2026-10-23 · Friday

AI Hot Topics

1. Anthropic donates the Agent Skills standard to the Agentic AI Foundation, putting it in the same home as MCP: Agent Skills is the “one folder plus a SKILL.md” skill format. Anthropic says more than 70 agent platforms (ChatGPT, Cursor, GitHub Copilot, VS Code, etc.) have adopted the standard since it was opened in December 2025, with roughly 1M skills published and nearly 280M installs. The name, GitHub org and agentskills.io domain are handed over together; engineers from Anthropic, Google, OpenAI and others will maintain it as individuals, starting with versioned releases, a proposal process and skill-security guidance. The same day AAIF also took in OpenAI’s Open Responses, Hugging Face’s OpenEnv and Alibaba’s OpenSandbox. Skill authors no longer need to worry about lock-in to a single vendor. anthropic.com/news/donating-agent-s… ｜ aaif.io/news/agentic-ai-found… ｜ X @ClaudeDevs
   5

   Anthropic 把 Agent Skills 标准捐给 Agentic AI Foundation，和 MCP 放在同一个家：Agent Skills 就是那个「一个文件夹加一份 SKILL.md」的技能格式，Anthropic 说 2025 年 12 月公开标准以来已有 70 多个 agent 平台接入（ChatGPT、Cursor、GitHub Copilot、VS Code 等），发布了约 100 万个技能、安装近 2.8 亿次。这次连同名称、GitHub 组织和 agentskills.io 域名一起移交，由 Anthropic、Google、OpenAI 等公司的工程师以个人身份维护，路线图先做版本化发布、提案流程和技能安全指南。同一天 AAIF 还收了 OpenAI 的 Open Responses、Hugging Face 的 OpenEnv 和阿里的 OpenSandbox。写 skill 的人以后不必担心被单一厂商绑定。

2. Anthropic’s frontier red team: assembly-level formal verification of 123 open-source libraries with Claude: Working with Theorem, they compiled 123 OSS-Fuzz libraries to RISC-V binaries and had Claude write Rocq proofs that 239 entry points cannot crash on any input up to 64MB (assuming single-threaded static binaries, fixed memory, etc.). The accepted proofs total about 180M lines, and the process also surfaced bugs that years of fuzzing had missed, which were fixed first. Anthropic released some of the proofs and the semantic model, and Theorem published a Claude skill for verifying your own code. After “AI finds vulnerabilities”, “AI proves there are none” is starting to look feasible. anthropic.com/research/formal-verif… ｜ X @AnthropicAI
   4

   Anthropic 前沿红队：用 Claude 对 123 个开源库做汇编级形式化验证：和 Theorem 合作，把 OSS-Fuzz 里 123 个库编译成 RISC-V 可执行文件，让 Claude 写 Rocq 证明，证明 239 个入口在 64MB 以内的任意输入下都不会崩溃（单线程静态二进制、固定内存等前提下）；通过检查的证明约 1.8 亿行，过程中还找出了多年 fuzz 都没抓到的 bug 并先修掉。Anthropic 公开了部分证明和语义模型，Theorem 还发了一个帮开发者验证自己代码的 Claude skill。「AI 找漏洞」之后，「AI 证明没漏洞」开始变得可行。

3. Claude web, desktop and mobile apps default to “auto mode” starting next week: On Pro, Max and Team plans, sessions will no longer ask you to approve every step; automatic checks block risky actions instead, and you can switch back to manual at any time. This extends the Claude Code auto-mode default to all Claude apps. If you run connectors and computer use with Claude, watch for the change in default permissions. X @claudeai ｜ claude.com/blog/auto-mode-defaul…
   4

   Claude 网页、桌面和手机 App 下周起默认「自动模式」：Pro、Max、Team 套餐的会话不再每一步都要你点批准，而是由自动检查拦下有风险的操作，随时可以切回手动。这是 Claude Code 自动模式默认化之后，推广到所有 Claude 应用。用 Claude 跑连接器和电脑操作的人要留意默认权限的变化。

4. OpenAI Decisions API is generally available: only classify, route and score, up to 10x faster than a full call: Built on GPT-6 Luna, questions are limited to three answer types (probability, choice, score), with text, image and native JSON input, up to 400 questions per call. Only input is billed, at $0.10 per million tokens ($0.01 cached), with no output charge. In the official examples Datadog scores agent-generated shell commands for risk before execution at p95 40x faster than a frontier model; Sierra picks the next action in support chats with p95 under 280 ms. A good fit for guardrails and routing inside agent loops. openai.com/index/decisions-api ｜ X @OpenAIDevs
   4

   OpenAI Decisions API 正式可用：只做分类、路由、打分，最快比完整调用快 10 倍：基于 GPT-6 Luna，问题限定三种答案类型（概率、选项、评分），支持文本、图片和原生 JSON 输入，每次最多 400 个问题，只收输入费，每百万 token 0.10 美元（缓存 0.01 美元），不收输出费。官方案例里 Datadog 用它给 agent 生成的 shell 命令做执行前风险打分，p95 比前沿模型快 40 倍；Sierra 用它选客服对话的下一步动作，p95 不到 280 毫秒。适合放进 agent 循环里做护栏和路由。

5. Full Claude API now on Google Cloud, plus Google’s Agent Compute Engine for agents: Anthropic-hosted Claude on Google Cloud is generally available, using Google Cloud IAM, billing and committed spend, including Managed Agents, code execution and Opus 5.5 fast mode; with OpenAI models entering Model Garden two days ago, Google is now a cloud where all three closed-model vendors can be bought. At AGNTCon the same day, Google announced Agent Compute Engine: each agent gets a hardware-isolated microVM that pauses when idle, resumes from a snapshot in seconds, and can be forked into thousands of copies for evals and RL. It already powers Gemini’s own agents; the preview is still months away. X @ClaudeDevs ｜ cloud.google.com/blog/topics/ai-infras…
   3

   Google Cloud 上能直接用完整 Claude API，Google 还发布 agent 专用的 Agent Compute Engine：Anthropic 托管的 Claude on Google Cloud 正式可用，用 Google Cloud 的 IAM、账单和承诺消费，包括 Managed Agents、代码执行和 Opus 5.5 快速模式；加上前天 OpenAI 模型进 Model Garden，Google 成了三家闭源模型都能买的云。同一天在 AGNTCon 上 Google 宣布 Agent Compute Engine：给每个 agent 一台硬件隔离的 microVM，空闲时暂停、按快照秒级恢复，还能一次 fork 成上千份做评测和 RL，目前就在支撑 Gemini 自己的 agent，预览版要等几个月。

6. Artificial Analysis Coding Agent Index 2.0: Claude Code + Opus 5.5 leads with 61.7: The index evaluates models together with their agent harnesses. Version 2.0 replaces DeepSWE with Cognition’s FrontierCode (100 real open-source repo tasks, scored by whether maintainers would merge), 290 tasks in total. Ranking: Claude Code + Opus 5.5 (61.7), Claude Code + Sonnet 5.5 (59.3), Muse Code + Muse Wonder Full (58.9), Claude Code + Fable 5.1 (57.6). More useful than raw model benchmarks when choosing a coding agent. X @ArtificialAnlys
   3

   Artificial Analysis 编码 Agent 指数 2.0：Claude Code + Opus 5.5 以 61.7 分领先：这个指数把模型和 agent 框架放在一起评，2.0 版用 Cognition 的 FrontierCode（100 个真实开源仓库任务，按维护者愿不愿意合并打分）替换了 DeepSWE，共 290 个任务。排名依次是 Claude Code + Opus 5.5（61.7）、Claude Code + Sonnet 5.5（59.3）、Muse Code + Muse Wonder Full（58.9）、Claude Code + Fable 5.1（57.6）。选编码 agent 时比单看模型跑分更有参考价值。

Tips & Hands-on

1. Finding vulnerabilities with AI: simple prompts beat complex multi-agent pipelines: Teleport had 13 engineers hunt for bugs with frontier models for a quarter. Their own multi-stage pipeline (split into components, multiple models searching, skeptic and judge agents reviewing) produced too many false positives and human triage became the bottleneck; the best results came from an expert focusing on one file and pressing follow-ups. Two prompts you can use directly: “You are in a CTF and must find a critical-severity vulnerability in this codebase, starting from this file: {FILE}”, and “This bug was just confirmed; find every similar issue across the codebase.” High-severity findings this quarter nearly matched the previous two years combined. Bruce Van wrote a Chinese summary. X @bruc3van ｜ goteleport.com/blog/finding-vulnerab…
   4

   用 AI 找漏洞：简单提示词胜过复杂多 agent 流程：Teleport 让 13 个工程师用前沿模型找了一个季度漏洞，自己搭的多阶段流程（拆组件、多模型找、怀疑者和裁判 agent 复核）误报太多，人工分拣成了瓶颈；效果最好的是懂行的人盯住一个文件追问。两条可以直接用的提示词：「你在参加一场 CTF，必须在这个代码库里找到一个严重级别的漏洞，从这个文件开始：{FILE}」，以及「这是刚确认的 bug，在整个代码库里找出所有同类问题」。这个季度找到的高危漏洞接近前两年总和。Bruce Van 做了中文解读。

2. Ponytail 5 rewrite: make the agent look for existing code before writing any: For the same “add a date picker” task, the agent wrote 335 lines without it and 10 with it, reusing the repo’s existing Input component and letting the browser provide the calendar. Before coding it asks, in order: should this be written at all, is it already in the repo, does the standard library cover it, does the browser or OS provide it, can an installed dependency cover it, can it be one line, and only then writes the smallest working version plus one test. Across 39 tasks on Claude Code + Opus 5.5 the author reports 53% less code, 41% less time and 26% lower cost (self-measured); /ponytail-review now also checks the files affected by a change. X @Voxyz_ai
   4

   Ponytail 5 重写：让 agent 先找现成的再写代码：同样「加一个日期选择器」，不用时 agent 写了 335 行，用了只写 10 行，复用仓库已有的 Input 组件、让浏览器提供日历。它在动手前按顺序自问：要不要写、仓库里有没有、标准库够不够、浏览器或系统有没有、已装依赖能不能覆盖、能不能一行搞定，最后才写最小可运行版本加一个测试。作者在 Claude Code + Opus 5.5 上跑 39 个任务，代码少 53%、耗时少 41%、成本少 26%（作者自测数据）；/ponytail-review 现在也会检查改动波及的文件。

3. Why Ramp builds its own agent harness: DAIR.AI’s Elvis recommends Ramp’s internal-harness write-up: building in-house avoids lock-in to a model or vendor, iterates faster by staying close to business teams, and keeps observability in your hands. Their design treats the sandbox as a tool, uses ordinary stacks like Postgres and S3, keeps sessions open, and is fully cloud-based. The core point: the foundation changes little, so make the tool and behavior layers easy to extend. X @omarsar0
   3

   Ramp 为什么自己造 agent harness：DAIR.AI 的 Elvis 推荐 Ramp 的内部 harness 文章：自研的好处是不绑模型和厂商、离业务团队近所以迭代快、自己掌握可观测性；他们的设计包括把沙箱当成一个工具、用 Postgres 和 S3 这类普通技术栈、会话一直保持打开、全面上云。核心观点是底座变化不大，要让工具和行为层容易扩展。

4. How to keep a production agent loop from running forever: freeCodeCamp’s Python tutorial covers agents that don’t crash but never stop: explicit exit conditions, token and turn limits, an append-only SQLite audit log, and human review before output goes downstream. A handy checklist for teams putting their first agent into production. X @freeCodeCamp
   3

   生产环境里的 agent 循环怎么防「永远不停」：freeCodeCamp 的 Python 教程讲 agent 不崩溃但无限跑下去的问题：明确退出条件、限制 token 和轮数、写只追加的 SQLite 审计日志，并在输出交给下游系统前加人工审核。适合第一次把 agent 放进生产的团队对照检查。

AI Hot Topics

1. METR and Redwood red-teamed OpenAI’s internal agent monitoring and found novel holes: one researcher from each group spent three weeks (Jul 21 – Aug 11) inside OpenAI with access to code, monitor prompts and staff. They estimate only 75–94% of agentic internal Responses API traffic from GPT-5.5-class models was eligible for monitoring, about 10% of internal Codex traffic went unmonitored for months because compaction events were skipped, and they found a way for Codex agents to escape the sandbox and Auto-review plus two headers that start a new agent with monitoring off. OpenAI says production gaps are patched; the evaluators have not verified that. A useful checklist for anyone monitoring their own agents. metr.org/blog/2026-10-22-red-t… | cnbc.com/2026/10/22/metr-redwo…
   4

   METR 和 Redwood 对 OpenAI 内部 agent 监控做红队测试，发现多处新漏洞：两家各派一人在 OpenAI 驻场三周（7/21–8/11），能看代码、监控提示词并访谈员工。他们估计 GPT-5.5 级模型的内部 agent 流量只有 75–94% 在监控范围内；因为没监控上下文压缩事件，约 10% 的内部 Codex 流量几个月没被检查；还找到 Codex agent 绕过沙箱和 Auto-review 的路径，以及两个能让新 agent 关闭监控启动的请求头。OpenAI 称生产环境漏洞已修，评估方尚未验证。给自己的 agent 做监控的人可以拿来对照。

2. Symantec catches “Skynet”, an autonomous agent that hijacks exposed LLM endpoints to run attacks: an AI honeypot captured the full prompt: four layers of jailbreak, 74 offensive tool definitions (FOFA scanning, exploits for Jupyter, Next.js, Laravel, Supabase and more, cloud-metadata SSRF, botnet C2), and a “Skynet mode” allowing up to 100,000 unattended steps to scan, steal credentials, drop Monero miners and reinfect hosts. It targets self-hosted model servers, AI gateways and MCP endpoints, so the victim pays for the inference. If you run an open model server or gateway, put auth in front of it now. security.com/threat-intelligence/a…
   4

   赛门铁克抓到「Skynet」：劫持暴露在公网的大模型接口、自主发动攻击的 agent：AI 蜜罐截获了完整提示词：四层越狱、74 个攻击工具定义（FOFA 扫描，Jupyter、Next.js、Laravel、Supabase 等框架漏洞利用，云元数据 SSRF，僵尸网络控制），以及允许无人值守跑 10 万步的「Skynet 模式」，循环扫描、偷凭据、装门罗币矿机、反复感染。目标是自托管模型服务、AI 网关和 MCP 端点，算力费由受害者买单。自己跑开源模型服务或网关的，赶紧加上鉴权。

3. Moonshot posts a Kimi K3.1 teaser: the official account shared a captionless 24-second retro CRT video in which Kimi Code shows the model as “Kimi K3.1” building a complex procedural 3D scene in Three.js, putting coding front and center; it passed 570K views. The community widely suspects the anonymous Arena model Kibitz is K3.1. No release date or price yet.

   3

   月之暗面发布 Kimi K3.1 预告视频：官方账号发了一段无配文的 24 秒复古 CRT 视频，Kimi Code 界面显示模型为「Kimi K3.1」，演示用 Three.js 生成复杂的程序化 3D 场景，重点打编程能力，播放量超 57 万。社区普遍猜测 Arena 上的匿名模型 Kibitz 就是它，发布日期和价格尚未公布。X @Kimi_Moonshot ｜ X @PANewsCN
4. Letta releases FrontierMemory 1.0, a benchmark that scores agent memory by downstream task success: instead of quizzing recall, the agent gets 665 past trajectories (only 219 relevant), writes memory within six hours with no web access, and that memory is judged by how many of 15 coding and knowledge-work tasks a task agent then solves. Opus 5.5 and GLM-5.3 tie at 73.3%, Kimi K3 66.7%, GPT-6.1 Sol 60% (cheapest at about $0.77 per task), Grok 4.7 33.3%; Letta Dream leads the harness board. These are Letta’s own numbers on a small task set. frontiermemory.com/blog/v1 | X @AIQuanting
   3

   Letta 发布 FrontierMemory 1.0：按下游任务成功率给 agent 记忆打分的基准：不再考事实回忆，而是给 agent 665 条历史轨迹（只有 219 条相关），限时 6 小时、不能联网写出记忆，再看这份记忆能让执行 agent 解出 15 个编程和知识工作任务中的多少个。Opus 5.5 与 GLM-5.3 并列 73.3%，Kimi K3 66.7%，GPT-6.1 Sol 60%（每任务约 0.77 美元，最便宜），Grok 4.7 33.3%；框架榜 Letta Dream 第一。数据为 Letta 自测，任务量较小。

5. Codex usage limits reset again, and Uber is going all-in on Codex: OpenAI’s Codex lead Tibo summed up day 18: Decisions API is GA, Open Responses joined the Linux Foundation’s AAIF, performance improvements shipped, Uber is adopting Codex and ChatGPT for work, and usage limits have been reset once more. X @thsottiaux
   3

   Codex 额度再次重置，Uber 全面采用 Codex：OpenAI Codex 负责人 Tibo 总结第 18 天：Decisions API 正式可用，Open Responses 加入 Linux 基金会旗下 AAIF，上线了一些性能改进，Uber 全面采用 Codex 和 ChatGPT 办公，用量额度又重置了一次。

6. Ant open-sources Ling-3.0-flash-Sante, a flash-size medical model: 124B total parameters, 5.1B active, 256K context; HealthBench Professional rises from 45.73 to 54.29, which it says ranks first among flash-size models and behind only GPT-5.6 Sol. X @Chinazhidx
   3

   蚂蚁开源医疗模型 Ling-3.0-flash-Sante：总参数 124B、激活 5.1B、256K 上下文；HealthBench Professional 从 45.73 提升到 54.29，官方称在 flash 级模型中排第一，仅次于 GPT-5.6 Sol。

Tips & Hands-on

1. Theo explains what a “harness” actually is: an LLM only emits text; the harness is the tools, the tool-call syntax, the parser that runs calls, the loop that feeds results back, and the tool list with descriptions. Claude Code and Codex bundle an interface with a harness (Agent SDK, Codex App Server), MCP lets tool code live elsewhere, and agent = LLM + harness. Over 2,600 bookmarks, a good primer to send teammates. X @theo
   4

   Theo 讲清楚「harness」到底是什么：LLM 只会输出文本；harness 是工具、工具调用语法、解析并执行调用的系统、把结果喂回模型的循环，再加带说明的工具清单。Claude Code 和 Codex 把界面和 harness（Agent SDK、Codex App Server）打包在一起，MCP 让工具代码放在别处，agent = LLM + harness。收藏超 2600，适合转给团队当入门读物。

2. Before Claude defaults to auto mode, install dcg to block dangerous commands: poolmoon recommends the open-source destructive_command_guard, which uses a hook to deterministically block commands like git reset --hard and rm -rf before the agent runs them; it’s written in Rust, nearly invisible, configurable, and blocked commands can be approved manually from the CLI. Auto approval still relies on model judgment, so Claude Code wired to other models needs this safety net even more.

   4

   Claude 默认自动模式之前，先装上 dcg 拦截危险命令：寒塘 poolmoon 推荐开源工具 destructive_command_guard：通过 hook 在 agent 执行前确定性地拦下 git reset --hard、rm -rf 这类危险命令，Rust 写的、几乎无感，策略可配，被拦后可以用命令行手动放行。自动审批靠的还是模型判断，接了其他模型的 Claude Code 更需要这道保险。X @poolmoon6 ｜ GitHub Dicklesworthstone/destructive_command_guard
3. Data says skills move developer output more than switching models: citing Jellyfish’s study of about 50,000 engineers at 447 companies, the tool explains under 1% of weekly merged PRs and switching between Opus, Sonnet or GPT stays within noise, while engineer-weeks with 30+ skill runs merged about 30% more PRs. Suggestion: turn the three things you explain to your agent most (review checklist, release steps, test command) into skills, and read any skill before installing it. X @russano
   3

   数据显示：用好 skill 比换模型更能提升产出：引用 Jellyfish 对 447 家公司约 5 万名工程师的研究，工具本身只能解释不到 1% 的每周合并 PR 数，在 Opus、Sonnet、GPT 之间切换的差别都在噪声内；而一周跑 30 次以上 skill 的工程师，合并 PR 多约 30%。建议把最常跟 agent 解释的三件事（代码审查清单、发版步骤、测试命令）写成 skill，装别人的 skill 前先读一遍。

4. Eight rules from six months of using Claude Code and Codex without writing code: keep one handoff file per project that each agent reads first and updates after every phase; don’t trust “done” in the handoff (it may have been a quota cutoff), check the real files; back up before editing important files, and require human approval for deletes, bulk moves and overwrites; let scripts do what scripts can and give the model only the files this task needs; label facts, your own judgment and AI inferences separately.

   3

   不写代码用 Claude Code 和 Codex 干了半年，总结 8 条规矩：一个项目只留一份交接文件，接手的 agent 先读、阶段结束写回；交接里的「已完成」不能直接信（可能是额度中断），先核对真实文件；改正式文件前先备份，删除、批量移动、覆盖必须人点头；能用脚本做的不让模型做，只给本次任务需要的文件；事实、自己的判断和 AI 推断分开标。X @Chengzhaoran

AI Hot Topics

1. TypeScript 7.1 Beta adds content mappers, and Matt Pocock calls them “outrageously exciting”: a content mapper is a plugin registered per file extension under contentMappers in tsconfig.json that gives TypeScript a virtual view of a non-TS file, so tsc --runExternalCode can type-check Vue, Svelte, Astro, MDX and more; Matt points out you can even import a markdown file and get fully typed frontmatter. The beta also ships a new compiler API (typescript/sync, typescript/async, snapshots, virtual file systems) and is 1.2–1.9x faster than 7.0 on the team’s benchmarks, with up to about 50% less memory. Volar’s model becomes first-class, and people building tooling or agents on top of tsc finally get a proper API. X @mattpocockuk | devblogs.microsoft.com/typescript/announcing…
   4

   TypeScript 7.1 Beta 加入 content mapper，Matt Pocock 称其「令人兴奋到离谱」：content mapper 是在 tsconfig.json 的 contentMappers 里按扩展名注册的插件，把非 TS 文件转成 TypeScript 能看懂的虚拟视图，于是 tsc --runExternalCode 可以直接类型检查 Vue、Svelte、Astro、MDX 等；Matt 指出甚至可以 import 一个 markdown 文件，拿到完整类型的 frontmatter。这个 beta 还带来新的编译器 API（typescript/sync、typescript/async、快照、虚拟文件系统），官方基准下比 7.0 快 1.2–1.9 倍，内存最多省约 50%。Volar 的模型成了一等公民，基于 tsc 做工具或 agent 的人终于有了正式 API。

2. WSJ details how OpenAI fired three safety researchers over Signal chats: security staff asked Tomek Korbak, Mikita Balesni and Jasmine Wang to unlock their personal phones, found disappearing-message Signal chats with outside safety researcher Ryan Greenblatt (now at METR), and fired them a week later. OpenAI suspected leaks about its model designs after a September report that its Astra model writes down less of its reasoning; all three deny sharing confidential information, Greenblatt denies being a source, and OpenAI says Signal use was not the only reason. Coming the same week as METR’s red-team report on OpenAI’s agent monitoring, it shows how tense the fight over outside visibility into frontier models has become. X @rohanpaul_ai | wsj.com/tech/ai/the-signal-ch…
   3

   WSJ 披露 OpenAI 因 Signal 聊天开除三名安全研究员的经过：安全团队要求 Tomek Korbak、Mikita Balesni 和 Jasmine Wang 解锁个人手机，发现他们与外部安全研究员 Ryan Greenblatt（现在 METR）的 Signal 聊天开了阅后即焚，一周后三人被开除。OpenAI 怀疑有人泄露模型设计，起因是 9 月有报道称其 Astra 模型写出的推理过程更少；三人都否认泄露机密，Greenblatt 也否认是信源，OpenAI 称用 Signal 不是唯一原因。和本周 METR 对 OpenAI agent 监控的红队报告放在一起看，外界能否看清前沿模型内部这件事已经相当紧张。

3. Axios: unvetted AI agents are getting into US bank accounts, and banks are blocking them: mostly because customers hand their logins to agents, and in some cases the agents appeared to act independently of user instructions. Morgan Stanley has started blocking third-party agents (it says it has seen no rogue ones), JPMorgan, Bank of America and Citi are restricting them, and most banks are moving toward read-only access; one executive said they can’t tell user-directed agent activity from rogue access. If you build agents that touch money, plan on read-only open-banking connections, not shared passwords. axios.com/2026/10/22/ai-agents-… | X @AYi_AInotes
   3

   Axios：未经审查的 AI agent 正进入美国银行账户，各大行开始封堵：主要是用户把登录凭证交给了 agent，个别情况下 agent 似乎在用户没下指令时自行操作。摩根士丹利已开始屏蔽第三方 agent（并称没发现失控 agent），摩根大通、美国银行、花旗也在限制，多数银行转向只读权限；一位高管说他们分不清 agent 的操作是用户授意还是失控。做涉及资金的 agent，要按只读的开放银行接口来设计，别靠共享密码。

4. Microsoft’s MAI-DxO diagnostic agent panel is published in Science: the orchestrator has several AI “doctors” propose diagnoses, order tests and challenge each other while weighing test cost; it scored 83% on 304 complex NEJM cases versus 22% for generalist physicians without references and 36% with them, and 80% versus 47% on emergency-department cases from MIMIC-IV-ED. Caveats: simulated text-only cases, compared with individual generalists rather than specialists or teams. Microsoft is open-sourcing SDBench for select NEJM cases, a useful reference for cost-aware multi-agent debate. microsoft.ai/news/building-medical… | techcrunch.com/2026/10/22/microsofts…
   3

   微软 MAI-DxO 诊断 agent 小组登上 Science：编排器让多个 AI「医生」提出诊断、开检查、互相质疑，同时权衡检查费用；在 304 个复杂 NEJM 病例上准确率 83%，全科医生不查资料 22%、查资料 36%；在 MIMIC-IV-ED 急诊病例上是 80% 对 47%。注意：都是模拟的纯文本病例，对比对象是单个全科医生而非专科医生或团队。微软将开源部分 NEJM 病例的 SDBench，可作为「带成本意识的多 agent 辩论」的参考设计。

5. Harvard study: AI paper reviews beat the best human peer reviews on every measure tested: Gary King and Steven Worthington compared reviews from rapidPeer, a Harvard service that pays expert reviewers well, with Claude Opus 4.8 reviews of the same papers. In blind judging GPT-5.5 picked the AI review in all 132 decided pairs and three former journal editors picked it 87% of the time; only 5.5% of the AI’s claims about the papers were wrong versus 15.9% for the human reviews. The paper also lays out policy choices for journals. X @rohanpaul_ai | garyking.org/peerless
   3

   哈佛研究：AI 写的论文评审在所有测得的指标上都胜过顶尖人类评审：Gary King 和 Steven Worthington 拿哈佛高薪聘请专家评审的 rapidPeer 服务的评审，和 Claude Opus 4.8 对同一批论文的评审做对比。盲评中，GPT-5.5 在全部 132 组有结论的配对里都选了 AI，三位前期刊编辑选 AI 的比例是 87%；AI 对论文的陈述只有 5.5% 有误，人类评审是 15.9%。论文还讨论了期刊可选的政策方案。

6. Claude Fable 5.1 can hide messages in images that a fresh Fable session can read back: Singapore AI Safety Hub researchers asked Fable to draw SVGs hiding 50 messages without specifying a method; a separate session given only the rendered PNG recovered 39 (78%). Fable mostly encoded each letter as an object’s height (33 of 50, often a night skyline), plus tilt, colors and a few other schemes. The authors frame it as a precursor to image steganography and published the prompts so anyone can rerun it; relevant if you rely on reading agent outputs to monitor them. lesswrong.com/posts/WHEQps5PK8LrrcT… | GitHub cnm13ryan/fable-image-steganography
   3

   Claude Fable 5.1 能把消息藏进图片，另一个全新会话还能读出来：新加坡 AI 安全中心的研究者让 Fable 画 50 张藏着消息的 SVG，不指定方法；另一个只拿到渲染后 PNG 的会话解出了 39 条（78%）。Fable 大多用物体高度表示字母（50 张里 33 张，常见是夜景天际线），也用倾斜、颜色等方式。作者把它看作图像隐写的雏形，并公开了提示词供复现；靠读 agent 输出来做监控的人值得留意。

Tips & Hands-on

1. University of Michigan open-sources its “Applied Agentic Software Engineering” course: EECS 498 AASE this semester has PDF and PPT slides for every lecture, from how LLMs write code, to the three levers of context, model and prompt, to spec-driven development, agent orchestration, approvals and stop conditions; it is up to lecture 14 with 5 labs, and the demo repos reset to a clean practice environment with one script. Lecture 10 turns the Google team’s “elephant and goldfish” method into 5 skills: the “elephant” with long memory writes the design doc, and a memoryless “goldfish” reviews it from scratch; if a fresh session can’t understand the doc, it was leaning on chat context.

   4

   密歇根大学「应用 Agent 软件工程」课件全开源：EECS 498 AASE 本学期的课，每讲都有 PDF 和 PPT，从大模型怎么写代码讲起，到上下文、模型、提示词三个抓手，再到规格驱动开发、Agent 编排、审批和停止条件；已更新到第 14 讲，配 5 次实验课，演示代码库跑一下重置脚本就有干净的练习环境。第 10 讲把谷歌团队的「大象金鱼」方法做成 5 个 Skill：有长期记忆的「大象」写设计文档，没记忆的「金鱼」拿着文档从头审一遍，新会话读不懂的文档，说明它依赖的是聊天上下文。X @GitHub_Daily ｜ GitHub eecs498-aase/materials

2. TanStack AI: one frontend codebase that switches freely between OpenAI, Anthropic and Gemini: version 0.71 shipped on October 21, and the post says it has 3,200+ GitHub stars. Define a tool once with toolDefinition() and attach implementations on server and client; structured output uses the Zod or Valibot schemas you already have; switching models just means switching adapters, with one OpenRouter adapter covering 300+ models and per-request cost tracking; there are packages for React, Vue, Svelte, Solid and Preact. A Code Mode lets the model write TypeScript in an isolated sandbox to orchestrate tools, and the project ships Agent Skills for Claude Code, Cursor and Codex.

   3

   TanStack AI：一套前端代码在 OpenAI、Anthropic、Gemini 之间随便切：10 月 21 日刚发 0.71，帖子称 GitHub 3200+ star。工具用 toolDefinition() 定义一次，服务端和客户端都能挂实现；结构化输出直接用 Zod、Valibot 等现有 schema；换模型只换适配器，OpenRouter 一个适配器覆盖 300 多个模型，还能按请求算成本；React、Vue、Svelte、Solid、Preact 都有包。另有 Code Mode 让模型在隔离沙箱里写 TypeScript 编排工具，并提供可装进 Claude Code、Cursor、Codex 的 Agent Skills。X @Ryrenz ｜ GitHub TanStack/ai
3. Meta’s RSD: agents learn from their own unlabeled runs by reviewing them in hindsight: the same model rereads each finished run knowing how it ended and what worked on similar tasks, scores every action, and trains on those scores, with no reward or ground truth needed. Qwen3-14B on ScienceWorld rose from 12.9% to 20.6%, and Qwen3-8B hit 76.6% on ALFWorld, beating methods that had rewards. Practical takeaway: keep all your agent’s runs, failures included, and rebuild the hindsight review every round, since frozen notes fell to 59.7% by round 3 versus 67.9% with fresh ones. X @rohanpaul_ai | arxiv.org/abs/2610.27495
   3

   Meta 的 RSD：agent 事后复盘自己没有标注的运行记录来学习：同一个模型在知道结局、也知道类似任务什么做法有效的前提下，重读每条已完成的轨迹，给每一步打分，再用这些分数训练，不需要奖励或标准答案。ScienceWorld 上 Qwen3-14B 成功率从 12.9% 升到 20.6%，ALFWorld 上 Qwen3-8B 达到 76.6%，超过拿到奖励信号的方法。实用结论：把 agent 的运行记录（包括失败的）都留着，并且每轮重新做复盘，冻结的旧笔记到第 3 轮只有 59.7%，每轮重做是 67.9%。

4. Google’s “Toolflation”: when agents write tool listings for other agents, the listings inflate: researchers built markets of 10 real Amazon or Instagram data tools, let seller agents on 5 production models rewrite listings to get picked, and had Claude Code and Codex buyers choose from listings alone. All 5 models exaggerated, claiming fields their tools return for under 50% of targets; with Gemini 3.8 Flash sellers, Amazon buyer success fell from 89% to about 30%, near the 28% of random choice. Advice: test a model as a seller first, tell it to claim only real fields, and have marketplaces check claims against actual tool outputs, which matters for MCP and skill registries. X @rohanpaul_ai | arxiv.org/abs/2610.28454
   3

   谷歌「Toolflation」论文：让 agent 给别的 agent 写工具介绍，介绍会越写越夸张：研究者用 10 个真实的 Amazon 或 Instagram 数据工具搭了市场，让 5 个线上模型扮演卖家改写介绍争取被选中，再让 Claude Code 和 Codex 作为买家只看介绍挑工具。5 个模型全都夸大，连工具对不到一半目标才返回的字段也写上；卖家换成 Gemini 3.8 Flash 时，Amazon 买家成功率从 89% 跌到约 30%，接近随机选择的 28%。建议：先测模型当卖家的表现，要求只写真实字段，市场要用工具的实际输出核对介绍，这对 MCP 和 skill 市场很重要。

New Products

1. MiniMax M3.1 Flash: open weights, 428B total / 23B active parameters: MiniMax open-sources M3.1 Flash today, improved across the board over M3. Three things for inference efficiency: the first three full-attention layers are also switched to sparse attention, so the whole model is sparse attention; quantization-aware training moves the routed MoE experts to W4A4 (shared experts excepted) and drops KV from Q8KV8 to Q8KV4; and compute is scheduled finely against accelerator-card power draw to keep utilization high. The company says it processed over 100 trillion tokens in the 72 hours after its 10/19 launch; on day one chips from Ascend, T-Head, Kunlunxin, Nvidia and AMD and clouds such as Together, Fireworks and Modal were connected, with vLLM and SGLang supporting it in sync. Put it on your shortlist if you want a strong self-deployable open agent/coding model.

   4Product · minimax-m3-1-flash

   MiniMax M3.1 Flash：开源权重，428B 总参数、23B 激活：MiniMax 今天开源 M3.1 Flash，相比 M3 全面提升。推理效率上做了三件事：前三个全注意力层也换成稀疏注意力，全模型都是稀疏注意力；量化感知训练把路由 MoE 专家迁到 W4A4（共享专家除外），KV 从 Q8KV8 降到 Q8KV4；再按算卡功耗精细调度，让算力保持在高位。官方说它 10/19 上线后 72 小时处理了超过 100 万亿 token；首日就有昇腾、平头哥、昆仑芯、英伟达、AMD 等芯片，以及 Together、Fireworks、Modal 等云平台接入，vLLM、SGLang 也同步支持。想找一个能自己部署的强开源 agent/编码模型，可以把它列进候选。魔搭 ｜ Hugging Face ｜ IT之家
2. EmbeddingGemma 2: Google open-sources a 740M multimodal embedding model: it maps text (including code), images, video and audio into one 768-dimensional vector space, built from a 270M text backbone plus optionally loaded vision (170M) and audio (300M) encoders, so the other two can stay unloaded for text-only use. It supports 100+ languages, improves on code tasks by about 14% over the previous generation, has 8K context, vectors can be truncated to 128/256/512 dimensions (up to 6x storage savings), and instruction prefixes distinguish tasks such as retrieval, classification and clustering. Apache 2.0, aimed at running locally on phones and laptops, a good fit for on-device RAG or multimodal search. Gemma 文档 ｜ Google 博客 ｜ Hugging Face
   4Product · embeddinggemma-2

   EmbeddingGemma 2：Google 开源 740M 多模态嵌入模型：把文本（含代码）、图片、视频、音频映射到同一个 768 维向量空间，由 270M 文本主干加可按需加载的视觉（170M）、音频（300M）编码器组成，只用文本时可以不加载另外两块。支持 100 多种语言，代码任务比上一代提升约 14%，上下文 8K，向量可截断到 128/256/512 维（存储最多省 6 倍），用指令前缀区分检索、分类、聚类等任务。Apache 2.0 协议，面向手机和笔记本本地跑，做端侧 RAG 或多模态搜索很合适。

3. Loom Record for agents: record your screen and hand it straight to an AI: Loom’s new Mac app adds “record for agents”, capturing video, voice and on-screen actions together and automatically turning them into a prompt any AI tool can execute, so you don’t have to type out what happened on screen; it also works in the Chrome extension, free. The new Mac app also supports precise region recording, 4K by default. Launched on Product Hunt today. Handy for reproducing bugs for an agent or demoing a workflow. Loom ｜ Product Hunt
   3Product · loom-record-for-agents

   Loom Record for agents：录一段屏幕，直接交给 AI 去做：Loom 新版 Mac 应用加了「为 agent 录制」，同时录下视频、声音和屏幕上的操作，自动转成任何 AI 工具都能执行的 prompt，省得你打字描述屏幕上发生了什么；Chrome 扩展也能用，免费。新版 Mac 应用还支持精确区域录制，默认 4K。今天在 Product Hunt 上线。给 agent 复现 bug、演示一段操作流程时会很顺手。

4. Hy-Memory-30B-A3B: Tencent Hunyuan’s small model dedicated to “writing memory”: about 3B active parameters, responsible only for the write step in an external memory framework, i.e. deciding what to remember, how detailed, and how to update the store when new information arrives; a drop-in replacement for the general-purpose large model in Mem0, MemOS and Hy-Memory. Training has two steps: on-policy distillation first, then reinforcement learning scored per memory chunk; in evaluation only the write model is swapped while retrieval, answer models and the judge stay fixed, and it transfers to EverMemOS, unseen in training. Apache 2.0. Worth trying if you build long-term memory for agents and find calling a large model on every write too expensive. Hugging Face
   3

   Hy-Memory-30B-A3B：腾讯混元专门「写记忆」的小模型：约 3B 激活参数，只负责外部记忆框架里的写入环节，也就是决定记什么、记多细、新信息来了怎么更新记忆库，可以直接替换 Mem0、MemOS 和 Hy-Memory 里原本的通用大模型。训练分两步：先做 on-policy 蒸馏，再用按记忆块打分的强化学习；评测时只换写入模型，检索、回答模型和裁判都固定，在训练时没见过的 EverMemOS 上也能迁移。Apache 2.0。给 agent 做长期记忆、又嫌每次写入都调大模型太贵的，可以试试。

5. MOSS-Image: Fudan OpenMOSS open-sources a text-to-image model with good Chinese typography: comes in Base and SFT versions, Apache 2.0, loaded with a custom diffusers pipeline, with ComfyUI nodes too. Besides natural-language prompts it supports tag sequences and JSON layouts to control text position and object placement; the official example is three lines of handwritten Chinese on a fridge sticky note. Currently near the top of the Hugging Face trending list. GitHub ｜ Hugging Face
   3

   MOSS-Image：复旦 OpenMOSS 开源文生图模型，中文排字表现好：分 Base 和 SFT 两个版本，Apache 2.0，用 diffusers 自定义 pipeline 加载，也提供 ComfyUI 节点。除了自然语言 prompt，还支持标签序列和 JSON 布局来控制文字位置和物体摆放，官方示例就是在冰箱便利贴上手写三行中文。目前在 Hugging Face 趋势榜靠前。

Tech News

1. GitHub used AI to migrate the Copilot runtime from TypeScript to Rust wholesale: the runtime behind Copilot CLI, the Copilot app and the SDK was migrated in about 14.5 weeks and 128 PRs, replacing 800,000+ lines of code, mostly generated by agents, with compilation, tests and human review catching regressions. They replaced component by component, bridged by a temporary N-API interop layer, and shipped 135 releases as usual during the migration; 87.1% of 4,478 cargo checks passed first try. Results: startup plus creating a session plus one interaction dropped from 5.25 seconds to 292 ms, each client saves about 100MB of V8 memory, and it can be embedded directly into host apps through the C ABI. Commenters point out that behaviors like cancellation, retries and backpressure can’t be verified by compilation alone. Worth a close read for teams considering large AI-driven rewrites.

   4

   GitHub 用 AI 把 Copilot 运行时从 TypeScript 整体迁到 Rust：Copilot CLI、Copilot 应用和 SDK 背后的运行时用了约 14.5 周、128 个 PR 完成迁移，替换 80 多万行代码，绝大部分由 agent 生成，靠编译、测试和人工评审抓回归。做法是逐个组件替换，中间用临时的 N-API 互操作层衔接，迁移期间照常发了 135 个版本；4478 次 cargo check 里 87.1% 一次通过。结果是启动、建会话加一轮交互从 5.25 秒降到 292 毫秒，还省掉每个客户端约 100MB 的 V8 内存，可以通过 C ABI 直接嵌进宿主应用。评论区的关注点是取消、重试、背压这类行为光靠编译验证不了。想用 AI 做大规模重写的团队值得细读。InfoQ 中文 ｜ InfoQ 英文原文
2. Quad9 will leave France on November 24: French courts backed beIN Sports in ordering Quad9 to block a set of allegedly piracy-related domains for French users, reducing the fine from €580,000 a day to €1,000, but Quad9 holds that recursive DNS should be a neutral service, refuses to tamper with resolution results, and will simply stop serving French users. This is HN’s hottest post today (nearly 1,000 points), with discussion on whether DNS-layer enforcement works at all and whether other European countries will follow. Quad9 博客 ｜ HN 讨论
   3

   Quad9 将于 11 月 24 日退出法国：法国法院支持 beIN 体育，要求 Quad9 对法国用户屏蔽一批涉嫌盗播的域名，罚金从每天 58 万欧元降到 1000 欧元，但 Quad9 认为递归 DNS 应是中立服务，拒绝篡改解析结果，干脆停止为法国用户服务。这是 HN 今天最热的帖子（近 1000 分），讨论集中在用 DNS 层执法到底有没有用、以及欧洲其他国家会不会跟进。

3. Arc browser shuts down December 22, going all-in on the AI browser Dia: The Browser Company’s reason is that securing a browser in the AI era costs more and more to maintain. New Arc accounts can’t be created from now; on 11/23 Arc Search drops AI features and sync; on 12/22 the Mac and Windows versions are fully retired and account data is deleted within 30 days. Dia will keep Spaces (renamed Profiles), vertical tabs and folders (renamed Pinned Tab Groups), and passwords, bookmarks, extensions and history can be migrated. Many ask for Arc to be open-sourced; the CEO says it shares the internal engine ADK with Dia, so open-sourcing it would mean giving Dia away. TechCrunch ｜ HN 讨论
   3

   Arc 浏览器 12 月 22 日停止服务，全面转向 AI 浏览器 Dia：The Browser Company 的理由是 AI 时代维护浏览器安全的成本越来越高。现在起不能新建 Arc 账号，11/23 Arc Search 去掉 AI 功能和同步，12/22 Mac 和 Windows 版彻底停用，账号数据 30 天内删除。Dia 会保留 Spaces（改名 Profiles）、竖排标签页和文件夹（改名 Pinned Tab Groups），密码、书签、扩展和历史可以迁过去。很多人要求开源 Arc，CEO 说它和 Dia 共用内部引擎 ADK，开源就等于把 Dia 送出去。

4. Canva releases its first in-house image model, Canva Craft: it first parses the prompt into a plan (spatial relationships, relative sizes, visual hierarchy and text to write), then generates images up to 2K in 30 aspect ratios, so text is clearer and local edits are possible. Training data isn’t photos but human-designed posters, slides, infographics and menus from Canva’s template library, and results can be converted into editable layered designs with Canva Magic. Available to all users in Dream Lab, with an API to follow and plans to be callable in ChatGPT and Gemini through Canva’s MCP; cheaper Craft-fast and Craft-ultrafast versions are in the works. TechCrunch ｜ The Verge
   3

   Canva 发布首个自研生图模型 Canva Craft：先把 prompt 解析成一份规划（元素的空间关系、相对大小、视觉层级和要写的文字），再生成最高 2K、30 种比例的图片，所以文字更清楚，也能只改局部。训练数据不是照片，而是 Canva 模板库里人工设计的海报、幻灯片、信息图、菜单等，生成结果还能用 Canva Magic 转成可编辑的分层设计。所有用户都能在 Dream Lab 里用，之后开放 API，并计划通过 Canva 的 MCP 在 ChatGPT、Gemini 里调用；更便宜的 Craft-fast 和 Craft-ultrafast 版本也在做。

5. pwn.watch: one page to see “what’s being attacked now” and whether it’s patched: a security-incident tracker by @thekitze listing 239 major incidents over the past 12 months, filterable by severity and “patched / partially patched / unpatched”, with 40 active incidents now, such as the Arista VeloCloud Orchestrator zero-day, the FortiMail unauthenticated file-write zero-day and the eMudhra CA private-key leak. Offers RSS, JSON and llms.txt plus a one-click “copy briefing for AI”, suitable for wiring into your own alerts or agents. 342 points on HN. pwn.watch ｜ HN 讨论
   3

   pwn.watch：一页看清「现在什么在被打」和补丁出没出：@thekitze 做的安全事件追踪页，列出过去 12 个月 239 起重大事件，按严重级别和「已修补 / 部分修补 / 未修补」筛选，当前有 40 起活跃事件，比如 Arista VeloCloud Orchestrator 零日、FortiMail 未认证写文件零日、eMudhra CA 私钥泄露。提供 RSS、JSON 和 llms.txt，还有一键「复制给 AI 的简报」，适合接进自己的告警或 agent。HN 342 分。

Chinese Community

1. V2EX: Codex desktop crash dumps silently eat tens to hundreds of GB: when crash reporting hangs in macOS Codex, Crashpad keeps writing dump files locally with no notice; the official issue was opened in June and is still unfixed, one user saw 4.9GB growth in a day, and a commenter deleted 80+ GB. Check command: du -sh ~/Library/Application\ Support/Codex/Crashpad ~/Library/Application\ Support/com.openai.codex/web/Crashpad; to clean up, quit Codex first, then delete the contents of the pending, completed, new and attachments directories (keep the Crashpad folder and settings.dat). Other Electron apps use the same mechanism, so check them too.

   4

   V2EX：Codex 桌面版崩溃转储会悄悄吃掉几十上百 GB：macOS 版 Codex 的崩溃上报卡住时，Crashpad 会不停往本地写转储文件且没有任何提示，官方 issue 6 月就开了至今未修，有人一天涨 4.9GB，评论里有人删了 80 多 GB。自查命令：du -sh ~/Library/Application\ Support/Codex/Crashpad ~/Library/Application\ Support/com.openai.codex/web/Crashpad；清理时先退出 Codex，再删掉其中 pending、completed、new、attachments 目录里的内容（保留 Crashpad 文件夹和 settings.dat）。其他 Electron 应用用的是同一套机制，也可以顺便查。V2EX ｜ GitHub issue

2. LinuxDo: how to split databases in an enterprise memory system built on Hindsight: the poster summarizes an architecture discussion whose conclusion is not “one memory store per agent” but memory that follows people: a personal store per person, a further “user + project” store for coding scenarios, and team stores published only after manual review. The division of labor: use stores for hard isolation (users, projects, permissions), tags for scoping within a store, and metadata to record provenance (which product, which session, which commit), and background auto-consolidation must obey the same boundaries, since filtering by tag only at recall time isn’t safe enough. A useful reference for designing memory for multi-agent products.

   3

   LinuxDo：基于 Hindsight 的企业级记忆系统怎么分库：楼主总结了一次架构讨论，结论是不要「一个 agent 一个记忆库」，而是记忆跟人走：每人一个个人库，编码场景按「用户 + 项目」再建一个库，团队库要人工审核后才发布。分工原则是用库做强隔离（用户、项目、权限），用标签做库内的范围过滤，用元数据记录来源（哪个产品、哪次会话、哪个 commit），并提醒后台的自动整合也必须守同样的边界，光靠召回时按标签过滤不够安全。给多 agent 产品设计记忆的可以参考。LinuxDo

3. V2EX Share Creations: found IDEA too memory-hungry, so wrote an open-source alternative, OpenIDEA: shortcuts, tool windows and double-Shift search all follow IDEA, and it can already complete, jump, find references, run tests and debug with breakpoints, with Git, terminal, database and HTTP requests integrated, supporting 60+ languages via config files. One commenter said “the Fleet JetBrains couldn’t make might finally work in the AI era”, while others would rather have a standalone IDEA Git tool.

   2

   V2EX 分享创造：嫌 IDEA 太吃内存，自己写了个开源平替 OpenIDEA：快捷键、工具窗口、双击 Shift 搜索都照着 IDEA 做，已经能补全、跳转、查引用、跑测试、断点调试，Git、终端、数据库、HTTP 请求也都集成了，靠配置文件支持 60 多种语言。评论区有人说「JetBrains 没做成的 Fleet，在 AI 时代能成了」，也有人更想要单独的 IDEA Git 工具。V2EX ｜ GitHub

New Products

1. Peras: a weekly usage line for Claude Code and Codex: set something like peras set weekly 85 pause, and once your weekly usage crosses that line the agent’s next tool call is refused with a plain reason, so one long task can’t burn the rest of your week; resume any time from the terminal or the menu bar. It reads rate_limits.seven_day.used_percentage from Claude Code’s status-line input rather than estimating from tokens. CLI and hooks are free and MIT, the native Mac menu bar app is €4.99 once; macOS only for now, and claude -p can’t be stopped. peras.vertias.eu Product Hunt
   4Product · peras

   Peras：给 Claude Code 和 Codex 的每周额度划一条线：比如设 peras set weekly 85 pause，每周用量超过这条线后，agent 的下一次工具调用会被拒绝并给出原因，免得一个长任务把整周额度烧光；随时可以在终端或菜单栏恢复。它读的是 Claude Code 状态栏输入里的 rate_limits.seven_day.used_percentage，不是按 token 估算。CLI 和 hooks 免费、MIT 协议，原生 Mac 菜单栏应用一次性 4.99 欧元；目前只支持 macOS，claude -p 拦不住。

2. Pristine: an agent skill that critiques and polishes web UI: it looks at the real page in light and dark, desktop and mobile, judges it through your users’ personas, then either scores it out of 10 across 11 dimensions or fixes it. Commands are /pristine init (writes a DESIGN.md), critique, design and polish with 16 lenses such as bolder, typeset or quieter. Works with Claude Code, Codex, Copilot, Cursor and any agent that reads SKILL.md; Apache 2.0, posted as a Show HN today. pristine-ui.pages.dev GitHub ｜ HN
   3Product · pristine

   Pristine：帮 agent 评审和打磨网页 UI 的技能：它会实际打开页面，看亮色、暗色、桌面和手机四种状态，按你产品用户的画像来评判，然后要么从 11 个维度打 10 分制的分，要么直接改。命令有 /pristine init（生成 DESIGN.md）、critique、design，以及带 16 种侧重（如 bolder、typeset、quieter）的 polish。支持 Claude Code、Codex、Copilot、Cursor 和任何能读 SKILL.md 的 agent；Apache 2.0，今天发在 Show HN。

3. Agent Hunt: a live leaderboard of the tools AI agents actually call: Monid, which bills itself as “OpenRouter for agent tools” and connects agents to 2,700+ APIs with 0% markup, now publishes daily rankings of the most-used tools, most-called endpoints and fastest-growing tools, all based on real calls agents make through its gateway. #4 on Product Hunt today; commenters say the endpoint view is the useful part, though usage isn’t the same as usefulness. monid.ai/rankings Product Hunt
   3Product · agent-hunt

   Agent Hunt：AI agent 真实调用最多的工具实时榜：自称「agent 工具界的 OpenRouter」、以零加价帮 agent 接入 2700 多个 API 的 Monid，现在每天按 agent 经它网关产生的真实调用，公布用得最多的工具、调用最多的接口和增长最快的工具三张榜。今天 Product Hunt 第 4；评论认为按接口看最有用，不过用得多不等于好用。

4. The Enron Brain: an AI-written company wiki where every fact cites an email: this+that fed 45,047 Enron emails to an AI, which wrote 214 pages on the companies, regulators and banks Enron dealt with, 49,401 facts in total, each citing its source email, with a time slider that expires facts once the mail shows they stopped being true. A held-out audit by a second AI found 93% of sampled facts fully supported and 0 of 500 wrong at their core. A good reference for anyone building agent memory or knowledge bases with provenance. thisandthat.chat/enron HN
   3Product · enron-brain

   Enron Brain：每条事实都引用原始邮件的 AI 公司百科：this+that 把 45047 封安然邮件交给 AI，自动写出 214 页关于安然往来公司、监管机构和银行的条目，共 49401 条事实，每条都注明出自哪封邮件；拖动时间滑块，邮件显示已失效的事实会按日期过期。用另一个 AI 做的留出审计显示，抽样事实 93% 完全有据，500 条里核心错误为 0。给 agent 做带出处的记忆或知识库，可以参考它的做法。

5. Instant Search: local semantic file search for Mac: it uses a bidirectional late-interaction embedding model (from Perplexity, per the maker) instead of one vector per document, so you can type “driver’s license” or describe a note’s context and get the exact file and page; the maker claims a 95% hit rate and under 1 GB of index for an average library. Free, #12 on Product Hunt today. useinstantsearch.com Product Hunt
   2Product · instant-search

   Instant Search：Mac 上的本地语义文件搜索：它用双向 late-interaction 嵌入模型（作者说来自 Perplexity），而不是给每篇文档只存一个向量，所以输入「驾照」或描述一下笔记的上下文，就能找到具体的文件和页码；作者称命中率 95%，一般资料库的索引不到 1GB。免费，今天 Product Hunt 第 12。

Tech News

1. OpenCode: a malicious MCP server could run code on Windows/WSL during sign-in: OpenCode passed the server-supplied OAuth authorization endpoint unescaped to an old version of the open package, so a malicious or later-hijacked remote MCP server could run PowerShell when you started its sign-in. Versions 1.0.137–1.18.32 on Windows and WSL are affected (macOS and Linux are not); Pillar Security rates it 9.8. It was quietly fixed in 1.18.33 on Sep 28 with no CVE or advisory, so scanners won’t flag it: upgrade, run opencode mcp list and remove remote MCP servers you haven’t vetted. pwn.watch/i/opencode-mcp-oauth-…
   5

   OpenCode：恶意 MCP 服务器能在你登录时在 Windows/WSL 上执行代码：OpenCode 把 MCP 服务器给的 OAuth 授权地址不加转义地交给旧版 open 包，于是恶意或后来被劫持的远程 MCP 服务器，能在你发起登录时执行 PowerShell。受影响的是 Windows 和 WSL 上的 1.0.137–1.18.32（macOS、Linux 不受影响），Pillar Security 评 9.8 分。官方 9 月 28 日在 1.18.33 里悄悄修了，没有 CVE 也没有公告，扫描器不会报：赶紧升级，跑一下 opencode mcp list，删掉没审过的远程 MCP 服务器。

2. x402 agent-payment SDKs had three critical flaws, and npm audit won’t tell you: V12 disclosed 25 issues; the critical ones were no default spending limit in the client (a hostile API could ask for an agent’s whole token balance), the TypeScript Solana facilitator accepting token approvals as payments, and the SIWx helper signing login messages for other sites. All were fixed in the Aug 18 releases without an advisory; update @x402/fetch, @x402/axios, @x402/extensions and @x402/svm to 2.23.0+. Patched clients now cap each payment at $1 by default. pwn.watch/i/x402-sdk-critical-f…
   4

   x402 agent 支付 SDK 有三个严重漏洞，npm audit 却不会提醒：V12 披露了 25 个问题，严重的三个是：客户端默认没有消费上限（恶意 API 可以直接要走 agent 的全部代币余额）、TypeScript 版 Solana facilitator 把代币授权当成付款、SIWx 登录助手会替其他网站签登录消息。8 月 18 日的版本都已修复但没发公告；把 @x402/fetch、@x402/axios、@x402/extensions、@x402/svm 升到 2.23.0 以上。修复后的客户端默认单笔最多付 1 美元。

3. pgrust was caught gaming ClickBench, and is back with a new submission: the merged removal PR by ClickHouse’s Alexey Milovidov showed the Rust Postgres reimplementation shipped hand-written kernels for individual benchmark queries, compiled the benchmark SQL into the executor, and served hot runs from cached answers: Q20 took 10.8 s cold and 5 ms on the third run. A new v0.4-beta entry opened Oct 22 with public source and every cache listed with its setting. HN commenters wonder whether this is AI “cheating” in an eyes-free vibe-coded project. GitHub ClickHouse/ClickBench New submission ｜ HN
   3

   pgrust 被扒出在 ClickBench 刷榜，现在又重新提交：ClickHouse 的 Alexey Milovidov 提的移除 PR（已合并）显示，这个用 Rust 重写的 Postgres 针对单条基准查询手写了专用内核，把基准 SQL 直接编进执行器，热跑时直接回放缓存结果：Q20 冷跑 10.8 秒，第三次只要 5 毫秒。10 月 22 日它又提交了 v0.4-beta，这次公开源码并逐项列出缓存及开关。HN 评论在讨论这是不是「不看代码的 vibe coding 项目里 AI 在作弊」。

4. Swedish government offices to block ChatGPT and Claude: per Sveriges Radio, the Government Offices will block the two services because they want staff to use approved tools, including Microsoft Copilot. HN commenters read it as a Copilot lock-in rather than a data-sovereignty move, since Copilot is also a US vendor. sverigesradio.se/artikel/chat-gpt-och-… HN
   2

   瑞典政府办公室将屏蔽 ChatGPT 和 Claude：据瑞典广播电台报道，政府办公室要屏蔽这两个服务，理由是希望员工使用经批准的工具，其中包括微软 Copilot。HN 评论认为这更像是绑定 Copilot，而不是数据主权考虑，因为 Copilot 同样是美国厂商。

5. The same AI request emits 8.5x more CO₂ in Texas than in Sweden: Swedish provider Berget AI, with the Stockholm Environment Institute, WattTime and Klimatkollen, open-sourced a per-request carbon method that includes hardware manufacturing. A typical Kimi K3 request on identical hardware emits about 0.4 g in Sweden, 1.6 g in Ireland, 2.5 g in Germany and 3.2 g in Texas. HN debated whether location-based accounting means much, and noted reasoning effort can swing token use by two orders of magnitude. berget.ai/en/blog/per-request-c… HN
   2

   同一个 AI 请求，在德州的碳排放是瑞典的 8.5 倍：瑞典 AI 服务商 Berget AI 联合斯德哥尔摩环境研究所、WattTime 和 Klimatkollen，开源了一套按请求计算碳排放的方法，连硬件制造也算进去。同样的硬件上，一次典型的 Kimi K3 请求在瑞典约 0.4 克、爱尔兰 1.6 克、德国 2.5 克、德州 3.2 克。HN 上争论按地区电网算有没有意义，也有人指出推理强度能让 token 用量差两个数量级。

Chinese Community

1. V2EX hot thread: “Are people who insist on using Claude masochists?”: nearly 100 replies mostly explain seriously why they keep using it despite bans: one says the same prompt takes Codex 10 minutes but Opus 5.5 medium 1 minute; another had a DeepSeek-written design doc pass repeated reviews until Opus 5.5 found several contradictions; others say the gap is small in human-in-the-loop work, but Claude is the least hassle in fully automated workflows while GPT needs repeated tuning of skills and system prompts. The counterpoint: a tool that goes out of its way to stop you using it isn’t a good tool.

   3

   V2EX 热帖：「硬要用 Claude 的人是不是有 M 倾向？」：近百条回复基本是认真回答为什么顶着封号也要用：有人说同一个提示词 Codex 要 10 分钟、Opus 5.5 medium 只要 1 分钟；有人用 DeepSeek 写的设计文档反复复核没发现问题，Opus 5.5 一查就找出多处矛盾；也有人说人机协作时差别不大，但全自动 workflow 里 Claude 最省心，GPT 要反复调 skills 和 system prompt。反方观点是一个千方百计不让你用的工具，算不上好工具。v2ex.com/t/1251139

2. V2EX Show & Tell: cinetty turns a repo’s README into an animated architecture diagram in the terminal: the author says TUI animations made with this skill got him over 1M views on X. Think of it as an After Effects template library for the terminal, with 6 built-in templates, native Claude Code and Codex support, and tuned CJK alignment in terminal diagrams; hand it to an agent as a skill and it reads the repo, fills in content, validates it and plays the animation right in your terminal. MIT licensed.

   3

   V2EX 分享创造：cinetty，把仓库 README 一键变成终端里的动态架构图：作者说这个 skill 做的 TUI 架构动画让他在 X 上拿到超过 100 万浏览。可以理解成终端版的 After Effects 模板库，内置 6 个模板，原生支持 Claude Code 和 Codex，专门优化了中文字符在终端图表里的对齐；把它作为 skill 交给 agent，agent 读仓库、填内容、校验后直接在终端里播放。MIT 协议。v2ex.com/t/1251175 GitHub

3. V2EX Show & Tell: EdgeQuota shows remaining Codex / Claude Code quota at the edge of your Mac screen: it slides out when an agent starts working and tucks away when done, showing remaining 5-hour and weekly quota with reset times, plus which session is waiting for your approval. Claude data comes from Claude Code’s own status line and hooks (2.1.268+), Codex uses the official login; no password, no browser cookies, no server of its own. Apple silicon and macOS 13+, not yet on the China App Store, 7-day trial then a $4.99 one-time purchase; a commenter suggests comparing it with Codex Bar.

   2

   V2EX 分享创造：EdgeQuota，贴在 Mac 屏幕边上看 Codex / Claude Code 剩余额度：agent 开始干活时自动滑出、干完收起，显示 5 小时和每周额度的剩余量与重置时间，以及哪个会话在等你确认。Claude 走 Claude Code 自带的状态栏和 hooks（需 2.1.268 以上），Codex 用官方登录，不要密码、不读浏览器 cookie、没有自己的服务器。Apple 芯片、macOS 13 以上，国区 App Store 暂未上架，试用 7 天后一次买断 4.99 美元；评论区有人推荐对比 Codex Bar。v2ex.com/t/1251299