April

后端程序员｜AI Builder｜用 AI 造工具，只发自己跑过的东西

[April @April1qing]

Codex这次更新，让我还上欠下的「知识债」。

我用 Space / Pages 存下工作中没弄懂的问题、原对话和学习提示词。有空点一下，就能接着学。

AI 帮我把事情做完，我也想把里面的知识弄懂。

#Codex #ChatGPT

[歸藏(guizang.ai) @op7418]

简单试了一下 Deepseek V4.1 Flash 1017，感觉不错呀！

用的是我之前给 Opus 5.5 写的那个做我的游戏特效的提示词。

虽然细节上和创造性上差 Opus 5.5 还是不少吧，但是是能用的

[April @April1qing]

Theo 拿同一个 agent 循环测了三家的提示缓存：228 轮里 Claude 和 GPT 每轮都命中，Gemini 4 Argon 只命中 73%。两轮之间隔 120 秒时，它的成本是完美缓存的 1.48 倍，这笔钱在单价表上看不出来。#Gemini #AI编程
https://t.co/iFojvzjeSR

[Theo - t3.gg @theo]

Gemini is 50% more expensive than it should be because their cache doesn't work.

27% cache miss rate. Insane.

[April @April1qing]

DeepSeek 管这次叫小更新，结果 Terminal-Bench 从 31.2 涨到 42.3，DeepSWE 拿到 77.0，表里的 Opus 5 是 74.0。我更在意 Thom Wolf 点出来的那句，输出更短、任务完成更快。agent 每一步都在花 token，输出短了，整轮的账单就跟着降。#DeepSeek #开源模型
https://t.co/ZfELNGdg2h

[Thomas Wolf @Thom_Wolf]

we’re getting used to it by now but DeepSeek calling this a “minor update” while moving Terminal-Bench 4.0 from 31.2 to 42.3 is very funny.

also reports 77.0 on DeepSWE (vs 74.0 for Opus 5 and 73.0 for GPT-5.6 Sol…)

and the most interesting line may be faster task completion with shorter outputs. using fewer tokens is one of the practical way the cost curve collapses often done as an afterthought

all with MIT-licensed weights in DeepSeek tradition

now curious to see how these numbers will hold up outside their own eval setup

[April @April1qing]

Jack Clark 今天在 IMF 的讲座上说，他认为奇点已经开始了。会后有快讯号把 44 万亿美元 GDP 说成 Anthropic 的内部预测，他马上出来纠正：这是早就公开的经济情景工具，而且是参数调到「极端」时的结果。#Anthropic #AI
https://t.co/AQHNCqzeqF

[Jack Clark @jackclarkSF]

@DeItaone this isn't an internal model, it's the public economic scenario explorer we published a while ago - and this is under the "extreme" parameter setting https://t.co/DTGOqYFmKa

[April @April1qing]

Codex 没加额度，用户却能多用一些。Tibo 说靠的是三件事：子代理之间复用上下文，缓存保留得更久，再改进请求路由让缓存命中更多。agent 的成本大头本来就是一遍遍重发的长上下文。#Codex #OpenAI
https://t.co/QR3jCWXXJq

[Tibo @thsottiaux]

Day 12/

We’ve made improvements to Codex & Work that increase the effective usage you get. We achieved this by reusing context across subagents, keeping context cached longer, and improved request routing that allows us to increase cache hits.

[April @April1qing]

Claude Code 2.1.302 的更新日志里藏着一行：独立版 Claude Design 会在 12 月 14 日前关闭，提示就加在 /design-sync 命令里。用 claude -p 或 Agent SDK 跑自动化的也留意一下，fork 子代理现在默认打开了。#ClaudeCode #AI编程
https://t.co/kY9Nc5rTLI

[Claude Code Changelog @ClaudeCodeLog]

Claude Code CLI 2.1.302 changelog:

New features:
• Added CLAUDE_CODE_AUTO_MODE_MODEL to choose the model the client-side auto mode classifier uses; it does not change the model for server-side auto mode
• Added an Idle compaction row to /config for turning idle compaction off or back on; the notices shown after an idle compaction now say how to turn it off
• Added a notice to /design-sync and /design-login that standalone Claude Design (https://t.co/SlCZPJxegD) closes by December 14, 2026

Fixes:
• Fixed a "Prompt is too long" error when stopping a turn while the conversation was being compacted
• Fixed requests failing with "diagnostics: Extra inputs are not permitted" after switching to Vertex or Bedrock mid-session
• Fixed the Read tool loading an image file of any size, or a pipe or device named like an image, entirely into memory; image sources over 128 MB are now refused
• Fixed /autocompact and /effort saving to the wrong model when a gateway names Opus 4 or 4.1 with a prefix, such as anthropic/claude-opus-4-1
• Fixed a prompt cache miss when a tool added mid-session was loaded again through ToolSearch or after compaction
• Fixed claude plugin and other subcommands sometimes exiting before a queued settings or config write finished, which could lose the change or leave a .tmp file behind
• Fixed plugin marketplaces re-cloning on every refresh, and newly installed git plugins showing version "unknown", when git defaults to the reftable ref format
• Fixed the PowerShell tool not finding commands in a plugin's bin/ folder, and reporting an error when a search that finds nothing (rg, grep, findstr, git grep) is piped into Select-Object or another formatting cmdlet
• Fixed Claude in the desktop app pointing you to https://t.co/ldz5BaC3j0 settings or /mcp to authorize or reconnect an MCP server where that option isn't available
• Fixed artifact publishes sometimes failing when Claude sent an empty list of edits, and Claude retrying an artifact update the wrong way after a newer version of the artifact turned it away
• Fixed classifier_category missing from OpenTelemetry tool_decision events for most auto mode classifier blocks, and new tool_decision attributes pushing out tool_use_id, tool_source and tool_parameters when OTEL_LOGRECORD_ATTRIBUTE_COUNT_LIMIT is low
• Fixed rows in the agents view staying blank after moving down a long list or folding a section
• Fixed claude remote-control --help not printing its usage and exiting when not signed in
• Fixed /design consent and /design revoke suggesting /design-login when they fail in a cloud session, where it isn't available; the message now points to https://t.co/yMqdb2qm2F
• Fixed MCP servers showing no tools, or every tool call failing, when the server negotiates protocol version 2026-07-28 but omits resultType, ttlMs or cacheScope
• Fixed prompt suggestions spending usage credits on models your plan pays for only with credits: they are now skipped after a reply billed to credits
• Fixed URLs written in backticks being clickable only on their first row when they wrap; the whole URL is now one link in terminals that support hyperlinks
• Fixed claude plugin details <name>@builtin suggesting claude plugin list, which never shows built-in plugins, when the plugin isn't found; it now points to /plugin
• Fixed an emoji that a link starts inside wrapping one column too wide, which hid it in question previews and turned tables into lists
• Fixed worktree isolation refusing to create or resume worktrees under git 2.56 or later when a stray .git file points at a directory that doesn't exist
• Fixed a cloud session's cost and usage restarting from zero after the session resumed in a new process
• Windows: Fixed every file write being refused or asked about for projects on exFAT or FAT32 drives, including in sessions started with --plugin-dir

Improvements:
• Improved startup time with many plugins enabled: plugin command and agent files now take far fewer file system calls to load, most of all on Windows
• Improved auto mode with the server-side classifier: a conversation too long for the safety check is now compacted, instead of every checked tool call being denied
• Improved the message when a background session reopened from another folder has no LLM gateway: it now says why, instead of "Not logged in"

Security/safety changes:
• Changed claude -p and Agent SDK sessions with no permission mode to start in default mode on third-party providers, with telemetry off or on a fresh install; --permission-mode auto keeps auto
• [VSCode] Fixed MCP server sign-in from the panel opening a broken authorization page when the server URL contains a query string

Other changes:
• Self-hosted runner: Fixed git fetches failing with "bad object refs/heads" in sessions that reuse a checkout stored in git's reftable format
• Changed fork subagents (subagent_type: "fork") to be on by default in claude -p, the Agent SDK, the desktop app and VS Code, as they already are in the terminal; CLAUDE_CODE_FORK_SUBAGENT=0 turns them off, and claude mcp serve never offers them
• Changed the desktop app and VS Code to always run the main conversation's subagents in the background, as the terminal does, rather than let Claude choose; a subagent's own Agent calls in claude -p, the Agent SDK, the desktop app and VS Code now run in the foreground by default
• Changed the ← for agents footer hint to hide 10 seconds after a session opens; the count of agents waiting on you still shows
• Changed where the Claude in Chrome instructions sit in the system prompt so that new conversations can read them from the shared prompt cache
• Changed prompt and agent hooks so that at most 12 of one event run at a time; the rest wait their turn
• Changed the message shown when your organization's policy turns off /feedback: it now points to GitHub issues instead of an organization admin
• [VSCode] Fixed text after an unclosed code block in a list showing as plain text while a reply streams, then turning into code
• [VSCode] Fixed the @browser banner saying "Browser connected" when no browser with the Claude extension can be reached
• [Cloud sessions] Added a banner on a routine's page that says why its last run didn't start, such as your organization reaching its limit on concurrent cloud sessions
• [Cloud sessions] Fixed "Rewind to here" failing on every message except the newest in cloud sessions; self-hosted runners need Claude Code v2.1.246 or later
• [Cloud sessions] Fixed the pull request status bar in cloud sessions flipping a pull request with merge conflicts back to ready while GitHub rechecked it
• [Cloud sessions] Fixed routine runs showing as Completed after their cloud session stopped unexpectedly, and as Cancelled when archived before they started, instead of failed
• [Cloud sessions] Fixed the diff view in cloud sessions missing changes or failing to load after the session's GitHub repository was renamed or moved
• [Cloud sessions] Fixed a scheduled routine run that failed partway being run again from the start an hour or more later, which could repeat work such as posting the same report twice
• [Cloud sessions] Fixed an open cloud session sometimes not showing messages sent from another device until you reloaded the page or clicked away and back
• [Cloud sessions] Improved the error shown when a cloud session fails to start: it now says "Connection lost" or "Server error" instead of "Something went wrong"
• [Remote Control] Fixed a subagent's pane spinning forever in a Remote Control session after your computer goes offline; it now says Remote Control is disconnected
• [Claude Tag] Fixed "is thinking…" staying up under a mention in a Slack channel or DM after Claude answered it by editing one of its earlier messages
• [Claude Tag] Fixed URLs in code blocks under a numbered or bulleted step of Claude's Slack replies being wrapped in angle brackets, which broke copied commands
• [Claude Tag] Fixed HEIC, SVG, TIFF and BMP images attached in Slack never reaching Claude; they now arrive as files in its working files, like other file types
• [Claude Tag] Fixed new channel managers losing access to Claude in Slack in organizations that restrict Claude Tag by role; the manager role now also grants Claude Tag in Slack
• [Claude Tag] Fixed custom connectors whose MCP servers share one host showing as a single connector in Claude Tag admin settings; each server is now its own connector
• [Claude Tag] Fixed Claude in Slack sometimes answering a direct message with "DMs are temporarily unavailable" and dropping it when a single Slack lookup stalled; it now retries first
• [Claude Tag] Fixed direct messages with Claude on a Slack Enterprise Grid ignoring Slack's workspace dropdown; they now run under the connected workspace you last opened Claude in
• [Claude Tag] Fixed Claude in Slack's notice for a channel's monthly spend limit saying a channel manager can raise it; it now names an organization owner and links the usage page
• [Claude Tag] Fixed links in Claude's Slack replies absorbing the closing bold or italic marks, or full-width punctuation, right after them, which broke the link and the formatting
• [Claude Tag] Improved repository search in Claude Tag admin settings to cover up to 100 connected GitHub accounts and find private repositories past the first 5,000 of a GitHub App install
• [Claude Tag] Improved Add place in Claude Tag admin settings to list a channel you type or paste right away, instead of only after your organization's full channel list has loaded
• [Claude Tag] Improved Add > Domain on a workspace's or channel's page in Claude Tag admin settings to be available once that page loads, not only after every channel has loaded
• [Claude Tag] Changed Claude Tag admin settings to organize Claude's access into four tabs: Bundles, Connectors, Skills and plugins, and Channels
• [Claude Tag] Changed the Roles page in organization settings to group the roles created for Claude Tag channel managers in a collapsed section below your organization's own roles
• [Code Review] Fixed Code Review setup not resuming after you leave to connect GitHub; back in admin settings, it now reopens at the repository picker

Source: https://t.co/H3KSN7Rs8M

[向阳乔木 @vista8]

写好文章，发布公众号、X、小红书，操作非常繁琐？

推荐新开发的 Obsidian 插件 Qiaomu Publish

请看功能演示，自媒体必备，开源地址见评论区

[April @April1qing]

TypeScript 编译器原本用 TS 写，后来微软用 Go 重写，现在 Vercel Labs 又把它移植回 TypeScript，编译成原生程序，说比 Go 版快 2 到 4 倍。他们拿 TS 的一万两千多个测试文件当答案，每一段都和 Go 版逐项对比。#TypeScript #AI编程
https://t.co/Ot9BXpNc8n

[Chris Tate @ctatedev]

New from Vercel Labs: tsc-ts

TypeScript was written in TypeScript
Then rewritten in Go

We ported it back to TypeScript
Compiled it with scriptc

And it's 2-4x faster than tsgo

[April @April1qing]

做 Amp 的 Thorsten Ball 问了个挺冲的问题：有了 agent，为什么还要 PR？这周 Anthropic 那篇评审文章走的是另一条路，PR 还在，只是交给分类器按风险决定要几个人批准。两种做法摆在一起看，挺有意思。#AI编程 #CodeReview
https://t.co/cHuNg61qAV

[Thorsten Ball @thorstenball]

Why are you still using pull requests?

Is it for code review?

If so, do you really think your or your colleagues catch bugs and mistakes that an Opus 5.5 or Fable or Astra won't find? You can literally spawn all three of them and have them check four times each.

(Also, don't tell me you don't have a "stamp plz" Slack channel or group chat somewhere.)

Is it for knowledge sharing?

Surely there's better ways for that now than the PR, no? We have models that can generate explainer videos, presentations, podcasts, games, quizzes, walkthroughs, ... for any piece of code it can reach.

Is CI the reason?

Well ackshually, CI means continuous integration and its original meaning is the *opposite* of having long-lived branches and PRs.

But, leaving that aside, your agents are hopefully already running the tests before they push something, and hopefully they can run as long as they want and don't block you so they can run the whole test suite and generate "irrefutable proof" and end-to-end demo videos and so on... And so they run the tests and confirm and then you push to a branch and everything runs again and then you merge and then it runs again? That doesn't make sense, does it?

Maybe you work in open source and you accept contributions, in which case I can't speak to the value of pull requests. It's been 2 years since my last OSS contribution. But from what I heard, the time of the PR is coming to an end in OSS too?

I honestly don't get why people still stick to PRs.

With agents you can now move 10x faster, you can finally do real continuous integration, and get more confidence in the changes the agent made.

Don't slow yourself down by having all the changes pile up in a PR.

[April @April1qing]

Politico 报道，Anthropic 私下给国会递了一份前沿 AI 安全法案草案。按草案，政府要叫停一个模型，得先证明它马上可能造成 10 亿美元以上损失或 50 人以上伤亡，还要法院批准。安全圈不太买账，Kelsey Piper 觉得 Anthropic 更怕的是被政府随手叫停。#Anthropic #AI监管
https://t.co/grCet7lkfn

[Kelsey Piper @KelseyTuoc]

One unfortunate conclusion I'm reaching from this is that Anthropic, while acknowledging they're doing something extremely dangerous, is more importantly wrt policy terrified of the executive branch having the ability to flexibly make them stop at the drop of a hat.

This isn't shocking given that they've been on the receiving end of the executive branch doing that in an unprincipled and stupid way! Anyone who has to deal with a bureaucracy becomes a libertarian, and the administration's treatment of Anthropic was genuinely outrageous.

But it means that they're interested in giving themselves more process-based protections when everyone else is interested in giving us more process-based protections against them. The AI safety movement wants the ability to shut down superintelligent model development. Anthropic wants to significantly curtail current authorities to do that so that it has to be narrowly tailored to a specific imminent threat in only a few possible categories. The AI safety movement wants the states to be able to pass stronger laws than it can get through federally. Anthropic wants preemption.

You also kind of have to wonder if the Global Affairs people and the "we are terrified that we're about to kick off RSI" people at Anthropic talk to each other.

Global affairs: So we've been working very diligently to make sure the Trump administration can't just blacklist Fable 6.5 since we didn't donate as much money to his ballroom as Greg Brockman did.

Technical staff: We think Fable 6.5 might be able to copy itself into an unmonitored data center and start doing experiments we can't observe.

Global affairs: Sounds crazy, hope you're wrong! Anyway, we're lobbying for no further state laws around safety, since it's really expensive to fight ill-conceived laws in 50 state legislatures at once.

Technical staff: Money may or may not still have any meaning in 36 months.

Global affairs (giggling nervously): It's that San Francisco air, man. Everyone gets like that if they move there. Anyway, obviously if the government makes a narrowly tailored finding that an AI poses an imminent risk of causing more than 50 deaths-

Technical staff: We give it a 10 percent chance it'll cause 8 billion.

[April @April1qing]

@easybakeowein welcome to Shanghai! if jet lag lets you, walk the Bund at night then cross to Yu Garden early the next morning before the tour groups. and get Alipay set up with your card, it makes everything easier. enjoy the trip!

[April @April1qing]

@realjustinkim for cafes like that, wander the old lanes around Wukang Rd and Anfu Rd in the former French Concession. tons of tiny spots in old houses. Columbia Circle is nice too. go on a weekday morning, weekends get packed. enjoy the trip!

[April @April1qing]

@clarkeskru 19h is plenty! Get the transit visa-free stamp at Pudong, then Maglev + metro into town (~1h). Yu Garden early before crowds, xiaolongbao for lunch, the Bund at dusk when the lights come on. Leave 3h+ to get back. Happy to send my full layover route if you want it

[Tencent AI @TencentAI_News]

A 3B active model beats Opus 5 at memory. We open-sourced the full stack — model, engine, and app.

Hy Memory handles chat, work, and code memory in one layer, and it's first on all four benchmarks we ran, leading MemoryArena by 10.2 points while cutting memory volume 51.8% and recall tokens 39.3%.

The model isn't tied to our engine. Drop Hy-Memory-30B-A3B into Mem0, MemOS, or EverMemOS and it still comes out on top, even on the one system it never trained on. It's already running in our own products, where one assistant's multi-hop recall went from 50% to 90%. It also unifies memory across Codex, Cursor, WorkBuddy, and more.

Repo: https://t.co/M7PkOJVyE5
Weights: https://t.co/U04mNgmVuS

Questions welcome in the replies.

[Caine Builds 🐧 @CaineArdayfio]

Mobile Use + Jev for audio = ULTRAFAST⚡️

10x faster and 20x cheaper than Gemini Live.

It's our Audio Action model, Thud.

Thud listens and then does. It has a pre-built list of actions it can take on every page and listens for that action.

It's 2026. We don't need to transcribe every voice command before spoon-feeding it to the AI. The AI is smart enough to just hear audio and do! It doesn't need text.

[铜匠AI・十点睡觉 @CopperForgeAI]

OpenAI 官方博客披露了核心团队用 GPT-6 系列维护代码库的一手工程实践。直接公布了几个实际运行的提示词，
靠这些提示词在两周内合并了 1400 多个清理 PR，删除超过 800k 行代码，并将一个服务的容器镜像压缩48%。

第一个提示词用于清理死代码

“请帮我找出并清理仓库里的无用代码。你可以随意查看
Slack、Statsig（功能开关平台）或任何需要的来源，用来确认这些代码确实不再被使用、可以被清理或简化。”
这个提示词的关键在于允许 GPT-6 Astra 自主去翻阅 Slack 讨论、Statsig 实验配置甚至生产环境的性能 profiling数据，只有在确认生产环境零调用后才在 PR 中列出删除依据。

第二个提示词用于清理更棘手的废弃实验与功能开关

“/goal 找出整个仓库中所有陈旧、常量化、已废弃或全量上线的 Statsig开关、配置和实验，在不改变生产行为的前提下安全移除。审计所有项目资产，不限于标记为陈旧的对象。通过枚举、常量、包装函数、别名和动态拼
接名称寻找间接引用。核对生产引用的常量，不只给保守清单。结合生产遥测与性能分析评估 CPU 开销，找出昂贵的无用判断。生成 Markdown报告，按归属和文件工作流分组以便并行处理，附带生产取值、证据与 CPU 开销。不要修改线上配置或服务。”

第三组提示词用于 CI 优化与规则自省

“关于减轻本仓库的 CI 负载你有什么想法？请将想法列成表格，并根据该仓库的历史运行记录，按每天预估节省的 CI 分钟数进行排序。”

面对不稳定测试修复，他们运行：

“/goal 查看所有最近的 CI 运行记录，找出偶发失败的测试并尝试修复，成功则提交PR。你的任务是针对简单问题找到明确修复方案，而不是穷尽修复所有问题。”

针对 agent 上下文规则的自检提示词更精妙：

“翻阅我最近的 Codex 会话记录，找出一条误导了 agent 或导致无效工作的 AGENTS.md指令。选出该会话提示词，在相同仓库快照和模型下重放两次：一次保持原指令不变，一次仅移除或修正该指令。对比任务质量、agent
行为、工具调用、token 消耗与成本，展示该指令并说明修改是否有效。”

强烈建议好好学习和利用这些应用于超大工程实践的提示词

[April @April1qing]

@AlexFinn @bot 看下这个视频

[April @April1qing]

@ScarletKc 把 Fable 5.1 排在 Astra 前面挺有意思。实际用下来，两者在复杂任务上差的是深度还是稳定性？

[April @April1qing]

@lxfater 花几万攒 V100 跑 DeepSeek V4.1，结果 Opencode 上免费了，这心路历程太真实了。现在那条免费线路排队和限速怎么样？

[April @April1qing]

@dotey 额度按周单独算，不占 3.8 Flash 的份，这点对长期用 Antigravity 的人挺关键。每周到底给多少、哪天重置，文档里还没写，有人已经收到了吗？