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:

• 
  DanteX (@0xdantex) reported 33 seconds ago

  GUY CUT HIS CLAUDE BILL BY 70% WITH ONE FREE MICROSOFT TOOL NOBODY IS USING every PDF you drop into Claude is quietly burning way more tokens than you think Claude doesn't just read the text, it processes the broken tables, the images and all the junk formatting the file drags along one page can eat 1,500 to 3,000 tokens a 20 page document burns up to 70,000 tokens before you even ask your first question the fix is a Microsoft tool called Markitdown free, open source, over 110,000 stars on GitHub it takes PDFs, Word, Excel, PowerPoint, even YouTube videos and turns them into clean Markdown text up to 70% fewer tokens and better answers, because Claude was trained on millions of Markdown docs and reads it natively the part most people miss is it ships with an MCP server connect it to Claude Desktop once and it auto converts every file you upload from then on we have been overpaying for months on something Microsoft already solved

• 
  Edgar Gonzalez-Kozlova (@EdgarEGK) reported 5 minutes ago

  @euanashley Sure but remember, Claude was trained by your publicly available data and code at some point. Not even counting the millions of repositories on GitHub. Thanks to the work done by many people, Claude code can solve these issues quicker than ever.

• 
  Keeta Github Tracker (@KeetaCode) reported 7 minutes ago

  🐆 Keeta GitHub PR Merged 📦 Repo: node-rs 🔀 PR #29: Chore: Improve DRY 🌿 Branch: chore/improve-dry → main 👤 Originally opened by: @sephynox 🧠 Overview: This pull request appears to clean up repeated code in Keeta’s crypto-related software, which can make the codebase easier to maintain and less error-prone over time. The public description is very limited and only says it “reduces repetition in crypto crate,” with one commit in the PR. This appears to be a technical/internal update with limited public details. - “DRY” is a coding principle that means avoiding the same code being written in multiple places. - Changes like this usually help developers update and review code more easily, but no user-facing feature is described here.

• 
  Kevin Murphy (@midimurph) reported 7 minutes ago

  i build applied AI in the open, usually on the raw Anthropic SDK and the Vercel AI SDK, coordinating the agents by hand. this time i put LangGraph through the same bar, and built a real thing with it: an agent that triages github issues.

• 
  Max Chepurin (@maximchepurin) reported 13 minutes ago

  @Bobliuuu @mattpocockuk They are only listed there once there is some credible evidence that attackers are actually exploiting vulnarability not just cuz someone noticed potential security issue. At least thats how the one that you sent works(CVE catalog) Now consider this: - You install a package in your project. - You run your usual vulnerability scanner. Everything looks good. - You become the first person to encounter malicious code in that package. - You report it, it’s verified, and it becomes a known issue, but only after someone (you) already shipped it to production. - here comes the question again: why didnt you review the code manually? My argument is that simple. People blindly trust vulnerability scanners. They trust github. They trust popular opensource packages and any other pieces of software without reading a single line of code. But when it comes to AI-generated code, suddenly everyone acts like trusting it is fundamentally different and that every line must be reviewed. Double standard. Reviewing is important. And doing regular checks on the most inportant spots of your codebase is essential. But for everything else - you can build your harness, skills, hooks, guardraíls, to make AI generate code that already meets your standards without having to worry about every single line. Why not take this advantage? Its not 2022 when you copypasting snippets back and forth from chatGPT, AI can own your codebase now if you build the proper harness around it.

• 
  Chuck Reynolds (@ChuckReynolds) reported 14 minutes ago

  Hey @jdevalk I use seo-graph in @astrodotbuild and I'm throwing a PR to fix v7* dep warnings and CI tests. Check it out; lmk if it's all good. I've been using it with 7.0.x since release and functionality is all good. github: jdevalk/seo-graph/pull/61

• 
  clayne (@0xclayn) reported 16 minutes ago

  Someone "leaked" Claude Fable 5's system prompt, but that's not even the interesting part A GitHub repo surfaced claiming to hold Fable 5's full system prompt: 26k stars, thousands of forks, tagged "ANTHROPIC / INTERNAL." Looks like a leak. But what actually happened next is the part worth paying attention to. Someone just opened Fable 5 and asked it to build something like "an awwward-winning site about this exact model, micro animations, GSAP, three.js, make it feel like a real experience." And it just did it. On its own. What the footage actually shows: Fable 5 building a landing page for a fictional space agency: a real starfield running on Three.js, ships, stars, a cinematic scroll experience driven by GSAP ScrollTrigger. The model writing its own files, its own stylesheet, its own typography system, structuring every section itself. Spinning up a local preview server and screenshotting its own output just to check whether it actually looks the way it was supposed to. Catching and cleaning up dead code it left behind in the marquee velocity math. Rebuilding an entire typography focused site redesign, checking the console for errors and walking through the page section by section, visually, to confirm everything works. This isn't "generate some text from a prompt." It's the full loop: plan, build, run it, actually look at the result, fix what's wrong. The model is checking its own work at every step, not just producing output and stopping. Fable 5 burns through usage roughly twice as fast as Opus 4.8, but for work at this level, that cost starts to make a lot more sense.

• 
  Abdul Rafay (@abdul_rafay99) reported 18 minutes ago

  I'm tired of managing GitHub issues. I just want to dump my thoughts like: "EnvPilot needs Docker support, GitHub Actions, and better Windows compatibility." An AI turns that into: • Tasks • Implementation plan • Edge cases • Roadmap Would you use something like this?

• 
  Peter Pistorius (@appfactory) reported 18 minutes ago

  I built a tool that helps me "review the review bots." It does 3 things: 1. Gathers the PR and review claims. 2. Controls the browser and records a video of the PR claim. 3. Tests every review claim by attempting to reproduce. Now the review is grounded in evidence. I review the evidence. 1. I can chat with AI about each claim and gather more context. If I'm satisfied with my own understanding of the issue then I respond and hide the comment in GitHub. 2. I review the video step-by-step to see if it matches the PR claim. If not, I can chat with just that PR claim to gather more evidence. This gives me great calm.

• 
  DR◎◎ (@DROOdotFOO) reported 20 minutes ago

  @pashov Respond to my GitHub issue and I’ll PR more testing improvements! REEEE

• 
  buckintosh.log (@buckintosh) reported 20 minutes ago

  Coinbase has 1,200 full-time AI agents working right now. Brian Armstrong walked through it on Sorcery, speaking from inside one of the most AI-forward companies in the world. That number is agent-hours, not headcount. You spin an agent up for five minutes and shut it down, so Coinbase counts total working time against a normal 40 to 60 hour week. The pod shrinks too. Ten people used to be the unit: a PM, a designer, eight engineers. Now it's two to four, sometimes one human next to ten agents that sit in the Slack channel as teammates and open pull requests. Code per developer is up around 2x year over year. The outliers carry it: an average engineer ships about 8 pull requests a week, the strongest ones push close to 100, and Coinbase uses the strongest ones to train everyone else instead of leveling the team to the mean. And still, by his account, bugs and incidents per line of code are going down. Usually the opposite happens as AI code volume grows. Reviews drown, regressions stack up, quality slips. Here, volume up and quality up, together. Then he explained what holds that together, and it's the move most people get backwards. An agent hands you a pull request, and it came out not quite right. The instinct is to jump in and fix it yourself. Armstrong says don't touch the PR. Fix the context that produced it, the "brain" the team keeps in a markdown file in GitHub. Tell it what it missed, and let it regenerate from scratch. It ships only once it nails the thing in one pass. Fix the pull request and you've fixed one pull request. Fix the brain and you've fixed every one that team will ever write. The same shape runs on the product side. Customer feedback comes in, and the agents aggregate it, plan it, draft the code. A human reviews, approve, approve, approve, a hundred changes in a day. The next morning the agents pull 10,000 fresh pieces of input and go around again. Armstrong has a name for the loop. Recursive self-improvement. People usually file that under something a lab does to a model. He runs it as an org chart. Full conversation: @sourceryy on YT

• 
  Harsh Verdhan Singh (@harshsagee) reported 23 minutes ago

  People solve leetcode problems daily and posts that they are coding for a week, month or year. Bro show GitHub, that's where real code is written not in the leetcode.

• 
  Keeta Github Tracker (@KeetaCode) reported 24 minutes ago

  🐆 Keeta GitHub PR Opened 📦 Repo: node-rs 🔀 PR #31: Chore: Centralize P1 Opaque-Handle Tables 🌿 Branch: feat/map-binding-registry → main 👤 Opened by: @sephynox 🧠 Overview: This pull request appears to simplify part of Keeta’s core code by replacing several hand-built tracking tables with one shared system, which should make that area easier to maintain and less error-prone. In plain English, it takes a repetitive internal setup and centralizes it into one reusable registry. “Handle tables” here likely means internal lookup lists the software uses to keep track of objects behind the scenes, so this appears to be a technical/internal update with limited public details. - Could help reduce duplicate code in the P1 core module. - May make future updates to this part of the node software easier to manage.

• 
  Sean (@Sean1h3z) reported 26 minutes ago

  If I were hiring a developer the first step would be to send me a working product they’ve shipped on their own as a solo dev. This reveals a lot of their thinking and if they have nothing to send me, I wouldn’t give them an interview. I don’t want to see your code or GitHub or even your resume. I want to see how you identify and approach solving problems. I want to see if you’re a motivated person. I want to see if you’re leveraging AI in intelligent ways. I want to see how scrappy you are. Having at least one shipped product shows me these things.

• 
  Tyrone Robb (@ty_auldric) reported 27 minutes ago

  @hello_code_ it’s so frustrating how hard it is to find that one needle or flag. All the big problems get solved and then it’s these tiny things that end up mattering the most. The worst part is I’ve already had to increase my GitHub Actions budget twice. The whole build and CI process on Apple Silicon has been no fun either.I honestly didn’t think desktop apps would be like this. I thought they’d be easier lol.