做 Agent 框架以来的这一年
一些纯古法手写的想法，表达我对不同 agent 场景的一些小观点与看法
Post Archive with Feature Rows
Codex 源码剖析：005. Claude Code、Codex 与行业形态
从实践压力出发，对照 Claude Code 公开源码样本、Codex 和 OpenClaw：先看 provider/platform contract，再看 source execution chain，最后提炼 coding agent harness 的可迁移规则。
Codex 源码剖析：004. App Server、多 Agent 与 Trace
从 App Server 的 JSON-RPC 生命周期出发，看 Codex 如何把 client、thread、turn、approval、多 Agent mailbox 和 rollout trace 放进同一套可路由的 harness。
Codex 源码剖析：003. 上下文、记忆与压缩
从 OpenAI Responses API 和 openai/codex 源码解释 Codex 如何把 base instructions、AGENTS.md、skills、apps、memory、tracked settings diff、compaction、rollout 和 trace 组织成可恢复的模...
Codex 源码剖析：002. 工具、审批与沙箱
从 ToolSpec、ResponseItem、ToolCall、ToolRegistry 到 approval/sandbox，追踪 Codex 如何把模型工具意图变成可审计、可恢复、可拒绝的受控执行。
Placeholder 1
This is some sample content that goes here with Markdown formatting.
Placeholder 3
This is some sample content that goes here with Markdown formatting.
Placeholder Image Left Aligned
This is some sample content that goes here with Markdown formatting. Left aligned with type="left"
Placeholder Image Right Aligned
This is some sample content that goes here with Markdown formatting. Right aligned with type="right"
Placeholder Image Center Aligned
This is some sample content that goes here with Markdown formatting. Centered with type="center"