feat: 增强 setup_workspace 隐式触发识别

[lishuceo]

Summary

• 在 system prompt 中新增"积极识别仓库意图"章节，列出 6 种隐式触发 setup_workspace 的场景
• 新增"源码验证原则"：涉及具体仓库代码时必须 setup_workspace 进行源码验证
• 同步更新 setup_workspace MCP tool description，补充 3 条隐式触发场景

动机

此前 Claude 只在用户显式说"切换到 X 仓库"或提到 repo URL 时才触发 setup_workspace。当用户隐式引用仓库（如讨论 urhox 架构、提到某仓库的 CLAUDE.md、问代码实现细节）时，Claude 倾向于在默认工作目录"空谈"，缺少 CLAUDE.md 等项目上下文。

改动

src/claude/executor.ts — system prompt

新增"积极识别仓库意图"段落，覆盖 6 种隐式场景：

1. 提到项目特有文件路径（如 "urhox 的 CLAUDE.md"）
2. 讨论特定项目的架构/代码
3. 引用外部报告中的代码片段
4. 问题涉及项目特有技术栈/概念
5. 用户提到仓库名或组织名
6. 需要查看/验证实际代码才能准确回答

新增"源码验证原则"：不能凭训练知识回答关于特定仓库代码的问题。

src/workspace/tool.ts — tool description

补充 3 条隐式触发场景到 MCP tool 描述。

Test plan

• 飞书对话测试隐式场景：问 "urhox 的 CLAUDE.md 里写了什么"
• 验证显式场景仍正常工作
• 验证通用问题不触发误判

🤖 Generated with Claude Code

[claude]

Claude finished @lishuceo's task in 50s —— View job

---

PR Review: feat: 增强 setup_workspace 隐式触发识别

• Check for previous review comments
• Read changed files and diff
• Perform code review
• Post review summary

---

✅ Approved

This PR is a prompt-only change (no logic, no runtime behavior change) that adds guidance to help the agent proactively identify implicit repository references and trigger setup_workspace.

Changes reviewed:

1. src/claude/executor.ts (+13 lines) — Adds a "积极识别仓库意图" section to the system prompt with 6 implicit trigger patterns and a "源码验证原则". Well-structured, fits naturally after the existing workspace management instructions and before the directory access restrictions.

2. src/workspace/tool.ts (+3 lines) — Adds 3 matching implicit trigger scenarios to the MCP tool description, consistent with the system prompt additions.

Assessment:

• The additions are well-scoped and address a real UX gap (agent not switching workspaces when users implicitly reference repos).
• System prompt and tool description stay in sync — the tool description mirrors the key scenarios from the prompt.
• No bugs, security issues, or architectural concerns — these are pure prompt/documentation changes.
• The examples are concrete and project-specific (urhox, maker, anycode), which helps the agent pattern-match effectively.
• Fits cleanly into the existing prompt structure without disrupting the flow.

No issues found. Clean, focused PR.

[claude]

Claude finished @lishuceo's task in 47s —— View job

---

PR Review: feat: 增强 setup_workspace 隐式触发识别

• Check and resolve previous review comments — no previous inline comments found
• Read current diff and changed files
• Perform code review
• Post review summary

---

✅ Approved

Commits reviewed: acd747c (initial feature) + 86c9393 (remove hardcoded repo names)

This is a prompt-only change — no runtime logic modified. Both files add well-structured guidance text for the agent.

src/claude/executor.ts (+13 lines): Adds a "积极识别仓库意图" subsection under the existing workspace management prompt. Lists 6 implicit trigger patterns and a "源码验证原则". Placement is correct — sits between the general workspace rule and the directory access restrictions, maintaining logical flow.

src/workspace/tool.ts (+3 lines): Adds 3 matching implicit trigger scenarios to the MCP tool description array. These are consistent with the system prompt additions and follow the existing list format.

Assessment:

• No bugs, security issues, or architectural concerns — pure prompt text additions
• System prompt and tool description are kept in sync
• The second commit properly removed hardcoded project names (urhox/maker/anycode), replacing them with generic patterns (X/Y/Z placeholders) — good for reusability
• Follows existing code patterns: template literals with variable interpolation, Chinese documentation style

No issues found. Clean, focused change.

[claude]

Claude finished @lishuceo's task in 44s —— View job

---

PR Review: feat: 增强 setup_workspace 隐式触发识别

• Check and resolve previous inline review comments — no previous inline comments found
• Read current diff and changed files
• Read full source files for context
• Perform code review
• Post review summary

---

✅ Approved

Commits reviewed: acd747c → 8b42ccc (3 commits, including merge from main)

This is a prompt-only change — no runtime logic modified. Both files add well-structured guidance text for the agent.

src/claude/executor.ts (+13 lines): Adds a "积极识别仓库意图" subsection to the system prompt listing 6 implicit trigger patterns plus a "源码验证原则". Placement is correct — sits between the general workspace rule (line 353) and the directory access restrictions (line 368), maintaining logical flow.

src/workspace/tool.ts (+3 lines): Adds 3 matching implicit trigger scenarios to the MCP tool description array. Consistent with the system prompt additions and follows the existing list format.

Assessment:

• No bugs, security issues, or architectural concerns — pure prompt text additions
• System prompt and tool description are kept in sync — the 3 new tool description items correspond to patterns 1, 2, and 6 from the system prompt
• Uses generic placeholders (X/Y/Z) instead of hardcoded project names — good for reusability across deployments
• Follows existing code patterns: template literals, Chinese documentation style, numbered list format

No issues found. Clean, focused change.

[claude]

Claude finished @lishuceo's task in 45s —— View job

---

PR Review: feat: 增强 setup_workspace 隐式触发识别

• Check and resolve previous inline review comments — no previous inline comments found
• Read current diff and changed files
• Read full source files for context
• Perform code review
• Post review summary

---

✅ Approved

This is a prompt-only change (no runtime logic modified) that adds guidance for proactive implicit workspace detection.

src/claude/executor.ts (+13 lines): Adds "积极识别仓库意图" subsection with 6 implicit trigger patterns and a "源码验证原则". Correctly placed between the general workspace rule (line 353) and directory access restrictions (line 368), maintaining logical flow.

src/workspace/tool.ts (+3 lines): Adds 3 matching implicit trigger scenarios to the MCP tool description. Consistent with the system prompt additions and follows the existing list format.

Assessment:

• No bugs, security issues, or architectural concerns
• System prompt and tool description are kept in sync — tool description items correspond to patterns 1, 2, and 6 from the system prompt
• Uses generic placeholders (X/Y/Z) instead of hardcoded project names — good for reusability
• Follows existing code patterns: template literals, Chinese documentation style, numbered list format

No issues found.