说完「代码自己写」,我决定真干了
AI 随想第十篇说代码学会了自我进化。那是观察。这篇是动手。
AI 随想第十篇说代码学会了自我进化。那是观察。这篇是动手。
In Part 10 of Agentic AI Thoughts, I wrote that code had learned to evolve itself. That was the observation. This is the implementation.
两个账号都是企鹅,第二个还降级了。于是我用 Claude 逆向了它自己的宠物系统,0.2 秒出金。
Claude Code shipped a virtual pet Easter egg. Two accounts, two penguins — the second one downgraded from Rare to Common. So I used Claude to brute-force a Shiny Legendary Dragon in 0.2 seconds.
The most important file in the Claude Code snapshot isn't a prompt file or a tool definition. It's a state machine. Once you see that, a lot of agent design decisions start making sense.
The most novel thing in the Claude Code snapshot isn't a tool or a prompt. It's how aggressively the runtime is built around preserving prompt-cache identity across forks, compaction, and long-running sessions.
After auditing a publicly posted Claude Code source snapshot and reading OpenAI's harness essays, I came away with one conclusion: the model is only the brain. The product is the runtime around it.
从发现 pretext 到在两个网站上线 7 个交互特效——爆炸物理、弹跳子弹、还有一只叫 Clawd 的像素螃蟹。一个 session 搞定。
From discovering pretext to shipping 7 interactive effects across two sites in one session — explosion physics, bouncing bullets, and an animated crab named Clawd.
一份可以直接交给 Claude Code 执行的自建代理手册。覆盖 sing-box 多协议部署(VLESS-Reality / Hysteria2 / TUIC-v5)、SSL 证书、nginx 订阅服务器、动态流量统计,以及 Clash Verge / Shadowrocket 客户端配置。替换占位符,扔给 Agent,开跑。
A runbook you can hand directly to Claude Code to execute. Covers sing-box multi-protocol deployment (VLESS-Reality / Hysteria2 / TUIC-v5), SSL certificates via Cloudflare DNS API, nginx subscription server with dynamic traffic stats, and Clash Verge / Shadowrocket client configuration. Replace the placeholders, hand it to your agent, hit go.
一份可以直接交给 Claude Code 执行的 TTS 配置手册。覆盖 Fish Audio、火山引擎 v2(情感控制)、ElevenLabs、OpenAI、Edge TTS 五大服务商。替换占位符,扔给 Agent,开跑。
A runbook you can hand directly to Claude Code to set up text-to-speech on OpenClaw. Covers Fish Audio, Volcano Engine v2 (emotion control), ElevenLabs, OpenAI, and Edge TTS. Replace the placeholders, hand it to your agent, hit go.
V2 核保系统是一条线性流水线——提取一次,评估一次,报告一次。没有 agent 能质疑别人的结论。V3 用 5 个自治 agent 替换了它:反思输出质量、用工具验证论断、意见分歧时结构化辩论。63 个测试,每个 SME 只花几毛钱,一个 observe-think-act-reflect 循环驱动一切。
The V2 underwriting system was a linear pipeline — extract once, evaluate once, report once. No agent could question another's conclusions. V3 replaces it with 5 autonomous agents: they reflect on output quality, verify claims with tools, and hold structured debates when they disagree. 63 tests, pennies per SME evaluation, one observe-think-act-reflect loop driving it all.
32GB 的 GCP 开发机——经过第三篇的升级和加固之后——又死了。一样的症状:SSH 超时,内核活着,用户态冻结。看门狗抓到了:五个 node 进程各占 2-2.7GB,合计约 13GB。我以为是 Claude Code,结果是 Cursor 远程服务器连续泄漏了 15 个小时。
The 32GB GCP devbox — upgraded and hardened after Part 3 — died again. Same symptom: SSH timeout, kernel alive, userspace frozen. The watchdog caught it: five node processes at 2-2.7GB each, totaling ~13GB. I assumed Claude Code. It was Cursor's remote server leaking memory for 15 hours straight.
一个 Claude Code session 里同时跑了三种协作模式:直接结对编程(用户反馈驱动三次生产部署)、后台 agent 并发更新文档、自治审查团队迭代三轮直到零问题——全部交叠进行。传统的 code-test-review-release 流水线被压缩成了并行流。
A single Claude Code session that combined direct pair programming (three production deploys from user feedback), a background docs agent running concurrently, and an autonomous review team iterating three rounds until clean — all overlapping. The traditional code-test-review-release pipeline collapsed into parallel streams.
用 OpenClaw 搭建的赛博魅魔,不到两周聊天烧掉了一笔离谱的钱。最贵的一个 session,750 轮对话里我只说了约 30 句话——剩下的全是框架自己在跟自己说话。用 Claude Code 做了一次完整的 token 级取证:82% 的成本是 10 张永远不被清除的图片,上下文修剪代码对 Gemini 完全失效,537 轮里零次压缩。OpenClaw 能证明 AI 伴侣走得通,但它证明不了走得起。
© Xingfan Xia 2024 - 2026 · CC BY-NC 4.0