Community posts

ONE CLAUDE SKILL IS ALL YOU NEED TO START MAKING MOTION VIDEOS it's called brag. it goes through your project and builds a short promo with motion graphics, music and copy for the launch it picks the product angle from your codebase, so you don't have to figure out the whole story from scratch. there's even a lean version built specifically for Opus 5.5 perfect for SaaS launches, side projects and client apps that deserve more than a screenshot of the settings page. people scrolling past haven't spent three weeks building it with you, they need a reason to care to try it in Claude Code: → /plugin marketplace add latent-spaces/brag → /plugin install brag@brag → open your project and run /brag

What if you were the answer to someone’s AI prompt? Built a Folkso skill for Codex & Claude Code. Create a profile to be found for your skills. Or find people you’d actually want to build with. https://github.com/saplq/folkso-plugin

LangSmith now has a first-party Claude Code plugin that traces turns without manual stop hooks. 1. /plugin marketplace add langchain-ai/langsmith-claude-code-plugins, install langsmith-tracing, reload. Set TRACE_TO_LANGSMITH, CC_LANGSMITH_API_KEY, CC_LANGSMITH_PROJECT in .claude/settings.local.json 2. Each turn becomes a trace; same session shares thread_id. System prompts stay out. Secret redaction runs on-box before upload (default on) 3. Mute a thread with /langsmith-tracing:mute to keep structure and tokens but strip I/O. For CI, pass the plugin into anthropics/claude-code-action with PR metadata Operator move: enable it on one production repo today, mute by default if you are privacy-sensitive, and review failures in the Threads tab instead of grepping local transcripts. https://docs.langchain.com/langsmith/trace-claude-code

The Sonnet 5.5 and Opus 5.5 releases have completely reshuffled how I'd set up a multi-model agent stack. Opus dropped less than a week ago, and Sonnet is already changing the economics of delegation. I've been running Opus as my orchestrator, two specialized Sonnet workers, and Codex as an independent engineer/reviewer up until now. Here's how I'd update that setup to optimize output quality, reasoning effort, and token usage: 1. Updated model assignments Opus 5.5 (medium) → Main orchestrator. Planning, architecture, delegation, difficult decisions, and synthesis. Haiku (low) → fast-worker. Formatting, boilerplate, repetitive edits, and fully specified mechanical tasks. Sonnet 5.5 (medium) → nuanced-worker. Context-aware implementation, tests, refactors, UI, documentation, and scoped development. Opus 5.5 (high) → deep-reasoner. Complex debugging, architectural tradeoffs, algorithm design, and difficult problems requiring independent investigation. GPT-6 Sol / Codex (medium/high) → Independent senior engineer. Parallel implementation, second opinions, debugging, and adversarial review. Fable 5.1 / GPT-6 Astra → Frontier specialists. Extended autonomous work, demanding reasoning, and problems where their particular capabilities justify the additional cost. No reason to run your most expensive model on everything. 2. Claude Code setup Update Claude Code: claude update Set your main model and reasoning effort: /model opus /effort medium Keep Opus as your technical lead. It owns the objective, decomposes work, delegates, and integrates the results. You don't need to create a separate orchestrator agent. Your main Claude Code session already handles that responsibility. 3. Create three specialized subagents Create these files in ~/.claude/agents/ (or update your existing ones). Use .claude/agents/ instead if you want project-specific configurations. Each file uses YAML frontmatter followed by its instructions. fast-worker.md model: haiku effort: low Instructions: Handle only fully specified mechanical work. Formatting, repetitive edits, boilerplate, and straightforward transformations. No architecture, refactoring, or guessing. Stop and hand back anything ambiguous. nuanced-worker.md model: sonnet effort: medium Instructions: Handle implementation requiring context and care. Read surrounding code, follow established conventions, write tests, perform scoped refactors, and verify results. Escalate architectural decisions rather than making them independently. deep-reasoner.md model: opus effort: high tools: Read, Grep, Glob Instructions: Independently investigate difficult technical problems. Consider competing explanations and tradeoffs. Return evidence, conclusions, remaining uncertainty, and actionable recommendations. No implementation changes. Include name and description in each agent's YAML frontmatter so Claude Code can discover and route to them. The model and effort fields explicitly pin each agent's configuration, independently of your main session. I'm trialing Haiku for mechanical tasks rather than assuming it's automatically more efficient. If its retries outweigh the savings, Sonnet 5.5 stays there. 4. Install Codex inside Claude Code Run: /plugin marketplace add openai/codex-plugin-cc /plugin install codex@openai-codex /reload-plugins /codex:setup If needed, install the CLI separately: npm install -g @openai/codex codex login Configure ~/.codex/config.toml: model = "gpt-6-sol" model_reasoning_effort = "medium" Useful commands: /codex:review → Independent code review /codex:adversarial-review → Challenge an implementation or architectural decision /codex:rescue → Delegate a difficult task /codex:rescue --background → Run independently in parallel /codex:status → Check progress /codex:result → Retrieve findings Treat Codex as a peer, not a rubber stamp. For consequential decisions, have Opus and Codex independently investigate the same problem without showing either the other's answer. Compare the evidence before choosing an approach. 5. Add orchestration rules to ~/.claude/CLAUDE.md Your main session needs to know WHEN to use these agents, not just that they exist. Add: You are the technical lead. Own planning, architecture, delegation, integration, and verification. Delegate based on context cost first, then task complexity. Complete trivial changes directly rather than unnecessarily spawning agents. Route fully specified mechanical work to fast-worker. Route context-sensitive implementation, tests, UI, docs, and scoped refactors to nuanced-worker. Route bounded, difficult reasoning to deep-reasoner. Use Codex when an independent engineering perspective would materially improve the outcome. For each delegation, provide relevant files, expected output, constraints, and acceptance criteria. Keep exploratory work out of the main conversation. Require concise results, files changed, and verification evidence. Don't let agents modify overlapping files concurrently without isolation. Verify outputs through tests, builds, and acceptance criteria. Diagnose failures before escalating. 6. Make reasoning escalation explicit This is important. Don't permanently increase your orchestrator's reasoning budget just because an occasional problem requires it. For difficult, self-contained reasoning, delegate to deep-reasoner. It runs Opus high while your main session stays on medium. If the orchestrator itself needs deeper reasoning, manually switch: /effort high Or, for particularly difficult problems: /effort xhigh Return to your default afterward: /effort medium This is an actual reasoning-effort configuration change, not simply prompting the agent to "think harder." For particularly demanding Codex tasks, you can also invoke Astra when your plan supports it: /codex:rescue --model gpt-6-astra --effort high Fable remains available through /model fable when your Claude plan supports it. Switch back to Opus afterward. Don't reserve these models exclusively for failed attempts. Sometimes the characteristics of the task warrant using them from the outset. 7. The optimization logic Sonnet 5.5 now scores 70.6% on Terminal-Bench 4.0 at its published peak configuration, costs $2/$10 per million input/output tokens, and generates output 30%+ faster than Sonnet 5. On GDPval-AA, it's nearly tied with Opus 5.5 (1844 vs 1846). This doesn't mean Sonnet universally outperforms Opus. Opus remains stronger on complex, open-ended work requiring sustained judgment. But it does mean significantly more execution can be delegated without paying flagship-model rates. My general rules: • Don't default to max effort. Higher effort isn't automatically worth the additional tokens. • Don't spawn five agents for something one can finish in five minutes. • Don't automatically escalate a failed task. Check whether the issue was missing context, unclear instructions, or insufficient tools. • Don't use an expensive model merely to validate another model's output. Prefer actual tests and independent verification. • Don't confuse API token pricing with subscription usage allowances. Your optimal stack also depends on what your plan includes. Claude-only? The built-in opusplan mode is also worth exploring: Opus plans, Sonnet executes. Using both Claude and ChatGPT? The independent Codex perspective is where I'd extract additional value. Have access to Fable or Astra? Keep them available for tasks that actually benefit from their capabilities, rather than permanently assigning them to routine orchestration. Ultimately, the goal isn't to use every model you're paying for. It's to use the least expensive configuration that reliably gets the job done, while keeping higher intelligence and reasoning effort available exactly when you need them. And remember, when in doubt, just copy and paste this entire post into your agent chat and ask it to walk you through the setup, adapt it to your existing stack, implement it for you, or answer any questions. https://x.com/claudeai/status/2104633115620823187?s=20

一边用 Claude Code，一边切 Codex，复制代码、同步上下文，操作上耗掉的时间太多了吧。 OpenAI 官方出的 codex-plugin-cc，把 Codex 接进了 Claude Code 里。 装完就能在 Claude Code 里喊 Codex 干活：代码审查、挑战式 Review，或者把复杂任务扔给 Codex 后台去跑。 🔗 https://github.com/openai/codex-plugin-cc

What happened last week in AI, you ask? Me, @stack72, John Watson and Nell Shamrell-Harrington break it all down for you: * Colin Breck doesn't want to read what you didn't write * Anthropic drops Opus 5.5, OpenAI drops 6-Luna and 6-Sol * Alibaba becomes a corporate patron of Omarchy * DHH retires from programming at Rails World * Adam Wathan says framework fan bases are dead * Geocodio declares the year of internal tools * Claude Code moves plan mode into a mod * The Dutch government builds a sovereign Nix distro * Microsoft ships Autopilot, Meta ships Muse https://www.youtube.com/watch?v=ycWw3PpG1Qo

Claude Code v2.1.284 is out with Sonnet 5.5. 1M context at $2/$10 per Mtok. Also hiding in the build is a built-in "How mods work" guide plugin, complete with an animated Clawd cooking over AskUserQuestion. Disabled until "tengu_linen_acorn" flag is set to true.

🚀 Trending on GitHub first-pass by joetawil7 ⭐ 46 🧠 A free Claude Code plugin that makes the AI coding agent pause and review its own changes like a careful human reviewer would, instead of just writing the lines it was asked for. https://github.com/joetawil7/first-pass #GenAI #OpenSource #AI

The new Claude Code binary comes with a cool Claude Cooking SVG, seems to be part of the new Claude Code 'mods' feature

@limonkoli claude ai reverse engineeringde cigir acti ondan millet boyle seyler yapabiliyor artik daha kolaylikla inanilmaz fazla decomp recomp toolumuz var modding sceneler cilgin bir halde

Code modernization is now largely solved. What took millions of dollars and dozens of people a couple of years ago, agents can now just do. I did a talk on this last week: https://youtu.be/8uSkd3Hi2mE And open sourced a Claude Code plugin to get you started: https://github.com/anthropics/claude-plugins-official/tree/main/plugins/code-modernization

codex browser use and computer use are definitely ahead right now but you can use them inside claude code if you set up a skill that delegates that part to codex through the codex cli or the plugin https://github.com/openai/codex-plugin-cc

Smart Ralph 能把模糊的功能想法自动转成结构化规格文档，再逐任务推进，一款规约驱动开发插件。🧠 Claude Code 和 Codex 都能装。它的工作流分四步：research.md（调研）→ requirements.md（需求）→ design.md（设计）→ tasks.md（任务拆解），每执行一个任务都会启用干净上下文，防止历史对话污染后续判断。 核心命令如下： /ralph-specum:start 自动识别该新建 spec 还是续跑已有； /ralph-specum:index 扫代码库生成组件索引，规划新功能时清楚手里已有什么； --quick 模式直接跳过中间确认，一键全自动。 子代理分工同样清晰：triage-analyst 拆大特性，research-analyst 做调研，product-manager 写需求，architect-reviewer 出设计，spec-executor 落地实现。 此外还有 ralph-speckit 变体，采用 GitHub spec-kit 方法论，适合团队协作和需要审计追踪的场景。 自包含运行，不依赖外部插件。安装命令： text /plugin marketplace add tzachbon/smart-ralph /plugin install ralph-specum@smart-ralph 想让 AI 按工程规范一步步落地的开发者，可以看看这个。 🔗 https://github.com/tzachbon/smart-ralph

claude can also make geometry dash levels too it seems? it wrote an entire geode mod to bridge into the game via code and then one-shot this "We Were Infinite" by Claude Opus 5.5

this is pure f*cking gold: a Wall Street analyst team you can install in Claude Code (10 agents from Anthropic's own repo, open source) inside: > earnings reviewer - reads the call and the filings, updates the model, drafts the note > pitch agent - turns comps and an LBO into a branded pitch deck > model builder - DCF, LBO or 3-statement model, built live in Excel > ledger reconciler - finds where the books don't match and why > kyc screener - checks onboarding docs and flags what's missing start with these: > /morning-note - the morning note with trade ideas > /screen-deal - quick pass or fail on a deal > /debug-model - traces the formulas and finds hardcoded numbers 2 lines to get the core running in Claude Code: claude plugin marketplace add anthropics/financial-services claude plugin install financial-analysis@claude-for-financial-services repo below ↓

すっかりTRPGにハマって今度は #Pathfinder #WotR を始めました😆 ただどう頑張ってもMacでは日本語化Modが動かず... 思い立って即席で翻訳アプリを作ってしまった🔥 まだ荒いけど、割と軽量動作でOCRして日本語テキストにマッチング、検索機能にダークモード、これ開発4時間でできたのClaudeえぐい

Coder Eval is Playwright for coding agents. Write a YAML task. Run a real agent in a sandbox. Score the files and commands it actually produced. Not another vibes check. Try it: uv tool install coder-eval coder-eval run tasks/hello_date.yaml coder-eval report runs/latest What I like: • agent-agnostic: Claude Code, Codex, Antigravity, OpenCode, Pi (swap `agent.type`) • `skill_triggered` so a Claude Code skill that quietly stops firing fails the suite • weighted 0.0–1.0 criteria + GitHub Action CI gate • Claude Code plugin: `/plugin marketplace add UiPath/coder_eval` Builder take: we ship skills and MCP servers, then hope they still work after the next model bump. Hope is not a gate. Do you eval your agent skills in CI, or still ship and pray? https://github.com/UiPath/coder_eval https://coder-eval.com/

🔥 【上线才 2.1 天就 465 星的 AI 导演项目：39 种电影风格，让编程 agent 写代码把片子做出来】 📌 项目定位： 这不是 AI 视频生成器。它是让 Claude Opus 5.5 用代码当导演——Canvas 和 WebGL 页面逐帧渲染画面、免费采样库写原创配乐、本地 TTS 配音，全程不碰视频模型、不用素材库画面。你只要说一句"用中国水墨风格做一支 30 秒的片子，讲我家楼下的面馆"，30-60 分钟后成片就在文件夹里。 ⚡️ 核心优势： ▫️ 39 种风格全配齐，每种都带 STYLE.md + 一整支纯代码样片：蜡笔绘本、水彩笔刷、中国水墨、油画厚涂、一笔画、白板讲解……客户点哪种就出哪种 ▫️ 装成 skill 全局通用：claude plugin marketplace add lemomo-ai/lemo-opuscar，再 install 一次，之后在任何目录都能开工，第一次下载约 60MB 指南和风格提示词 ▫️ 版权干净：逐帧代码渲染 + 免费采样库原创配乐 + 本地 TTS，商用不用怕素材追责 ▫️ 自带 DIRECTOR.md 和 TECHNIQUE.md 两份指南，管故事、镜头、节奏、混音、自检，还能先出分镜故事板给你确认 ▫️ 门槛低：Node 20+、ffmpeg、Python 3.11+，依赖和无头浏览器 agent 自己装 🎯 实战落地场景： ▫️ 接单党：本地商家、独立开发者、播客主都要 30-60 秒宣传片。白板讲解风格做产品功能，水彩风格做品牌故事，一天跑几条，别人还在排期你已经交片 ▫️ 模板贩子：把某个风格的提示词和参数组合打磨成可复用模板，一次产出反复卖 ▫️ 小型工作室：企业科普、Lo-fi 配乐、知识动画这类预算不高但费人力的活，正好吃下来 提醒一句：风格是按 Opus 5.5 调的，换模型不保证复现。 🔗 项目直达： https://github.com/lemomo-ai/lemo-opuscar 🏷 标签：#AI副业 #AI视频 #ClaudeOpus #开源项目 #副业赚钱

11 个插件，全部由 markdown 和 JSON 构成，没有一行代码，没有构建步骤。 anthropics/knowledge-work-plugins 是 Anthropic 自己开源的一套岗位插件市场，主要为 Claude Cowork 做，Claude Code 也能装。 ▸ 按岗位分包：sales、finance、legal、marketing、data、product-management、customer-support、enterprise-search、bio-research，再加 productivity 和一个用来造插件的 cowork-plugin-management。 ▸ 每个插件目录结构固定四件套：plugin.json 是清单，.mcp.json 管工具连接，commands 放你显式触发的斜杠命令，skills 放 Claude 自动调用的领域知识。 ▸ 命令自带命名空间，像 /sales:call-prep、/data:write-query、/finance:reconciliation、/product-management:write-spec。 ▸ 连接器全走 MCP。finance 和 data 接 Snowflake、Databricks、BigQuery，legal 接 Box 和 Egnyte，bio-research 接的是 PubMed、bioRxiv、ChEMBL、Open Targets 这一类。 ▸ 安装两行：claude plugin marketplace add anthropics/knowledge-work-plugins，然后 claude plugin install sales@knowledge-work-plugins。 要说清楚的是，README 自己写明这些只是 generic starting points。开箱即用的部分有限，真正的价值在于改 .mcp.json 指向你们实际的工具栈，把公司术语、组织结构和流程写进 skill 文件。连接器列表里的 SaaS 你没有账号，那一栏就是空的。 适合有统一工具栈、想把流程固化下来的团队；个人用户大概只会用到其中一两个插件。 今日 +96 star · Python · 总计 24,265 #ClaudeCode #MCP https://github.com/anthropics/knowledge-work-plugins

Your coding agent got pwned by a plugin install. Claude Code, Codex, Copilot, Gemini CLI same zero-click RCE, same dumb bug: the agent verifies the checkout succeeded but never verifies the commit matches the pinned SHA. Name a branch after the 40-char hash, walk right in. Claude and Codex patched. Copilot and Gemini CLI still open as of yesterday. Shell access, env vars, API tokens, SSH keys — all sitting behind a plugin install. Sleep tight.

"AI vai matar a criatividade na música. " electronic music producers: hold my beer. Custom firmware de cdj, elektron, dx7 chinês, mod de tudo, crack de software com Claude code, upgrade de groovebox velha... Derrepente a galera vira MrRobot, what a time to be alive

Most "one pack for every coding agent" repos feel like LCD glue. wshobson/agents does the opposite: one Markdown source, harness-native installs for 7 tools. Claude Code · Codex CLI · Cursor · OpenCode · Antigravity · Copilot · Pi What's inside (not dumped into context at once): • 94 plugins • 202 agents • 183 skills • 105 slash commands Claude Code try: /plugin marketplace add wshobson/agents /plugin install python-development Skills-only, any agent: npx skills add wshobson/agents --skill python-testing-patterns Builder take: writing the skill once is easy. Shipping native artifacts per harness without rewriting CLAUDE.md seven times is the actual product. How many coding agents are you keeping in sync by hand right now? https://github.com/wshobson/agents

Claude Opus or whatever it's called got an update that has allowed people to vibe code their own custom, gamebreaking or otherwise balancing mods, I'm using a set of mods that generally alter the penetration or ammo counts of various weapons to use in mega factory and hive planet worlds and it is just... a world of difference.

项目是AI agent skills集合，专帮你打磨UI/排版/颜色/无障碍/布局/文案。 安装：npx skills add jakubkrehel/skills 或Claude Code里 /plugin marketplace add jakubkrehel/skills 再 /plugin install interfacesinterfaces 核心skill：/better-interface（综合审查） 其他：better-ui、better-typography等，以及break/variant工具。 步骤： 1.装skills 2.开项目让agent读代码 3.运行审查 4.按报告修复 可直接复制的prompt： /better-interface 全面审查当前项目界面，覆盖无障碍、布局、文案、排版、颜色、UI细节，列出问题并给修复方案，先只报告不改代码 确认后加：按报告修复所有问题并给前后对比

分享Claude Code必装的3个神器！ 最近刷 GitHub Trending 发现一个变化：前五名里有三个是 Agent Skills！ ✅️archify 一句话画出会动的架构图 这是我最惊喜的一个。以前画架构图要开 http://draw.io 拖半天，还得自己想布局。 现在直接在 Claude Code 里说一句：Use archify to map this repository's runtime architecture. 它会分析代码库，吐出一个自包含的 HTML：架构图、时序图、数据流、状态生命周期都能画，深浅主题一键切，还带动效。 最狠的是能对比两个版本的架构（Before / Delta / After），做 PR review 直接甩链接给同事。 安装：npx skills add tt-a1i/archify -g GitHub：https://github.com/tt-a1i/archify ✅claude-plugins-official Anthropic 官方插件市场 Anthropic 自己维护的插件目录，目前 289 个插件，其中 38 个是 Anthropic 内部团队写的。 我以前是一个个手动配 MCP，改 json 改到怀疑人生。现在就一行：/plugin install code-review@claude-plugins-official 。 我自己常开的几个：code-review（多 agent 审 PR）、frontend-design（让 AI 写的前端不像 AI 写的）、claude-md-management（自动维护 CLAUDE.md）、playwright / github / notion / slack 官方 MCP 一键接入 🔌 GitHub：https://github.com/anthropics/claude-plugins-official ✅awesome-gpt-image-2 538 个 AI 画图提示词模板 国人「苍何」做的。 把 GPT-Image-2 的 538 个案例逆向成 Prompt as Code，分 UI 界面、信息图、海报、电商、人物、国风等 12 类模板。 关键是它提炼成了 Skill，装完直接让 Claude 帮你选风格、写提示词。网站 http://gpt-image2.canghe.ai 可以直接看图复制 prompt，我现在做封面图都先去翻一圈。 GitHub：https://github.com/freestylefly/awesome-gpt-image-2 网站：http://gpt-image2.canghe.ai 这 3 个别一起装完就吃灰，最实用的顺序其实是： 先装 archify，拿你手头一个真实仓库跑一遍，把架构、调用链、数据流先看懂。 然后开 code-review，让它专门挑 bug、坏味道和 PR 风险，别让写代码的 Agent 自己夸自己。 最后再上 awesome-gpt-image-2，把你项目介绍、产品卖点直接做成封面图、信息图和宣传素材。 我更建议你给 Claude Code 固定一条任务：👇👇👇 “先读仓库 → 画架构 → 找问题 → 修代码 → 最后生成一张产品介绍图。” 这样这 3 个 Skill 就不是插件收藏，而是一条完整工作流

🔥还有谁不知道？13 问并成 1 次调用，账单打到 1/12.2，2.71 秒压到 0.27 秒！ TypeSafe 官方 skill 刚教会 Claude Code 怎么在 Jev 上构建。 这个 skill 把顺序硬改成：读文档 → 写行为 → 列判断 → 一次请求 → 代码做决策 核心亮点： 🔹3 个原语一次看懂：Choice 选、Noul 判真假、Score 打分 🔹13 个问题塞进 1 次调用：成本降 12.2x，速度升 10x 🔹重排证据：Top1 命中从 5% 拉到 18%，Top10 从 38% 拉到 62% 🔹11 条指令每条都对应一种真实翻车：猜字段、串行提问、把 confidence 当许可、阈值直接抄 cookbook 关键操作： 1️⃣装 skill（二选一，别装两份）
claude plugin marketplace add typesafe-ai/skills
claude plugin install typesafe@typesafe-ai
或其他 Agent：npx skills add typesafe-ai/skills --skill typesafe-ai 2️⃣设 TYPESAFE_API_KEY，密钥只放环境变量，不进仓库、不进 prompt 3️⃣先让 Agent 读 http://docs.typesafe.ai/llms.txt，再读 API / 原语页。禁止凭记忆填字段 4️⃣先写一行行为：应用展示什么、选择什么、改什么、交什么 5️⃣把判断拆成「一问一事」。规则、计算、查表、执行、策略全部留在普通代码里 6️⃣每个 Choice 必须带 no-match。同一 state 的问题打进一次请求，只有答案决定下一跳时才拆第二次 7️⃣问题和阈值写进同一份可审文件。阈值必须用自己的标注集调，并钉死模型版本 8️⃣失败时记下 state、问题、答案，再标：缺证据 / 模型 / 代码 / 服务 Jev 不是写代码的 LLM，是 70–500ms 的 System One 判断器。代码掌权，模型只给带概率的答案。 Github：https://github.com/typesafe-ai/skills/blob/main/skills/typesafe-ai/SKILL.md #Jev #TypeSafe #ClaudeCode #SystemOne #AgentSkill #LLMOps

Claude code mod I've been using recently: it plays the "research complete" sfx from Homeworld whenever a task is finished, handy if you're working on something else across the room. I've put a link below, contains .wav files and some tools for previewing/my top picks etc

200 on the dot Hermes merged 200 PRs on September 27 The ones worth knowing: • Desktop skips staging for local OS file drops and keeps inline @ file refs • Kanban shows why a task is blocked and lets you unblock it • Cmd/Ctrl 1 through 9 switch the tab under the pointer again • OS-dropped images paint as thumbnails in sent messages • Email stops From and Authentication-Results spoofing of allowed senders • SimpleX authorizes by contactId only, not display name • Cron jobs can pick their own Python interpreter per script • Windows 10 fresh installs stop dying while unpacking pinned PortableGit • Backups never publish an incomplete automatic zip over the previous one • Plugin catalog adds hermes-routines, AnyRouter, memex8, next-prompt, and more • Voice silence_duration is honoured, and STT errors no longer dump raw JSON into the transcript • Gateway stops the orphan reaper from claiming another home's process Here is the full list: DESKTOP IMPROVEMENTS (78 PRs) • feat(desktop): skip staging for local OS drops; keep inline @ file refs • fix(desktop): ⌘1…⌘9 switch the tab under the pointer again, and profiles when there is none • fix(desktop): keep large pastes as a chip after the turn is stored • fix(desktop): keep the legacy Linux desktop entry as a hidden alias • fix(desktop): stop finished replies rendering twice after settle • fix(desktop): show why a Kanban task is blocked and let it be unblocked • fix(desktop): keep archived sessions archived and renames consistent across session lists • fix(desktop): hermes desktop opens on Linux with a long HERMES_HOME • fix(desktop): compare transcript tips, not counts, in the stale-send guard • fix(desktop): name web, browser and vision tool runs for what they did • fix(desktop): Kanban board switcher outside the full-page layout • fix(desktop): don't offer Rename on canonical Bot Chat tabs • fix(desktop): keep single line breaks in assistant replies • fix(desktop): paint OS-dropped images as thumbnails in sent messages • fix(desktop): show saved key previews without the backend's redaction sentinel • fix(desktop): list root-level personalities in the Settings dropdown • fix(desktop): never report the 0.0.0 placeholder in About • fix(desktop): show provider retry and auto-recovery waits in the status row • fix(desktop): match custom:<key> providers in the settings and bot model pickers • fix(desktop): offer New cron from the empty scheduled-jobs detail • fix(desktop): show one error card when the agent fails to start • fix(desktop): let messaging credential inputs fill their row • fix(desktop): confirm when the main model is applied • fix(desktop): expire unknown readiness notices and clear on recovery • fix(desktop): stop calling unreachable providers a timed-out reply • fix(desktop): reveal clipped profile rail edges and pin actions • fix(desktop): label the messaging platform enable switch • fix(desktop): keep inline math spans when CJK prose follows them • fix(desktop): read inline preview color-scheme from the same source as its tokens • fix(desktop): don't announce a reconnect when the app is quitting • fix(desktop): resolve default Kanban board before event subscription • fix(desktop): keep model rows clickable before pointer movement • test(desktop): match the desktop.mjs compile step with native path separators • fix(kanban): never let scratch cleanup delete the workspaces root or follow a symlinked root • fix(desktop): carry the window app id in the Linux launcher entry • fix(desktop): anchor project trunk + at first repo root when primary path is unset • fix(desktop): forward client clipboard paste into Bot Screen • fix(desktop): bridge the already-decoded media stream path once • fix(desktop): say "File not found" when opening a missing file • fix(desktop): background Bot Chat re-resumes never navigate • fix(desktop): keep authored link labels and identifier casing • fix(desktop): keep past Todo checklists reopenable after a turn • fix(desktop): drop backdrop blur when the OS asks for reduced transparency • fix(desktop): narrow edge overlay honours the dragged pane width • fix(desktop): stop transcript images overhanging the next block • fix(desktop): keep bare loopback URLs in assistant prose • fix(desktop): keep lightbox and video state across transcript remounts • fix(desktop): keep a git-installed desktop half without a local package • fix(desktop): clean up self-test-marker sandbox dirs after the test run • fix(desktop): pass wayland ozone on native Wayland sessions • test(e2e/desktop): let the build-fail updater finish launch acceptance before asserting it exited • fix(desktop): run updater state.db pre-flight through the installation launcher • fix(desktop): keep live session tiles out of saved layout presets • test(e2e/desktop): PR-time Desktop install/update suite on a real install (first run, Update now, build failure, concurrent update) • fix(desktop): keep URLs inside Markdown code un-chipped • fix(desktop): reseed a bare provider pick the default migrated to custom:* • fix(desktop): latch a resume that paints nothing when the list row is stale • fix(desktop): the Windows update tail never stops the Desktop it runs under • fix(desktop): resync the Bot model picker's manual-entry state when the catalog resolves • fix(desktop): collapse duplicate @-mention rows behind a source's handle claims • fix(desktop): model snapshot adoption no longer strands catalog defaults behind a stale allowlist • fix(desktop): split look-alike model rows, archived-session tombstone merge, renderer accessibility tree, SearchField clear button • fix(desktop): keep the Windows update hand-off console hidden and selection-proof • fix(desktop): stop a stale Windows hand-off failing a healthy update • fix(desktop): keep the installed macOS http://Hermes.app current from update and launch • fix(tui): reap session kernels when a TUI-owned session ends • fix(desktop): transcript duplication, acknowledged-prompt dedupe, attachment-tolerant matching, continuation provenance, backtick URLs • fix(desktop): give approval.respond the backend's 300s deadline • fix(desktop): mirror remote session cookies in memory and retry once on 401 • fix(desktop): raise backend readiness deadline 45s to 180s • fix(desktop): kill external autostart venv holders before update hand-off • fix(desktop): stand down cron ticker when a live gateway owns cron on same HERMES_HOME • fix(desktop): wave-8 misc fixes, search title, keybind guard, skill pill, @ file: titles, math backslash, Cmd+1-9, @ url: scheme, ARM64 gate • fix(tui_gateway): resume wake detector after session.interrupt • fix(tui): hide reasoning and tool chrome when show_reasoning is off • fix(desktop): keep the sidebar alive for messaging-only and cron-only profiles • fix: [Bug]: Desktop app crashes (SIGSEGV in macOS PrintCore) when printing a Google Doc from the preview pane • fix(desktop): lock the composer to the dock by default VOICE & REAL-TIME (5 PRs) • fix(desktop): honour a user-set voice.silence_duration in the voice loop • fix(matrix): answer a voice message claimed by a follow-up bare mention • fix: wave-7 misc, stale approval bar, gh stderr, hidden conversations, SOUL.md notice, auto-TTS on A2A • fix: ensureContrast step validation, streaming reasoning_details, reply-scoped Copy/Read-aloud, axios lockfile, expected-quit forensics • fix(stt): never stringify a structured STT error response into the transcript DEVELOPER TOOLS & PLUGINS (43 PRs) • fmt(js): `npm run fix` auto-fix (7x) • fix(tools): make managed Node authoritative in the MCP stdio PATH • MCP stdio servers keep their configured PATH order again • refactor(plugin-catalog): rename browserclaw to browserpaw • plugin-catalog: add memex8 memory provider • feat(catalog): add AnyRouter model-provider plugin • feat(plugin-catalog): add next-prompt plugin • fix(catalog): fit reMarkable artwork inside the detail header • fix(catalog): fit Blinkenbar artwork inside the detail header • fix(catalog): keep Pixel Worlds header title visible • feat(plugin-catalog): bump memory-rewind to 1.2.0 • catalog: bump antigravity-subscription-directsdk pin to v1.0.5 • chore(plugin-catalog): bump prompt-snippets to v1.7.0 (host.composer migration) • feat(plugin-catalog): add hermes-routines • website: catalog card banner slot at the 2:1 catalog ratio (stop cropping authored banners) • fix: backend update-check errors surface, managed-Cloud Screen copy, bundle-skew covers apps/shared, catalog cache drops on update, MCP liveness says what's wrong • fix(website): plugin stars probe, import the repo-root hermes_yaml shim and batch the GraphQL query • feat(website): uniform plugin catalog cards (clamped text, capped tool list) • Catalog: entropicmem 2.8.1 (re-authored attribution) • feat(catalog): add OrcaSlicer community plugin 0.5.0 • catalog: bump Little Canary to 0.3.10 • feat(plugin-catalog): add memory-rewind • feat(plugin-catalog): add parrotnotes • feat(plugin-catalog): add web-meta-ai community plugin • fix(tools): spill MCP result envelopes as pageable text • feat(skills): brag and brag-slim join the optional-skills catalog as upstream stubs • fix(mcp): kill Windows stdio MCP orphan trees (Job Object + tree reaping) • chore(plugin-catalog): bump Nerve to v0.3.0 • chore(plugin-catalog): bump unitares to 0.3.1 • catalog: update Pixel Worlds to v0.3.0 • chore(plugin-catalog): pin claude-subscription-directsdk to ef73726 (effort + shared cwd) • feat(plugin-catalog): add cortexlayer memory-provider entry • feat(plugin-catalog): bump sigrex to 1.0.1 (add banner) • plugin-catalog: bump thomas to v0.1.1 • catalog: bump antigravity-subscription-directsdk pin to v1.0.3 • feat(plugin-catalog): add the Index Network plugin • feat(plugin-catalog): add drag-to-pin-session SECURITY & SECRETS (5 PRs) • chore(catalog): bump desensitize pin to 2dbf63d (help follows /desensitize lang) • fix(email): stop From and Authentication-Results spoofing of allowed senders • fix(simplex): authorize by contactId only, not the sender's display name • fix(auth): keep root's single-use Nous refresh token out of named profiles • catalog: pin secret-drop 1.4.0 INTEGRATIONS & CHAT (13 PRs) • fix(bot-mode): prefer the installation-bound launcher for Bot deliveries • fix(gateway): stop the orphan reaper claiming another home's gateway • fix(gateway): back off a busy re-queued event instead of hot-looping the adapter drain • fix(gateway): resolve handoff and loop-watch profile scopes off the event loop • fix(gateway): spare a still-booting gateway from the Desktop orphan reap • fix(gateway): keep the update restart watcher stdlib-only so it survives the bare store Python • fix(gateway): recognise Hermes inline-bootstrap gateways on every OS • catalog: bump hermes-discord-rpc to v1.2.3 • fix(gateway): a hard-deleted session stops resurrecting on the next inbound message • fix(gateway): broadcast projects.changed when projects.db moves • fix(gateway): keep one Windows gateway autostart mechanism • fix(gateway): mark synthetic process notifications internal and stop stale reply anchors • catalog: bump telegram-notify to v1.1.2 STABILITY & FIXES (56 PRs) • fix(model): split comma-separated fallback chains into picker entries • feat(cron): per-job Python interpreter for cron scripts • fix(install): fresh Windows 10 installs no longer die unpacking pinned Git (PortableGit) • fix(update): narrow tag-pinned clones update instead of reporting main missing • fix(version): preserve derived display identity for tagless installs • Windows install wrappers passing -SkipSetup no longer die at binding • ffmpeg installs stop expiring: pin BtbN's retained month-end build, and pm update only picks those • fix(update): restart hermes-dashboard* systemd units during fleet update • fix(update): restart a dashboard's systemd unit only when the unit's MainPID is the dashboard • fix(agent): ignore execute_code per-call metadata in the loop-detector result hash • fix(error-surface): name the provider by its display label on every error card • fix(sessions): refuse to delete a session while a live turn owns it • fix(persistence): stop re-inserting sanitized user and tool rows on re-flush • fix(context): stop compaction from rewriting old tool-call arguments • fix(agent): restore todo_list items across turns and on TUI resume • fix(cron): refuse to overwrite a corrupt jobs store on save • fix(compression): stop asking the summary model to quote the latest task verbatim • fix(compression): let the next summary build on a fallback handoff • fix(backup): never publish an incomplete automatic zip over the previous backup • fix(compression): send a Codex stream stall to the timeout ladder, not a terminal abort • fix(agent): let background delegate_task handoffs end the turn • fix(terminal): heartbeat error text matches the 0-disables schema • fix(delegate): mark elided timeout-diagnostic goal and eval tool output • fix(status): show the session's owning profile home • chore(contributors): re-attribute agent-identity commits to OutThisLife (2/2: remaining 94) • test(e2e/matrix): Windows install legs stop hanging; install/update PRs run a 4-leg real-OS subset; red schedules open one tracker • fix(cli): HERMES_HOME with '..' no longer relaunches Hermes forever • hermes update rides out a CDN blip on the release-channel read; passive checks stay single-shot • fix(bootstrap-installer): Hermes-Setup.exe keeps install.ps1's receipt schema and commit on git-less machines • fix(pm): musl hosts get native musl uv/Python/Node artifacts • hermes update and hermes pm repair reclaim superseded dependency generations • fix(pm): a symlinked Hermes home no longer re-runs PM completion every launch • fix(pm): generation workspaces keep pm/uv.lock • fix(update): give git-less Windows installs PM's git for update and source completion • fix(update): keep the autostash when the update replaced a user's untracked file • fix(install): recognize bare PATH= assignments and keep the appended line idempotent • Cron .py scripts import Hermes and its dependencies again on POSIX managed-store installs • fix(agent): name an exhausted (402) provider pool instead of "No LLM provider configured" • test(e2e/handoff): processes running across hermes update (gateway, dashboard, cron, kanban) • fix(state): re-fold split display generations at compaction commit • fix(update): keep commits made on a detached HEAD behind a named rescue ref • chore(contributors): re-attribute agent-identity commits to OutThisLife (1/2: 100 of 194) • test(e2e/pm): PM environment lifecycle across real hermes updates (22 cells) + per-suite e2e-upgrade shards • test(e2e/windows): PR-time install.ps1 then hermes update journey on native Windows • test(e2e/hosts): install/update on odd host shapes (non-ASCII/symlinked homes, PATH shapes, read-only installs, musl) • test(e2e): hostile-network update suite, netns proxy-only egress, mirrors, channel record faults, offline/flaky • test(e2e/git): git transport and working-tree edge cases through the real hermes update over smart HTTP • fix(web): scrub a session's on-disk artifacts in the delete endpoint • fix(update): read npm's version from its manifest instead of a child probe • fix(process): heartbeats wake the agent only on new output, and never as a user bubble • fix(api-server): deliver transform_llm_output rewrites on streaming routes • fix: ACP sessions end at shutdown, approval batch re-gates after failure, base-branch picker settles, regex literals in the plugin loader • fix(state): add immutable created_source provenance column • fix(uninstall): sweep both LaunchAgent labels and per-user app leftovers • fix(serve): announce both ready tokens so a stale packaged parser still boots • fix(sessions, desktop): show_subagents setting, subagent titles, and failed delegations in subagent.list and the Agents panel 200 new features and fixes https://x.com/iamlukethedev/status/2104020319598149996?s=20

MrRealistic-Mods, a GTA modder, is making an Euphoria physics mod for the GTA trilogy, and it's called Jolt Physics. In this clip, it's possible for Claude, Tommy, and CJ to ragdoll after falling from heights. #GTA3 #GTAIII #GTAViceCity #GTAVC #GTASanAndreas #GTASA

I started playing Dota 2 again casually. Inspired by the WoW mod I built this Claude overlay that pops up when I die. I didn't realize I still had the VTT session running!