I build local-first, auditable tools for AI coding agents. My open-source work focuses on OpenAI Codex skills, agent memory and context engineering, evidence-backed knowledge pipelines, deterministic evaluation, and reusable developer workflows.
中文：我专注于构建本地优先、可审计的 AI Agent 系统，包括 Codex Skills、长期记忆与上下文工程、基于证据的工程知识提炼，以及可复用的开发工作流。
- Agent skills: reusable capabilities with explicit triggers, bounded procedures, and deterministic validation.
- Agent memory: layered long-term memory, retrieval, write gates, provenance, and governance for coding agents.
- Engineering knowledge systems: pipelines that turn commit-pinned source evidence into reusable patterns.
- Local-first developer tools: inspectable workflows that keep users in control of their data and artifacts.
| Project | Use it when you need | What makes it useful |
|---|---|---|
| Codex Custom Skills | Production-grade skills for OpenAI Codex and compatible coding agents | Trigger contracts, evals, validators, schemas, and portable workflows |
| Codex Long-Term Memory System | A practical architecture for persistent agent memory and context engineering | Layered retrieval, controlled writes, sanitization, governance, and executable checks |
| GitHub Pattern Knowledge | Evidence-backed learning from real GitHub repositories | Commit-pinned provenance, deep finalization gates, and reusable engineering patterns |
| Recode | A public learning archive for building in the AI era | Next.js, TypeScript, MDX, project notes, and an evidence-driven digital garden |
- Looking for reusable Codex skills? Start with Codex Custom Skills.
- Designing long-term memory for AI agents? Read Codex Long-Term Memory System.
- Turning repositories into auditable engineering knowledge? Explore GitHub Pattern Knowledge.
- Treat tests, schemas, provenance, and source traceability as product features.
- Separate canonical source, generated projections, runtime evidence, and private data.
- Make agent triggers, stopping conditions, and verification explicit.
- Prefer local-first systems that remain inspectable, portable, and recoverable.
TypeScript, Python, React, Next.js, FastAPI, GitHub Actions, Markdown knowledge systems, OpenAI Codex, AI agents, context engineering, and developer tooling.