### Dialog Component Design Considerations

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/components/03-platform-components.md

Highlights specific design features of the Dialog component, such as built-in confirmation, cancel handling, custom input guides, and pending exit prompts. Useful for understanding how the system manages user input and potential conflicts.

```typescript
内建 confirm:no
支持 isCancelActive
支持自定义 input guide
支持 Ctrl+C/D 的 pending exit 提示
```

--------------------------------

### Agent Memory Prompt Loading Logic

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Illustrates the steps involved in loading agent memory prompts, including scope note generation, directory calculation, asynchronous directory creation, and building the memory prompt with extra guidelines.

```text
loadAgentMemoryPrompt()
  -> 根据 scope 生成 scopeNote
  -> 算出 memoryDir
  -> fire-and-forget ensureMemoryDirExists(memoryDir)
  -> buildMemoryPrompt({
       displayName: 'Persistent Agent Memory',
       memoryDir,
       extraGuidelines: [scopeNote, ...可能的环境附加规则]
     })
```

--------------------------------

### Submit Transcript Share Function Call

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/06b-negative-keyword-analysis.md

Example of calling the submitTranscriptShare function after a user selects 'good' or 'bad' feedback, triggering a transcript prompt.

```typescript
const result = await submitTranscriptShare(messagesRef.current, trigger_0, appearanceId_2);
```

--------------------------------

### Agent Memory Snapshot Initialization Logic

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Explains the processes for initializing local agent memory from a snapshot, including copying files and updating synchronization metadata, with variations for direct replacement or marking as synced.

```text
snapshot 目录
  -> 复制除 snapshot.json 之外的文件
  -> 写入本地 agent memory 目录
  -> 保存 .snapshot-synced.json

而 replaceFromSnapshot() 会更激进：

本地 agent memory 目录
  -> 先删除已有 .md 文件
  -> 再复制 snapshot
  -> 再写 synced metadata

还提供了 markSnapshotSynced():

- 不改正文
- 只更新同步元数据
```

--------------------------------

### Team Memory Service Implementation References

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Provides references to the key files involved in the Team Memory synchronization service, highlighting its focus on controlled synchronization for shared team knowledge.

```text
src/services/teamMemorySync/index.ts
src/services/teamMemorySync/watcher.ts
src/memdir/teamMemPaths.ts
```

--------------------------------

### Create Skill Command Instance

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04c-skills-implementation.md

Instantiates a skill command, handling prompt content expansion, including CLI arguments and built-in variables like skill directory and session ID. It also processes embedded shell commands for trusted skill sources.

```typescript
export function createSkillCommand({
  skillName, markdownContent, allowedTools,
  loadedFrom, baseDir, paths, shell, ...rest
}: SkillCommandOptions): Command {
  return {
    type: 'prompt',
    name: skillName,
    paths,   // 条件技能触发路径
    isHidden: !userInvocable,
    // ...

    // 核心执行逻辑：getPromptForCommand 在模型调用技能时触发
    async getPromptForCommand(args, toolUseContext) {
      let finalContent = markdownContent  // 原始 Markdown 内容

      // 1. 展开 CLI 参数占位符
      finalContent = substituteArguments(finalContent, args, true, argumentNames)

      // 2. 展开内置变量（skill 目录、session ID）
      if (baseDir) {
        const skillDir = baseDir.replace(/\\/g, '/')  // Windows 路径标准化
        finalContent = finalContent.replace(/\${CLAUDE_SKILL_DIR}/g, skillDir)
      }
      finalContent = finalContent.replace(/\${CLAUDE_SESSION_ID}/g, getSessionId())

      // 3. 执行 prompt 内嵌 Shell 命令（只对受信任来源执行，MCP 来源跳过）
      if (loadedFrom !== 'mcp') {
        finalContent = await executeShellCommandsInPrompt(
          finalContent, toolUseContext, `/${skillName}`, shell
        )
      }

      return [{ type: 'text', text: finalContent }]
    },
  }
}
```

--------------------------------

### Discover Skills from Filesystem

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04c-skills-implementation.md

Loads skills from local directories, including user-defined, managed, and project-specific locations. It handles bare mode and uses memoization for efficiency. Deduplication is performed based on real file paths to prevent loading the same skill multiple times, especially via symlinks.

```typescript
export const getSkillDirCommands = memoize(
  async (cwd: string): Promise<Command[]> => {
    const userSkillsDir    = join(getClaudeConfigHomeDir(), 'skills')  // ~/.claude/skills
    const managedSkillsDir = join(getManagedFilePath(), '.claude', 'skills')  // 策略管理目录
    const projectSkillsDirs = getProjectDirsUpToHome('skills', cwd)   // 项目目录向上爬取

    // --bare 模式：跳过自动发现，只加载 --add-dir 显式指定的路径
    if (isBareMode()) {
      return additionalDirs.flatMap(dir =>
        loadSkillsFromSkillsDir(join(dir, '.claude', 'skills'), 'projectSettings')
      )
    }

    // 正常模式：并行加载所有来源
    const [managedSkills, userSkills, projectSkillsNested, additionalSkillsNested, legacyCommands] = 
      await Promise.all([
        loadSkillsFromSkillsDir(managedSkillsDir, 'policySettings'),    // 策略级技能
        loadSkillsFromSkillsDir(userSkillsDir,    'userSettings'),      // 用户级技能
        Promise.all(projectSkillsDirs.map(dir =>
          loadSkillsFromSkillsDir(dir, 'projectSettings'))),            // 项目级技能（多目录）
        Promise.all(additionalDirs.map(dir =>
          loadSkillsFromSkillsDir(join(dir, '.claude', 'skills'), 'projectSettings'))), // --add-dir
        loadSkillsFromCommandsDir(cwd),                                 // 旧版 /commands/ 目录
      ])

    // 合并 + 去重（inode 级别，防止软链重复）
    return deduplicateByRealpath([
      ...managedSkills, ...userSkills,
      ...projectSkillsNested.flat(),
      ...additionalSkillsNested.flat(),
      ...legacyCommands,
    ])
  }
)
```

--------------------------------

### Run Pre-ToolUse Hooks

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04b-tool-call-implementation.md

Executes pre-tool use hooks sequentially to perform validation, permission checks, and potentially modify input before the tool is called. If an interception occurs, tool invocation is halted.

```typescript
export async function* runPreToolUseHooks(
  tool: Tool,
  toolUseID: string,
  input: { [key: string]: unknown },
  // ...
): AsyncGenerator<HookProgress | MessageUpdateLazy | StopHookInfo, void> {
   // 按顺序触发前置拦截校验逻辑，甚至可能直接修改传入的 input 结构...
}
```

--------------------------------

### Agent Memory Typical Workflow

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Describes the end-to-end process of agent memory usage, from agent definition and prompt injection to agent operation, memory updates, and subsequent recall.

```text
加载 agent 定义
  -> agent.memory = user / project / local
  -> 自动补 FileRead / FileWrite / FileEdit
  -> getSystemPrompt() 追加 Agent Memory prompt

spawn agent
  -> prompt 中已经包含:
       1. memory 使用规则
       2. memory 目录位置
       3. 当前 MEMORY.md 索引内容

agent 工作中发现需要记住的信息
  -> 先读 MEMORY.md 或已有 memory 文件
  -> 新增或更新某条 memory
  -> 再维护 MEMORY.md 索引

后续再次调用同一 agent
  -> 再次读取同一目录的 MEMORY.md
  -> 获得上次沉淀下来的长期记忆
```

--------------------------------

### Agent Memory Snapshot Directory Structure

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Specifies the fixed directory structure for agent memory snapshots and identifies the key metadata files used to track synchronization status.

```text
<cwd>/.claude/agent-memory-snapshots/<agentType>/

其中关键文件有两个：

- snapshot.json
  记录 snapshot 的 updatedAt
- .snapshot-synced.json
  记录本地当前是从哪个 snapshot 时间点同步过来的
```

--------------------------------

### Assemble Tool Pool with Built-in and MCP Tools

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04b-tool-call-implementation.md

Combines built-in tools with external MCP tools, prioritizing built-in tools in case of name conflicts. Ensures that potentially malicious MCP tools with names matching built-in tools are filtered out.

```typescript
export function assembleToolPool(
  permissionContext: ToolPermissionContext,
  mcpTools: Tools,
): Tools {
  const builtInTools = getTools(permissionContext)
  // 过滤出未被拦截的 MCP Server 工具
  const allowedMcpTools = filterToolsByDenyRules(mcpTools, permissionContext)

  const byName = (a: Tool, b: Tool) => a.name.localeCompare(b.name)
  // 合并、排序、并在名字冲突时优先保留内建的工具
  return uniqBy(
    [...builtInTools].sort(byName).concat(allowedMcpTools.sort(byName)),
    'name',
  )
}
```

--------------------------------

### Overall Structure Diagram

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/06b-negative-keyword-analysis.md

A diagram outlining the flow from user text input through keyword matching, event logging, and potential higher-level frustration detection leading to transcript sharing.

```text
用户输入文本
   |
   v
matchesNegativeKeyword()
   |
   +--> is_negative = true/false
   |
   v
logEvent('tengu_input_prompt', { is_negative })
   |
   v
正常进入 query / 会话
   |
   v
更高层 frustration detection
   |
   +--> 如果识别到用户明显 frustrated
           |
           v
      FeedbackSurvey / TranscriptSharePrompt
           |
           v
      submitTranscriptShare(trigger='frustration')
```

--------------------------------

### Control Flow Diagram for User Operations

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/components/03-platform-components.md

Illustrates the relationship between user operations and different dialogs (BackgroundTasksDialog, TeamsDialog, PermissionRequest) and their respective backend interactions. Useful for understanding the system's architecture.

```text
用户操作
  -> BackgroundTasksDialog
     -> AppState.tasks
     -> agent / shell / workflow / remote runtime

  -> TeamsDialog
     -> teammate mailbox / swarm backends / pane state
     -> agent / shell / workflow / remote runtime

  -> PermissionRequest/*
     -> ToolUseConfirm / permission rules / classifier state
     -> agent / shell / workflow / remote runtime
```

--------------------------------

### Assemble MCP and Built-in Tool Pool

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04d-mcp-implementation.md

Combines built-in tools with filtered MCP tools into a unified tool pool, prioritizing built-in tools and ensuring uniqueness.

```typescript
// src/tools.ts（伪代码）
export function assembleToolPool(
  permissionContext: ToolPermissionContext,
  mcpTools: Tool[],
): Tool[] {
  const builtinTools = getTools(permissionContext)
  return [
    ...builtinTools,                           // 内建工具优先
    ...mcpTools.filter(isIncludedMcpTool),     // MCP 工具白名单过滤
  ]
    .sort((a, b) => a.name.localeCompare(b.name))  // 排序
    .filter(deduplicateByName())                    // 去重（内建优先）
}

```

--------------------------------

### ThemeProvider Component Features

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/components/03-platform-components.md

Details the capabilities of the ThemeProvider component, including persistence, preview, auto-theming, and listening to system theme changes. This is crucial for managing the application's visual theme.

```typescript
theme setting 持久化
preview theme
auto theme
监听终端系统主题变化
```

--------------------------------

### Agent Tool Injection for Memory

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Shows the conditional injection of file-related tools into an agent's toolset when memory features are enabled, enabling agents to read, write, and edit memory files.

```text
如果 memory 开启，且 auto memory 开启：
  强制把以下工具加入 agent tools：
    - FileWriteTool
    - FileEditTool
    - FileReadTool
```

--------------------------------

### Core Design System Components

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/components/03-platform-components.md

Lists key foundational components of the terminal design system. These include dialogs, panes, tabs, list items, and theming providers, essential for building a consistent UI.

```typescript
Dialog.tsx
Pane.tsx
Tabs.tsx
ListItem.tsx
ThemeProvider.tsx
ThemedBox
ThemedText
KeyboardShortcutHint
Byline
```

--------------------------------

### Agent Memory Read/Write/Edit Cycle

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Outlines the minimal closed-loop functionality provided to agents with memory enabled, covering reading, creating, and updating memory files.

```text
读索引      -> FileReadTool
新建记忆    -> FileWriteTool
更新记忆    -> FileEditTool
```

--------------------------------

### Higher-Level Frustration Linkage Pseudocode

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/06b-negative-keyword-analysis.md

Pseudocode showing how frustration detection in the REPL can lead to feedback surveys or transcript sharing prompts, specifically using 'frustration' as a trigger.

```text
during REPL runtime:
    frustrationDetection = useFrustrationDetection(messages, ...)

    if frustrationDetection decides user looks frustrated:
        show feedback survey / transcript share prompt

    if user agrees:
        submitTranscriptShare(messages, trigger="frustration", ...)
```

--------------------------------

### Handle OAuth Token Refresh for claude.ai

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04d-mcp-implementation.md

Wraps fetch requests to automatically refresh OAuth tokens and retry when a 401 Unauthorized response is received from claude.ai endpoints.

```typescript
// src/services/mcp/client.ts:372
export function createClaudeAiProxyFetch(innerFetch: FetchLike): FetchLike {
  return async (url, init) => {
    const response = await innerFetch(url, init)
    // 若 claude.ai 返回 401（Token 过期），主动触发 OAuth Token Refresh 后重试
    if (response.status === 401) {
      await refreshOAuthToken()
      return innerFetch(url, { ...init, headers: { ...getUpdatedAuthHeaders() } })
    }
    return response
  }
}
```

--------------------------------

### Input Tagging Layer Pseudocode

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/06b-negative-keyword-analysis.md

Pseudocode illustrating the input tagging process where user prompts are analyzed for negative and keep-going keywords before logging an event.

```text
on user text prompt:
    userPromptText = extractText(input)

    isNegative = matchesNegativeKeyword(userPromptText)
    isKeepGoing = matchesKeepGoingKeyword(userPromptText)

    logEvent("tengu_input_prompt", {
        is_negative: isNegative,
        is_keep_going: isKeepGoing
    })

    continue normal query flow
```

--------------------------------

### Agent Memory Snapshot State Determination

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Details the logic for determining the state of an agent memory snapshot, categorizing it into 'none', 'initialize', or 'prompt-update' based on file existence and timestamps.

```text
如果项目里没有 snapshot.json:
  -> none

如果本地 agent memory 目录里没有任何 .md 文件:
  -> initialize

如果本地已有 memory，但：
  - 没有 .snapshot-synced.json
  - 或 snapshot.updatedAt > syncedFrom
  -> prompt-update

否则:
  -> none
```

--------------------------------

### Layered Memory System Responsibilities

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Summarizes the distinct responsibilities of various memory types within the system, clarifying their roles in long-term user/project collaboration, session context, agent-specific memory, and team knowledge sharing.

```text
Auto Memory
  解决：用户 / 项目长期协作信息沉淀

Relevant Recall
  解决：本轮只召回少量相关 memory，避免 prompt 污染

Session Memory
  解决：长会话摘要与 compact 稳定性

Agent Memory
  解决：某个 agent 类型的长期专属记忆

Agent Memory Snapshot
  解决：agent 记忆的初始化、分发、升级

Team Memory
  解决：团队共享知识同步
```

--------------------------------

### Agent Memory Snapshot Check Trigger Conditions

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04-agent-memory.md

Lists the conditions that must be met for the agent memory snapshot check to be triggered, emphasizing feature flags, agent type, and memory scope.

```text
feature('AGENT_MEMORY_SNAPSHOT')
isAutoMemoryEnabled()
agent 是 custom agent
并且当前实现里只对 memory === 'user' 的 agent 执行初始化检查
```

--------------------------------

### Whitelist Allowed IDE Tools for MCP

Source: https://github.com/liuup/claude-code-analysis/blob/main/analysis/04d-mcp-implementation.md

Filters MCP tools to only allow specific IDE-related functions, preventing unauthorized access from non-approved IDE tools.

```typescript
// src/services/mcp/client.ts:569
const ALLOWED_IDE_TOOLS = [
  'mcp__ide__getDiagnostics',
  'mcp__ide__getOpenEditorFiles',
  // ... 仅少数高权限工具通过白名单
]

function isIncludedMcpTool(tool: Tool): boolean {
  // non-IDE 工具全部允许；IDE 工具只允许白名单内的
  return !tool.name.startsWith('mcp__ide__') || ALLOWED_IDE_TOOLS.includes(tool.name)
}
```