jarvis-agent-factory 4.7.85 → 4.7.87

package/README.md

@@ -165,7 +165,6 @@ jarvis remove claude -g --engine --force # 全局级：用户目录全部 Jarvis
 | | `/skill-flow` | [skill-flow.md](docs/flows/skill-flow.md) | Skill化 |
 | | `/task-design` | [task-design.md](docs/flows/task-design.md) | 任务分解 |
 | 发布 | `/publish` | [publish.md](docs/flows/publish.md) | 一键发布 |
-| 移动 | `/mobile` | [mobile.md](docs/flows/mobile.md) | 统一编排 |
 | 框架 | `/flutter` `/expo` `/swift` `/kotlin` `/taro` `/miniprogram` `/uni-app` `/react` `/vue` | 各框架同 | A→E |
 
 ## License

package/dist/package.json

@@ -1,6 +1,6 @@
 {
   "name": "jarvis-agent-factory",
-  "version": "4.7.85",
+  "version": "4.7.87",
   "description": "Jarvis Agent Factory CLI — Claude Code 多智能体 AI 编程助手配置安装器 | Multi-agent AI coding assistant config installer for Claude Code",
   "keywords": [
     "jarvis",

package/dist/src/templates/platforms/claude/commands/ask.md

@@ -3,7 +3,7 @@ name: ask
 description: 需求探询指令——4模式(K0需求摄入→K1信息收集→K2分析综合→K3交付产出)，文档驱动，Team/Subagent调度硬约束。支持 --* 标志位指定模式。
 model: inherit
 argument-hint: [--interview|--direct|--consensus|--review] <需求/想法/计划/指令>
-tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "Agent", "AskUserQuestion", "WebFetch", "WebSearch", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status"]
+tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "Agent", "AskUserQuestion", "WebFetch", "WebSearch", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # 需求探询（4 模式自适应 · 文档驱动 · Flag 可控）
@@ -59,12 +59,12 @@ Skill("spec-driven-development")
 
 ### 步骤
 
-0. **澄清前并行探索（需求摄入前，同一消息同时发出）**——spawn `code-explore-expert` + `docs-research-expert`：
-   - `code-explore-expert`：项目全景——技术栈、目录结构、已有模块/页面/组件、路由、状态管理模式
-   - `docs-research-expert`：项目 AGENTS.md/README.md/架构决策记录、历史需求文档
-   - 探索结果整理为"项目上下文摘要"，用于后续模式选择和需求澄清
+0. **加载会话上下文（摄入前第一步，不可跳过）**：
+   - `mcp__jarvis-engine__session_context()` — 加载历史会话摘要，检查是否有与此需求相关的未完成事项或历史决策
+   - `mcp__jarvis-engine__jarvis_priority_context({ action: "get" })` — 读取项目关键约束（如"API 端口 3456""用 SQLite 不用 PostgreSQL"）
+   - 若历史会话有相关上下文，整理为"历史参考摘要"，用于后续模式选择和需求澄清
 
-1. **解析 Flag**——从用户输入中提取 `--*` 标志位：
+1. **快速解析用户输入**——扫描 `--*` 标志位并初步评估输入清晰度：
 
    ```
    解析规则：扫描 ARGUMENTS，按第一个匹配的 flag 确定模式。
@@ -77,6 +77,8 @@ Skill("spec-driven-development")
 
    匹配到 flag 后，将其从提示词中移除，剩余部分作为需求文本。**直接跳到步骤 3 进入对应模式，跳过步骤 2 的自动判定。**
 
+   同时提取用户输入中的关键主题词（如"用户认证""数据库迁移""性能优化"），用于步骤 2.5 的靶向探索。
+
 2. **（仅无 flag 时）解析用户输入**——评估输入清晰度，自动判定模式：
 
    | 信号 | 含义 | 倾向模式 |
@@ -89,20 +91,37 @@ Skill("spec-driven-development")
 
    输出模式判定及一句话理由。
 
-3. **按模式执行 K0 摄入**：
+2.5. **靶向并行探索（模式确定后、摄入前，同一消息同时发出）**——根据实际需要 spawn 探索 Agent（按需派出，最多 10 个）：
+
+   必选（始终派出）：
+   - `code-explore-expert`：基于用户输入中的关键主题词，针对性探索相关代码模块、现有实现模式、依赖关系、可复用组件（而非全项目盲目扫描）
+   - `docs-research-expert`：项目 AGENTS.md/README.md/架构决策记录、与主题相关的历史需求文档
+
+   按需（根据主题判断是否派出）：
+   - `external-resource-expert`：涉及外部技术选型/框架/库/API 时，搜索相关最新文档和最佳实践
+   - 额外 `code-explore-expert`：主题跨多个不相关模块时，每个模块独立派出一个 Agent
+   - 额外 `docs-research-expert`：涉及多个独立子系统/服务时，每个派一个 Agent 调查对应文档
+
+   - spawn 前调用 `mcp__jarvis-engine__gate_check({ operation: "read" })` 确认 Gate K0 允许读取
+   - **靶向原则**：探索范围由步骤 1 提取的关键主题词限定，避免泛泛的全项目扫描；若无明确主题词（如用户只说"帮我看看"），则使用项目全景摘要（技术栈、目录结构、核心模块）
+   - **上限 10 个**：根据主题复杂度和涉及模块数量决定派出几个 Agent，最多 10 个，宁可多探索也不要遗漏关键上下文
+   - 探索结果整理为"靶向上下文摘要"，用于后续模式特定摄入
+
+3. **按模式执行 K0 摄入**（基于步骤 2.5 的靶向上下文）：
 
    **Interview**——苏格拉底式追问（每轮 1-2 问，AskUserQuestion）：
+   - 基于靶向上下文，聚焦关键澄清点
    - 这个功能最终要达成什么目标？
    - 谁会使用？在什么场景下？
    - 有哪些已知约束？
    - 什么明确不在范围内？
    - 若用户拒绝回答或回答模糊，记录为"待澄清"并继续
 
-   **Direct**——直接解析用户输入，提取关键需求点、约束、验收条件
+   **Direct**——直接解析用户输入，结合靶向上下文提取关键需求点、约束、验收条件
 
-   **Consensus**——加载用户提供的现有计划/方案，确认审查范围
+   **Consensus**——加载用户提供的现有计划/方案，结合靶向上下文确认审查范围
 
-   **Review**——加载待优化的指令/流程编排，确认优化目标
+   **Review**——加载待优化的指令/流程编排，结合靶向上下文确认优化目标
 
 4. **产出 K0 文档**（必须写入，作为 Gate 检查依据）：
    - Interview → `.jarvis/YYYY-MM-DD/requirements/problem-space.md`
@@ -120,17 +139,14 @@ Skill("spec-driven-development")
 
 ### Interview 模式——深度代码探索
 
-1. spawn `code-explore-expert`（subagent，只读）：
-   - 当前项目的技术栈和架构
-   - 相关现有模块和代码路径
-   - 可能受影响的区域
-   - **失败处理**：重试 1 次；仍失败则编排者自行 Read 关键文件，记录"探索不完整"标记
-
-2. spawn `external-resource-expert`（subagent，只读，可选）补充领域知识
+1. 根据 K0 澄清结果和需求范围，按需派出探索 Agent（最多 10 个）：
+   - `code-explore-expert`（subagent，只读）：当前项目的技术栈和架构、相关现有模块和代码路径、可能受影响的区域。涉及多个独立模块时，每个模块派出一个独立 Agent
+   - `external-resource-expert`（subagent，只读，按需）：若涉及外部技术选型/框架/库/API，搜索相关最新文档
+   - **失败处理**：单个 Agent 失败重试 1 次；仍失败则编排者自行 Read 关键文件，记录"探索不完整"标记
 
-3. 两个 subagent 可并行 spawn（同一消息中并发 Agent 调用）
+2. 所有探索 Agent 在同一消息中并发 spawn（无依赖关系）
 
-4. 产出 `.jarvis/YYYY-MM-DD/requirements/exploration-report.md`
+3. 产出 `.jarvis/YYYY-MM-DD/requirements/exploration-report.md`
 
 ### Direct 模式——快速上下文确认
 
@@ -372,8 +388,8 @@ Skill("spec-driven-development")
 
 | Gate | 引擎策略 | 调度方式 | 约束 |
 |------|---------|---------|------|
-| K0 | `subagent_only` | 编排者直接执行 | 不 spawn Agent，编排者做模式判定+追问 |
-| K1 | `subagent_only` | `Agent` 工具直接 spawn subagent | 只读，1-2 个 code-explore-expert 可并行；external-resource-expert 可并行 |
+| K0 | `subagent_only` | 编排者主导 + subagent 靶向探索 | 步骤 2.5 按需 spawn 探索 Agent（最多 10 个，只读）；编排者做模式判定+追问 |
+| K1 | `subagent_only` | `Agent` 工具直接 spawn subagent | 只读，按需派出（最多 10 个），同一消息并发；external-resource-expert 按需 |
 | K2 Consensus | `prefer_team` | 首选 TeamCreate + Agent(team_name)；回退 subagent | Architect 并行审查，各只读；Critic 由编排者担任 |
 | K2 其他 | `prefer_team` | 编排者主导，subagent 辅助 | code-explore-expert 只读确认 |
 | K3 | `subagent_only` | 编排者直接产出 | 不 spawn Agent |

package/dist/src/templates/platforms/claude/commands/auto.md

@@ -113,10 +113,12 @@ mcp__jarvis-engine__session_join({
 若路由到 Gate C-impl / Gate C2 / Gate D（需要写代码或审查的 Gate），在 gate_jump 之前先并行收集信息：
 
 ```
-# 同一消息同时发出（Gate A 允许 spawn_impl）
+# 根据实际需要派出（最多 10 个），同一消息同时发出
 Agent(code-explore-expert, "探索相关代码区域，收集现有实现模式、文件结构和依赖关系")
 Agent(external-resource-expert, "搜索相关技术最新文档和最佳实践")  
 Agent(docs-research-expert, "调研项目现有文档和架构决策记录")
+# 若涉及多个独立模块，每个模块派出独立 code-explore-expert
+# 若涉及多个外部技术栈，每个技术栈派出独立 external-resource-expert
 ```
 
 探索 Agent 返回后，整理上下文摘要，再 gate_jump 到目标 Gate。这样进入目标 Gate 时已有完整上下文，后续 Agent 工作效率大幅提升。
@@ -133,12 +135,12 @@ Agent(docs-research-expert, "调研项目现有文档和架构决策记录")
 
 **🔴 设计类任务硬约束**：路由到 `full`/`frontend`/`backend` 流水线 Gate A 入口的任务，**必须先与用户确认需求再产出文档**。跳过确认直接写需求文档 → 违反红线。
 
-**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出）**
+**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出，按需派出，最多 10 个）**
 
-spawn `code-explore-expert` + `external-resource-expert` + `docs-research-expert`（spawn 前 `gate_check("read")`）：
-- `code-explore-expert`：项目全景——技术栈、目录结构、已有模块/页面/组件、路由、状态管理模式
-- `external-resource-expert`：相关技术最新文档、版本变更、最佳实践
-- `docs-research-expert`：项目 AGENTS.md/README.md/架构决策记录、历史需求文档
+spawn 前 `gate_check("read")`，根据实际需要派出探索 Agent：
+- `code-explore-expert`（必选）：项目全景——技术栈、目录结构、已有模块/页面/组件、路由、状态管理模式。若项目涉及多个独立子系统，每个子系统派一个独立 Agent
+- `external-resource-expert`（按需）：涉及外部技术选型/框架/库/API 时，搜索相关最新文档、版本变更、最佳实践
+- `docs-research-expert`（按需）：项目 AGENTS.md/README.md/架构决策记录、历史需求文档
 
 探索结果回来后，整理为"项目上下文摘要"，用于后续需求澄清。
 
@@ -148,11 +150,11 @@ spawn `code-explore-expert` + `external-resource-expert` + `docs-research-expert
 - 模糊时加载 `Skill("idea-refine")`
 - 产出需求文档到 `.jarvis/YYYY-MM-DD/requirements/<topic>.md`，标注 `REQ-XXX`
 
-**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出）**
+**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出，按需派出，最多 10 个）**
 
-spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
-- `code-explore-expert`：需求涉及的特定模块、相关代码路径、依赖链路、可复用组件
-- `external-resource-expert`：需求相关的库 API 文档、兼容性注意事项、技术方案参考
+spawn 前 `gate_check("read")`，根据确认后的需求范围和涉及模块数量派出 Agent：
+- `code-explore-expert`（必选）：需求涉及的特定模块、相关代码路径、依赖链路、可复用组件。涉及多个独立模块时每个模块独立派一个 Agent
+- `external-resource-expert`（按需）：需求相关的库 API 文档、兼容性注意事项、技术方案参考。涉及多个外部技术栈时每个派一个 Agent
 
 探索结果整理为"靶向上下文摘要"，注入 Gate B 任务分解和 Gate C 实现规划。
 
@@ -219,12 +221,14 @@ Skill(skill="<从skill-assignment文档复制>")
 
 Lint → Type-check → Build → Deps Audit，全部通过后推进。
 
-### Gate C1.5：视觉验证（条件性）
+### Gate C1.5：视觉验证（前端任务强制执行，不询问）
 
-**触发条件**：涉及前端页面/组件/样式变更、UI 框架开发、移动端开发。纯后端/逻辑/算法/配置任务 → 跳过。
+**触发条件**：涉及前端页面/组件/样式变更、UI 框架开发、移动端开发 → **直接执行，不询问用户**。纯后端/逻辑/算法/配置任务 → 跳过。
+
+> **设计意图**：开发调试本身就需要视觉验证，暂停询问是多余的中断。前端任务默认执行 C1.5，不做"要不要验证"的判断。
 
 **条件**：
-- 预览服务器已启动（通过 Playwright MCP 连接浏览器）
+- 预览服务器已启动（dev server 运行中）
 - 修改前/后对比截图已附
 - 响应式三视口截图已附（mobile 375x812 / tablet 768x1024 / desktop 1280x800）
 - 关键样式属性已验证（布局/颜色/字体/间距）
@@ -234,7 +238,19 @@ Lint → Type-check → Build → Deps Audit，全部通过后推进。
 
 ```
 spawn 前 gate_check({ operation: "spawn_test" }) 确认 Gate C1.5 允许测试
-└── spawn browser-test-expert（浏览器视觉验证：截图+响应式多视口+样式检查）
+└── spawn frontend-debug-expert（主力：Chrome DevTools MCP 深度视觉验证）
+    ├── Chrome DevTools MCP：navigate → snapshot → screenshot（含多视口）
+    ├── 性能追踪（performance_start_trace）+ 控制台/网络诊断
+    ├── 响应式三视口截图（mobile/tablet/desktop）
+    ├── 修改前/后对比截图
+    ├── 关键样式属性验证（evaluate_script 读取 computed styles）
+    └── 产出视觉验证报告到 .jarvis/YYYY-MM-DD/testing/<topic>-visual-verification.md
+```
+
+**兜底（CDP 不可用或仅需快速页面验证时）**：
+
+```
+spawn browser-test-expert（兜底：agent-browser snapshot + Playwright MCP）
     ├── 模式：agent-browser snapshot 精确获取页面 + Playwright MCP 稳定执行截图
     ├── 响应式三视口截图（mobile/tablet/desktop）
     ├── 修改前/后对比截图
@@ -245,8 +261,8 @@ spawn 前 gate_check({ operation: "spawn_test" }) 确认 Gate C1.5 允许测试
 **通过**：`advance_gate({ gate: "Gate C2" })`
 
 **不通过**：
-1. **证据缺失** → 退回 browser-test-expert 补充截图/样式验证数据
-2. **布局问题**（溢出/重叠/错位）→ spawn `frontend-debug-expert`（Chrome DevTools：元素定位/样式追踪/布局分析）定位根因 → spawn 原实现 Agent 修复源文件 → 重新走视觉验证
+1. **证据缺失** → 退回执行 agent 补充截图/样式验证数据
+2. **布局问题**（溢出/重叠/错位）→ frontend-debug-expert 定位根因（Chrome DevTools：元素定位/样式追踪/布局分析）→ spawn 原实现 Agent 修复源文件 → 重新走视觉验证
 3. 修复后重新过 Gate C1.5，最多 2 轮；仍不通过 → 标记 `BLOCKED`，附最新截图证据向用户报告
 
 ### Gate C2：测试验证

package/dist/src/templates/platforms/claude/commands/backend.md

@@ -3,7 +3,7 @@ description: 后端开发生命周期——需求→任务→计划→实现→
 name: backend
 model: inherit
 argument-hint: [后端需求描述]
-tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__file_claim_check", "mcp__jarvis-engine__file_claim_register", "mcp__jarvis-engine__file_claim_release", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce"]
+tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__file_claim_check", "mcp__jarvis-engine__file_claim_register", "mcp__jarvis-engine__file_claim_release", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # 后端开发生命周期
@@ -90,7 +90,7 @@ Gate C-impl:
 
 ### Gate A：需求澄清
 
-**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出）**
+**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出，最多 10 个）**
 
 spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
 - `code-explore-expert`：项目全景——技术栈、目录结构、已有服务/路由/中间件、数据层 Schema、API 契约
@@ -104,7 +104,7 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 - 模糊时加载 `Skill("idea-refine")`
 - 产出需求文档到 `.jarvis/YYYY-MM-DD/requirements/`，标注 `REQ-XXX`
 
-**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出）**
+**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出，最多 10 个）**
 
 spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
 - `code-explore-expert`：需求涉及的特定服务/模块、相关代码路径、数据模型、API 依赖链路
@@ -118,7 +118,7 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 
 | Gate | 可并行操作 |
 |------|-----------|
-| Gate A 通过后 | `code-explore-expert` × N（多目录并行探索，spawn 前 `gate_check("read")`）+ `external-resource-expert` × N（多库并行搜索） |
+| Gate A 通过后 | `code-explore-expert` × N（多目录并行探索，最多 10 个，spawn 前 `gate_check("read")`）+ `external-resource-expert` × N（多库并行搜索，最多 10 个） |
 | Gate B：DDD→BDD→TDD | `task-design` 按顺序 spawn（DDD→BDD→TDD），不可并行但可连续 |
 | Gate B1 | `backend-architect` + `database-architect`（如需架构评审，二者可并行） |
 | Gate C 实现 Batch | 按 `parallel_batches` 执行，同 Batch 内并行 |

package/dist/src/templates/platforms/claude/commands/browser.md

@@ -3,7 +3,7 @@ description: 浏览器自动化——agent-browser + Playwright MCP 混合模式
 name: browser
 model: inherit
 argument-hint: "[URL 或功能描述]"
-tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__gate_jump"]
+tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__gate_jump", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # 浏览器自动化

package/dist/src/templates/platforms/claude/commands/bug-fix.md

@@ -3,7 +3,7 @@ description: Bug 修复闭环——浏览器复现→定位根因→修复→浏
 name: bug-fix
 model: inherit
 argument-hint: [Bug 描述、URL 或复现步骤]
-tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_jump", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__advance_gate", "mcp__playwright__browser_navigate", "mcp__playwright__browser_snapshot", "mcp__playwright__browser_click", "mcp__playwright__browser_type", "mcp__playwright__browser_press_key", "mcp__playwright__browser_take_screenshot", "mcp__playwright__browser_console_messages", "mcp__playwright__browser_network_requests", "mcp__playwright__browser_wait_for"]
+tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_jump", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__advance_gate", "mcp__playwright__browser_navigate", "mcp__playwright__browser_snapshot", "mcp__playwright__browser_click", "mcp__playwright__browser_type", "mcp__playwright__browser_press_key", "mcp__playwright__browser_take_screenshot", "mcp__playwright__browser_console_messages", "mcp__playwright__browser_network_requests", "mcp__playwright__browser_wait_for", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # Bug 修复闭环（浏览器复现 → 修复 → 验证）
@@ -28,8 +28,9 @@ Skill("browser-testing")
 
 产物输出目录: `.jarvis/YYYY-MM-DD/bug-fix/`
 
-### 步骤 0：并行信息收集（同一消息同时发出）
+### 步骤 0：并行信息收集（同一消息同时发出，最多 10 个）
 Agent(code-explore-expert, "根据Bug描述探索相关代码区域：定位可能的根因文件、分析调用链、检查相关测试用例和最近的git变更")
+# 若涉及多个独立模块，每个模块派出独立 code-explore-expert
 
 ## 步骤 1：收集 Bug 信息（不可绕过）
 向用户确认（如未提供）：

package/dist/src/templates/platforms/claude/commands/consult.md

@@ -3,7 +3,7 @@ description: 架构师直接对话——前端/后端/算法三大领域技术
 name: consult
 model: inherit
 argument-hint: "[--expert frontend-architect|backend-architect|algorithm-expert] [问题描述]"
-tools: ["Read", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce"]
+tools: ["Read", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 > ⚡ 此命令仅用于架构方案讨论与技术咨询，不参与流水线编排。流水线中架构评审由编排者在 Gate B1 自动 spawn 对应架构师 Agent。

package/dist/src/templates/platforms/claude/commands/debug.md

@@ -3,7 +3,7 @@ name: debug
 description: 调试诊断指令——D0收集信息→D1复现用例→D2调试会话→D3交互诊断→D4输出报告，5Gate诊断流程
 model: inherit
 argument-hint: [异常描述或Bug报告]
-tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status"]
+tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # 调试诊断

package/dist/src/templates/platforms/claude/commands/deepinit.md

@@ -3,7 +3,7 @@ description: AI 驱动的分层架构文档初始化——扫描目录树，并
 name: deepinit
 argument-hint: "[--force | --quick]"
 model: inherit
-tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__plugin_oh-my-claudecode_t__deepinit_manifest"]
+tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__plugin_oh-my-claudecode_t__deepinit_manifest", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # DeepInit — AI 驱动的分层架构文档初始化

package/dist/src/templates/platforms/claude/commands/evaluate.md

@@ -3,7 +3,7 @@ name: evaluate
 description: 技术评估指令——E0定义标准→E1生成原型→E2收集指标→E3生成报告，4Gate评估流程
 model: inherit
 argument-hint: [评估对象，如"是否采用Bun替代Node.js"或"React vs Vue技术选型"]
-tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "AskUserQuestion", "Agent", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status"]
+tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "AskUserQuestion", "Agent", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # 技术评估
@@ -30,9 +30,10 @@ Skill("source-driven-development")
 **Gate 检查条件**：评估标准文档已产出，含评估维度+权重+用例清单
 
 ### 步骤
-### 步骤 0：并行信息收集（同一消息同时发出）
+### 步骤 0：并行信息收集（同一消息同时发出，最多 10 个）
 spawn code-explore-expert → 扫描当前项目技术栈，识别与评估对象相关的现有代码、依赖和架构模式
 spawn external-resource-expert → 搜索所有备选方案的最新对比数据（性能基准、社区活跃度、已知问题）
+# 若涉及多个独立模块或备选方案，分别派出独立 Agent
 
 ### 步骤 1：定义评估标准
 1. 明确评估对象和备选方案：
@@ -60,7 +61,7 @@ spawn external-resource-expert → 搜索所有备选方案的最新对比数据
 
 5. 输出 `.jarvis/YYYY-MM-DD/evaluation/evaluation-criteria.md`
 
-### 步骤 2：澄清后靶向探索（评估标准确认后，同一消息同时发出）
+### 步骤 2：澄清后靶向探索（评估标准确认后，同一消息同时发出，最多 10 个）
 
 spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
 - `code-explore-expert`：各备选方案在当前项目中的集成点、受影响模块、迁移成本估算

package/dist/src/templates/platforms/claude/commands/expo.md

@@ -3,7 +3,7 @@ description: Expo 开发生命周期——需求→任务→计划→实现→
 name: expo
 model: inherit
 argument-hint: "[Expo需求描述]"
-tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__file_claim_check", "mcp__jarvis-engine__file_claim_register", "mcp__jarvis-engine__file_claim_release", "mcp__playwright__browser_snapshot", "mcp__playwright__browser_take_screenshot"]
+tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__file_claim_check", "mcp__jarvis-engine__file_claim_register", "mcp__jarvis-engine__file_claim_release", "mcp__playwright__browser_snapshot", "mcp__playwright__browser_take_screenshot", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # Expo 开发生命周期
@@ -95,7 +95,7 @@ Gate C-impl:
 
 ### Gate A：需求澄清
 
-**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出）**
+**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出，最多 10 个）**
 
 spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
 - `code-explore-expert`：项目全景——技术栈、目录结构、已有页面/组件、路由配置、状态管理模式、原生模块配置
@@ -109,7 +109,7 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 - 模糊时加载 `Skill("idea-refine")`
 - 产出需求文档到 `.jarvis/YYYY-MM-DD/requirements/`，标注 `REQ-XXX`
 
-**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出）**
+**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出，最多 10 个）**
 
 spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
 - `code-explore-expert`：需求涉及的特定页面/组件、相关代码路径、依赖链路、Expo 模块依赖
@@ -290,7 +290,7 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 
 | Gate | 可并行操作 |
 |------|-----------|
-| Gate A 通过后 | `code-explore-expert` × N（多目录并行探索）+ `external-resource-expert` × N（多源并行搜索） |
+| Gate A 通过后 | `code-explore-expert` × N（多目录并行探索，最多 10 个）+ `external-resource-expert` × N（多源并行搜索，最多 10 个） |
 | Gate B：DDD→BDD→TDD | `task-design` 按顺序 spawn，不可并行但可连续 |
 | Gate B1 | `mobile-architect`（架构评审，条件性触发） |
 | Gate C 实现 Batch | 按 `parallel_batches` 执行，同 Batch 内并行 |

package/dist/src/templates/platforms/claude/commands/flutter.md

@@ -3,7 +3,7 @@ description: Flutter 开发生命周期——需求→任务→计划→实现
 name: flutter
 model: inherit
 argument-hint: "[Flutter需求描述]"
-tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__file_claim_check", "mcp__jarvis-engine__file_claim_register", "mcp__jarvis-engine__file_claim_release"]
+tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__file_claim_check", "mcp__jarvis-engine__file_claim_register", "mcp__jarvis-engine__file_claim_release", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # Flutter 开发生命周期
@@ -93,7 +93,7 @@ Gate C-impl:
 
 ### Gate A：需求澄清
 
-**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出）**
+**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出，最多 10 个）**
 
 spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
 - `code-explore-expert`：项目全景——技术栈、目录结构、已有 Widget/页面、路由配置、状态管理模式、平台适配代码
@@ -107,7 +107,7 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 - 模糊时加载 `Skill("idea-refine")`
 - 产出需求文档到 `.jarvis/YYYY-MM-DD/requirements/`，标注 `REQ-XXX`
 
-**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出）**
+**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出，最多 10 个）**
 
 spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
 - `code-explore-expert`：需求涉及的特定 Widget/模块、相关代码路径、依赖链路、可复用组件
@@ -278,7 +278,7 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 
 | Gate | 可并行操作 |
 |------|-----------|
-| Gate A 通过后 | `code-explore-expert` × N（多目录并行探索）+ `external-resource-expert` × N（多源并行搜索） |
+| Gate A 通过后 | `code-explore-expert` × N（多目录并行探索，最多 10 个）+ `external-resource-expert` × N（多源并行搜索，最多 10 个） |
 | Gate B：DDD→BDD→TDD | `task-design` 按顺序 spawn，不可并行但可连续 |
 | Gate B1 | `mobile-architect`（架构评审，条件性触发） |
 | Gate C 实现 Batch | 按 `parallel_batches` 执行，同 Batch 内并行 |

package/dist/src/templates/platforms/claude/commands/frontend.md

@@ -3,7 +3,7 @@ description: 前端开发生命周期——需求→任务→计划→实现→
 name: frontend
 model: inherit
 argument-hint: [前端需求描述]
-tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__file_claim_check", "mcp__jarvis-engine__file_claim_register", "mcp__jarvis-engine__file_claim_release", "mcp__playwright__browser_snapshot", "mcp__playwright__browser_take_screenshot"]
+tools: ["Read", "Write", "Edit", "Bash", "Glob", "Grep", "Skill", "AskUserQuestion", "Agent", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__file_claim_check", "mcp__jarvis-engine__file_claim_register", "mcp__jarvis-engine__file_claim_release", "mcp__playwright__browser_snapshot", "mcp__playwright__browser_take_screenshot", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # 前端开发生命周期
@@ -94,7 +94,7 @@ Gate C-impl:
 
 ### Gate A：需求澄清
 
-**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出）**
+**Step 1：澄清前并行探索（需求澄清前，同一消息同时发出，最多 10 个）**
 
 spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
 - `code-explore-expert`：项目全景——技术栈、目录结构、已有页面/组件、路由配置、状态管理模式
@@ -108,7 +108,7 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 - 模糊时加载 `Skill("idea-refine")`
 - 产出需求文档到 `.jarvis/YYYY-MM-DD/requirements/`，标注 `REQ-XXX`
 
-**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出）**
+**Step 3：澄清后靶向探索（需求确认后，同一消息同时发出，最多 10 个）**
 
 spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check("read")`）：
 - `code-explore-expert`：需求涉及的特定页面/组件、相关代码路径、依赖链路、可复用模块
@@ -122,7 +122,7 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 
 | Gate | 可并行操作 |
 |------|-----------|
-| Gate A 通过后 | `code-explore-expert` × N（多目录并行探索，spawn 前 `gate_check("read")`）+ `external-resource-expert` × N（多库并行搜索） |
+| Gate A 通过后 | `code-explore-expert` × N（多目录并行探索，最多 10 个，spawn 前 `gate_check("read")`）+ `external-resource-expert` × N（多库并行搜索，最多 10 个） |
 | Gate B：DDD→BDD→TDD | `task-design` 按顺序 spawn（DDD→BDD→TDD），不可并行但可连续 |
 | Gate B1 | `frontend-architect`（架构评审，条件性触发） |
 | Gate C 实现 Batch | 按 `parallel_batches` 执行，同 Batch 内并行 |
@@ -145,7 +145,7 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 ### Gate C1.5：视觉验证（强制，不可跳过）
 
 **前端任务必须过此门。** 条件：
-- 预览服务器已启动（通过 Playwright MCP 连接浏览器）
+- 预览服务器已启动（dev server 运行中）
 - 修改前/后对比截图已附
 - 响应式三视口截图已附（mobile 375x812 / tablet 768x1024 / desktop 1280x800）
 - 关键样式属性已验证（布局/颜色/字体/间距）
@@ -155,7 +155,19 @@ spawn `code-explore-expert` + `external-resource-expert`（spawn 前 `gate_check
 
 ```
 spawn 前 gate_check({ operation: "spawn_test" }) 确认 Gate C1.5 允许测试
-└── spawn browser-test-expert（浏览器视觉验证）
+└── spawn frontend-debug-expert（主力：Chrome DevTools MCP 深度视觉验证）
+    ├── Chrome DevTools MCP：navigate → snapshot → screenshot（含多视口）
+    ├── 性能追踪（performance_start_trace）+ 控制台/网络诊断
+    ├── 响应式三视口截图（mobile 375x812 / tablet 768x1024 / desktop 1280x800）
+    ├── 修改前/后对比截图
+    ├── 关键样式属性验证（evaluate_script 读取 computed styles）
+    └── 产出视觉验证报告到 .jarvis/YYYY-MM-DD/testing/<topic>-visual-verification.md
+```
+
+**兜底（CDP 不可用或仅需快速页面验证时）**：
+
+```
+spawn browser-test-expert（兜底：agent-browser snapshot + Playwright MCP）
     ├── agent-browser snapshot 精确获取页面结构
     ├── Playwright MCP 执行截图（browser_take_screenshot）+ 快照验证（browser_snapshot）
     ├── 响应式三视口截图（mobile 375x812 / tablet 768x1024 / desktop 1280x800）
@@ -167,8 +179,8 @@ spawn 前 gate_check({ operation: "spawn_test" }) 确认 Gate C1.5 允许测试
 **通过**：进入 Gate C2
 
 **不通过**：
-1. **证据缺失** → 退回 browser-test-expert 补充截图/样式验证数据
-2. **布局问题**（溢出/重叠/错位）→ spawn `frontend-debug-expert`（Chrome DevTools诊断：元素定位/样式追踪/布局分析）定位根因 → spawn 原实现 Agent 修复源文件 → 重新走视觉验证
+1. **证据缺失** → 退回执行 agent 补充截图/样式验证数据
+2. **布局问题**（溢出/重叠/错位）→ frontend-debug-expert 定位根因（Chrome DevTools：元素定位/样式追踪/布局分析）→ spawn 原实现 Agent 修复源文件 → 重新走视觉验证
 3. 修复后重新过 Gate C1.5，最多 2 轮；仍不通过 → 标记 `BLOCKED`，附最新截图证据向用户报告
 
 ### Gate C2：测试

package/dist/src/templates/platforms/claude/commands/hotfix.md

@@ -3,7 +3,7 @@ name: hotfix
 description: 紧急热修复指令——H0紧急声明→H1最小化修复→H2快速验证+回滚→H3事后审计，4Gate紧急流程
 model: inherit
 argument-hint: [故障描述或issue编号]
-tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status"]
+tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "AskUserQuestion", "Agent", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status", "WebFetch", "WebSearch", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # 紧急热修复

package/dist/src/templates/platforms/claude/commands/improve.md

@@ -3,7 +3,7 @@ name: improve
 description: 自主迭代改进——IM0目标定义→IM1研究分析→IM2计划制定→IM3执行验证→IM4评估迭代
 model: inherit
 argument-hint: [改进目标/优化方向]
-tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "Agent", "AskUserQuestion", "WebFetch", "WebSearch", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status", "mcp__jarvis-engine__gate_jump"]
+tools: ["Read", "Glob", "Grep", "Bash", "Write", "Edit", "Skill", "Agent", "AskUserQuestion", "WebFetch", "WebSearch", "TeamCreate", "SendMessage", "TeamDelete", "mcp__jarvis-engine__session_join", "mcp__jarvis-engine__pipeline_guide", "mcp__jarvis-engine__gate_check", "mcp__jarvis-engine__advance_gate", "mcp__jarvis-engine__gate_enforce", "mcp__jarvis-engine__report_status", "mcp__jarvis-engine__gate_jump", "mcp__jarvis-engine__session_context", "mcp__jarvis-engine__jarvis_priority_context"]
 ---
 
 # 自主迭代改进
@@ -33,7 +33,7 @@ Skill("refactoring")
 
 ### 步骤
 
-0. **澄清前并行探索（目标定义前，同一消息同时发出）**——spawn `code-explore-expert` + `external-resource-expert`：
+0. **澄清前并行探索（目标定义前，同一消息同时发出，最多 10 个）**——spawn `code-explore-expert` + `external-resource-expert`：
    - `code-explore-expert`：项目全景——技术栈、目录结构、已识别性能热点/复杂模块/代码质量基线
    - `external-resource-expert`：改进方向相关的行业基准数据、最佳实践、工具链更新
    - 探索结果整理为"改进上下文摘要"，用于后续目标定义