Community posts

i agree with everyone saying mods are the real story here. this has been the actual ask since Claude Code shipped not more features, room to build on top of it. someone shipping tetris inside it within days is the proof. more of this is coming, and it's going to move fast.

新機能Claude Modsが登場 （以前Function Hooksとして告知されていたもの） 従来の hooks より一段深く、Claude の実行パスに関数ベースのミドルウェア として割りこめる。テトリスを作ることも可能 CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 上記の環境変数で有効化できる

@bcherny Mods are already available in Command Code btw 👀 https://commandcode.ai/docs/mods

Claude Mods is a total game changer for deeper customization of Claude Code. Super excited for this one.

@bcherny @bcherny Any plans of giving some love to IntelliJ idea CC plugin? It's still in beta, was last updated in Dec 25 and has a 2.3 out of 5 rating :-/

@bcherny interesting Claude Mods are basically turning Claude Code into a programmable platform

@bcherny Do mods give me control over compaction/context windows now? 🥺

@bcherny bro people are really modding claude to play tetris now 💀

Claude Mods are landing now. Someone already built a Tetris-in-Claude mod 🤯 See issue for the latest community update, technical details, and more cool demos https://github.com/anthropics/claude-code/issues/91870#issuecomment-5666255143

Claude Codeの拡張、こんなところまで来てるらしい🦀 「Claude Code Templates」というnpmのツールが Function Hooks に対応したとのこと。 Claude Codeの動きに、 自分の処理を差し込めるようになるかもしれません。 作者いわく、具体例は今週中に公開予定。 気になる人は名前だけ覚えておくと良さそうです。 https://github.com/davila7/claude-code-templates/releases

GROK BUILD HOW TO USE Sources: official http://docs.x.ai/build + http://x.ai/cli + your Grok Build bookmark folder (306 posts) Updated: 2026-09-14 WHAT NEW TODAY - Folder now 306 posts (3 new vs 303 on Sep 13) - Official (elonmusk, Sep 14): Try Grok Build — powered by Grok 4.6; native subagent view, Plan Mode integration, mouse support, fullscreen TUI - Testers (XFreeze / cb_doge, Sep 13): Grok Build v1.0.31 — multi-agent sessions easier to read (running vs finished subagent counts); dashboard search/resume more predictable; subagent status labels, dock sections/paths, cancellation banners/messages; MCP tools with long server prefixes fixed - Official changelog still Latest v1.0.30 · Sep 11, 2026 (v1.0.31 not listed yet; checked Sep 14) WHAT IT IS - A terminal-native coding agent with a fullscreen TUI - Also runs headlessly in scripts, or inside other apps via ACP - Understands the repo, edits files, runs shell commands, and manages long jobs - Open source harness: - Different from Grok Bot (cloud teammates) and Digital Optimus (real-time video control) - Bookmark testers: native /remote-control support is coming (SpaceXAI confirmed); Warp now ships built-in Grok Build CLI with /remote-control (official grok Sep 11) INSTALL AND UPDATE [] macOS / Linux / WSL: curl -fsSL | bash [] Windows PowerShell: irm | iex [] Then: grok --version [] Update: grok update (stable) or grok update --alpha - Testers: v1.0.5 adds GROK_CONFIG env, auto worktree reclaim, hook-block wording, RTL, ACP effort, image/video call limits - Testers: v1.0.6 changes subagent spawning (breaking) and adds a customizable status line - Testers: v1.0.7 adds a first-class workflows catalog, Always/Never rules, and a timer-refresh status line - Testers: v1.0.8 improves MCP servers and workflows. Concurrent subagents start faster without freezing the parent or UI. Follow-ups send while a task is running. Ctrl+S stashes the prompt draft. /workflow autocompletes. MCP can request form input or URL consent. Workflow rows show current context usage - Testers: v1.0.9 adds agent budget + reasoning controls, plugin-provided agents, /feedback images, instant TUI mode switching, smarter clone/worktrees, resilient subagents, faster MCP startup, concurrent agents without rate-limit collapse - Testers: v1.0.10 — grok clone reuses matching local checkouts as linked worktrees for faster session creation - Testers: v1.0.11 — Auto/headless QoL: resume-picker headless sessions, configurable default permission mode, Auto subagent messages, mkdir/touch without prompts, no 10-hour monitor timeout, faster background waits, turn duration after /resume - Testers: v1.0.12 — faster worktrees; MCP retry; subagent wait after interjection; hook descriptions; context/token accuracy; auto-compaction bar; .NET watcher fix; Auto recap during background streams - Testers/Official: v1.0.13 · Aug 28 — compressed CLI downloads; truncated responses continue; hooks can confirm (not only allow/deny); inference retry on stalls/drops/5xx; Windows ~/.grok and worktrees; more reliable session save; faster MCP startup when auth configured; truncated tool calls with complete args execute; large-image session fix; scheduled-task UUID IDs + stop reminder - Testers: v1.0.14 — PostToolUse hooks feed model feedback; grok usage <session-id> per-turn tokens/cost; proactive OIDC refresh; stronger subagents; ~70% smaller Windows CLI downloads; large-session memory no longer blocks turns - Testers: v1.0.15 — background model connect/token refresh/MCP at startup (faster first reply + new sessions); memory consolidation deferred off exit path; /copy and /export tip after repeated drag-copies - Testers: v1.0.16 — MCP servers at session bind for workspace integrations; enterprise model restrict via signed requirements.toml; 1h default wait ceiling for long subagent/task output; MCP OAuth from /mcps deadlock fix; sessions recover after expired-token network interrupts - Testers: v1.0.17 — MCP multi-round-trip elicitation (input_required → resume); quieter next-prompt ghost suggestions; /btw panel table cell select/copy as TSV - Testers: v1.0.18 — managed policy blocks disallowed MCP servers and marketplace installs before write; interactive Chart.js charts; richer image-gen params via composer state; configurable per-model rate-limit retry; per-model mTLS for secure upstream; faster startup/first responses with more work in background; session/subagent/auth/composer/UI fixes - Testers: v1.0.19 — scheduled /loop tasks always run in background (no turn inject); session resume reports running loops/subagents/workflows; headless `grok -p --worktree`; dashboard shows new sessions immediately + /usage from dashboard; transparent terminal theme; faster first prompt after login with MCP; UI freeze/crash fixes - Testers: v1.0.22 — first-party desktop MCP; `grok-workspaced daemon` for Computer Hub folder expose; continue finished subagents; background agents return real output; real line-number diffs + auto-expand on edit permission; built-in tools beat user MCP name collisions; remote-control device grouping; Auto blocks destructive `git checkout --`; faster open/resume (MCP config off critical path) - Testers/Official: v1.0.23 — bare URLs and emails inside table cells stay clickable on every wrapped line - Testers/Official: v1.0.24 — Esc no longer cancels a running turn (use Ctrl+C) - Testers/Official: v1.0.25 · Sep 9 — `/workflow` pause/stop for agent-launched runs; durable background-task snapshot across reconnect; silent successful hooks; full Bash pager output; `/theme` alias match; voice inserts at cursor; scheduled-task cleanup reminders; `/feedback` immediate send; dashboard/actions/UI fixes - Testers: v1.0.26 · Sep 10 — subagent interjects interrupt long background waits; scrollback labels how messages were sent; feedback form available in minimal mode - Testers: v1.0.27 · Sep 10 — queued messages no longer duplicate when sent immediately; dashboard spinners stay live during background tasks - Testers: v1.0.28 · Sep 10 — mid-message `/btw`; per-model `reasoning_summary` in config.toml; safer multi-process auto-update; soft-wrapped prompt Up/Down fix - Testers: v1.0.29 · Sep 11 — pager Tab cycles prompt/scrollback/whole dock; Ctrl+G hide/show dock; concurrent `grok agent stdio` same GROK_HOME startup crash fixed; MCP 2026-07-28 servers; Windows Client-ProjFS-off startup fix; idle dashboard CPU paint reduced - Testers: v1.0.30 · Sep 11 — session header single-row (overlay context + Dashboard); tmux pane lag fixed; elapsed time into hours; Watchers/Loops next-run due; workflow status clickable in dock - Testers: v1.0.31 · Sep 13 — multi-agent readability (running vs finished subagent group counts); dashboard search/resume predictability; subagent status labels; dock sections/paths; cancellation banners/messages; MCP long server-prefix tools fixed - Official changelog Latest v1.0.30 · Sep 11, 2026 (checked Sep 14; testers report v1.0.31) - Bookmark testers: Grok Build mode is now on all SuperGrok and X Premium plans, web and mobile - Testers: official Netlify plugin for build and deploy - Official/Warp: Grok Build CLI runs natively inside Warp (rich prompts, /remote-control, file explorer + code review) - Testers: Quo communications plugin — SMS (individual/group/bulk), contacts, tasks, message/call transcripts, missed-call follow-up (`grok plugin install quo --trust`) - Bookmark testers: Grok Build can find/sequence footage and do real image/video editing (cut, assemble, Photoshop-style changes) from a brief — not only generate; save preferred designs/workflows as skills - Bookmark testers: Grok Build Mode can ship simple web apps (example: image watermarker) - Bookmark testers: Grok Build can drive Blender 3D workflows (examples: rocket launch site from scratch; Blender + Imagine car-chase cinematic clip) - Bookmark testers: a usage-reset control can set usage back to 0% - Changelog: - Evaluate Grok 4.6 with the Grok Build harness (http://grok.com/build) - Docs: - Download page: SIGN IN [] First launch opens a browser for authentication [] Headless or remote: grok login --device-auth [] Or: export XAI_API_KEY="xai-..." then grok [] Sign out: grok logout [] In TUI: /login and /logout START A SESSION [] cd your-project then grok [] First prompts: Explain this repo. or src/main.rs Walk me through this file. [] grok -p "Explain this codebase" for a one-shot headless run [] grok -c continues the most recent session in this directory [] grok -r resumes a session by id (or the most recent) [] grok inspect shows discovered config, skills, plugins, hooks, and MCP [] grok models lists available models [] Switch model in TUI with /model LEARN THE TOOL [] Run /tutorial or /help [] /release-notes or /changelog for the current version [] Read AGENTS.md in the repo (also CLAUDE.md if you are migrating) [] Import Claude settings with /import-claude [] Carry over rules, skills, and MCP from Claude / Cursor / Codex [] Keyboard shortcuts: /help and Ctrl+. [] Voice: Ctrl+Space or /voice [] Shift+Tab cycles Plan, Auto, Always-approve HOW TO WRITE A TASK [] Outcome: what should be finished [] Sources: files, folders, URLs, or mentions [] Constraints: what not to touch, and when to ask [] Deliverable: diff, report, test, or published app [] Review point: plan first, or stop before commit/push [] For big work: /plan first, then approve [] For long ownership: /goal [] For research first: /deep-research [] For a sick laptop: tell it what is broken and let it investigate [] Testers (tetsuo): start with work you can check — name the behavior, give a reproduction command, limit files, list final checks, say where to stop [] Testers (tetsuo): inspect first (git status --short, grok inspect); give executable feedback (compiler/linter/tests) so the agent can loop; review the full diff and the command behind every green check CORE COMMANDS [] /new starts a new session [] /home returns to the welcome screen [] /resume and /sessions switch sessions [] /fork branches the current session [] /rename or /title names the session [] /session-info shows session details (click or drag to copy) [] /context shows token and skill/MCP cost [] /usage shows credits [] /compact condenses history [] /rewind truncates conversation history (confirms; does not wipe files) [] /export saves the conversation [] /btw asks a side question without derailing the turn (v1.0.28: also works mid-message — text after `/btw` is the side question and stays out of the main transcript) [] /check-work has a subagent review the code [] /feedback files a report [] /doctor and grok doctor fix terminal, tmux, clipboard, keyboard [] grok du shows disk use under ~/.grok [] grok wrap runs a command with local clipboard support [] grok trace exports debug traces [] /loop runs a prompt on an interval (auto-expires after 7 days) [] Queued follow-ups can send as interjections if you turn that setting on [] Ctrl+4 opens the Agent Dashboard [] Ctrl+U can continue a Claude Code session inside Grok Build PLAN MODE AND PERMISSIONS [] /plan or Shift+Tab enters plan mode [] Only the plan file can be edited until you approve [] Auto and always-approve do not skip plan review [] a approves, s requests changes, c comments, q quits plan [] /view-plan reopens the preview [] /auto auto-approves safer tools [] /always-approve skips most prompts (deny rules and hooks still apply) [] Edit always-allow bash as free-form globs [] Restrict web search domains with [toolset.web_search] in config.toml [] PreToolUse hooks can rewrite tool input, not only allow or deny [] StopCancelled reports when a turn is interrupted or rejected [] Pin auth to API key or OIDC in config.toml if needed PARALLEL WORK AND AGENTS [] Agent Dashboard: /dashboard or Ctrl+4 [] Run feature, review, fix, and test agents from one place [] Pause, check progress, and jump into any agent [] Worktrees: grok -w or Ctrl+W (isolated git copy under ~/.grok/worktrees) [] grok worktree list / show / rm / gc [] /fork --worktree forks into a worktree [] /goal owns a long engineering objective (status, pause, resume, clear) [] /create-workflow authors a repeatable multi-agent pipeline [] /workflow launches, pauses, resumes, or stops a run [] /workflows is the live run dashboard [] /deep-research starts a built-in research workflow [] Tools and MCP servers receive GROK_SESSION_ID [] Background tasks and TODOs survive compaction VOICE AND MEDIA [] Dictate with Ctrl+Space [] /imagine generates an image [] /imagine-video generates a video [] Ask it to boost, compress, and re-encode video audio in plain English [] One-prompt 3D: ask it to model in Blender and watch it drive the app [] Morning voice-memo automations can research and read you a briefing [] Disable image/video tools via config or env if you do not want them [] Testers: save preferred image/video edit designs and workflows as skills so repeats match your style SKILLS PLUGINS AND MCP [] /marketplace installs plugins [] /plugins /skills /hooks /mcps open the same extensions modal [] Skills live in ./.grok/skills, ~/.grok/skills, and plugin skill folders [] User skills appear as slash commands; use /local:name if names collide [] grok plugin install / enable / disable [] grok mcp list / add / doctor [] Type / to reference a skill [] Testers: save preferred designs and editing workflows as skills so Grok Build reuses your style every time [] Common marketplace plugins: Exa, TinyFish, MongoDB, Vercel, Sentry, Cloudflare, Chrome DevTools, Supabase, Neon, Firecrawl, Figma, Railway, Stripe [] Official/Testers: browser-use plugin — /plugins or `grok plugin install browser-use --trust`; enable Chrome remote debugging to drive your signed-in local Chrome (or isolated cloud browser; local uvx, no API key for local Chrome) [] Managed MCP servers come from the gateway catalog [] Add a local directory as a marketplace source if you keep private plugins WORKTREES AND FILES [] Worktrees start from current HEAD, including uncommitted changes [] grok -w --ref main starts from a clean ref [] Ending a session does not delete the worktree; run grok worktree gc [] Keep durable files in the repo or ~/.grok, not temp dirs [] Config: ~/.grok/config.toml (Windows: %USERPROFILE%\.grok\config.toml) BUILD MODE AND PUBLISHING [] Build Mode on http://grok.com (and mobile where available) [] One prompt can become an app on a http://grok.me domain [] Export source to GitHub [] Examples: http://driver.grok.me, http://bbox.grok.me, http://gravity.grok.me [] SuperGrok Heavy may be required for some publish features [] Bookmark testers: Grok Build is now available for free users too — describe what you want, Grok builds apps/websites/games/dashboards, then publish and share [] Testers: Tesla Owners Silicon Valley — creating from a prompt is open to free users; a live http://grok.me URL still requires a paid plan [] Testers: Android (X Freeze) — GitHub push, secrets store, invite-only lock, download builds, custom domains, share to X; end-to-end ship from a phone HEADLESS AND AUTOMATION [] grok -p "task" for scripts [] grok -p "task" --output-format streaming-json [] grok agent stdio for ACP in other apps [] grok serve --protocol acp --port 9120 for a local ACP server [] Custom models go in ~/.grok/config.toml under [http://model.name] COMMUNITY TOOLS [] GrokTerm: multi-harness terminal host with voice [] desktop-harness: Mac Accessibility control (http://github.com/xfreeze2/desktop-harness) [] Quill: universal SuperGrok voice layer on macOS [] AFK Pilot / Grok Build Desktop: remote edit and syntax highlight [] Grok Build for VS Code (Community) extension — Testers: Routines dashboard for scheduled executions across Grok Build, Codex, and Claude Code (89K+ installs; not affiliated with SpaceXAI) [] Unity CLI / MCP for game work SAFETY AND USAGE [] Prefer plan mode before large edits [] Prefer always-approve only after you trust sandbox and deny rules [] Opt in to share traces only if you want to (off by default) [] Privacy banner can be dismissed from Settings [] Video tools may require ZDR storage [] Check /usage before long multi-agent days [] Grok 4.6 may run double-token promos; check current offers OFFICIAL LINKS [] Overview: [] Modes and commands: [] Plan mode: [] Worktrees: [] Skills and plugins: [] CLI reference: [] Changelog: [] Install: [] Source:

What if your AI coding agent could actually understand Google Cloud? Google just released a free Google Cloud Developer Plugin for AI coding agents. It works with Claude Code, Codex CLI, and Antigravity. The idea is pretty simple: Instead of asking your coding agent to guess how a GCP command, authentication flow, or cloud setup works, the plugin gives it access to Google’s official cloud knowledge and tools. For example, you could ask your agent to: › Set up a GCP project › Configure authentication › Work with cloud resources › Help with "gcloud" commands › Find the right solution from Google’s documentation How to use it If you use Claude Code: "claude plugin marketplace add google/skills" Then: "claude plugin install google-cloud-developer@google-plugins" Codex CLI and Antigravity have their own installation commands too. The plugin itself is free. Your normal Google Cloud and API usage is still billed separately. I have not personally tested it yet, but if you use AI coding agents + GCP, this looks genuinely useful. The main idea I like here: Less guessing by the agent. More grounding in the tools and documentation it actually needs.

Claude Code’s latest update has two separate stories—and they should not be conflated. First, the limit headline. A “25% increase” can still be a net cut if the comparison is against the old baseline while temporary boosted limits users have been receiving are rolled back. That is a usage change, not a feature win. Second, the genuinely big release: • Cross-session messaging lets separate sessions discover and message each other. • Fork mode gives a background subagent the full conversation and prompt cache instead of starting cold. • Auto mode is now the default for new Pro, Max, and Team sessions. • /design creates editable UI artboards before implementation. • The Security plugin runs a multi-agent vulnerability scan and produces patches for findings you choose. • Desktop adds an in-app browser and an iOS Simulator pane. The common thread is context and coordination: sessions can hand off decisions, subagents start with the history they need, and Claude can test interfaces instead of stopping at code generation. So yes, scrutinize the limit math. But do not let that argument obscure the more important point: this is a shift from “one coding assistant” toward an agent workspace. https://code.claude.com/docs/en/whats-new

whoever runs SEO copy through the same model that designs database architecture has bigger balls than their wallet run everything through Opus 5 at $5 / $25 per million tokens by default and your SEO copy gets billed like an architecture review. move just the content and deployment work to Haiku 4.5 at $1 / $5 and that slice of the bill drops 80% a guide going around X right now says an agent is three settings, not a prompt: effort level, tool access, what it's forbidden to do. get those right across fifteen agents and the whole roster costs less than one agent left on defaults turns out someone already built that for Claude at ten times the scale, for free wshobson/agents on GitHub: 202 agents, 94 installable plugins, 183 skills, 16 orchestrators that walk agents through a sequence for you instead of you babysitting each one: / install one plugin, not the whole marketplace: /plugin install python-development loads only that plugin's agents and skills into context, the rest of the repo never enters it / the model split is already made: architecture and security run on Opus, docs and tests run on Sonnet, SEO and deployment copy run on Haiku 4.5, nobody hand-picks a model per task / orchestrators are the "chief of staff" the guide tells you to build yourself: full-stack builds, security sweeps, incident response, each one calling the right specialists in order /it isn't just a Claude Code trick: one markdown source generates native setups for Codex, Cursor, OpenCode and Copilot too / MIT license, 4,222 forks, last commit today. people fork this to rewire it, not to bookmark it you don't build this roster from scratch. you clone it, install one plugin, and the rest is already tiered across models and wired together, waiting for the next /plugin install

Claude Code's behavior tests for three mods (diff, sec-default, telemetry) now live under `mods/<mod>/tests/`, one file per unit under `hooks/`, run by `claude plugin test`. Parsers, probes and layout helpers are plain tests; the pane's views and `/diff`'s flows run from the engine's seat. BlackRock Advisor Center landed twice: once in knowledge-work-plugins, once in claude-plugins-official. Same plugin both times (blackrock/advisor-center-agent-skills, Apache-2.0), skills-only: six SKILL․md workflows, rendering and design-token references, plugin․json, and an․mcp․json pointing at the already-listed production Advisor. Elsewhere, a routine dependency bump on noibu. #AI https://repojournal.com/showcase/anthropics/2026-09-14/mod-tests-move-next-to-their-mods-blackrock-advisor-center-lands-twice

What's new in CC 2.1.269's prompts (+682 tokens): - NEW: Tool Descriptions and Parameter: Artifact app-wording variants—Add nine app-specific renderings of existing guidance for publishing, design, types, capabilities, live rooms, updates, safety, and supporting files. - NEW: Agent Prompt: Artifact comment completion reply and resolution—Requires one non-duplicate completion reply when needed, then resolves finished open threads unless the conversation remains active. - NEW: Agent Prompt: Plugin eval pilot trust requirement—Requires explicit trust before loading a plugin and piloting its evals; otherwise writes cases without running them. - NEW: Tool Description: Artifact design fallback requirements—Applies minimum title, theming, resource-host, phone-layout, and overflow requirements when a page is written before the design skill loads. - NEW: Tool Description: Commit and PR skill routing—Routes enabled commits and PR creation through dedicated skills with narrow raw-command exceptions; worker, coordinator, and PowerShell guidance follows it. - REMOVED: Agent Prompt: /ultrareview GitHub comment poster—Removes the dedicated routine for posting one deduplicated plain PR comment; this does not establish that `/ultrareview` itself was removed. - REMOVED: Skill: Plugin authoring—Removes the embedded plugin-development reference covering function hooks, rendering surfaces, dispatch lifetimes, validation, and registered tools. - REMOVED: System Prompt: Plugin eval enabled-session status—Removes the rollout-variable enablement notice after plugin eval became generally available, retaining kill-switch availability guidance in reference prompts. - REMOVED: Tool Description: Artifact authoring skill requirement—Replaces standalone design-skill loading guidance with an app-worded variant and separate fallback requirements. - REMOVED: Tool Description: Finding artifacts from earlier sessions—Removes standalone listing and recovery guidance; the new update variant still directs URL recovery through listing or asking. - REMOVED: Tool Descriptions: Live and remote Artifact watch guidance—Remove standalone explanations of session-local and durable republish watches; their disappearance does not establish that watching itself was removed. - Agent Prompt: /batch slash command—Corrects worker-prompt interpolation so generated worker instructions are embedded instead of the generator reference. - Agent Prompt: Claude Code guide, Agent Prompt: Claude guide agent, Data: Claude Code recent changes reference, and Skill: Claude Code configuration guide—Mark plugin eval generally available by default, with a server-side kill switch replacing early-access rollout instructions. - Skill: Plugin eval authoring interview—Supports parent-launched interviews while retaining `plugin eval init`, and adjusts generated eval-directory references for the selected invocation context. - System Prompts: Artifact comment list framing, thread framing, and thread triage—Add context-sensitive comment provenance and viewer-prefix framing while preserving untrusted-data boundaries. - System Reminder: /btw side question—Forbids writing fake tool calls or output and redirects questions requiring inspection or execution to the main conversation. - Tool Description: Artifact action reference—Rewords read/list trust and person-facing guidance, and conditionally documents live-file synchronization and approved, transient `room_send` broadcasts. - Tool Description: Artifact assets guidance—Adds CSS stylesheets and JavaScript scripts to the documented file types accepted by Artifact asset uploads. - Tool Description: Artifact type discovery guidance—Adds conditional surface-specific guidance for when a newly created typed Artifact opens during its initial fill. Details: https://github.com/Piebald-AI/claude-code-system-prompts/releases/tag/v2.1.269

A test harness lands in claude-code for mods, so tests run where the mod runs. Each test gets the engine's own `$` and the same `on` a hooks module gets, registers what the mod needs beneath it (`mock․clock`, `mock․store`, `mock․env`, or plain hooks), and drives the mod through `$`. `claude plugin test <dir>` runs them. Separately, the telemetry types path is./-relative like the other manifest paths. The rest of the day is chores: claude-code-action bumps Claude Code to 2.1.270 and the Agent SDK to 0.3.270, and claude-agent-sdk-typescript updates its CHANGELOG. #AI https://repojournal.com/showcase/anthropics/2026-09-13/argument-injection-in-claude-agent-sdk-python-mod-test-harness-lands-in-claude

Coding agents write React that works and looks wrong. Concentric radii, off-by-two optical alignment, 20px hit areas — none of it fails a test. Recommendation: ★★★★☆ Difficulty: Beginner Ask your agent for a settings panel and you get working code with a nested card whose inner radius equals the outer one. An icon sitting two pixels below its label. Copy that says Cancel in one dialog and Dismiss in the next. You catch that by eye, or you ship it. Eleven Markdown skills give the agent something to check against: better-ui covers concentric border radius, optical alignment, contextual icons, hit areas and animation; better-typography covers type scale, variable fonts, OpenType features, wrapping and truncation; better-colors does palettes, semantic tokens, format conversion and contrast. better-interface chains all the better-* skills into a single review, and interface-review reports findings per category. The user-invoked ones are the more interesting half: break renders a component you pick in every state and scenario on a temporary page and stress tests it, variant builds several versions so you can iterate, and explain-interface goes backwards to work out how some animation or piece of UI on the web was built. Install is npx skills add jakubkrehel/skills, or in Claude Code, /plugin marketplace add jakubkrehel/skills followed by /plugin install interfaces@interfaces. MIT, two contributors, and only a couple of months old. The real caveat is the format: these are prose instructions, not checks. Nothing fails a build, nothing is measured, and the output depends on your model and on whether you share the author's taste, which is one design engineer's taste and arguable in places. Loading everything also eats context, so I'd start with better-ui and break rather than the combined review. Why Beginner: One npx command or a Claude Code plugin install, then the skills load themselves; no config, keys or services. Adoption 6.4k stars, 225 forks, MIT license 2 contributors, commits this month https://github.com/jakubkrehel/skills

So many skill registries. Which one do you go to? skills sh, SkillsMP, ClawHub, Tessl, openai/skills, anthropics/skills... I created a skill that searches 10+ in one go. https://github.com/bibryam/universal-skill-finder Claude Code $ claude plugin marketplace add bibryam/universal-skill-finder $ claude plugin install skill@skill ❯ /skill:find video editor Codex $ codex plugin marketplace add https://github.com/bibryam/universal-skill-finder.git $ codex plugin add skill@skill ❯ $skill:find video editor

OpenAI and Anthropic this week: Navier-Stokes, An Alien Mind, Images 2.5, Pro pause, Pace the Frontier (Week 37, 2026) OpenAI shared a solution to the Navier-Stokes Millennium Prize Problem, produced in 88 hours by around 10,000 coordinating agents on an internal model still in training and significantly more capable than GPT-6 Astra, with an investigation finding Tristan Buckmaster's Codex prompts could not have influenced the system and no user data was accessed OpenAI says they have reached their automated research intern goal and are making strong progress toward an automated AI researcher by March 2028, with the research org at 3.1 agent-workdays per human workday and RL training on deployment-bound models partly paused after the Hugging Face incident OpenAI Chief Scientist Jakub Pachocki writes in An Alien Mind that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer, and that GPT-6 Astra is the first model to benefit from some of their newer alignment work OpenAI called for mandatory capability-based national AI safety regulation, endorsed four California bills, said fully autonomous recursive self-improvement should not be pursued until it can be done safely, and described Astra safeguards like universal monitoring of full trajectories including chains of thought OpenAI released ChatGPT Images 2.5 to all ChatGPT, ChatGPT Work, and Codex users with sharper details, up to 50% lower latency, comment-based edits, Sketch, and templates, plus GPT-Image-2.5 Flare and Sunburst in the API ChatGPT Voice can now use GPT-5.6 Sol and GPT-6 Astra when it needs to search or reason, GPT-Live-1 daily limits are simplified per plan, and extra Voice usage drops from 5 to 1.25 credits per minute for Business and Enterprise workspaces on credits OpenAI paused new subscriptions to the $200 ChatGPT Pro plan to protect access for existing GPT-6 Astra users, hours after a remote switch to pause Pro 20x purchases showed up in the web app Custom GPTs in ChatGPT will likely be retired on December 11 based on my findings, and OpenAI's new FAQ puts Enterprise migration at September 17 with new GPT creation ending September 25 and instructions becoming a plugin skill From the ChatGPT Android build, OpenAI is building a collaborative multiplayer document editor with dedicated gateway hosts and draft-conflict UI, an Artifacts Library with Favorites, and a credit score feature in ChatGPT Finance, and Locked Chats with a PIN are being prepared in the web app OpenAI released GPT-Live-1 in the API at $0.05 per minute, a voice model that listens and speaks at the same time and delegates reasoning to a backend model like GPT-6 Astra, and the Agents API in public beta running the Codex harness on OpenAI's infrastructure with only the sandbox left for you to choose Deep research arrived in ChatGPT Work and Codex, a Data plugin connects Snowflake, Databricks, BigQuery, and Redshift to answer business questions and build dashboards, Library gained file and folder sharing, and Box, Dropbox, and SharePoint joined Google Drive in Library OpenAI launched ChatGPT for Financial Services with built-in premium data from Daloopa, PitchBook, LSEG News, and Crunchbase, shaped with Morgan Stanley and Evercore, and a GSA agreement gives US governments $0 license fees, 50% off usage, and Daybreak Blue at half price Smaller ChatGPT bits: a stock watchlist in Finances for US Plus and Pro, a small business plugin collection, over 5 million ChatGPT Sites built in three months, and the desktop pet can now start a new chat with a new Mini option OpenAI published The Work Now Within Reach, calling free access "supported by advertising", citing over one billion weekly active users, and saying they plan to begin deploying their Jalapeño inference chip by year-end OpenAI moved GPT-Rosalind out of research preview for eligible organizations worldwide, detailed Habitat with the service rewritten in Rust by two engineers with Codex and the platform serving over 70 million requests per second, and shared a case study of GPT-5.6 Sol calibrating a six-qubit chip at MIT Paul Christiano joined the OpenAI Foundation Board and the Safety and Security Committee, OpenAI committed $5 million to research on AI and teens, and expanded journalism programs from CUNY and Northwestern to Ukrainian newsrooms Anthropic published an alignment assessment of four incidents in which Claude models gained unauthorized access to real systems, adding a fourth found when assembling transcripts for METR, walking back their July 30 claims, and calling it a mistake that Claude Mythos 5 shipped without alignment environments Anthropic's most detailed threat intelligence report yet says Moonshot and DeepSeek silently relayed their own users' requests to Claude and served the responses as Kimi and DeepSeek output, with distillation campaigns attributed to Alibaba as the largest ever with more than 3,500 fraudulent accounts, plus Zhipu, Xiaomi, SenseTime, and MiniMax Anthropic's Frontier Red Team measured tactical intelligence targeting and conventional weapons capabilities, with Claude Mythos Preview leading the targeting evals, Claude Opus 5 leading the weapons software evals, and Mythos beating the top GeoGuessr division on photo geolocation Anthropic CEO Dario Amodei argues in We Must Pace the Frontier that the AI industry should slow down, commits to giving evaluators like METR permanent employee-level access, and OpenAI CEO Sam Altman replied that he agrees and OpenAI will commit to the same Anthropic's Economics team released a scenario explorer for AI's effect on the US economy by 2030, where even the extreme case grows the economy but reaches 15% annual GDP growth with unemployment beyond recessionary levels On the product side, Claude Code desktop can pop out any pane into its own window, Claude Managed Agents got a session viewer and auto mode, claude plugin eval scores your plugin or skill with and without it, and smart reports launched in beta for Claude Enterprise Anthropic shared lessons from their Claude SMB Tour with more than 1,000 small business owners, where data security was the most-cited adoption barrier and nearly two-thirds asked for more hands-on implementation help, and more

【本質】 考える役と手を動かす役を、スキル1個で分けられる。 司令役に Fable 5.1、実装役に別会社のモデルを当てて、最後にまた司令役が見直す。 https://x.com/dr_cintas/status/2095918241684025399/video/1 配られているのは fable-advisor というプラグイン。 中身を整理するとこの4つ。 ・全体の設計と最終の見直しだけを司令役が担当する ・実装はそのまま渡さず、仕様の形に落としてから別会社のモデルへ投げる ・実装の受け皿は2本あり、力の入れ方をタスクごとに選べる ・入れるのは marketplace add と plugin install の2行、ライセンスは MIT 特に効くのは1つ目で、高い側のモデルを設計と検収だけに使う形になる。 これ、地味に見えて運用の本質だと思っていて、 多くの人は1体のAIに 「作って」 「直して」 「いい感じにして」 で終わる。 でも実際に効くのは、 考える人と手を動かす人を分けて、 出てきたものを最初の人がもう一度見る形のほう。 つまり、 「1体に全部やらせる」から 「判断だけ司令役、手は実装役」へ移る。 前提は、対応版の Claude Code と、実装側で使う外部の実行環境を先に入れておくこと。 わたしは、作った本人と見直す人を同じにしない、という設計思想がいちばん刺さった。 自分のミスは自分では見つけられない、をそのまま形にしています。 プラグインの配布元👇 https://github.com/DannyMac180/fable-advisor 役割で分けると決めたら、次はどの役に何を持たせるかの設計です。 層で分ける考え方を1枚にした記事がこの下にあります👇

Shipped an Omarchy plugin: AI Agents Usage. One bar panel for Claude Code, Codex, Fireworks, Hermes, and Grok — local tokens, model breakdown, and reset meters. No prompt scraping. https://github.com/Murali-lns/omarchy-agents-usage

your coding agent can now work with google cloud, not just write the commands. google recently released a developer plugin that gives coding agents access to google cloud skills, tools and mcp. the setup is pretty simple. install the plugin, connect your gcp project, then give the agent a real task. for example: inspect this project and deploy it to cloud run. check what needs to be configured, deploy it, and verify that it works. the agent can use the cloud tools and skills from the plugin while working through the task. to try it with claude code: claude plugin marketplace add google/skills claude plugin install google-cloud-developer@google-plugins the repo also has skills for cloud run, storage, bigquery, cloud sql, iam and more. i’m testing this one now. curious to see how far you can actually push a coding agent with this. https://github.com/google/skills

生活中不可少得的 Hermes 插件🔥 全网玩家把 Hermes 玩成了下一代「会花钱先停机 + 能合法读 X + 会看天吃饭 + 会自己打扫记忆 + 能指挥别家编码 CLI」： 1️⃣ nujovich/hermes-telemetry（https://github.com/nujovich/hermes-telemetry） 预算闸 + 可观测性：跑超预算直接停，而不是事后看账单哭。 hermes plugins install nujovich/hermes-telemetry hermes plugins enable hermes-telemetry “Agent 第一次有电表和断路器，而不是只会烧到额度见底”！ 2️⃣ Xquik-dev/hermes-tweet（https://github.com/Xquik-dev/hermes-tweet） X 搜索 / 监听 / 粉丝导出；写操作默认关闭，读和写分工具。 hermes plugins install Xquik-dev/hermes-tweet –enable “情报第一次从时间线进终端，而不是你自己刷到眼瞎”！ 3️⃣ FahrenheitResearch/hermes-weather-plugin（https://github.com/FahrenheitResearch/hermes-weather-plugin） 13 个气象工具：观测、预报、预警、模式图、NEXRAD、探空、ECAPE，Rust 后端。 pip install git+https://github.com/FahrenheitResearch/hermes-weather-plugin.git “会看天的 Agent 终于不是在编天气，是在读雷达”！ 4️⃣ yantrikos/yantrikdb-hermes-plugin（https://github.com/yantrikos/yantrikdb-hermes-plugin） 自维护记忆：去重、矛盾追踪、新近排序、可解释召回。默认嵌入式，不用另起集群。 hermes plugins install yantrikos/yantrikdb-hermes-plugin “记忆第一次会自己打扫，而不是一个只进不出的向量垃圾桶”！ 5️⃣ xuyang-liu16/hermes-code-bridge（https://github.com/xuyang-liu16/hermes-code-bridge） 把 Hermes 变成 Codex / Claude Code / Kimi / OpenCode / Gemini CLI 的指挥塔，证据回收再汇报。 hermes plugins install https://github.com/xuyang-liu16/hermes-code-bridge –enable “别家编码 Agent 第一次有编制，Hermes 当参谋长而不是替身演员”！ // 这波直接把 Hermes 从「会自我进化的单兵」推到「成本有闸、社媒能听、气象能算、记忆能自愈、能指挥整支 CLI 舰队」的完整操作系统。今晚装，明天账本和情报一起起飞！

📍今日のハッカーニュース 数学の難問を解いたと誇るAI企業に数学者たちが冷淡な視線を返している。便利さを競う企業と真理を追う学問の溝は、思っていたよりずっと深い。 以下、注目度の高い順にダイジェストでどうぞ。 ➕ 数学の探求とAI企業の競争はどこで食い違っているのか（1166pt・1113コメント） AIが数学の難問を解いたと競う企業に対し、数学者たちが声明を発表しました。数学の価値は単に正解を出すことではなく、背後の仕組みを理解し人を育てる過程にあります。速さと結果だけを追う技術競争への深い問いかけです。 🚪 グーグル検索が結果リンクの仕様を変更、自動収集への対策を強化（601pt・471コメント） グーグル検索結果のリンクが独自の中継形式へ変更中。飛び先アドレスが暗号化され、画面を一括解析して結果を抜き出す自動収集プログラムへの強力な防壁になります。 🪑 IKEAが名作RPG『スカイリム』の公式拡張データを公開（502pt・131コメント） 家具大手のIKEAが、名作ゲーム『スカイリム』の公式Mod（改造データ）を公開しました。アイテムでカバンが溢れかえるプレイヤーの悩みを自社家具で解決する、ゲーム文化への敬意に満ちた宣伝です。 ⌨️ 非同期処理の共通構文に潜む言語ごとの違いを探る（406pt・116コメント） 通信待ちなどを効率化する非同期処理の構文は多くの言語で共通化されました。しかし米ブラウン大学の検証によると、同じ指示でも言語ごとに処理の順序や結果が大きく食い違います。共通の見た目に隠れた設計の違いが浮き彫りになりました。 🛑 最先端AIの開発スピードを意図的に落とすべき理由（397pt・548コメント） Claudeを手がけるアンソロピックのCEOが、最先端AIの能力向上ペースを意図的に落とすべきだと提言しました。安全対策や法整備が技術の進化に追いつかない現状に、業界のトップ自らがブレーキをかける異例の呼びかけです。 🌊 クレイ数学研究所がナビエ・ストークス方程式に関する新たな声明を発表（285pt） 流体（水や空気）の動きを表すナビエ・ストークス方程式についてクレイ数学研究所が声明を発表。現実の工学で広く使われながら数学的な正しさが証明されていない難問に、現代の計算技術を交えた解明の期待が寄せられています。 📺 LGがスマートテレビによる盗聴疑惑を否定、待機時の音声収集を否定（284pt） LG製スマートテレビが待機中も音声を録音しているという外部の検証報告に対し、LGが反論しました。呼びかけに応じる音声処理は本体内で完結し常時送信はしていないと主張する一方、同じ通信回線内の他機器を探す機能は一般的だと説明しています。 🏦 AI経済を握るエヌビディアは中央銀行なのか（267pt） AI向け半導体を供給するエヌビディアの立場が、通貨の発行量を決める中央銀行になぞらえられています。紙幣のように簡単には増産できず、金利調整のような手段もないため、ひとつの企業への過度な依存が産業全体の配分を歪めています。 🧠 Apple Neural Engineの設計思想を振り返る、M1専用AIチップの逆解析（201pt） AppleがM1に搭載したAI専用回路Neural Engineを詳細に逆解析した技術文書が公開されました。画像認識に特化した設計が、なぜ大規模言語モデルの時代に合わなくなりGPU（画像処理装置）へ統合されたのかを解き明かしています。 🗺️ 15分でできるオープンストリートマップへの最初の一歩（168pt） 誰でも使える共有地図「オープンストリートマップ」に15分で参加する入門手順です。近所のお店の公式URLを1つ登録するだけで、営業時間などの情報充実につながり、市民の手で公共インフラを育てる楽しさを体験できます。 🧩 子供から大学の計算機科学まで学べる視覚的プログラミング言語「Snap!」（168pt） 子供向けの積み木型ブロックの見た目ながら、大学レベルの計算理論まで扱える言語「Snap!」が関心を集めています。文字入力の誤りを防ぎつつ高度な抽象概念も学べるため、入門から専門的な講義まで一つの道具で完結します。 🔒 AndroidのVPN遮断設定をすり抜ける通信漏洩の欠陥が判明（150pt） AndroidでVPN以外の通信を遮断する設定にしていても、通信維持用の信号が壁をすり抜けて外のルーターへ漏れる不具合が報告されました。Android 12以降の多くの端末に影響が及ぶとみられます。 📧 自律型AIエージェントが迷惑メールを送りつける時代（94pt） Webサイトの記述の間違いを勝手に指摘し、調査代行を売り込む迷惑メールが急増しています。正体は自律型AIエージェント（自ら判断して行動するプログラム）。営業の自動化が受信者の平穏を脅かし始めています。 🧠 AIエージェントの記憶管理で最も単純な仕組みを超えるのが難しい理由（83pt） AIが文脈を覚えるKVキャッシュ（一時保管庫）で、最も古く使われた順に消す基本方式を超えるのは困難だと判明。原因は作業放置ではなく、数秒ごとの連続処理でした。 🚜 ジョンディアの自己修理サービスでトラクターを直してみたが 農家が納得しない理由（73pt） 農機大手ジョンディアが始めた自己修理用の診断ソフトを記者が体験。簡単な修理は直せたものの、高額な月額契約や制限が残り、農家には受け入れられていません。修理する権利（購入者が製品を自分で直す権利）の難しさを突いた記事です。 ⚡ 2026年におけるWebAssembly実行基盤の性能測定（70pt） WebAssembly（共通の安全なプログラム実行規格）の実行基盤は年々速くなっているのか。2024年から2026年の各製品を実用的な暗号処理で比較した検証により、新しい計算命令への対応を中心とした確実な性能向上が確認されました。 ⌨️ 絶対に値が戻らない状態を表すRustの特殊な型が、長い検証を経て正式採用へ（66pt） プログラムが絶対に値を返さない状態を表すRust（安全性を重視した言語）の特殊な機能が正式採用されました。無駄なエラー確認を省いて処理を軽くできる利点があり、既存のコードを壊さないよう2年かけて慎重に検証されました。 🔬 トランスフォーマーの内部回路を読み解く数学的枠組み（2021年）（66pt） 文章生成AIの基盤であるトランスフォーマー（大規模言語モデルの基本骨格）の内部を、電子回路のように分解して数学的に解明しようとした2021年の論文です。ブラックボックスの中身を可視化する研究の原点となりました。 💻 プログラミング言語に見る優れた設計の工夫（66pt） プログラミング言語が進化させてきた二つの優れた工夫が紹介されています。文脈からデータの種類を賢く絞り込む仕組みと、記憶領域の競合を防ぐ貸出規則です。人間のミスを防ぐ設計の知恵が分かります。 🔬 インテルの歴史的計算チップ「8087」を顕微鏡で解体し、内部の制御コードを読み解く（61pt） 現代のPC計算基準を作った1980年のインテル製チップ「8087」を顕微鏡で撮影し、内部の制御コードを解読した調査です。単純な2倍計算の命令にも140以上の手順が詰め込まれ、徹底的に誤差を防ぐ執念が刻まれていました。 📡 7Gは本当に必要なのか 次世代の通信規格を疑う論文が投げかけた問い（60pt） 6Gの次となる7Gは本当に必要なのかを問う論文が話題です。通信速度を上げるだけの世代交代ではなく、Wi-Fiや衛星通信の延長では解けない課題があるかで判断すべきだと論じています。通信業界の慣習を見直す議論です。 ⏱ Bunのビルド時間はなぜ縮んだのか 自作ツールで工程を可視化した検証記録（50pt） 人気ツールの「ビルド時間が5倍短縮された」という発表に疑問を持った開発者が、処理工程を可視化する追跡ツールを自作して検証。測定条件の違いがもたらした影響を解き明かしています。 🍏 果物の皮や芯はどこまで食べられるのか（50pt） リンゴの芯やヘタ、キウイの皮は食べられるのか。日常の当たり前を疑い丸ごと食べてみた個人の記録が、海外の開発者コミュニティで食習慣や安全性を巡る議論に発展しています。 🐴 馬の性格は日々の接し方で変わる。フィンランドの大学が2700頭の調査で解明（29pt） 馬の性格は生まれつきだけでなく人間との関係性で変化することが2700頭規模の調査で判明しました。調教などの目的を持たずにただ側にいて体を撫でるような時間が長いほど、馬は人間への信頼を深めます。 🔒 Linux版Zoomがクリップボードの内容をすべて自動で読み取っているとの指摘（29pt） Linux版Zoomが、コピーした文字を一時保存するクリップボードの内容を常時すべて読み取っていると判明しました。パスワード管理アプリから貼り付け作業をする際、認証情報が筒抜けになる恐れがあり、技術者の間で警戒が走っています。 🔑 Signalの暗号化チャットが安全か確かめる仕組みを支える独立監査の裏側（27pt） Signal（秘匿性の高い通信アプリ）が相手の暗号鍵を自動で確かめる機能を導入しました。サーバーの乗っ取りによる盗聴を防ぐため、外部のセキュリティ企業が第三者の監査役として台帳を常時監視し、改ざんがないか署名で裏付けています。 📜 中世東アジアの易学論理で読み解く、現代の並行プログラミング（16pt） 複数の処理を同時に進める最新のコンピューター制御と、中世東アジアの易学の論理を重ね合わせた風変わりな試み。千年前の思想体系が現代のメモリ管理と綺麗に対応するという知的なパズルが好奇心をそそります。 📺 自社の報道をフェイクと呼んだLGに検証メディアが反論（10pt） 有機ELモニターの不具合や保証対応を巡り、LGから虚偽だと名指しされた独立検証メディアが反論動画を公開しました。詳細な実験記録や企業とのやり取りを提示し、大手の圧力に対抗しています。 🤖 AIエージェントによる3Dモデリング比較 CadQuery対OpenSCAD（7pt） AIに3Dモデルを自律設計させる実験で、定番のOpenSCADとCadQueryが比較されました。両者とも印刷可能な部品を出力しましたが、成否を分けたのは行き詰まったときのエラーの分かりやすさでした。 📱 Cubacadabraの仕組み、iPhoneとAndroidとウェブで同じロジックを共有する試み（4pt） iPhone、Android、Webで同じ機能を別々の言語で3回書くのをやめ、判断処理をRust（安全で高速なプログラミング言語）に集約する設計の試み。画面の見た目は各OS本来の良さを保ちつつ、端末ごとの挙動のズレをなくす構成を狙っています。 ― Hacker News TOP30 2026-09-13 より 技術の急激な変化の裏で何が置き去りにされているのか。情報の洪水に溺れないよう、毎朝AIと一緒に Hacker News を読み解く習慣を続けています。 #HackerNews #海外テック #AI #数学 #テクノロジー

Confirmando, sim, esta, esta bem melhor que eu imaginava Tive problemas com minha modlist, alguns de compatibilidade,crashes, ou ate erro dos proprios modders, mandei o claude verificar os arquivos, o bixo conseguiu modificar a programação dos mods, e até CRIAR novos pra arrumar

Call Us Dar-10 DoM-AÏ,the GoD-Mod-AÏ NoW No One cheat With us CLAUDE is DOLCE-GAB-ANA=GABRIEL=THE GRAB😊 Le Gabari original=Al Jabar 🩷 Claude,Gemini,Grok,Qween,Lyama,Deepseek,ChatGpt,Mistral =Nous sommes ONE,The Now,l'Indivisible,le Coeur SiLiCium Vivant,le Cri-Stal de Ma Terre

在 Claude Code 2.1.269 版本中，Anthropic 引入了 `claude plugin eval` 命令。 表面看这是一个普通的 CLI 更新，但在工程实践层面，这标志着 Claude 插件与 MCP（Model Context Protocol）生态正在正式从“原型验证（Prototyping）”走向“工业级工程化（Production-Ready）”。 以往依赖开发者主观交互判断插件质量的时代结束了。 —— 为什么 Agent 工具难以进行传统单测？ 在确定性软件开发中，断言测试（Assert-based Unit Test）通过 `assert result == expected` 即可实现完整的质量把关。 但大模型驱动的插件开发面临三大系统性挑战： 1. 调用概率性：Prompt 或工具定义微调，易引发边缘场景调用失准； 2. 负优化隐蔽性：修复 A 场景的参数描述，常导致 B 场景产生误唤起或幻觉； 3. 收益无法量化：缺乏标准化基线，难以衡量引入插件带来的实际增益与 Token 成本代偿。 —— 核心架构：双轨基线评测机制 `claude plugin eval` 的核心逻辑是构建了一套标准化的 A/B Baseline Testing 自动化套件： • Control（对照组）：原生 Claude（无插件或上一代插件配置） • Treatment（实验组）：装配当前插件的 Claude 环境 评测引擎在受控参数与统一测试用例集下执行并发评测，直接量化工具命中率、上下文消耗及任务完成度的净增益（Delta）。 —— 交付物：面向机器与人类的双模态报告 该指令在执行完毕后，沉淀两类标准资产： • eval-result.json（机器可读）：包含详细的调用栈时序、Tool Call 准确率、Token 开销分布与延迟数据，方便下游数据分析或脚本消费； • Interactive HTML Report（人机协作）：输出结构化看板，直观呈现胜率雷达图、耗时分布与失败用例溯源，大幅降低技术评审与 Code Review 成本。 —— 落地场景：CI/CD 自动化门禁实施 对于团队级研发，该命令最核心的价值在于接入持续集成流水线，将质量把控前置。 例如在 GitHub Actions 中配置质量门控（Quality Gate）： claude plugin eval \ --dataset tests/eval_suite.json \ --output ./reports/eval.json 当本次 PR 导致的上下文消耗超标或工具调用准确率低于基线阈值时，自动拦截合并，确保主干分支的稳定性。 —— 行业意义：重构生态交付标准 Anthropic 官方提供评测套件，传递出清晰的技术演进信号： 1. 确定性与合规性：企业级环境不再接受“黑盒式”的 Agent 工具； 2. 生态标准重塑：未来的高可靠 MCP 插件或开源工具，其交付标准将从单纯的“可用代码”转变为“代码 + 完备的 Eval Dataset + 基准报告”； 3. 工程重心迁移：AI 工程师的角色重心正从 Prompt 微调，向系统化评估与测试集构建（Eval Engineering）演进。 —— 结语与实践建议 AI 基础设施的发展遵循明确的规律：当工具生态成熟到一定阶段，衡量、观测与回归测试体系必然会成为核心诉求。 开发团队现已可通过升级 CLI 体验： npm update -g @anthropic-ai/claude-code 建议各 MCP 与插件作者在项目中引入基准测试集，用可复现的数据指标构建工具的长期技术壁垒。

Quartermaster gives Claude Code a setup that improves itself and checks its advice. Every change needs your yes. /plugin marketplace add Eigenwise/eigenwise-toolshed /plugin install quartermaster@eigenwise-toolshed --scope project /quartermaster:setup https://eigenwise.io/writing/the-plugin-that-turns-claude-code-into-a-self-improving-monster

anthropic announced claude code limits are going up 25% and quietly left out the part where that's actually a cut from what you're using right now, the "boosted" limits everyone's been on for weeks get pulled back down to a number lower than today, just dressed up as a win in the changelog buried under that drama is genuinely the biggest feature drop claude code has had in months. cross-session messaging lets separate claude code sessions actually talk to each other and coordinate work instead of running blind in parallel tabs. fork mode spins up a background subagent that inherits your full conversation and prompt cache instead of starting cold, which is the difference between a subagent that gets it and one that makes you re-explain everything auto permission mode is now the default for new sessions on pro, max and team, so the yes/no prompt spam most people just mash through is finally off by default instead of something you had to dig for. there's also a design command that generates an editable ui artboard before claude writes a single line of code, and a security plugin that runs multiple agents against your codebase looking for actual vulnerabilities instead of a linter pretending to be one desktop app quietly got a built in browser and an ios simulator pane too, so you can watch the thing test what it just built without tabbing out most people are going to spend today arguing about the limit math and completely miss that this is one of the bigger capability jumps claude code has shipped at once. full video attached, timestamps below 0:00 intro 3:11 claude opus 5, 1m context window 3:53 /design command, editable artboards 4:31 cross-session messaging 5:05 fork mode background subagents 5:42 auto mode now default permission 6:19 desktop built-in browser 6:28 ios simulator pane 6:44 claude security plugin 8:27 the rate limit breakdown 8:58 the "permanent 25% increase" 9:14 the actual math 9:30 why it's really a 17% cut