AI coding agents, source code walkthroughs, and reflections on writing code.
agent-desktop（Vercel Labs，Rust，Apache-2.0）深度解析：无障碍树快照与 ref、观察-行动-验证循环、CLI/FFI 到 PlatformAdapter 再到 AX/UIA/AT-SPI 的分层、STALE_REF 与严格重识别、headless 默认与 --headed、内置 skills 与 MCP、以及和截图派 computer-use 的对比。含四张架构与流程配图。
Read articleOrca v0.5.14 到 v0.5.16 的结果协议复盘：D-7 只读工作区暴露假成功，report_blocker 将受阻保存为 blocked 和退出码 5，附带一个问题并支持 resume；writes_applied 汇总子代理写入，通过任务树和逻辑轮次基线处理暂停恢复。含三张飞书画板图、JSONL 示例和真实模型验收范围。
记录 Cleanzo 0.3.0 从网页面板到 mygo 原生 UI 的完整开发历程。包括用 AI Agent 写代码的真实体感、菜单栏面板的架构决策、多显示器定位 bug 的诊断过程、以及性能调优的完整链路。适合对 macOS 原生开发、AI 辅助编程、或 Cleanzo 技术内幕感兴趣的读者。
搭一套能力评测的完整过程：axis A 用 2× / 4× 预算把 16 道 0 分题拆成 time-bound、环境和能力三类；三层评测（L1 契约 / TB2 外部基准 / A–H 能力阶梯）各回答一个问题；场景是数据（scenario.json + fixture + verify + prepare 钩子），执行可换后端（host / docker / harbor），验证器跑完才注入隐藏测试，失败按 harness → environment → product_bug → time_bound → verification_strictness → capability_gap 的顺序归因；以及踩过的几个坑（workflow 异步启动、全过时 L50 不可拟合、docker 只读挂载冲突、goal 只能走 TUI）。附录是八条轴在 v0.5.14 上的实测结果。
一次本机 OCR 开发复盘：用搜索词召回率和字符错误率评测 macOS Vision，处理大图小字的空结果，以解码一次的分块方式缩短长截图耗时，拆分 SQLite 图片与识别文字，用 epoch 丢弃关闭重开后的旧结果，并用针对错误的测试守住行为。
留白 0.6.7 的开发复盘，从摘要搜索的范围缺口出发，讲 Go 内存扫描与索引的取舍、正则热点的定位、三个 Markdown 解析器共享双链样本，以及搜索导航取消后前后端状态不一致的真实 API 回归。
Understand a coding agent through its source code.
Chapters are in Chinese; titles and summaries are available in English.
Understand a production-grade agent through its source code
A source-based guide to the open-source openai/codex repository. Follow a request through the model loop, tool execution, approval and sandbox boundaries, conversation persistence, and recovery. The chapters explain why these mechanisms exist and what breaks when their assumptions do not hold.
Open guideFrom agent loop to a complete runtime
A source-based guide to earendil-works/pi at a pinned version. Explore how provider adapters, an agent loop, tool execution, a session tree, extensions, and a terminal interface fit together. The guide distinguishes behavior implemented in the repository from the separate Harness v2 design.
Open guideFrom pressing Enter to a complete agent platform
A source-based guide to the official DeepSeek Harness repository at a fixed commit. Start with a terminal command or a browser message, follow it through the inbox and agent loop, and inspect how execution events support model context, persistence, and user interfaces.
Open guideTurning ideas into code, and refining them over time.
A coding agent built for DeepSeek, with a Rust runtime and terminal interface.
A desktop browser for ACP agents, with MCP-driven browsing, approvals, manual takeover, and session recovery.
A CLI coding agent with 20+ built-in tools, MCP, and support for multiple models.
An automated BMAD development pipeline, coordinating multiple agents across planning, development, testing, and delivery.
A TypeScript SDK for agent applications, with sessions, tool calls, and model adapters.
A macOS cleanup tool with storage analysis, large-file search, similar-photo detection, and app uninstallation.
