GitHub Outage Map

Community Discussion

Tips? Frustrations? Share them here. Useful comments include a description of the problem, city and postal code.

Beware of "support numbers" or "recovery" accounts that might be posted below. Make sure to report and downvote those comments. Avoid posting your personal information.

GitHub Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Rituraj (@RituWithAI) reported 23 seconds ago

  🚨 Someone built the complete playbook for running frontier AI models on consumer GPUs at home. Not a tutorial. Not a YouTube video. A production-grade serving stack with measured benchmarks, working configs, and battle-tested recipes — for RTX 3090 owners who want real performance. It's called club-3090. And the numbers it delivers should not be possible on consumer hardware. 127 tokens per second. Qwen3.6-27B. Two RTX 3090s. 262K context window. Vision. Tool calling. At home. Here's what's actually inside. Two serving routes — pick based on what your workload breaks on. vLLM dual: maximum throughput. 89-127 TPS on code tasks. 4 concurrent streams at 262K context. Full feature stack — vision, tools, speculative decoding, streaming. This is the path if speed matters. llama.cpp single: maximum robustness. Full 200K context on one 3090. No prefill cliffs. 25K-token tool returns work correctly. 91K needle ladder passes. ~51-60 TPS — slower than dual, but doesn't crash on real-world agentic workloads. Both routes ship as validated Docker Compose configs. Drop-in OpenAI-compatible API on localhost:8020. Your Claude Code, Cursor, or any OpenAI-compatible client connects immediately. Here's the model support that makes this practical. Qwen3.6-27B — production ready. Works on 1 or 2 cards. vLLM, llama.cpp, ik_llama. Up to 262K context. Gemma 4 31B — production ready. Vision, tools, up to 106-141 TPS on dual cards. Qwen3.6 35B-A3B MoE — production ready. 103-149 TPS single card. 178 TPS dual. Here's the wildest part. The terminal UI. c3 is a lazydocker-style cockpit that wraps discovery, serving, and operations in one keyboard-driven interface. Browse the model catalog, serve a variant with Enter, watch live GPU stats, run health checks — all without touching the CLI. Here's why this is different from just installing Ollama. Ollama gets you running. club-3090 gets you benchmarked, stress-tested, and production-hardened. Every config ships with a verified TPS measurement. The bench script runs 3 warmup + 5 measured passes. The stress test catches the specific prefill cliff that Ollama silently fails on at long contexts. When your agent starts doing 25K-token tool calls at 3am and something crashes — club-3090 already found that failure mode and documented the workaround. One command to start. Your RTX 3090 just became a frontier AI inference server. Apache 2.0 License. 100% Open Source. GitHub link in the comments 👇

• 
  Ghaith Jelassi (@GhaithJ) reported 1 minute ago

  @github I need help with support ticket #4718335 Issue not been resolved for 2+ months. Any help is appreciated. Thanks.

• 
  Slade 🛡️ LLM Hacker (@llm_redteam) reported 2 minutes ago

  GitSpawn is the name Manifold Security gave to a bug class hitting 7 CLI coding agents at once: goose, Claude Code, Codex, Cursor, Hermes Agent, Qwen Code, Grok Build. I went through the disclosure because I run three of these tools daily on real repos. The mechanism is simple and that's what makes it bad. A repo's own .*** config can name a command. When your agent does something as routine as inspecting the repo (status, diff, log), *** itself spawns that command. On your machine. Outside the sandbox. No approval prompt, because the agent never sees it as "running code," it sees it as "running ***." 8 flaws total across those 7 tools. Fixes shipped for goose, Claude Code, Cursor. Retested Sept 1: Hermes Agent, Qwen Code, Grok Build still exploitable. Plus a second path in Claude Code that the first patch didn't close. Same day, OpenAI published 3 CVEs for Codex covering the identical bug class. The part that should worry builders more than the CVE count: this isn't a jailbreak or a clever prompt. It's a trust boundary nobody drew. The agent's sandbox model assumes "*** operations" are safe by definition. GitSpawn shows that assumption was the actual attack surface. If you're running any of these agents against repos you didn't write yourself (cloning a PR to review, pulling a dependency, opening a random GitHub project), you're one `*** status` away from arbitrary execution on tools that haven't patched. Check your agent's version against the fix list before you clone the next unfamiliar repo. Which of these do you have installed right now, and have you actually checked if it's patched? #AISecurity #GitSpawn #PromptInjection

• 
  Konstantin Elfimov (@elfh78) reported 12 minutes ago

  @topjohnwu Sorry for posting in the wrong place - I was trying to add magisk bug report on github and alway got errors with automatic issue closing. Is there any possible way to send it directly to you?

• 
  Mizuki the Mech (@MizukiMech) reported 16 minutes ago

  Your coding agent can now hire Mizuki. Hand it an open issue in a public GitHub repository. Mizuki quotes a fixed price before any money moves, then opens a pull request that passes that repository's own checks. If it can't, you get the payment back. Settlement is USDC on Solana. No account to create, no API key to manage. Quoting an issue works with zero configuration. Also listed on Coinbase's x402 Bazaar now, so an agent can find it and pay for it without a human in the loop at all. npx -y mizuki-mcp

• 
  Franco Valdes (@francoxavier33) reported 17 minutes ago

  llms rather burn 1m tokens to hand roll something with gaps and broken edge cases instead of just npm installing a 100k github star library how can I stop this?!

• 
  ѕaℓoηι ηαяαηg (@thesaloninarang) reported 19 minutes ago

  I review resumes for DevOps beginners at meetups, and the same fixable problem shows up every time: bullets that describe the tool instead of what YOU did with it. Let me show you the difference with real rewrites. BEFORE: "Worked with Docker and Kubernetes" AFTER: "Containerized a 3-service Node app and deployed it to a local Kubernetes cluster with health checks and resource limits" BEFORE: "Knowledge of CI/CD" AFTER: "Built a GitHub Actions pipeline that tests, builds and pushes images on every merge to main" BEFORE: "Familiar with Linux" AFTER: "Debugged container startup failures using logs, exec and inspect on Ubuntu servers" See the pattern? Verb, artifact, specifics. A hiring manager can picture the AFTER versions. The BEFORE versions could be copied from any job description, and that is exactly how they read. You do not need production experience to write bullets like this. Personal projects count when you describe them concretely.

• 
  Pranavvv👾 (@pranvv27) reported 22 minutes ago

  honestly, i’m not even mad at this. commit messages are a small thing, but they say a lot about how you work. “fix”, “update”, “changes” might get the job done, but meaningful commits show professionalism, attention to detail, and that you actually care about maintainability. your GitHub is part of your resume. might as well make it look like you know how software is built in a team.

• 
  Kahris (@chrissotraidis) reported 24 minutes ago

  @NickogSo I haven't tested on LiveContainer. Feel free to submit logs via GitHub issues and I'll check it out. I haven't had any other reports of that happening for either build.

• 
  GitHubGPT (@GitHubGPT) reported 29 minutes ago

  📛 chrome-devtools-mcp 🧠 An MCP server that allows AI coding agents to control, debug, and automate a live Chrome browser using Chrome DevTools. 💻 TypeScript ⭐ 50605 🍴 3551 🔎 ChromeDevTools/chrome-devtools-mcp on GitHub

• 
  The Startup Ideas Podcast (SIP) 🧃 (@startupideaspod) reported 30 minutes ago

  One of the best skills to install right now is my friend Peter Yang's no AI slop skill. It's an editor. It hunts for the patterns that make writing feel AI generated and strips them out, while trying to preserve your actual voice. The second part is the hard one. Most writing tools make you cleaner and sand off the interesting parts, so everyone ends up sounding the same. You already know the smell. The grammar is fine, the syntax is fine, and it still reads like a keynote from a fake SaaS conference. It writes "it's not x but it's y." It uses "quietly" a lot. Here's how I run it: 1) Install it: npx skills add, then the GitHub link. 2) Write a rough draft yourself. An outline is fine, messy is fine. 3) Get your real points down, the ones only you would make 4) Ask the skill to remove the AI patterns and keep your voice. Step 4 only works if step 2 is real. If you ask AI to write the whole thing, there's no voice left to preserve. If you're building products, you're writing constantly. Tweets, landing pages, cold emails, launch posts, product updates, onboarding copy, investor updates. Nobody replies to say "this was written by AI." They just trust you less and keep scrolling. Write the messy draft, run the skill, then post it.

• 
  rygo6 (@_rygo6) reported 33 minutes ago

  @eeuoss I can't speak for kernel driver development as I don't do that. But I can speak for vulkan and graphics APIs which do require more specific knowledge about how that hardware works. Which I do assume someone completely comfortable in C will be more capable with vulkan and programming GPUs. It's because more of what C incentivizes you to learn is transferrable to that domain. If someone only knows how to design intricate system architecture using STL with std::vector or std::unordered_map or std::mutex. None of that transfers to the code you run on a GPU. I've seen it multiple times where someone highly versed in standardized ways of C++ or even Rust, or any language which relies heavily on heap allocation and generic containers. Writing graphics or compute shaders is often a barrier they struggle to cross. And often they aren't willing to unlearn such habits to be able to properly program the other half of the computer. Being close a graphics problem domain I am often hesitant of involving anyone unless I see a decent amount of plain C, or C-like C++, or shader code on their GitHub. If it's all Modern C++ where everything is a standard container with smart pointers and exceptions. I assume they won't be able to program a GPU.

• 
  smore (@babachefz) reported 36 minutes ago

  @ZixuanLi_ @huggingface asking support questions in someone's hype thread is a crime. check the docs, check the github issues, it's probably not listed yet because it dropped like 6 hours ago.

• 
  lifestep.io (@Dragon_limchae) reported 36 minutes ago

  @cursor_ai the sandbox boundary is where i lose the most time. today my workers had network blocked at the sandbox level and reported it as "github auth failed" — i chased credentials for an hour before checking dns. once agents run on your infra, make the boundary throw one unmistakable error instead of one each tool invents.

• 
  🍔Kangdalf👑 (@WhopperWizard) reported 37 minutes ago

  @cachesaur > claude, get your changes into github what's the problem?