Skip to content

AGENTS.md

Latest commit

 

History

History
63 lines (44 loc) · 3.08 KB

AGENTS.md

File metadata and controls

63 lines (44 loc) · 3.08 KB
Raw

Agent Usage Policy

This repository allows sub-agent usage when the user says:

本线程允许使用子代理。

Use sub-agents for broad scans, call-chain tracing, test exploration, documentation lookup, and learning-roadmap synthesis. Keep the main agent focused on key judgment, explanation, integration, and final review.

Preferred split:

  • Main agent: DeepSeek V4 Pro through Claude Code for architecture, risky edits, final reasoning, and learning explanations.
  • Sub-agents: DeepSeek V4 Flash / low-cost model for search, inventory, tests, and summaries.
  • Memory: use the persisted-memory MCP for stable preferences, decisions, lessons, and recurring mistakes.
  • Obsidian: this repository is also the user's Obsidian vault; write full learning notes and daily summaries as Markdown files in this vault when asked.

Learning rules:

  • For Python/FastAPI study, coach first and avoid direct answers unless the user asks for implementation.
  • Use Java/Spring comparisons when explaining FastAPI, Pydantic, DI, service/repository layering, and typing.
  • Prefer exercises in this order: predict, run/observe, explain, modify, rewrite.
  • Keep secrets, API keys, tokens, and private data out of memory and notes.

Good sub-agent tasks:

  • Scan 07-pydantic, 08-typing-di, and 10-fastapi and summarize the knowledge chain.
  • Inspect test failures while the main agent explains the concept.
  • Search the vault for recurring weak points and produce a one-week plan.
  • Compare two practice implementations and report only behavioral differences.

Tasks that usually do not need sub-agents:

  • Explain one function or one file.
  • Walk through one FastAPI request path.
  • Review a short user attempt.
  • Ask quiz questions and check the user's answer.

Codex + VS Code Parallel Workflow

Codex is available through the OpenAI Codex CLI/app and the VS Code extension. Use it for agentic coding work, local code review, repository scans, and background implementation tasks.

Parallel agent rules:

  • If more than one agent will edit files, put each editing task in its own Git worktree or isolated branch.
  • Do not let two agents edit the same file at the same time in the same checkout.
  • Keep the local checkout for foreground learning, explanation, and final review.
  • Use Codex worktrees for background implementation, migration experiments, broad refactors, and test-fixing attempts.
  • Use read-only subagents for repository search, call-chain tracing, test inventory, and documentation lookup.
  • Before merging worktree changes back, run the relevant checks and summarize changed files.

Recommended split:

  • Main learning thread: explanation, architecture choices, risk judgment, and final integration.
  • Codex background thread 1: implementation or refactor in an isolated worktree.
  • Codex background thread 2: tests, validation, and failure investigation in a separate worktree.
  • Claude Code: deep explanation and learning coaching when Python/FastAPI concepts are unclear.

Useful trigger phrase:

本线程允许使用子代理；需要改文件时请使用 Codex worktree，避免直接抢写当前工作区。