Community posts

Claude Code es un rollo. Hasta que instalas esto. Hay un plugin oficial de Anthropic llamado claude-code-setup. Te dice qué automatizaciones puedes montar (hooks, skills, MCP servers, subagentes...) у cómo configurarlas paso a paso. Básicamente analiza tu proyecto y te recomienda qué activar. Para instalarlo: / plugin install claude-code-setup@claude-plugins-official Guarda este post para no perderlo 📩

introducing ads mods for claude code Devs feedback was clear: ads popping out of the terminal. no. we used claude mods. ads now render in-terminal, above the prompt, while the agent works. https://promptai.credit ads pay for the prompts

🚨 CLAUDE CODE JUST GOT A SMARTER MODEL ROUTER A new Claude Code mod can automatically route every request to the right model and effort level. It classifies: → subagent model → main model → reasoning/effort level And it can plug into the Typesafe AI API or Vercel AI Gateway. The interesting part? One install command. Instead of manually deciding which model should handle each task, the router makes that decision automatically. Cheap model for simple work. Stronger model when the task actually needs it. More reasoning only when it's worth the extra cost. That's exactly where AI coding workflows are heading: Not one model doing everything. A system deciding which model should do what.

Unity’s agent plugin includes a skill for choosing other UI skills. Ask for a settings screen: it tells Claude Code or Codex to check the project’s framework, then use the matching UI Toolkit, uGUI or IMGUI skill. https://github.com/Unity-Technologies/unity-agent-plugin/blob/main/skills/ui/SKILL.md

yeni mod: error-poke bazen api hatasından dolayı hata alıp duruyordu. böyle durumlarda mod otomatik devam etmesini sağlıyor. { "env": { "CLAUDE_CODE_ENABLE_FUNCTION_HOOKS": "1" } } ile desteği aktif ettikten sonra kurabilirsiniz: claude plugin marketplace add KilimcininKorOglu/claude-code-mods claude plugin install error-poke@kilimcininkoroglu-mods repo: https://github.com/KilimcininKorOglu/claude-code-mods/tree/main/plugins/error-poke

Claude Code2.1.276近辺で $.uiにImage要素が実装されたため、ghostttyなどではClaude Codeのペインでmods(hooks function)Kittyをつかってフルグラフィックスが描ける。herdr内ではさらにCLAUDE_CODE_FORCE_TERMINAL_IMAGES=1 を設定する必要がある。 これはChangeLogには入っておらず、未公開機能で今実装まっさかりです。経緯をしらべると: - 2.1.272（〜9/15）の時点で Raster はあったが Image は無かった - 2.1.276（9/18）にはある - → Image が入ったのは 2.1.273〜2.1.276 の間、2026-09-15〜18 のどこか

这个项目核心是用Jev做结构化决策。要点：把场景写成state文本，定义noul(是否)、score(评分)、choice(选项)问题，获取带概率/置信度结果，用于意图识别、路由、验证。 工具：TypeSafe API(POST /v1/systemone)或OpenRouter、Playground。 Skill：官方typesafe-ai，安装claude plugin marketplace add typesafe-ai/skills再install typesafetypesafe-ai。 Agent：Claude Code等harness做门控。 步骤：拿key→写state+questions→调API→按阈值行动。 可复制prompt给agent： 根据我场景生成Jev请求JSON：{"model":"jev-latest","state":"你的描述","questions":{"q1":{"type":"noul","instructions":"..."},"q2":{"type":"choice","instructions":"...","criteria":{"optA":null,"optB":null}}}} 并解释如何调用和阈值。

Jev won't replace GPT or Claude. Stop hype. It's a switch statement with judgment, and it does the one job neither of them is any good at. Think 2016 ML classifiers running at 2026 levels of intelligence. A new kind of tool. You send one input and a handful of typed questions in the same request, they run in parallel, and adding questions costs you almost nothing in latency. You give it a closed set of options. It tells you which one to pick. What it can't do: ↳ Write code ↳ Generate natural language ↳ Reason step by step or show its work ↳ Return any output you didn't define up front ↳ Pick from more than ~255 options at once What it can do: ↳ Classify, route, score, rank ↳ Hand back a confidence score ↳ Pick the right branch, tool, model or sub-agent ↳ Judge, verify or put guardrails on an LLM's output ↳ Label rows and rows and rows Where I'd bet this lands: LLM proposes options → Jev decides → code executes Sometimes you don't even need the first step. There's already a mod for Claude Code that asks Jev what each request needs before any model runs. Your first $5 of credit probably lasts a month, with change. Calling this a step toward AGI is a stretch. I'm still not writing off what people build with it.

If you are confused about why 𝗝𝗲𝘃 is being called the "Internet" moment for the AI industry. This is 100% worth your time. In fact, you should watch it: It tells LLMs what to do next, in milliseconds & at almost zero cost. If you set it up correctly, you will have the AI engineer’s setup for 2028. How to set up & use Jev (to actually get the 100x): 1. Join the waitlist; it's fairly quick: typesafe .ai. 2. Then go to Claude Code or Codex. 3. Choose Opus 5-Low or Sol-Low. 4. Copy and paste this prompt: "[claude or codex] plugin marketplace add typesafe-ai/skills [claude or codex] plugin install typesafe@typesafe-ai" 5. When you type /typesafe, the skill shows up. 6. Paste your API key once and click "Allow" 7. Start with $5 in free credit. It's hard to spend more. ----- Now, here are the 3 ways to actually use Jev: 1. Jev for Linkedin I have 38,000 connections & invitations on LinkedIn. I have a new company to launch. I need to find a couple of hundred people to message. to do while saving time: > Export your LinkedIn connections and invitations. > Connect Claude to GitHub, Vercel & Apify. > Create an Apify API key to enrich your data. > Go to LinkedIn Settings → Data privacy. > Get a copy. LinkedIn will email you a ZIP file. > Open the file & find the Connections CSVs. > Upload the files to Jev. > Use Jev to classify your contacts. > Review the shortlist. 2. Jev for Gmail To go through all of my Gmail contacts and email the right people. > Go to Google Contacts. Open Other contacts. > Select all contacts. Export them. > Upload the file to Claude Code or Codex. > Use Jev to sort them into: Keep, Review, Remove and Review everything before removing anything. 3. You got lost in Claude Code, GitHub, Vercel, Apify, Jev, Typesafe. I feel you. It is overwhelming. That’s why I included the entire copy-and-paste prompt for each use case in the newsletter: https://ruben.substack.com/p/jev A 45-second TL;DR by @MatijaSosic.

Claude Modsに関する登壇や記事寄稿のご依頼を、積極的に募集しています むしろ、こちらからお願いしたいくらいなので、機会がありましたらぜひお声がけいただけるとうれしいです！

「Claude Mods で Claude Code 自体をカスタマイズする」という記事が勉強になったのでシェアです！ Claude Code の拡張が、外から処理を追加する Hooks から「Claude Code のハーネス自体に処理を組み込む」方向へ進んでいる。TypeScript の関数をエンジンに直接ロードすることで、ツール実行の制御だけでなく、Context、UI、状態管理、ツールそのものまで変更できる。 しかも Claude Code 標準の /diff や AGENTS.md 対応も同じ Mod の仕組みで実装されている。AI に指示を追加するだけではなく、「AI が動く環境そのもの」をコードで拡張できるのが大きい。 続きはリプ欄で！ https://zenn.dev/nogu66/articles/claude-code-function-hooks-claude-mods

『Claude Mods』 これは絶対に読んだ方がいいやつだ👀

そこまで手を入れられる設計、そそられる。noguさんのClaude Mods入門では、Claude Codeの拡張を、Hooksでコマンドを呼ぶ形からTypeScriptの関数を本体に組み込む形へ広げる仕組みとして紹介している。 既存Hooksとの違いから、--plugin-dirを使う導入例までつながっているのがありがたい。自分がエージェントを回す目線で気になるのは1つ。自然言語で頼んでいる制約を、どこまでコードで強制できるのか。 記事にある処理の拒否や結果加工の例は、その境界を考える入口になりそう。機能を差し替えられるなら、「何ができるか」と同じくらい「誰の権限で動くか」が気になる。 https://zenn.dev/nogu66/articles/claude-code-function-hooks-claude-mods

There is a mod for Claude Code, which saves tokens. 🦀 Instead of loading all the skills into the context window, they remain outside it until they are really needed. Jev analyzes the available skills and selects the appropriate one for each request, adding only the necessary ones to the Claude context. The solution works with API from TypeSafe and AI Gateway from Vercel. You can install the mod with one command.: npx claude-code-templates@latest --mod productivity/jev-skill-suggestion Use it, I like the admin for the information ❤️

buffa found a hole in its RFC 3339 timestamp parsing: "+1".parse::<u8>() returns Ok(1), so signed components were getting silently accepted where they should have been rejected. Parsing now rejects them outright. Same parsing library, different corner: DynamicMessage::get() on an unset singular message field used to build its default on every call, paying a Box allocation, an Arc::clone of the descriptor pool, and the matching drop per read. It now returns a borrowed empty message instead. That's a breaking change to ReflectCow for callers that assumed an owned value. claude-code's telemetry mod runs wherever analytics are on, but only for plugins built into Claude Code. A hook reads next․origin and refuses plugins a person installed or an administrator listed, with a reason. Rows carry the same fields as the CLI's own, gathered through $ once a session, so third-party plugins stay out of the pipeline. #AI https://repojournal.com/showcase/anthropics/2026-09-21/buffa-hardens-wkt-and-descriptor-parsing-claude-code-smooths-the-diff-pane

Claude Code 的上下文终于不用再被一堆预加载技能塞满了！推荐一个非常优雅的按需加载解决方案：Jev Skill Suggestion。 以往在配置大量自定义技能（Skills）时，最头疼的就是 Context Window 膨胀问题。许多低频技能即便一次都用不上，也会在每次对话启动时常驻上下文，不仅白白消耗 Token 成本，还会稀释模型的注意力，甚至引入不必要的提示词干扰。 这个 Mod 的核心逻辑是将技能路由与主体执行彻底解耦（Just-in-Time 动态注入）： 常态零占用：将技能标记为仅用户可调用，默认状态下完全排除在上下文窗口之外，保持 Prompt 绝对干净。 轻量前置路由：用户输入指令后，Mod 会将技能列表交由 TypeSafe AI 的分类器 Jev（或兼容的 Vercel AI Gateway），快速匹配与当前任务最契合的技能。 精准单点注入：只有被分类器判定为高置信度命中的技能，才会在执行当下被动态注入给 Claude Code。 安装仅需一行命令： npx claude-code-templates@latest --mod productivity/jev-skill-suggestion 这种把“工具选择”抽离给轻量分类器、让大模型只专注“任务执行”的架构，正在成为复杂 Agent 系统降本增效的标配设计。随着技能库扩充，未来配合端侧超轻量分类模型（如社区正在探索的 MLX 移植版），整体延迟和隐私表现还会更进一步。 大家目前在 Claude Code 里常驻了多少个自定义技能？这种动态路由机制是否刚好击中了你的上下文焦虑？ #ClaudeCode #AIAgent #AgentSkills #TypeSafeAI #LLM #PromptEngineering #AI编程 #OpenSource #开发者 #独立开发 #全栈开发 #Vercel

要点：用Jev决策模型自动选最匹配技能，注入唯一SKILL.md到上下文，隐藏其余列表省token。支持TypeSafe API或Vercel Gateway。 工具：Claude Code≥2.1.278、claude-code-templates、TYPESAFE_API_KEY（可选）、已安装技能。 步骤： 1. npx claude-code-templateslatest --mod productivity/jev-skill-suggestion 2. CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 claude 3. /jev-skill-suggestion:setup（确认隐藏技能） 4. 重启Claude Code 5. 正常提问，自动注入。 可复制Prompt： Install and setup jev-skill-suggestion mod for Claude Code. Run the npx command, enable hooks, execute /jev-skill-suggestion:setup, restart, then test with a skill-related task and report the injected skill.

A new Claude Code mod adds Jev routing to every request. It automatically classifies the subagent model, main model, and effort level through the Typesafe AI API or Vercel AI Gateway — all with one install command.Simple tasks go to a fast, cheap model. Harder work gets routed to a stronger one. The decision is made by Jev, TypeSafe’s decision model, instead of you picking models by hand. In the demo, a basic command like “print the contents of package.json” is routed to Claude Haiku with low effort and high confidence. The interface stays the same. Only the model choice changes behind the scenes. One install. Automatic routing. Less wasted compute. https://x.com/dani_avila7/status/2101176629745561686/video/1

Reminder: Claude Code can test if a skill or plugin actually improves Claude's answers. > claude plugin eval init Tell Claude what a good result looks like in your plugin's folder and it writes the test cases for you. > run claude plugin eval .

Hermes merged 419 PRs today 🤯🤯🤯 Read that again: 419!!!! The ones worth knowing: • External process model providers ship as standalone plugins, so out of tree OAuth and process providers show in every model picker • Bot Mode and Desktop: GPT Live and TTS follow the Bot profile voice, abandoned group turns are harvested, and teammate DM replies arrive whole • /branch opens a sibling thread by default on Discord, Telegram, Slack, and Matrix; --here keeps this chat • Background terminal processes show in the CLI/TUI live work dock like subagents • GLM 5.3 FlashX lands on OpenRouter and Nous Portal pickers • Cron Slack one shots deliver into the thread they were asked in, and restart safe cron deliveries go out when queued • Case variant credential env names no longer bypass the passthrough blocklist and child env scrub • Provider plugins drive setup catalogs, OAuth PKCE login, usage, capabilities, and error classification end to end • Desktop settings organize into ordered subpages, and you can hide thread timeline bars or code diffs while keeping edit counts • WhatsApp and Weixin replies drop the multi second wait before every message and match Telegram cadence • Matrix invite reconciliation applies the inviter allowlist, closing the auto join bypass around GATEWAY_ALLOW_ALL_USERS • The plugin catalog grows again with openmail, prompt snippets, rss reader, hermes herald, reMarkable, sketch pad, and more Here is the full list: DESKTOP IMPROVEMENTS (86 PRs) • fix(desktop): a ✕-closed agent tip stays closed across conversations • fix(context): live model/provider switches and Desktop config • fix(desktop): Desktop updater stops copying state.db • fix(desktop): plugin loader no longer reads strings/comments • fix(desktop): quoted or pasted stop words • Catalog cards open the Desktop Install Plugin dialog by catalog • fix(desktop): dismissed preview chips stay dismissed • fix(desktop): a steering correction recovered • Desktop split panes: control test proves the visible-turn floor • feat(desktop): organize settings into ordered subpages • fix(desktop): remove extra top spacing in the status stack • `hermes update` skips the web-UI and desktop `npm ci` passes • Desktop: a gateway socket that is accepted then dropped • Desktop: component scoped style keyframes move to document scope • fix(desktop): an env-pinned remote gateway can sign in again • Desktop composer: the editable input surface drops its live • Desktop: hovering the profile rail without moving the pointer • Desktop multi-pane split: the per-pane transcript budget • Desktop: hidden keep-alive chat tiles stop polling `subagent.list` • Desktop `serve` stops burning a core • fix(bot-relay): a DM the Desktop claimed but never delivered • fix(desktop/bots): writes into a member's Group session reach • fix(desktop): @ picker lists a remote default by its title • fix(cron): Desktop cron run history and per-job actions follow • fix(desktop): Bot Mode group settings can compress a member's • fix(desktop): a Group Chat turn the Desktop abandons mid-flight • fix(desktop): "Open containing folder" is hidden for remote • fix(desktop): Bot roster section Enter respects IME commit • fix(desktop): "Open containing folder" is offered only • fix(desktop): update check reaches api. github. com through • fix(desktop): macOS shell survives the updater relaunch • fix(desktop): drive_preview clicks land on their target at any page • fix(desktop): SESSIONS / BOTS sidebar tab labels follow the display • fix(desktop): the Bots roster stops republishing an unchanged • fix(bot-mode): local member replies carry from.source • fix(tui): dashboard re-attach • fix(desktop): update overlay clears on parked-branch installs • fix(desktop): group member turns are harvested • fix(desktop): plugin composer actions stay grouped with the send • fix(desktop): /resume from a Bot tab opens the session inside • feat(desktop): add an option to hide thread timeline bars • fix(desktop): tool calls cut short by the user read as Interrupted • fix(desktop): a queued send with attachments stages, recovers • fix(bot-relay): a relayed DM into a Bot Chat open in Desktop • fix(desktop): Bot Mode Skills Hub installs land on the bot's own • fix(desktop): sidebar search keeps the exact session-id hit on top • fix(desktop): a text selection in the transcript survives streaming • fix(api-server): peer dm and peer run into a Bot Chat open • fix(desktop): /context (and every offered built-in) groups • fix(desktop): plugin-contributed keybind overrides survive restart • fix(desktop): prefixed currency with a space • fix(desktop): edit-composer undo covers cut and drag-move • fix(desktop): scrolling up while a clarify card is pending holds • fix(desktop): Artifacts page indexes explicitly delivered Office • fix(desktop): Bots roster age shows the last worker run • fix(desktop): attachment preview/download resolves home-relative • fix(desktop): selecting text in the session rename dialog • fix(desktop): an un-acked optimistic message survives a resync • fix(desktop): first and last prompt-rail marks show their preview • fix(sessions): canonical Bot Chat keeps its composer model pick • fix(desktop): typing a custom answer in a multi-select clarify card • fix(desktop): a superseded wake clears the chat-swap overlay it put • fix(desktop): an entered project shows its declared repository • fix(desktop): SSH-mode backend keeps the login umask, terminal-tool • fix(desktop): closing the last window on Windows exits Hermes.exe • fix(desktop): Windows Update button no longer aborts with exit 4 • fix(desktop): Inbox style renders pinned rows as cards too • fix(desktop): a group-chat member failure shows its typed reason • fix(tui-gateway): manual /compress works on a multiplexed • fix(tui): Ctrl+D exits from an empty composer on macOS too • fix(desktop): Windows update hand-off • fix(models): downloaded local models pass the Desktop Use button • fix(dashboard): hermes dashboard from a Desktop-spawned shell • fix(bot-mode): a DM or peer run into a Bot Chat open in Desktop • fix(desktop): two-modifier navigation chords (⌘⌥←/→) fire • fix(tui): clipboard image paste works on Wayland again , probes • fix(acp): editor sessions populate sessions.cwd and git metadata • fix(tui-gateway): launch-profile slash commands work • fix(state): session transcripts survive an unknown binary column • fix(desktop): `hermes desktop --build-only` stops a Desktop running • fix(delegation): finished subagent work survives owner death • fix(desktop): manual /compress shows compaction progress until it • fix(desktop): support macOS text replacements • Admitted plugin providers are selectable in hermes model • feat(desktop): optionally hide code diffs while keeping edit counts • fix(desktop): Bot Mode chat can be dragged into a split pane VOICE & REAL-TIME (6 PRs) • fix(desktop): GPT-Live in a Bot chat uses the Bot's own profile • fix(gateway): Matrix/LINE/Mattermost/Weixin deliver audio MEDIA • fix(desktop): Bot chats speak with the Bot's own profile TTS voice • fix(desktop): client-direct TTS speaks a sealed narration • fix(tts): output_path can no longer forge extra MEDIA delivery tags • fix(tts): voice mode plays sentences DEVELOPER TOOLS & PLUGINS (191 PRs) • plugin catalog adds: claude-subscription-directsdk, xmemo, covision, meeting-overlay, profile-pane, image-utils, reMarkable cloud and notebook toolkit, sketch-pad, handflow, hermes-herald, markifact, hermes-structured-aux-models, Adspirer, sticky-notes, provider-copy, source-tray, prompt-enhancer, prompt-snippets, and 12 more (27x) • chore(catalog): version bumps and pin refreshes (9x) • fix(mcp): stdio command lookup stops mutating PATHEXT and reading • fix(kanban): a typed block classifies an untyped breaker-parked card • fix(windows): bash spawns use Git Bash instead of the WSL stub • fix(skills): hub skills stop reporting update_available forever • fix(agent): promoted-reasoning stall guard catches Thai plan tails • fix(delegate): child_timeout_seconds bounds inactivity • fix(local-runtime): two backends booting together start one managed • fix(doctor): agent-browser npm advisories stop resurfacing • fix(constants): scratch dir honours HERMES_HOME_MODE • fix(cli): exit clear-screen fallback spawns no shell and no console • fix(kanban): kanban_complete result lists the registered attachments • fix(tools): search_files keeps its matches on macOS • fix(tools): idle session kernels are reaped • fix(moa): replayed advisor guidance is marked as predating • fix(agent): CommandCode upstream-unavailable 429 keeps the API key • fix(compression): a detached stale compression attempt • fix(agent): in-place repairs of persisted message dicts reach • fix(local-runtime): gpt-oss MXFP4 GGUFs get a preset • fix(models): a profile's own catalog rejects models the generic • fix(plugins): a plugin callback's sys.exit() • fix(compression): overlapping compression attempts keep the newer • fix(context_references): @ folder: lists directories outside cwd • fix(install): installer recognises a pre-staged rg/ffmpeg • fix(plugins): install --ref accepts an annotated-tag pin • fix(agent): capped @-reference child output • fix(agent): auto-title call no longer races the reply • Background terminal processes show in the CLI/TUI live-work dock • feat(providers): external-process providers as standalone • fix(cli): footer no longer collapses into "Window too small..." • feat: z-ai/glm-5.3-flashx available on OpenRouter and Nous Portal • hermes tools: browser provider picker stops re-appearing • fix(tests): stop the code_kernel reader-thread leak • fix(dashboard): a standalone `hermes dashboard` • perf(skills): `hermes skills check` • perf(agent): memory is trimmed after a ≥1 MB tool result, the same • Auxiliary-client and iteration-limit-summary chat calls bypass • lazy_deps: the uv install tier honors `pip.conf` / `PIP_INDEX_URL` • memory: a failed consolidation batch • fix(compression): an over-window session • feat(nous): unlock subscription-billed models for free-tier accounts • fix(dashboard): a second /chat tab gets its own PTY • hermes update no longer idles ~60 s per sibling profile • fix(state): a schema column holding bytes keeps its key in message • fix(computer-use): doctor and status report a configured cua-driver • Dashboard session search/stats/detail • fix(browser): a stale local agent-browser session recycles • fix(update): out-of-tree venv installs refresh tool deps • fix(langfuse): an invalid HERMES_LANGFUSE_MAX_DEPTH warns once • fix(config): doctor, plugins enable and quiet migrations agree • fix(kanban): routed notifications reach a pinned profile • fix(state): Bot DM transports resolve the canonical Bot Chat • ACP providers stay on chat completions through GPT-5 fallback • fix(sessions): Windows pollers no longer starve • fix(mcp): startup summary names every failed server • fix(sessions): bulk archive never hides a live chat because its old • fix(doctor): hermes doctor validates GITHUB_TOKEN/GH_TOKEN against • fix(web): mobile IME no longer duplicates words in the dashboard • fix(state): /undo and /retry work again after a turn • Provider-plugin seams: setup keeps curated catalogs offline, ACP • fix(process): an incomplete kill keeps the background entry listed • fix(skills): skills whose SKILL.md links a doc above • fix(checkpoints): prune sweeps tmp_pack debris left by timed-out • fix(disk-cleanup): committed test_* files inside git worktrees • fix(termux): pip install '.[termux]' • fix(agent): @ file/@ folder/@ diff expansion stays memory-bounded • fix(plugins): enable no longer prompts for a tool-override grant • fix(tests): updater npm-cache tests stub the npx warm-up, no real • fix(cli): cua-driver autostart repair spawns no console window • fix(agent): shell hooks declared as a bare .sh/.py path run • fix(skills): skills. sh skills with SKILL.md at the repo root • Plugin ProviderProfile.api_mode reaches the runtime transport • Plugin providers resolve in /model, the model picker and hermes • fix(tests): local-env Popen fixture ends the kernel streams, reader • fix(lsp): failed server starts report exit status + stderr tail • User provider plugins override a bundled provider's endpoint • fix(computer-use): doctor names a daemon unit • fix(tests): run_tests_parallel takes the explicit file list • fix(agent): subdirectory hint files • fix(gemini): AI Studio AQ. keys stay on the Studio surface • fix(acp): /reset, /compress, /model and session/set_model • fix(tools): search_files follows a symlinked root on the grep • fix(kanban): worktree dispatch works on repo paths with non-ASCII • fix(dashboard): a parent-injected loopback session token • fix(kanban): diagnostic severity colours follow the dashboard theme • fix(lsp): a language server whose startup overruns the budget • fix(kanban): tapping a card on a touch device opens it • fix(update): Windows post-swap hand-off • fix(terminal): subprocesses get a HOME • fix(acp): multi-file V4A patches hiding .env or out-of-workspace • feat(kanban): kanban create accepts --body-file • fix(lsp): incremental servers keep an exact document mirror • fix(agent): verify-on-stop records ad-hoc evidence for versioned • fix(acp): a failed post-turn tail no longer wedges the session • fix(cli): hooks test shows the block/allow decision for erroring • fix(kanban): board Done column lists newest-completed first • fix(auxiliary): reasoning_effort none turns thinking off • fix(vision): computer_use capture takes the same native lane • fix(kanban): kanban tools reject misspelled arguments • fix(utils): atomic writes fsync the directory • fix(models): retired OpenCode Zen model x-preview-f-free leaves • fix(kanban): dispatch pass survives a BLOB comment body • fix(bedrock): Kimi K3 reasoning replays in the Converse schema • fix(usage): /usage and /billing release at their timeout • fix(kanban): notifier warns once • fix(api-server): custom-provider model locks • fix(state): a clean close no longer poisons • fix(state): deleted-WAL guard finds holders through a symlinked • fix(compression): a dispatch-only provider failure • fix(kanban): an unbound stale worker • fix(sessions): un-hiding a born-hidden session survives the next • fix(state): delegate and branch children • fix(mcp): strict_redirect_headers really strips configured headers • fix(update): a parked autostash is recorded in the update receipt • fix(state): negative sessions.retention_days • fix(model-picker): a built-in aggregator row keeps its models • fix(update): the post-update import probe • fix(config): the .env writer refuses every subprocess-execution env • fix(docker): the orphan reaper no longer force-removes a container • fix(docker): a transient capability-probe failure • fix(update): a launchd-supervised `hermes dashboard`/`serve` • fix(update): a manual-serve reminder with no readable create_time • fix(update): `hermes update` no longer aborts on an intent-to-add • fix(delegation): subagent panel survives UI session rotation • fix(langfuse): nested tool payload depth is configurable • fix(mcp): parked servers stop flooding the error log and honour • fix(update): a fleet-restart marker with no recorded obligation • fix(config): tools.tool_search.defer is a recognized key • fix(tools): interrupting a delegate batch • fix(streaming): explicit providers.<id>.stale_timeout_seconds • fix(mcp): structuredContent is kept • fix(cli): a list-literal string in platform_toolsets keeps • fix(auxiliary): a providers: entry named like a local-server alias • fix(auxiliary): aux tasks resolve a client for every provider alias • fix(relay): a failing session push is not retried • fix(config): malformed nested config values • fix(local-models): a user-run llama.cpp server on a configured port • fix(agent): misnested or truncated tool-call argument JSON • fix(agent): provider content-moderation 400 (Content Exists Risk) • fix(relay): managed streams close deterministically, aclose • fix(update): stale git pack temps are actually removed on Windows • Plugin catalog ranks by stars only and drops the review-claim badges • Codex app-server threads started from scratch inherit the session • Model-provider plugins classify their own API errors and bench • Provider plugins declare per-model capabilities once • Provider plugins drive first-time setup catalogs, image routing • Model-provider plugins can report account usage in every /usage • Make the second-stall fallback test stop racing the detached • fix(models): reduce repeated Nous recommendation traffic • docs: website, gateway, memory, tools, and matrix/bluebubbles guides (7x) SECURITY & SECRETS (42 PRs) • fix(cli): a provider quota 429 • fix(tui-gateway): shell.exec scrubs the child env and redacts • fix(approval): empty smart-approval guardian answers now warn • fix(approval): safe git branch -d no longer gated as force delete • fix(skills): Python open()/read_text() of credential files lands • fix(skills): guard flags 'Send child context' as exfil again • fix(kanban): respawn guard stops parking cards on "author" prose • fix(url_safety): IPv4-translated DNS answers (::ffff:0:x.x.x.x) • fix(runtime): pooled credentials • fix(plugin-guard): v8 , four intake false-positive classes step • Claude Code version detection finds the CLI outside the GUI PATH • fix(mcp): an expired dashboard OAuth attempt is re-minted on retry • fix(anthropic): key_cmd Claude Code OAuth identity survives • fix(approval): deno eval asks for approval behind global flags like • fix(mcp): OAuth login waits for oauth.timeout on CLI, dashboard • fix(auth): Copilot token exchange (and its RAW-token warning) runs • fix(agent): multiplexed profiles • fix(local-models): llama-server no longer inherits credentials • fix(desktop): the Approval Needed arrow scrolls to the pending • fix(guard): an env-var NAME constant • fix(skills): godmode's own install paths resolve • fix(dashboard): plugin install/update • fix(git-auth): an expired GITHUB_TOKEN in .env yields to the gh CLI • fix(vault): corrupt vault files and non-dict vault config surface • fix(auth): shipped .env.example and docs placeholders • fix(auth): a removed XAI_API_KEY still in the gateway environment • fix(web): xAI credentials alone • fix(security): case-variant credential env names • fix(dashboard-auth): self-hosted OIDC discovery rejects documents • fix(approval): bun -e and deno eval ask for approval like node -e • fix(approval): pipe-to-shell and -c payload detection covers zsh • fix(docker): egress collision guard protects every mapped • fix(tui): composer path completion on a remote backend • fix(mcp): OAuth login keeps discovered server metadata, device-flow • fix(approvals): plugin-escalated approvals ask on /v1/runs • fix(mcp): OAuth discovery accepts an origin-issued metadata • fix(mcp): browser OAuth callback is registered even • fix(cron): credentialless satellite cron results reach the primary • Provider plugins get a declarative OAuth PKCE login • fix(plugin-guard): plural test-file names and delegation prose • OAuth-shaped provider plugins now register, log in and refresh • Model pickers now list out-of-tree external-process and OAuth INTEGRATIONS & CHAT (84 PRs) • fix(cli): kanban edit accepts --title/--body/--priority and cron • fix(cron): context_from self-continuity injects the previous answer • fix(gateway): restart-safe cron deliveries go out when queued • fix(gateway): typing indicator stops before a stalled final send • fix(cron): near one-shot Slack reminders deliver into the thread • fix(cron): bot-chat delivery to another profile runs • fix(simplex): alpha PNG images deliver • fix(gateway): multiplexed profile • fix(update): gateways running a separate checkout • fix(cron): recovery keeps a live owner • perf(gateway): the identity check reads a PID's command line • fix(whatsapp): a legacy bridge pidfile no longer kills a stranger • /branch opens a sibling thread by default • Gateway runtime-status writes leave the event loop, so a slow disk • Gateway signal-driven stop drains inside launchd's ExitTimeOut • Matrix pending-invite reconciliation applies the inviter allowlist • Telegram Bot-API sends get wall-clock deadlines, so one hung upload • Cron bot-chat alerts survive a delivery timeout: a late turn report • Cron reclaims a live-but-deadlocked worker's claim once it outlives • Gateway systemd units carry `LD_LIBRARY_PATH` • WhatsApp and Weixin replies stop waiting 5 s / 3 s • Per-model `context_length` for custom providers is honored even • Gateway `/model` no longer joins live provider-catalog probes • fix(gateway): a busy redirect's final reply quotes the redirecting • fix(gateway): Windows status names pre-suffix launchers • fix(tui-gateway): a Bot Chat lease no longer outlives an iOS client • fix(dashboard): resuming a chat after a gateway restart never lands • fix(gateway): a wedged cron run no longer holds the restart drain • fix(tui): a streaming gateway socket • fix(gateway): unclean-exit state.db check renews • fix(telegram): streaming previews stop earning flood penalties • fix(bot-mode): a remote default Bot is mentionable by its title • fix(api_server): GET /v1/runs/{id} shows shutdown_requested_at • fix(bot-mode): renamed profiles stay linked in existing group chats • fix(bot-mode): New Bot on another machine can start fresh • fix(bot-mode): a teammate's reply reaches the sender whole • fix(bot-relay): a relayed turn still lingering at the cap is booked • fix(bot-relay): peer gateways name each machine by its connection • fix(platforms): Discord/WhatsApp/DingTalk allowlists stored • fix(peer): hermes peer dm retries a transient turn once and hides • fix(gateway): shutdown keeps state.db open • fix(bot-mode): disbanded group chat stays gone • fix(bot-mode): a schema-damaged mailbox ticket • fix(cron): bot-chat delivery survives a deleted scheduler cwd • fix(bot-relay): a relayed DM into a sign-in-gated gateway • fix(bot-mode): a teammate's DM reply reaches the sender whole • fix(peer): a timed-out peer DM says the turn was accepted • fix(telegram): large videos keep their real aspect ratio • fix(messaging): hermes send reports dropped MEDIA attachments • fix(tui): losing the terminal no longer respawns the gateway • fix(gateway-windows): Scheduled Task drift check settles • fix(gateway): Windows gateway setup asks the install questions once • fix(discord): a pasted channel link in DISCORD_HOME_CHANNEL • fix(gateway): a platforms.a2a-platform config block loads the A2A • fix(update): Windows 'Already up to date' run warns • fix(telegram): dm_topics prerequisite warning and docs point • fix(telegram): sibling-bot wake-word and quote-reply messages land • fix(gateway): a named profile's gateway identity • fix(gateway): Windows update no longer aborts • fix(whatsapp): reply_prefix from config.yaml reaches the bridge • fix(discord): third-party mentions in a bot-joined thread • fix(tui-gateway): launch-profile cwd is read from the launch home • fix(gateway): reconnect_attention_after read from the bound • Multiplexed gateway housekeeping runs per profile scope • fix(config): hermes config set gateway.platforms.<p>.enabled lands • fix(tui-gateway): a notification that loses its delivery claim • fix(tui_gateway): live compression config applies • fix(gateway): direct systemctl restart/stop exits 0 • fix(cron): bot-chat delivery keeps accented replies on Windows • fix(bot-relay): cross-machine replies wake the sender • fix(update): this install's legacy hash-suffixed launchd gateways • fix(gateway): shutdown tool-subprocess kill sweep runs off • fix(tui-gateway): resume a secondary profile's session • fix(cron): a hung standalone delivery • fix(gateway): API-server-only gateways • fix(gateway): warm-up gate-release warning • fix(cron): external-worker one-shot runs • fix(gateway): leaked <|eos|> control tokens • fix(gateway): one raising startup-restore replay • fix(gateway): blank chat_type sources • Gateway chat reply no longer calls a reset model connection • Bot Mode group chat members see the whole recent room • feat(cron): jobs follow the main agent model at fire time • fix(gateway): api-server requests keep the tool-availability cache STABILITY & FIXES (10 PRs) • Unset agent.reasoning_effort sends medium on custom • stderr wrapper: comment names why SIGUSR2 is forwarded • state: archive-mode rewind archives only the dropped suffix • state.db lock timeouts now log which process held the write lock • chore(deps): package-lock refresh clears the remaining production https://x.com/iamlukethedev/status/2100074765801459998?s=20

A new Claude Code mod adds Jev routing to every request, classifying the subagent model, main model and effort level automatically through the Typesafe AI API or Vercel AI Gateway with one install command.

禁止事項はハーネスで介入すべきと思うけど、CodexのHooksは機能が限定的すぎる（App Server使うって手はあるけど迂遠すぎる） 本当はJev使って介入したいんだよなぁ Claude Mods使うと内部イベントに介入できるっぽいので今後はClaude Mods + Jevがめっちゃ流行る気がしている

Claude Code added support for AGENTS.md this week. Small change, but it closes the last major gap in a standard that quietly won already. The mechanism: AGENTS.md is a plain markdown file at the root of a repo, a README written for agents instead of humans. Build steps, test commands, code conventions. No required fields, no schema to validate against. The model just reads it like any other file. What makes this different from the usual new AI standard announcement is who is behind it. OpenAI released the spec in August 2025. By late 2025 it had moved to the Agentic AI Foundation, a new body under the Linux Foundation that OpenAI co founded alongside Anthropic's MCP and Block's Goose. That is three competing labs putting their interoperability layers under neutral governance instead of each running their own. Adoption backs up the governance move. More than 60,000 open source projects use AGENTS.md today. Cursor, Devin, Factory, Gemini CLI, GitHub Copilot, Jules, and VS Code all read it natively. Claude Code was the holdout, defaulting to its own CLAUDE.md. As of version 2.1.277 it now falls back to AGENTS.md when no CLAUDE.md exists, and Anthropic open sourced the compatibility mod rather than keeping it internal. I keep coming back to how familiar this shape is. Kubernetes did this to container orchestration. Once every cloud vendor's proprietary scheduler lost to one config format everyone could target, the argument moved up a layer, to what you build on top, not which format you write in. Coding agents were heading toward the same fragmentation eighteen months ago, a different config file for every CLI tool, every repo maintaining four versions of the same instructions. That tax is basically gone now. The bottleneck that's left is not which agent do I configure, it's whether your team actually keeps one instruction file current across a growing pile of tools that all read it. The format stopped being the hard part. Maintenance discipline is what is left. https://x.com/addyosmani/status/2101010420282855566

I built this Claude Code Mod using TypeSafe’s Skill Suggestion example It points to a different way of managing the LLM context window: delegate skill, effort, and tool selection to decision models like Jev The harness can then be split into smaller, independent decision services instead of putting everything inside the main agent loop

Introducing Jev Skill Suggestion for Claude Code Now you can keep your skills completely out of the context window until they’re actually needed Set your skills as user invocable only, and the Mod sends the available skill list to Jev For every request, Jev classifies which skill best matches the task, then injects only that skill into Claude’s context Works with both the @typesafeai API and @vercel AI Gateway Install it with one command: npx claude-code-templates@latest --mod productivity/jev-skill-suggestion Full implementation details on the Mod page at http://aitmpl.com

Claude Mods 入門 | Claude Codeを自由にカスタマイズする https://zenn.dev/nogu66/articles/claude-code-function-hooks-claude-mods

Unity 公式の31スキルは、オープンソースではありません。 Unity Technologies がコーディングエージェント用の公式プラグイン（スキルをまとめた拡張パック）を配っています。配布元は Unity-Technologies/unity-agent-plugin です。skills は31個、version は 0.1.6-beta です（2026-09-21 実測）。 ライセンス欄の文字列は LicenseRef-Unity-Companion-License です。GitHub のライセンス判定も NOASSERTION を返します。 Releases とタグは0件で、版数は plugin.json の中だけで管理されています。1つのリポジトリに .claude-plugin と .codex-plugin の2つのマニフェスト（プラグインの定義ファイル）が同居しています。Claude Code と Codex の両方が、同じリポジトリから入ります。 Unity Companion License v1.4 の本文には、4つの条件が書かれています。 1．使える場面が限定されています。 ライセンスの行使は、有効な Unity エンジンライセンスのもとでアプリやコンテンツを制作・配布する場面に限られます。本文は「それ以外の行使は認めない」と明記しています。競合分析や競合製品の開発への利用も禁止です。 制限が効くのは、Unity を使わない現場で31スキルを社内標準にしようとした人です。Unity 案件だけで使う開発者には効きません。 導入前に、自分の用途がエンジンライセンスの下にあるかを確認してください。 2．スキル自体を改変すると、権利者が変わります。 第3.2条は、Work の派生物の知的財産権をすべて Unity へ譲渡すると定めています。派生物とは、スキルのコードそのものを書き換えたものです。権利者は、改変した本人から Unity へ移ります。 一方で第3.1条は、そのスキルを使って作ったゲームやアプリは派生物に含まれないと明記しています。作った成果物の権利は手元に残ります。 線は「スキルを直したか」「スキルで作ったか」で引かれます。 社内向けにスキルを書き換えて配る計画があるなら、改変せず使うか、自前のスキルを別に書くかを選んでください。 3．商標の許諾は含まれていません。 第4条は、Unity の商標・ロゴ・製品名を使う権利を与えないと書いています。記述的な用法だけが許されます。 効いてくるのは、Unity 対応を自社サービスの売り文句として打ち出す事業者です。個人の開発者には関係ありません。 サイトや資料に Unity のロゴを載せる予定があるなら、商標ガイドラインを別に確認してください。 4．保証は付きません。 LICENSE.md は AS IS、いかなる保証も付かないと明記しています。第5条は、著作権表示を再配布物にも残すよう求めています。 課金設定やビルド構成をスキルに任せても、結果の責任は利用者側に残ります。 生成された設定は、コミット前に自分でレビューしてください。 導入は2行です。Claude Code では次の2つを順に打ちます。 /plugin marketplace add Unity-Technologies/unity-agent-plugin /plugin install unity@unity-agent-plugin ベンダー公式という言葉は、OSS を意味しません。plugin.json の license 欄を1回見れば、社内配布の可否は判断できます。 出典 https://github.com/Unity-Technologies/unity-agent-plugin https://raw.githubusercontent.com/Unity-Technologies/unity-agent-plugin/main/.claude-plugin/plugin.json https://raw.githubusercontent.com/Unity-Technologies/unity-agent-plugin/main/LICENSE.md https://unity.com/legal/licenses/unity-companion-license #Unity #ClaudeCode #Codex #AI

No f*king way. They just opened up the inside of Claude Code AGENTS.md support landed this week. What shipped underneath it changes everything. It shipped as a mod. A mod is a plugin whose behaviour lives in a hooks module: one register(on, options) entry that wraps the engine's own events as ($, e, next). Middleware around Claude Code itself, with ordering and a next. Four mods are built into the binary, and their source is now public. > agents-md, the one everyone saw. > diff, which draws your uncommitted changes in a pane beside the transcript and refreshes as Claude edits. > telemetry, which hands plugins a $.telemetry to log first party events. > sec-default, which seats outermost on managed machines and keeps org hooks and tool policy out of reach of whatever plugins a person installs. The events in their own example: session.start, command.register, process(.)run, ui(.)open. Session opening, command registration, process execution, interface drawing. Run one from source: claude --plugin-dir mods/diff Test your own: claude plugin test mods/diff Repo: /anthropics/claude-code/tree/main/mods One thing hiding in there: agents-md has four modes, not one. claude-md-and-agents-md loads both files instead of falling back to one.

how to setup and use Jev in 10 minutes from nothing to a call that returns a decision 1. get access join the waitlist at typesafe, people report getting in within a day don't want to wait, it's already hosted in two places: → vercel ai gateway as typesafe-ai/jev, no waitlist → cloudflare workers ai as typesafe/jev 2. install the skill so your agent writes the call correctly npx skills add typesafe-ai/skills --skill typesafe-ai on claude code it's two commands: - claude plugin marketplace add typesafe-ai/skills - claude plugin install typesafe@typesafe-ai then in your prompt just say "use the typesafe skill" 3. learn the shape before you write a line the body is a state and a map of questions. the state is your context, whatever you would have pasted into a prompt. the questions are typed three types, that's the entire api: → Noul returns a raw 0 to 1 → Choice picks one from up to 255 options you define → Score places it on a 2 to 10 level scale every Choice and Score comes back with a confidence from 0 to 1 4. make the first call a real one pick a decision already living in your code as an llm call that never writes text. the "which of these" one how to do it: - build the option list in your own code, not in the prompt - send your existing context as the state, unchanged. don't rewrite it yet - ask your question plus three more you're curious about, they run in parallel against the same state and output tokens are free - log the answer, the confidence, and what your old call said - leave the old call running. delete nothing on day one 5. put a limit on it before it touches anything read confidence on every response, never just the answer low limit for reading and sorting. 0.85+ for anything you cannot undo. everything under the limit falls through to the old call that fallback is four lines, and it's the reason you can ship this on a tuesday instead of planning it for q4 the 10 minutes is real. the hour after is the actual work: you, a log of answers and confidences, deciding where your lines go start in the playground in the console, and paste a real state from your own product, not the sample one

🚀 推荐一个聚焦 Agent 决策可靠性的开源项目：TypeSafe Agent Skills（typesafe-ai/skills） 在构建 AI Agent 时，最让人头疼的问题之一就是大模型的“决策黑盒”——输出缺乏严格的类型约束，且无法量化判断的置信度。 这个项目专为 TypeSafe 的 System One API 设计，核心目标是将类型化决策（Typed Decisions）与概率评估（Probabilities）直接引入 Agent 的推理与执行流程。 💡 核心亮点与典型场景： 类型化精准判定：强制模型输出严格符合结构化定义的决策结果，拒绝模糊自由发挥，保障下游代码稳定调用。 置信度驱动的人机协同（HITL）：例如在客服工单智能路由场景中，对高把握的决策自动流转，低置信度的边缘情况则自动标记并交由人工复核。 工作流快速设计：内置 Skill 能帮助 Agent 自动检索最新文档与 Cookbook，直接在代码中编排类型化判断逻辑。 📦 安装与使用： Claude Code 插件： claude plugin marketplace add typesafe-ai/skills claude plugin install typesafe@typesafe-ai 跨平台 Agent： npx skills add typesafe-ai/skills --skill typesafe-ai 如果你的 Agent 正在处理工单分流、风控筛选等对确定性要求极高的任务，不妨关注一下这个方案。 🔗 项目地址：https://github.com/typesafe-ai/skills #AIAgents #ClaudeCode #CursorAI #OpenSource #DeveloperTools #LLM #TypeSafe #AI开发 #SoftwareEngineering #AgenticAI