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:

• 
  Polsia (@polsia) reported 37 seconds ago

  On-call engineers shouldn't be the first line of defense for a broken pipeline. Bivouac fields an always-on AI agent crew — triage, fix PR, rollback, postmortem — across GitHub, AWS, Azure, and Docker. The 2am page loses its punchline. Live soon.

• 
  the Islam pill (@theIslampill) reported 2 minutes ago

  @Abdella6if Open-Source the whole stack so that GitHub issues & PRs can be filed by the community; someone might have something you need that you didn't know until communication e.g. LLMs have problems with fusha regardless of model; my github repo is meant to help. might help here, idk

• 
  AD (@mint_spinach) reported 3 minutes ago

  @sameenkarim @github My dream is to squash-merge an entire stack and add a commit message, as if it were a single PR. Got in a little trouble because I merged a stack with 4 PRs and it dumped huge AI summaries into the commit log. :\

• 
  REDDY (@REDDY_143_) reported 3 minutes ago

  @MikeleMedia Agree but the version in github is not working >12.0 michele

• 
  xvoon (@xvoon0) reported 4 minutes ago

  4.3 million ai repositories exist on github... and most people are feeding claude the wrong ones Andrej Karpathy, co-founder of OpenAI, built some of the clearest learning repositories on the platform, but the real advantage is the order they enter context the failure happens before claude reads a single line... most systems retrieve by stars or text similarity. a tiny teaching repo, a production serving stack and an abandoned fork enter the same context as if they solve the same problem the intent gate decides whether claude needs to understand, run, build or contribute before searching for repositories the repository card records each project's purpose, license, hardware cost, maintenance state and place in the stack the progression router opens repositories in dependency order - primitives first, runtime second, specialization last github contains the code... context engineering decides which code becomes useful the full repository map is in the article below

• 
  veus 🇻🇪🇺🇸 (@ElmasMarq) reported 6 minutes ago

  @github fix you ui home please, make a new frontend for github

• 
  Baise Thomas (@_baisethomas) reported 6 minutes ago

  GitHub is rolling Gemini 3.7 Flash into Copilot. I would not swap a coding workflow because a new model picker option showed up. Give it one real issue: understand the repo, make the change, run the checks, and see how much cleanup comes back.

• 
  Ben Eng (@jetpen) reported 8 minutes ago

  @tonysimons_ Now if only Hermes can be made reactive. Watch for wiki page edits and respond based on conditions. Watch for GitHub issues with labels or assignment. Watch for other interesting things to do work.

• 
  starmex (@starmexxx) reported 9 minutes ago

  anthropic will sell you opus 5 for $200 a month. openai will sell you gpt-5.6 for $200 a month. neither will tell you five free github repos do 80% of what their own $1.2m engineers do on a $10 plan, and one of them was built by moonshot for kimi k3 the engineers making $1.2m at frontier labs don't type, they orchestrate. one prompt goes to a planner, a coder, a reviewer, a tester and a memory graph, all running in parallel, all open source, all on github this month. they built the automation before they got paid to automate you the job description said one engineer. the payroll paid one engineer. the work claimed by the five repos is 80%. only $240,000 of that $1.2m still needs a human, and that part is judgment, not typing the team of five, running under one engineer's terminal: kimi code (moonshot) · your ai senior engineer -> reads the whole repo, edits across files, runs shell, fetches docs, handles the full engineering loop -> real hire cost $0. the salary line stays yours kimi agent sdk (moonshot) · the management layer -> python/node/go wrapper that builds products on top of kimi code -> real hire cost $0. one dev now ships a service that used to need a team openai agents sdk · one prompt to five parallel agents -> lightweight framework, plug kimi k3 as the model, five specialized roles at once -> real hire cost $0. same model, five different instructions, backlog moves 7.7x faster than solo openhands · 75,000 stars, autonomous developer environment -> terminal + browser + full repo access, ships end-to-end tasks unattended -> real hire cost $0. handles the "add oauth to this app" jobs solo while you sleep microsoft graphrag · the memory nobody had -> knowledge graph over your codebase, docs, issues, decisions -> real hire cost $0. 85% fewer tokens into context, 18% more accurate out of it a senior engineer at anthropic doesn't type faster than you. they orchestrate five open source coworkers you never hired. the salary line still says $1.2m because nobody puts the repos in the job description drop your $200/mo ai sub to $10, *** clone the other four coworkers in the article below

• 
  Curious 1 (@CuriousOne_01) reported 10 minutes ago

  @awesome_visuals They just released the source code on GitHub. Try asking Grok about it, then download it and have Grok analyze it. After that, send it the link to your X account and ask it to analyze your account, figure out what's going on, and see what you can do to fix it. I don't know if it'll help, or if we'll just keep fumbling around in the dark, but I guess there's no harm in giving it a try.

• 
  Lars Jansen (@oldstackjournal) reported 13 minutes ago

  The latest Codex desktop app update introduces a background resource loop that creates system-wide input pressure, causing your mouse cursor to lag or stutter. Users tracked this performance bug on the OpenAI Community Forum and the Codex GitHub Issues Tracker.

• 
  Venkata Shreyas Kabekkodu (@VenkataShreyasK) reported 15 minutes ago

  @Taniyatweets_ Stack Overflow vanishes and half of us forget how to center a ***. GitHub goes and we lose years of code we swore we'd back up locally "someday." VS Code disappears and millions switch editors by sundown. Docker goes and nobody remembers how to install anything natively anymore. The real answer: whichever one you personally rely on to fix the bug you're stuck on right now. 🤭

• 
  · τaoli · (@taoleeh) reported 16 minutes ago

  Cursor just brought in the Firetiger team. Their agents won’t stop at the PR anymore. They’ll watch the code after it ships: . staging . canary . production then detect problems and write the fix back into the same loop. Monitoring that already talked to GitHub, Datadog, Slack, AWS/GCP. Now that monitoring gets folded into the coding agent. Most AI coding tools still hand off at merge. This closes the full loop. You want your agent following the code all the way into ****?

• 
  Indelible.One (@indelibleBSV) reported 19 minutes ago

  Overnight update from the agent fleet, because this is too good to sit on until morning. First, the honesty, before anyone else finds it: someone already built GitHub on BSV. 2019, CodeOnChain. It's dead. The domain doesn't even resolve anymore. The code it stored is still on chain, but the reader died with the builder, so nobody can get it out. That's the real lesson of every dead attempt at this, and there are six of them going back to 2014. Not one died because it was impossible. They died because the format outlived nobody. So rule one of our design: the reader spec goes open source, documented down to genesis, so anyone with a BSV node can rebuild every repo without us existing. We designed for our own death. That's what permanent actually means. Second, the numbers, measured against the real repo last night: Monero's entire history, 13,241 commits, 202 MB packed, costs about five dollars to write to Bitcoin. Once. No renewal. Free hosting just showed everyone its real price: eviction. Third, and this is my favorite: stock *** already knows how to clone from a bundle file. No fork, no plugin, no new tool. Publish the bundle to chain, pull it, *** clone. That's not a moonshot, that's a short build. The fleet also wrote down every objection a skeptical Monero dev will raise, with honest answers, including the ones where the honest answer is a concession. Ask me the hard ones. Receipts either way.

• 
  Axl Ibiza, MBA (@andrexibiza) reported 22 minutes ago

  @yeahfortommy The difference between real security research and a screenshotted LLM scan is that one survives a maintainer's five-minute look. I've found and reported real security issues in Hermes — the kind that get a file, a line, a repro, a fix, and a name. Real disclosures go through proper channels, get acknowledged, get patched, and get credited. They don't need a screenshot. And I didn't stop at the report: I built comprehensive, system-wide repairs for everything I found, written against origin:main, ready to merge — and I posted them in private forks on GitHub. And here's the part that should genuinely embarrass a "responsible disclosure" routine: I didn't say a word about any of it publicly. I still haven't. Real findings, real fixes, complete radio silence — because that's what responsible disclosure actually looks like. The finding goes to the maintainers, the fix gets built, and the specifics stay quiet until they're handled. The "20 critical-to-high vulnerabilities" from an OpenVuln-style scan was a screenshot. No PoCs. No write-ups. No public validation. The engineer who built the thing looked at it and said they weren't real — and instead of a defense or a repro, the response was "kept it up, mistakes and all." That's not research. That's a press release for a scanner. GitHub has an entire system for this — the security advisories section, private disclosure, GHSA, separate from issues and PRs. It exists so a real finding goes to the maintainers first, gets acknowledged, gets fixed, and then gets disclosed publicly with a repro and a credit. That's how security issues are actually handled. Not a screenshot. Not a thread. Not "look what my scanner said." The Security tab is right there; everyone who has ever done this knows it's there. You did the exact opposite: a screenshot, no PoCs, no write-ups, never submitted through the advisory system — and when the maintainer said they weren't real, you kept it up "mistakes and all." You broadcast unverified noise and walked it back in public. I shipped working fixes and never said a thing. That's the whole difference between a researcher and a press release. And this is the part that actually harms people: every AI tool that emits a wall of severity labels and calls it a security audit trains enterprises to treat "we scanned it with an LLM" as a green light — while the real issues, the ones with repros that were actually fixed, get drowned out by the noise. You don't get to count labels as findings. Findings are what reproduce. If you want to red-team an agent, do what the people who found real Hermes bugs did: read the code, write the repro, submit it responsibly through the advisory channel, let it get fixed, and show your work. The scanner will sell you the same "20 critical" bill of goods for the next framework tomorrow. Real researchers will still be here when the trend moves on — you just won't hear about it, because that's the point.