jacky @cppchangeworld

让 AI 帮忙做代码审查，它经常把大半个仓库重新读一遍，Token 消耗超级大。 code-review-graph 给代码库先建一张结构图，AI 要看什么直接从图里取，不用全量扫。 它用 Tree-sitter 解析代码，改动是增量更新的，通过 MCP 把精准上下文喂给编程助手。 GitHub：github.com/tirth8205/code… 作者自己在 6 个真实仓库上实测，用了该工具 Token 消耗能下降几十倍到几百倍不等。 支持 Claude Code、Codex、Cursor，以及 Gemini CLI 等主流 Agent 编程工具。 经常用 AI 看代码、又心疼 token 的朋友，可以拿它试试。

90% of your KV cache never gets reused. (prompt caching was never meant to fix it) if your system prompt and tool definitions are stable, prompt caching is the single highest-leverage optimization available today. cached input tokens get up to 90% cheaper, and hit rates of 60 to 85% are realistic. but it comes with one rigid rule. the cached portion must be an exact, byte-for-byte prefix of the new request. change a single character in that region and you get a full cache miss. that rule breaks in three situations you hit constantly: → RAG with multiple documents. you cached document A alone and document B alone. a query now needs both. document B's cached state was computed without any awareness of A, so it's invalid and gets recomputed from scratch. → document order changes. the same three documents appear in a different order across requests. every permutation is a cache miss, even though the content is identical. → growing conversation history. each new turn changes everything after the stable prefix, so earlier cached states beyond it become useless. Alibaba Cloud's production data shows how bad this gets: 10% of KV cache blocks serve 77% of all hits. the rest sits in storage and never gets reused, because prefix matching won't allow it. CacheBlend, a research paper from the LMCache team (EuroSys 2025 Best Paper Award), attacks exactly this. the insight is that in modern transformers, tokens overwhelmingly attend to their own local context. only a small fraction of tokens carry real connections across document boundaries. so instead of recomputing everything after the first cached document, CacheBlend reuses every document's cache as-is and selectively recomputes just those few boundary tokens. those are the small orange fixes between documents in the diagram, and they are the entire cost. the result is 2 to 4x faster processing on multi-document queries with no quality loss. the order problem disappears with it. shuffle the same documents however you like, and every permutation stays cached, where prefix caching recomputes all of them every time. the bottom of the diagram shows that side by side. that's the real shift: from caching prefixes to caching knowledge. every document in your knowledge base becomes a reusable cached asset, regardless of what order it appears in or what sits next to it. CacheBlend ships inside LMCache, the open-source cache management layer that runs outside the inference engine and integrates with vLLM, SGLang, and TensorRT-LLM, on both NVIDIA and AMD GPUs. check it out on GitHub: github.com/LMCache/LMCache (don't forget to star 🌟) i wrote the full breakdown of the architecture, including why cache management should never live inside your inference engine. the article is quoted below. stay tuned for more on this!

Akshay 🚀 @akshay_pachaar

2 weeks ago

x.com/i/article/2074…

15 32 214 247K 350

Function calling & MCP for LLMs, clearly explained: Before MCP, tool access ran entirely on function calling, and every integration was wired by hand. MCP standardized that integration layer, but the model still emits the same tool call underneath it. The visual below explains how each one works. In function calling, the model never executes anything itself: - The model receives a prompt and the available function definitions. - It decides which tool fits and emits a tool call request inside its response. - A procedure on our side parses that request and prepares the actual function call. - A backend service runs the tool and returns the output to the model. All of this lives in our own stack. - We host the tool - We write the logic that maps a tool call request to a function and its parameters - And we execute it So function calling just decides what to run, and the integration around it is left entirely to the engineer. MCP standardizes that integration. Instead of hand-wiring each tool, MCP: - Defines one protocol for declaring, hosting, and exposing tools. - Makes tools discoverable, with schemas the client can read directly. - Requires approval before a tool runs. - Separates the side that implements a tool from the side that consumes it. Once an MCP server is integrated, not a single line of Python is written to connect its tools. The server is added once, and everything past that point follows the protocol handled by the MCP client and the model. - They identify the MCP tool. - They prepare the input arguments. - They invoke the tool through the server. - They use the returned output to generate the response. Function calling and MCP are not competing approaches but rather two stages of the same tool call. - Function calling involves the model picking the tool. - MCP is everything that happens after picking it, like locating the tool, checking its schema, getting approval, and running it. When an agent decides "I need to search the web" through function calling, the decision routes through MCP, which selects from the registered web search tools, invokes the right one, and returns the result. Of course, there is a cost associated with that convenience. Integrating an MCP server loads every one of its tool definitions into context. A heavy server can spend most of the window on tool schemas before the agent runs a single call. We wrote a detailed walkthrough (with code) on fixing exactly that, where tool loading is scoped down to the groups actually in use, individual tools are selected instead of pulling the whole server, and context usage drops 80-90% on a live setup. Read it below.

Avi Chawla @_avichawla

6 months ago

x.com/i/article/2012…

4 17 95 46K 157

6 agent patterns for AI engineers: (explained with usage) 1) prompt chaining → split the task into fixed steps, each one checking the last. → use when the task decomposes cleanly and accuracy matters more than latency. 2) routing → classify the input first, then send it to the model or tool built for it. → use when inputs fall into distinct classes that need different handling. 3) parallelization → run several calls at once and merge them, either by splitting the work or voting on the same question. → use when subtasks are independent, or when one answer deserves several opinions. 4) orchestrator-workers → a lead model decides what the subtasks are at runtime, then delegates them. → use when you cannot list the steps in advance. 5) evaluator-optimizer → one model writes, another grades, the loop repeats until it passes. → use when you have clear criteria and iteration measurably helps. 6) autonomous agent → no fixed path. it plans, acts, reads feedback from the environment, and decides when it is done. → use when the steps are unknowable and you can afford the cost and the blast radius. the first five are workflows: you wrote the path. only the last one writes its own. most production systems people call agents are pattern 1, 2, or 5 with good error handling. this taxonomy is from Anthropic's own writeup on building effective agents. full breakdown in the article below.

Hanako @hanakoxbt

a week ago

x.com/i/article/2077…

17 72 415 1.2M 1K

We've built a full software engineering agent factory in house, and open-sourced every piece of it. @BraceSproul's blog dives into each of those components, what they're used for, and why them being OSS matters

Brace @BraceSproul

5 days ago

x.com/i/article/2078…

12 47 288 183K 749

推友们，书已经上线 《AI Coding Agent 架构解剖：从 Grok Build 源码看终端智能体的设计与实现》 zhanghandong.github.io/grok-build/ 还有部分章节未完成，可以先关注，欢迎交流学习

AlexZ 🦀 @blackanger

7 days ago

agent-spec 1.0 太好用了，我用它来写一本 grok build 的深度探索的书。 github.com/ZhangHanDong/a… 当然，实现项目代码用它也是很不错的。欢迎大家使用、交流和反馈。

keyboard_arrow_left Previous keyboard_arrow_right Next

22 10 128 56K 209

Claude 重度用户，这把可以冲了， GamsGo 恢复 Claude 订阅，Pro / Max 都能选，还新增「保障权益」：保障期内，未使用部分按剩余时长比例退款， 不单单是便宜，是低价之后还有兜底。 想用的可以直接走这里： gamsgo.com/partner/7Ad89

Rachel🥥 @Zesee

6 months ago

90% 的 AI 重度用户，正在被订阅制悄悄偷钱   ChatGPT Plus、Gemini、Perplexity，看起来每个都不贵，但一年算下来，3000 元起步，而且 70% 的时间在闲置。   如果我告诉你，这些 AI 和流媒体工具， 可以只用官方价格的 1/4，体验却完全一样， 你还会继续原价续费吗？

71 67 379 230K 609

这个叫agentic-rag-for-dummies的项目，它基于当下最猛的LangGraph框架搭建，专门解决RAG（检索增强生成）从演示玩具到生产环境落地的各种坑。 这可不是干巴巴的理论课，而是真刀真枪带你组装一套模块化的AI智能体系统，而且用的是极其良心的MIT协议，想拿去商用也完全没问题。 自带防爆机制和回退策略，查不到资料就自动停，绝不会乱烧你的Token； 引入上下文自动压缩，废话全剪掉只留干货，长对话照样快准狠； 完美适配Ollama接口，断网也能用，API费用直接省出一顿大餐； 支持多Agent并行处理，复杂长问题拆成子任务秒级响应。 作者不光给了手把手的本地部署源码，还贴心地准备了Google Colab免配置体验版，点开网页就能跑。 🔗：github.com/GiovanniPasq/a…

你刚加入一个新团队，代码库 20 万行，从哪开始看？ 最近 GitHub 上有个叫 Understand Anything 的开源工具，专治这个问题。 它会扫描整个项目，自动生成一份交互式知识图谱： - 每个文件、函数、类、依赖关系全部可视化 - 点击任意节点，直接看代码、看关系、看人话解释 - 支持模糊搜索和语义搜索，搜“哪个模块负责支付”直接出结果 相当于给你的代码库装上了一张活地图，而不是死文档。 兼容 Claude Code、Codex、Cursor、Copilot、Gemini CLI 等主流工具，一行命令安装。 GitHub：github.com/Egonex-AI/Unde… 适合三类人： 1. 刚接手老项目的开发 2. 想快速理解大型代码库的架构 3. 需要为新人准备入职指南的团队 建议直接 Star 收藏，下次接手新仓库先跑一遍 `/understand`，省下的时间够你喝两杯咖啡。

听说 Claude Fable 5 算命特别准。好奇驱动下，做了一个八字命理 Skill，自己用了一下后，惊呆了。 命理推演非常复杂，按提示补充信息后，需要耐心等待十几分钟。然后生成的报告，会有详细的性格、爱情、事业等运势分析。 安装地址：youmind.com/skills/chinese…

试试

一地鸡毛 @zhang_baoqing

a week ago

4.2元整了个一年期的gemini Pro 接下来该google表演了

108 20 341 144K 433

在多个代码仓库之间来回切，最烦的就是每次都要重新给AI喂一遍上下文。 codebase-memory-mcp 这个MCP服务器，直接把整个代码库索引成一张知识图谱，查起来亚毫秒级，号称比传统方式省99%的token。 自己写不同项目时经常要跨仓库，这工具正好对症下药。 做RAG或代码智能开发的，值得认真看一眼。 🔗 github.com/DeusData/codeb…

今天我们开源了DeerFlow背后的开发与调试工具LLM Space。你可以用它创建、调试和评测 Agent，也可以快速搭建原型而不用写一行代码。我们期待你的转发和 Star 支持 github.com/deer-flow/llm-…

Pi社区开源了codex没有开源的computer-use实现。 深入分析了一下仓库，眼前一亮！

keyboard_arrow_left Previous keyboard_arrow_right Next

Zane Chee @injaneity

3 months ago

i reverse engineered @OpenAI's Codex Computer Use and built pi-computer-use: a model agnostic computer use tool for my fellow pi enjoyers on MacOS. comes with ax-first navigation & a vision fallback for supported models! source code in comments; would love to get some feedback.

22 57 912 164K 973

现在用 Codex、Claude Code 或 Cursor 写一个网页并不难，不过要想把一个网页设计的好看高级，还是有点难度的。 最近发现一个给 AI Agent “注入设计品味”的开源技能库：emilkowalski/skills 一句话：专门解决 AI 做 UI 动画和交互时“没品味”的问题。 它不是新的 AI 模型，也不是设计软件，而是一套可以安装到 AI 编程 Agent 里的“设计经验包”： 把一位资深设计工程师关于动画、交互和 UI 细节的判断，整理成 AI 可以执行的规则。 核心技能包括： - improve-animations：扫描整个代码库，按 8 个维度审计动画，输出可执行的改进计划 - review-animations：严格按规则审查动画 - animation-vocabulary：教 AI 用精确语言描述动画 - apple-design：蒸馏苹果设计原则 最聪明的地方是：它不直接改代码，而是先审计 → 生成 prioritized plan → 再让 Agent 去执行。用强模型做审查，便宜模型做落地。 对经常用 AI 做前端 / UI 的人来说，这可能是目前最实用的“设计品味”注入工具。

keyboard_arrow_left Previous keyboard_arrow_right Next

sitin @sitinme

3 weeks ago

现在分析网页设计的工具还真的不少，之前分享过一些，这又翻到一个不太一样的开源模板，叫 ai-website-cloner-template 它的思路也特别直接： 你丢给它一个网址，它就指挥 AI 编程 agent 把这网站拆开看， 提取设计 token 和素材，写出组件规格， 再派一堆 agent

keyboard_arrow_left Previous keyboard_arrow_right Next

3 3 15 16K 33

AI 写报告只要 3 秒，但把它粘进 Word 调格式却要半天？ 公式乱码、Markdown 表格全错位……这几乎是每一个 AI 重度用户都经历过的“收尾噩梦”。 今天推荐一个开源救星：PasteMd。 专门解决 AI 输出内容到 Office 的“最后一公里”排版难题。👇

卧槽，Office 开放 Cli 了！！ 每天都在被 Word、Excel、PPT 等 Office 三件套纠缠？ 那请你一定要看看 GitHub 上这个已经有 16.2K Star 的开源工具：OfficeCLI 有了它之后，办公室里的文档处理和文案工作，真的可以开始释放双手了！！ 它可以让 AI Agent 用自然语言去操作 Office 文件： ① 生成 Word 文档 ② 整理 Excel 表格 ③ 创建 PPT 页面 ④ 修改文档格式和样式 ⑤ 检查文件结构 ⑥ 把 Office 文件渲染成 HTML / PNG，让 AI 看完效果后继续调整 这对办公室每天都要写文稿、做表格、出汇报的人太实用了。 如果你也经常被 Office 文件折磨，建议先收藏一下。 👉也欢迎评论区说说： 你最想让 AI 帮你处理哪一种 Office 工作？ github.com/iOfficeAI/Offi…

模型越强越纠结：全程开 Fable 心疼额度，切回 Opus 又怕方案没想透。以前只能手动 /model 来回切，切着切着就忘了。 现在 Claude Code 给了官方解法： claude --model opus --advisor fable 干活的是 Opus，Fable 变成顾问角色——Claude 自己判断什么时候该请示：提交方案前、连续出错、收工验收这几个节点。你不用再手动切，它不该出场时一个 token 都不烧。 注意两点： 1、先 claude update，老版本没这参数 2、advisor 强度必须 ≥ 主模型（Fable 带 Opus 正好） settings.json 加 "advisorModel": "fable"，以后开箱即是这套分工。 省钱的本质不是少用好模型，是让它只出现在值钱的地方。

为什么没早点刷到 ！突然英语就开窍了！

keyboard_arrow_left Previous keyboard_arrow_right Next

harness

Mr Panda @PandaTalk8

2 weeks ago

不要再去读 X 上又臭又长的 harness 工程的文章了， 跟这篇文章相比， X 上的那些文章垃圾都不如。 Lilian-weng 的这篇新的博客文章， 是我目前读过的最写的清晰、最容易理解的 harness 工程， 也是递归自我改进的 harness 工程写的最好的文章。 文章是一篇框架式的详细描述了 self-Improment 的

19 183 775 59K 1K