EvoSkills

Self-evolving skill systems for AI coding agents (Claude Code + Codex).

---

What Is This?

A curated collection of operational skills, comparison guides, and source-level analysis for getting the most out of AI coding agents -- particularly Claude Code and OpenAI Codex CLI. Focused on the "command center" use case: one human orchestrating multiple AI agents.

Contents

Codex Skills Handbook

docs/codex-skills.md -- Complete 12-section guide covering:

1. Three usage modes: MCP Tool, tmux interactive, non-interactive exec
2. MCP Tool invocation (recommended): structured, 30-second results
3. Prompt four elements: Goal, Context, Constraints, Done When
4. Codex vs Claude Code capability comparison: when to use which
5. Common stuck states and fixes: auth issues, sandbox failures, rate limits
6. Command center best practices: MCP Tool, tmux session, AGENTS.md configuration
7. Self-evolution loop: review -> fix -> test -> record skills -> update rules
8. Real-world practice records: 16 rounds of A-grade code audit output
9. MCO multi-CLI orchestration: parallel multi-model review
10. Codex Plugin deep dive: setup, commands, review gate, internal skills
11. Plugin internals: architecture, design principles, when to use what

Codex Source Code Study

docs/codex-source-study.md -- Analysis of the Codex CLI Rust codebase:

1. Architecture: 80+ crate workspace, three runtime modes (TUI, Exec, MCP Server)
2. Sandbox system: Why bwrap fails on certain servers, three-layer sandbox (bubblewrap + seccomp + Landlock), troubleshooting guide
3. MCP Server integration: As client (connecting to external servers) and as server (exposing codex/codex-reply tools)
4. Approval system: Four policies, granular approvals, guardian subagent
5. Network proxy: Built-in HTTP + SOCKS5 proxy, domain allowlist/blocklist
6. config.toml reference: Complete configuration with profiles, model providers, permissions
7. Session management: Resume, fork, ephemeral modes
8. Operator reference: Quick-reference tables for exec parameters, review usage, sandbox troubleshooting

Key Concepts

Task Allocation Principle

• Give to Codex: Code review, independent bug fixes, test writing, large-scale refactoring, parallel tasks
• Give to Claude Code: Multi-step coordinated tasks, browser automation, MCP tool chains, video production
• Both together: Claude Code plans + Codex executes; Claude Code reviews Codex output

Self-Evolution Loop

Round N:
1. Codex reviews code -> finds issues
2. Codex exec fixes issues
3. Run tests to verify
4. Record Skills (what worked / what failed)
5. Update AGENTS.md rules
6. -> Round N+1

Prompt Engineering (Four Elements)

Goal: What to change, why
Context: Related files, error messages, screenshots
Constraints: Standards, architecture requirements, security red lines
Done When: Completion criteria (tests pass / behavior changes / bug doesn't reproduce)

Sandbox Troubleshooting Quick Reference

Symptom	Cause	Fix
Can't mount proc	Container lacks proc mount permission	Codex auto-fallbacks, or add --no-proc
bwrap not found	bubblewrap not in PATH	apt install bubblewrap (vendored fallback exists)
Operation not permitted on namespace	Kernel blocks unprivileged user namespaces	sysctl kernel.unprivileged_userns_clone=1
All sandbox fails in container	Docker restrictions	Use --sandbox danger-full-access

sleep2agi — 专注 AI Agents 编排、群体智能、视频生成，帮助普通人跨越到 AGI。

License

MIT