Copilot Chat Agent / Tool / Prompt 源码细节

0. 真实边界

Copilot Chat 的 VS Code 扩展客户端已经开源，本文的类名、文件路径、工具数量和工具描述来自本次抓取的 main 分支快照。GitHub Copilot 的服务端模型路由、鉴权、配额、远端索引和推理基础设施仍不是该仓库公开内容。

所以本页能深入到 ToolRegistry、ToolsService、ToolCallingLoop 和 prompt-tsx 文件，但不会把服务端模型行为写成源码事实。

1. Chat Participant 与 Intent

Chat surface
  |
  +-- github.copilot.default
  +-- github.copilot.editingSession
  +-- github.copilot.editsAgent
  +-- github.copilot.notebook
  +-- github.copilot.vscode
  +-- github.copilot.terminal
  |
  v
chatParticipants.ts
  |  vscode.chat.createChatParticipant(...)
  v
ChatParticipantRequestHandler
  |  sanitize ignored references
  |  rebuild Conversation
  |  infer location and intent
  v
DefaultIntentRequestHandler
  |  confirmation + hooks + intent invocation
  v
DefaultToolCallingLoop
  |  PromptRenderer + AgentPrompt + available tools
  |  model turns + tool rounds + transcript
  v
final markdown / edits / progress / confirmations

2. 工具注册与调用架构

package.json contributes.languageModelTools
  |
  v
ToolName / ContributedToolName mapping
  |  src/extension/tools/common/toolNames.ts
  v
ToolRegistry.registerTool(...)
  |  concrete tool classes under tools/node and tools/vscode-node
  v
ToolsContribution
  |  vscode.lm.registerTool(getContributedToolName(name), tool)
  v
VS Code Language Model Tool API
  |
  v
ToolCallingLoop offers filtered tools to the selected model
  |
  v
ToolsService.invokeTool(...)
  |  telemetry span + debug args/result + vscode.lm.invokeTool
  v
LanguageModelToolResult2 returns to ChatToolCalls prompt section

3. 37 个工具逐项明细

4. Prompt 细节：公开源码中的提示词装配

AgentPrompt.tsx
  |
  +-- base SystemMessage
  |     +-- expert programming assistant identity
  |     +-- Copilot identity rules
  |     +-- SafetyRules
  |     +-- MemoryInstructionsPrompt
  |
  +-- model-specific customizations
  |     +-- default / OpenAI / Anthropic / Gemini / family H / minimax
  |
  +-- dynamic context
        +-- custom instructions
        +-- mode instructions
        +-- terminal state
        +-- workspace structure
        +-- notebook summary
        +-- memory context
        +-- tool call rounds and results
        +-- summarized conversation history

源码中可以直接看到默认 agent prompt 的核心规则：要求收集上下文再行动、读文件时优先大块读取、可并行调用多个工具、搜索优先使用 search_subagent，多数执行任务优先使用 execution_subagent，编辑时优先选择更精确和更省轮次的编辑工具。

5. ToolCallingLoop 的状态机

Start request
  |
  v
run SessionStart / UserPromptSubmit hooks
  |
  v
getAvailableTools
  |  request tools + tool picker + model overrides + MCP + config
  v
buildPrompt
  |  AgentPrompt / history / memory / tools / edited files
  v
send model turn
  |
  +-- text only ----------------------> Stop hooks -> final result
  |
  +-- tool calls
        |
        v
  execute each tool through ToolsService
        |
        v
  append toolCallRounds and toolCallResults
        |
        +-- hit limit / cancelled ----> cancellation or confirmation
        |
        +-- continue -----------------> next model turn
        |
        +-- autopilot complete -------> task_complete

ToolCallingLoop 还维护 transcript、token usage、debug logger、OpenTelemetry span、subagent trace context 和 stop hook reason。它不是一个简单 while 循环，而是 VS Code UI、模型 endpoint、工具系统和可观测性的协调点。

6. 与主页面的关系

主页面解释 Copilot Chat 的总体 VS Code 插件架构；本页按源码快照补足每个工具和 prompt/agent 编排文件的细节。

返回 Copilot Chat 架构总览