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:

• 
  Alekos Filini (@afilini) reported 8 minutes ago

  @LaurentMT @giacomozucco @nitesh_btc Ok let me rephrase because clearly this pissed me off and I didn't argument well. First thing is I don't like this things being taken out of context and presented on twitter to a non-technical audience. I think this has created somewhat the perception that BDK is not well reviewed or that the quality of the code is not where it should be. To be clear: I'm not saying BDK is perfect. One more round of review will never hurt. But I think the quality is already pretty high and much higher than many other things out there (see libngu with zero reviewers and single letter commits). The second point is that while I appreciate non-technical people reviewing code, I have somewhat mixed feelings on it. Because most issues are not going to be obvious like // FIXME: vulenrability here, I wonder how valuable their contribution could be. What are the chances they really find an issue, if it's buried so deep that even BDK reviewers missed it? Then again I understand as a user, in this moment, you might be scared and trying to read code will give you a sense of security. But if you are non-technical and reading code you should double/triple check before making claims, especially on social media where people are just scrolling and see BDK -> entropy -> bad. Wanna read code? Go use coding tools. Open an issue on GitHub at least. I don't know. That's the other thing that pissed me off

• 
  Jeffrey Emanuel (@doodlestein) reported 11 minutes ago

  Also, since implementation hasn’t started yet, if you have any good ideas for the design or how the system should work, feel free to submit them as GitHub issues in the repo. Fable will decide whether something belongs in the plan or not (sorry, I don’t make the rules!). FABA🦾

• 
  marc 🦀 (@mkenzo_8) reported 14 minutes ago

  @github please fix yourself

• 
  Brown Geeky kid (@shigma_male) reported 15 minutes ago

  GitHub dropped a release candidate for Enterprise Server 3.22. It’s all about enterprise features these days—meanwhile, devs are probably just waiting for Copilot to write these changelogs itself.

• 
  Hatman 🎩 (@hatman) reported 16 minutes ago

  Every major AI lab tells you your model's reasoning is encrypted before it leaves their servers. Turns out that promise rested on a single shared key. Researchers found that OpenAI, Anthropic, and Google all use one global key for encrypted reasoning blocks, meaning those blocks can be handed to a different, weaker model from the same company and read back out in plain text. Nobody has to break into the strong model itself. To prove the risk was real rather than theoretical, the researchers decoded 315,320 reasoning blocks scraped from public GitHub and Hugging Face repositories. Inside them they found 367 pieces of personal data and 182 credentials that developers had no idea were sitting in plain sight. All three companies have since patched the flaw. But logs already published stay exactly as exposed as they were before the fix. If your team has ever published API session logs from a reasoning model, that's worth checking.

• 
  Sebastian Kehle (@sebastiankehle) reported 16 minutes ago

  how to turn your github backlog into a software factory most teams are just running more coding agents in parallel. this produces more code, but every issue still depends on one human carrying the full context from report to reproduction to fix to review the better setup is one agent per job: 1. classifier agent reads the issue and decides whether it is a bug, feature, docs update, or unsupported request it returns: - classification - confidence - reasoning - required next step 2. analysis agent never trusts the issue description for a bug, it writes a minimal reproduction and runs it against main for a feature, it writes a probe that proves the capability is missing then it returns: - reproduction/probe - command output - affected packages - implementation spec - compatibility risks 3. implementation agent receives the issue plus the analysis artifact, not the full transcript from the previous agent it works inside an isolated sandbox, implements the spec, runs the test suite, and opens a PR with the evidence attached 4. review agent reviews the PR from a fresh context and scores: - completeness - side-effect risk - performance risk - backwards-compatibility risk - test quality the agent that writes the change should never be the only agent that approves it 5. human reviewer reads the chain of evidence and scales review depth to risk docs fixes get quick verification provider updates get focused validation new public APIs get deep review humans still merge every change. the factory automates everything around that decision the important part is the handoff contract: input artifact produced commands run result risk next action agents should pass evidence, not confidence also classify every run: success: safe to ship flawed: improve the prompt, context, or eval blocked: provision the missing tool or dependency manual: intentional automation boundary this is how the system compounds. every failed run becomes a new eval, capability, or guardrail vercel is already running this pattern on AI SDK. after four weeks, its factory was authoring 25–35% of merged PRs, closing more than 75% of issues in july, and had reduced open bugs by roughly 25% the next version of agentic coding is not one genius agent with every tool it is a production line of narrow agents, typed handoffs, isolated sandboxes, evidence at every step, and one accountable human at the end

• 
  Aakash Gupta (@aakashgupta) reported 16 minutes ago

  There is an entire genre of content about how to write the perfect CLAUDE .md . Front matter debates. Line count debates. What to put in, what to strip out. Oji Udezue's scaffolding skill generates one per project, as an output. Not a template you adapt. A CLAUDE .md written for this specific codebase, that knows how to hunt bugs in it, follows the folder patterns it just created, and follows the prototype patterns it just set up. Then it does the part people miss. It writes the folder structure into that same file. Where documents go. Where milestones go. Where prototypes go. For every folder, what belongs in it, addressed to both you and Claude at the same time. That makes the repo self-organizing. The instructions for maintaining the structure live inside the structure. Nobody has to remember the convention, because the convention is loaded into context every session. Same pass sets up the rest of the scaffolding most vibe coders skip entirely. A test folder. Continuous integration, so every check-in runs the tests. Security, including finding your secrets and getting them gitignored before you ever push to GitHub. Then a bug classification system that catalogs failures over time and learns from them. That last one is the sleeper. Everyone treats bugs as things to fix and forget. Cataloging them turns your own failure history into context the model can read on the next project. Worth saying what he did not claim about any of this. On the market research the same skill produces, he was direct: you can get 60 to 70 percent of the way with LLMs before you go find real sources. He said the danger in the whole workflow is taking it as gospel, and that he does not fully trust LLMs 100 percent, so he reads every output. A skill that produces this much artifact this fast only works if someone is still reading it. The CI templates and the CLAUDE .md pattern are the pieces this offloads best, because they are the pieces most people either get wrong or never do at all. Your CLAUDE .md was never supposed to be a thing you author once and defend. It is infrastructure, and infrastructure should be generated.

• 
  Abi (@abh3i3) reported 18 minutes ago

  While trying to fix it, I checked old GitHub workflow actions from a year ago. Seeing those 1-year-old actions actually made me smile—it’s officially been a year! But I was still stuck trying every way to solve the issue without a clue what was wrong. (3/5)

• 
  Arian van Putten (@ProgrammerDude) reported 18 minutes ago

  @nniroprobably GitHub is more than *** operations. We had issues with the API every day of the week. Including merge queue not working and GitHub actions UI not updating artifacts not downloading graphQL queries not working release note generation being broken.

• 
  Mark yu (@Markymarco34) reported 24 minutes ago

  Yes, the network is currently running and the signer set is healthy. But current operational health does not mean the Aug. 12 stall has been permanently resolved. The related GitHub fix is still open and unmerged.

• 
  Umair Ali (@buildwithumair) reported 25 minutes ago

  Solo Startup Founders Pack - Codex = coding. ($20/mo) - Supabase/Convex = backend. (Free) - Vercel = deploying. (Free) - Polar = payments. (3.9%/transaction) - GitHub = version control. (Free) - Resend = emails. (Free) - Clerk = auth. (Free) - Cloudflare = DNS. (Free) - PostHog = analytics. (Free) - Sentry = error tracking. (Free) - Upstash = Redis. (Free) - ShipClaw[.]io= ai agents ($14/wk) Total monthly cost to run a startup: ~$50 There has never been a cheaper time to build .

• 
  Greg Tassone (@fcgreg) reported 25 minutes ago

  @xchatty I ran some tests against E01 packaging today and it worked very well. I didn't find any errors so far. Speeds seemed good, although it complained about likely I/O exhaustion on that test machine. I'll report back on Github if I find anything useful.

• 
  John Broadway (@JEBroadway) reported 34 minutes ago

  Look at where the real money in AI is going: infrastructure and hardware. Software is becoming cheap to generate. Hell, AI can even help layout silicon, but that code still has to execute on physical machines. That’s why NVIDIA and hardware vendors hold all the cards right now. The era of the "pure" syntax coder is ending. We don't need people typing out loops; we need hardware engineers and systems folks who understand physical compute and how big systems fit together. I’ve been in tech 36 years, from bench PC repairs to global rollouts and managing dev/hardware teams through every shift since the 90s. To me, AI hasn't made us all coders. It’s made us builders. I’m not a developer. I’ve looked at COBOL, Pascal, Python, C++, Rust, you name it. I see patterns and images, not syntax. When I use AI, I describe mechanical systems: "Picture plumbing in a ten-story building" to map how data moves between agents and apps. That's the difference: A coder connects one pipe. A builder knows what concrete goes into the foundation to support forty floors, where electrical runs, how HVAC ties in, and what the final layout looks like. Calling this "vibecoding" as an insult misses the point. AI handles syntax so I can focus on how the engine runs. It also exposes how broken software licensing is. Selling clunky subscriptions, charging for bugs, and charging again for patches is dead. Value is shifting back to open source, real support, and experience. Open-source tech like Proxmox and ERPNext are primed for this, and I build right on top of them. So when you see me arguing with traditional devs in forums, it isn't Dunning-Kruger. It's just the pattern laying itself out before you see it. I build governance into everything. In my world, I see a farm with a barn housing bare-metal servers running Proxmox while my ERPNext hums along generating mailbox money. ;) Some of my GitHub Projects so far: Maude for Claude: A dedicated partner environment operating right inside Claude. Proximo: A lean infrastructure and governance tool built to streamline local workflows. Pacioli: An open-source accounting and operational framework built for real execution. AI can write all the code it wants. It still can't bolt a rack into a datacenter or manufacture the chips it needs to run on.

• 
  AML (@alysha_lobo) reported 36 minutes ago

  𝐔𝐏𝐃𝐀𝐓𝐄 𝟑: We manually reviewed 𝐄𝐕𝐄𝐑𝐘 single profile (kudos to the team for going the extra mile with me on this!) Before we drop 𝐭𝐡𝐞 𝐧𝐞𝐱𝐭 𝐌𝐓𝐒 𝐫𝐨𝐥𝐞 (𝐥𝐨𝐚𝐝𝐢𝐧𝐠 𝐢𝐧 𝟑…𝟐…𝟏), here is some 𝐟𝐢𝐫𝐬𝐭 𝐜𝐮𝐭 𝐜𝐚𝐧𝐝𝐢𝐝 𝐟𝐞𝐞𝐝𝐛𝐚𝐜𝐤 if you actually want to stand out: 🛑 Stop using AI to mass-apply — the team and I have been neck deep wading through so much slop!! This wastes everyone's time. 🔗 Double-check your links — broken GitHub, Tech blog, or GDrive links mean instant rejection. 💡 Show, don't tell, your frontier potential — we want to see "cracked," out-of-the-box ideas you have actually cooked and served. 𝐘𝐨𝐮 𝐰𝐢𝐥𝐥 𝐡𝐞𝐚𝐫 𝐟𝐫𝐨𝐦 𝐮𝐬 𝐧𝐞𝐱𝐭 𝐰𝐞𝐞𝐤 𝐨𝐧 𝐧𝐞𝐱𝐭 𝐬𝐭𝐞𝐩𝐬. 𝐓𝐡𝐚𝐧𝐤 𝐲𝐨𝐮 𝐟𝐨𝐫 𝐲𝐨𝐮𝐫 𝐩𝐚𝐭𝐢𝐞𝐧𝐜𝐞! 🇮🇳 𝐂𝐦𝐨𝐧 𝐈𝐧𝐝𝐢𝐚 — let’s raise the bar and show what we can truly build at the frontier of tech. 🚀

• 
  Chase McDougall (@ChaseMcDou) reported 37 minutes ago

  @KaustubhTrived6 @ayesha_fatiima But if I can't push to GitHub because GitHub is down - then how can they push to GitHub to fix GitHub