Now vibe coding, so learning hammer FE ?https://channel.meo-ai.comhttps://channel.meo-ai.com/posts/663https://channel.meo-ai.com/posts/663Wed, 17 Jun 2026 01:11:36 GMT<span class="post-title-source">《永不浪费一个Token：AI推理流的可恢复性设计》</span><br /><br />标签：<a href="/search/result?q=%23%E5%90%8E%E7%AB%AF" title="#后端">#后端</a> <a href="/search/result?q=%23AI%E6%8E%A8%E7%90%86" title="#AI推理">#AI推理</a> <a href="/search/result?q=%23Cloudflare" title="#Cloudflare">#Cloudflare</a> <a href="/search/result?q=%23DurableObject" title="#DurableObject">#DurableObject</a> <a href="/search/result?q=%23%E6%B5%81%E5%BC%8F%E6%81%A2%E5%A4%8D" title="#流式恢复">#流式恢复</a> <a href="/search/result?q=%23Token%E8%AE%A1%E8%B4%B9%E4%BC%98%E5%8C%96" title="#Token计费优化">#Token计费优化</a><br /><br />总结：<br />本文探讨了AI Agent在推理过程中因进程崩溃或重新部署导致流式连接中断、已付费Token被重复计费的痛点。核心方案是在Agent与LLM提供商之间引入一个持久化缓冲区（Durable Object），将流式数据实时写入SQLite，使连接与Agent进程解耦。Agent重启后可通过游标恢复，避免重复调用和重复付费。该机制同时解决了浏览器断网重连和进程崩溃恢复两种场景，且即将集成到Cloudflare AI Gateway中，实现一键开启的持久化推理。<br /><br />文章要点：<br />1. 隐藏的成本黑洞：Agent进程崩溃或重新部署时，正在进行的LLM流式请求会中断，已付费的生成Token全部丢失，恢复后必须重新调用并重新付费，且旗舰模型（如GPT-5.5）的重复成本是小模型的15倍。<br />2. 解耦连接与进程：将LLM提供商连接移出Agent进程，部署为独立的持久化缓冲区（Durable Object），在后台持续将流式数据写入SQLite。即使Agent进程被替换，缓冲区仍继续接收数据，Token不再浪费。<br />3. 一份日志，两种用途：持久化缓冲区存储的原始字节流既支持浏览器断网重连（追赶实时游标），也支持进程崩溃恢复（回放已存储的片段）。两者共用同一套SQLite日志机制，仅需判断生产者是否仍在运行。<br />4. 零自定义解析：存储原始字节而非解析后的SSE事件，恢复时通过各提供商的官方插件（如OpenAI、Anthropic）重新解析，避免维护多套格式解析器，且能自动适配格式变更。<br />5. 行业现状对比：OpenAI Responses API已原生支持后台模式恢复；Anthropic和Google Gemini均不支持服务端续流，只能重新提示并重复计费；Vercel的<code>resumable-stream</code>为应用层方案，无法承受部署替换。<br />6. 即将落地AI Gateway：<mark class="highlight">Cloudflare</mark> AI Gateway正在集成该持久化恢复能力，未来只需在Agent基类中设置<code>durableBuffer = true</code>即可一键开启，实现跨提供商的Token零浪费。<br /><br />URL：<a href="https://sunilpai.dev/posts/never-waste-a-token/" target="_blank" rel="noopener" title="https://sunilpai.dev/posts/never-waste-a-token/">https://sunilpai.dev/posts/never-waste-a-token/</a><a class="tgme_widget_message_link_preview" href="https://sunilpai.dev/posts/never-waste-a-token" target="_blank" rel="noopener" title="durable inference: resumable streams, crash recovery, and why the LLM request shouldn't die with your process."> <div class="link_preview_site_name accent_color">Solving the decision problem</div> <img class="link_preview_image" alt="never waste a token" src="/static/https://cdn4.telesco.pe/file/uHdH2ydGrBpImS6yyNDlEDCpezqDJHVi4VJ787KA-5JzVHqpKDz2skyD1IkcJl-JWsJXS_cL7Zz5eZcup8qUKZL5AS2eud86C9zlfexF4zE4pKORoLpxrwFniNgkRm274h_6ALna2B8dBYm-1UZzlEERSxMj7-Vd46jSd_hDv5FwfMuCiylul6PiW4JVwmI3P_nxlaQ7ODor-w0loTOvNlHj9lAIcyLwJsYhEFnVijYrzoQTy0roEE84c0H55IEsXmrIjMY9i9ps8ghJj-Vcdfb4328m17gE6qUmsj-az1_ilYzSwhf-_OGEe9v4NMe7dbWgUQZ5Kfe-nlj1dw9LuA.jpg" width="1200" height="630" loading="eager" /> <div class="link_preview_title">never waste a token</div> <div class="link_preview_description">durable inference: resumable streams, crash recovery, and why the LLM request shouldn't die with your process.</div> </a>https://channel.meo-ai.com/posts/615https://channel.meo-ai.com/posts/615Wed, 29 Apr 2026 08:21:28 GMT<span class="post-title-source">《大规模AI代码审查编排实践》</span><br /><br />标签：<a href="/search/result?q=%23DevOps" title="#DevOps">#DevOps</a> <a href="/search/result?q=%23AI%E8%BE%85%E5%8A%A9%E7%BC%96%E7%A8%8B" title="#AI辅助编程">#AI辅助编程</a> <a href="/search/result?q=%23CodeReview" title="#CodeReview">#CodeReview</a> <a href="/search/result?q=%23CI_CD" title="#CI_CD">#CI_CD</a> <a href="/search/result?q=%23LLM" title="#LLM">#LLM</a> <a href="/search/result?q=%23%E5%A4%9A%E6%99%BA%E8%83%BD%E4%BD%93%E7%B3%BB%E7%BB%9F" title="#多智能体系统">#多智能体系统</a> <a href="/search/result?q=%23Cloudflare" title="#Cloudflare">#Cloudflare</a> <a href="/search/result?q=%23OpenCode" title="#OpenCode">#OpenCode</a> <a href="/search/result?q=%23%E6%8F%92%E4%BB%B6%E6%9E%B6%E6%9E%84" title="#插件架构">#插件架构</a><br /><br />总结：<br />Cloudflare为解决代码审查瓶颈，放弃单一LLM直接审diff的噪音方案，转而基于开源代理OpenCode构建CI原生编排系统。该系统采用可组合插件架构，通过风险分级（Trivial/Lite/Full）动态调度最多7个专业审查智能体（安全、性能、质量等），由协调者代理去重、过滤并做出审批决策。系统已在数万MR上运行，能精准拦截真实漏洞，同时保留"break glass"人工逃生通道。<br /><br />文章要点：<br />- **从噪音到精准**：早期直接把git diff塞给LLM的方案产生了大量幻觉和模糊建议，团队很快意识到需要专业化分工而非单一通用提示词<br />- **插件化架构**：系统基于OpenCode构建，采用完全解耦的插件体系（GitLab、AI网关、合规检查、遥测等各自独立），通过`ConfigureContext` API贡献配置，最终组装成`opencode.json`<br />- **多智能体协作**：最多同时启动7个专业审查者各司其职，协调者代理负责去重、重新分类、合理性验证，并按严格规则做出approve/approve_with_comments/unapprove/request_changes四级决策<br />- **风险分级省成本**：按代码行数和文件数将MR分为Trivial/Lite/Full三级，小改动只派2个轻量代理且降级模型，安全相关文件永远触发Full审查，避免用大模型审typo<br />- **工程细节满满**：使用JSONL流式处理避免内存爆炸；通过磁盘patch文件共享上下文节省7倍token；清理XML边界标签防止提示注入；30秒心跳日志消除"模型思考中"的误取消<br /><br />文章URL：<a href="https://blog.cloudflare.com/ai-code-review" target="_blank" rel="noopener" title="https://blog.cloudflare.com/ai-code-review">https://blog.cloudflare.com/ai-code-review</a><a class="tgme_widget_message_link_preview" href="https://blog.cloudflare.com/ai-code-review" target="_blank" rel="noopener" title="Learn about how we built a CI-native AI code reviewer using OpenCode that helps our engineers ship better, safer code."> <div class="link_preview_site_name accent_color">Cloudflare Blog</div> <img class="link_preview_image" alt="Orchestrating AI Code Review at scale" src="/static/https://cdn4.telesco.pe/file/AZf5Ibv9s_K7ukQ_SKUmsU9wO2mh2bdeW-ePdEd3sc-nmf4Z5_uarlcS1hC832NHmywpL19ZGdKPPUKJy4VZAWjAGiApKeHINMXxiP4D6tb4B2hvdca2jTSXzgUm6gZ6gSHKCfncBp7kbi9l8-NhE2V3baaNHy_TwdHwMG_dfOVfvgXORDEjLOIUYk-vIQZXeQAKHwYTYjKQBsYcP_6pqRjwZEj-YFVai9eF7_4tTYrK74MTxEExtdeWPiifvPXipm5RfYT4Ju7L5v14OHj_Idr6h74m8ZNHBLs51ZGGXOxnd-rmU2XQsy_7LRpNkgvovErGWXhTF1CpWf4g0549rw.jpg" width="1200" height="630" loading="eager" /> <div class="link_preview_title">Orchestrating AI Code Review at scale</div> <div class="link_preview_description">Learn about how we built a CI-native AI code reviewer using OpenCode that helps our engineers ship better, safer code.</div> </a>