PetGPT: Instructions file for Claude Code

PetGPT

Communication

• 与用户交流时使用中文

Tauri v2 desktop app (Rust backend + React/Vite frontend). An AI pet that participates in group chats (QQ/WeChat via MCP) with a social agent system.

Build & Dev

npm run tauri:dev        # dev mode (macOS/Linux)
npm run tauri:build      # production build
npm run dev              # frontend only (Vite)
npm run lint             # ESLint
npm run typecheck        # tsc over the TypeScript files
npm test                 # node --test src/**/*.test.{mjs,ts}

Architecture

Frontend (src/)

• src/pages/ — React pages. SocialPage.jsx is the main social agent UI.
• src/utils/socialAgent.js — Core social agent loop: Intent (eval/plan) + Reply (send) + Observer + Fetcher layers per target (group/friend).
• src/utils/socialPromptBuilder.js — Builds system prompts for Intent eval and Reply LLM calls.
• src/utils/workspace/socialToolExecutor.js — Executes builtin tools (write_intent_plan, sticker, social_read/edit/write, history, etc.).
• src/utils/llm/ — LLM abstraction. index.js routes to Rust backend (llmCall/llmStream) or JS adapters.
• src/utils/mcp/toolExecutor.js — callLLMWithTools loop: handles tool calls, builtin vs MCP dispatch, stopAfterTool.
• src/utils/tauri.js — Tauri invoke wrappers (workspaceRead, workspaceWrite, llmCall, etc.).

Rust backend (src-tauri/)

• Handles LLM API calls, file I/O, MCP server management, and system tray.

Key Patterns

Social Agent Loop

Each target (group/friend) runs independent loops:

• Fetcher — polls MCP for new messages, wakes Intent via _wake()
• Intent loop — evaluates situation, calls social_edit then write_intent_plan(actions=[...])
• Reply loop — woken by replyWakeFlag, calls send_message via MCP (stopped after first call via stopAfterTool)
• Observer loop — periodic background analysis

Intent State

• Stored per-session in social/{group|friend}/INTENT_{targetId}.md (workspace file)
• Format: 【我刚做了】... 【群里情况】... 【我的判断】...
• LLM writes it directly via social_edit before calling write_intent_plan
• Read back at next eval via readIntentStateFile in system prompt