuln = await response.json();
@@ -626,8 +853,8 @@ async function downloadVulnerabilityAsMarkdown(id, event) {
         if (event && event.target) {
             const button = event.target.closest('button');
             if (button) {
-                const originalTitle = button.title || '下载Markdown';
-                button.title = '下载成功！';
+                const originalTitle = button.title || vulnT('vulnerabilityPage.downloadMarkdownTitle');
+                button.title = vulnT('vulnerabilityPage.downloadOkTitle');
                 setTimeout(() => {
                     button.title = originalTitle;
                 }, 2000);
@@ -635,7 +862,7 @@ async function downloadVulnerabilityAsMarkdown(id, event) {
         }
     } catch (error) {
         console.error('下载失败:', error);
-        alert('下载失败: ' + error.message);
+        alert(vulnT('vulnerabilityPage.downloadFailed') + ': ' + error.message);
     }
 }
 
@@ -645,5 +872,12 @@ window.onclick = function(event) {
     if (event.target === modal) {
         closeVulnerabilityModal();
     }
-}
+};
+
+document.addEventListener('languagechange', function () {
+    const page = document.getElementById('page-vulnerabilities');
+    if (page && page.classList.contains('active')) {
+        loadVulnerabilities();
+    }
+});
 

web/static/js/webshell.js

@@ -2881,17 +2881,6 @@ function runWebshellAiSend(conn, inputEl, sendBtn, messagesContainer) {
                     } else if (_et === 'warning') {
                         appendTimelineItem('warning', '⚠️ ' + (_em || ''), '', _ed);
 
-                    // ─── Eino recovery ───
-                    } else if (_et === 'eino_recovery') {
-                        var runIdx = _ed.runIndex != null ? _ed.runIndex : (_ed.einoRetry != null ? _ed.einoRetry + 1 : 1);
-                        var maxRuns = _ed.maxRuns != null ? _ed.maxRuns : 3;
-                        var recTitle = wsTOr('chat.einoRecoveryTitle', '') ||
-                            ('🔄 工具参数无效 · 第 ' + runIdx + '/' + maxRuns + ' 轮（已追加提示）');
-                        if (typeof window.t === 'function') {
-                            try { recTitle = window.t('chat.einoRecoveryTitle', { n: runIdx, max: maxRuns }); } catch (e) { /* */ }
-                        }
-                        appendTimelineItem('eino_recovery', recTitle, _em, _ed);
-
                     // ─── Tool calls ───
                     } else if (_et === 'tool_calls_detected' && _ed) {
                         var count = _ed.count || 0;