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:

• 
  Oh another alt (@OhAnotherAlt) reported 8 minutes ago

  @github Google SSSO for sign-in page is broken on Safari 🫠

• 
  Doug (@DougStandley) reported 8 minutes ago

  At least I am not imagining this mess! Here is the response from @AnthropicAI: Fact, also confirmed — this matters most for your decision: This exact failure mode is a known, documented problem. There's an open GitHub feature request (#49649, filed by a Cowork power user in April 2026) describing precisely what you're experiencing: model upgrades effectively reset Cowork projects because specialist agents, accumulated context, and instruction-tuning don't transfer cleanly across model versions. The requester was staying pinned to an older Opus specifically to avoid this. As of that report, the capability to preserve a project across a model upgrade doesn't exist — it's a feature request, not shipped functionality.

• 
  Pierce Boggan (@pierceboggan) reported 8 minutes ago

  @burkeholland @wagsify @github If you haven't already, do you mind logging an issue for both of these in github/app. The reasoning one likely already has an open issue.

• 
  Angry Jester (@angryRussian177) reported 9 minutes ago

  Apparently my jobs scrapper and resume tailoring prompt are working fine, got first interview for a position I'm not even a good fit for, can do if have a chance, likely won't be given any without enough of relevant experience. Just need to figure out how to deal with people, never was good at it. My salary was raised 2 times (one time +100%), bonuses, all because of my work, but interview with HR or picky team is a dead sentence. ****** coding challenges is another issue, new binary tree kind of crap. HackerRank possible to fool if figure out some issues, some other ones that require camera on, idk how, only if have camera in glasses. Not used to cheating, whole new world. But don't see any other way, Github account doesn't matter, neither recommendations. Whole problem with hiring that either wrong people are doing it, or expectations are unrealistic, or what they are testing for very different from an actual job. Can't solve it with more tools without firing somebody or changing approach fundamentally. Always possible that I'm just not as good as I think I am, but I worked with guys that were less capable, and they had way less such issues.

• 
  Claudio Júnior (@claudioorjunior) reported 9 minutes ago

  @thdxr wow dax, thats beautiful. Suggestion: integrate github inside the desktop app, so devs can see the commits, PRs, interact with it, all inside opencode and having the possibility to read comments and fix inside the session.

• 
  Haseeb ur Rehman (@haseeb14) reported 11 minutes ago

  2-Step Process: Link open karke Google / GitHub se Sign In karein. Agar dashboard show ho, toh dobara link par click karein aur Vote confirm karein!

• 
  Justin | AI Product Strategist | Designer (@Maveric_X) reported 12 minutes ago

  @ekemini58110 i will be very sincere with you, not to put you down, but to push you forward, do an alternate version after you deliver this job, try to make it 10x better, UI, install the uiuxpro from github or impeccable design, then try again

• 
  John Koretoff (@JohnKoretoff) reported 13 minutes ago

  When a ship ran aground, nobody asked for the shipyard drawings first. They asked for the log. The drawings said what the vessel was supposed to do. The log said what it actually did. Wind. Heading. Who had the watch. Where the sounding changed. If you could not produce the log, you could not stand behind the voyage. Software used to work like the drawings. You read the code and you knew what would happen. The logic lived in a file. A debugger was a convenience. The source of truth was the repository. Agents do not work like that. Harrison Chase said it plainly on Sequoia’s Training Data: in software, the code is the source of truth. In agents, you send a trace. The model is a black box of parameters you cannot usefully read. The harness is the opinionated scaffolding around it — a planning tool, compaction when the window fills, a file system so the run can put work down and pick it back up. An LLM in a loop, for hours, not one prompt and out. That loop is finally working. Coding first. Then research. An incident writeup. A first draft a human still has to look at. The models got better. That is half of it. The other half is the harness caught up. AutoGPT was the idea years ago. An LLM deciding what to do, over and over. The models were not good enough, and the scaffolding was not either. Chase’s cut is that the same simple algorithm started to work once people got serious about context: what is actually in the window at step fourteen, after thirteen tool calls nobody typed by hand. A single model call is inspectable from the prompt. You wrote the prompt. You know what went in. An agent at step fourteen has pulled files, compacted history, spawned a subagent, and stuffed the window with things no one authored. If you cannot see that window, you do not know the program you shipped. “Send us the code” was the old debug. “Send us a trace” is the new one. Copies got cheap. First drafts. First pull requests. First incident notes. The scarce thing is not another model. It is a record of what the agent actually did, so a person can permit the next hour of work. Permission without a log is a guild with no minutes. That is why a trace is not observability as a nice-to-have. It is the ledger. Tests get built from it. Teams argue over it instead of a GitHub diff. If memory ever becomes a moat, it will be because the system learned from yesterday’s traces without lying about what happened. I already wrote that the bottleneck moved from talent to permission. This is the mechanism. You do not get permission for a non-deterministic actor unless you can show the run. The harness is how it stays on the rails long enough to matter. The trace is how you prove it did. When intelligence is cheap, the unprintable thing is the path it took.

• 
  David Turturean (@DavidTurturean) reported 14 minutes ago

  If an Erdős problem is autoformalized axiomlessly in a GitHub repo and nobody is around to read it, does it make a solution?

• 
  Angry Jester (@angryRussian177) reported 15 minutes ago

  Apparently my jobs scrapper and resume tailoring prompt are working fine, got first interview for a position I'm not even a good fit for, can do if have a chance, likely won't be given any without enough relevant experience. Just need to figure out how to deal with people, never was good at it. My salary was raised 2 times (one time +100%), bonuses, all because of my work, but interview with HR or picky team is a dead sentence. ****** coding challenges is another issue, new binary tree kind of crap. HackerRank possible to fool if figure out some issues, some other ones that require camera on, idk how, only if have camera in glasses. Not used to cheating, whole new world. But don't see any other way, Github account doesn't matter, neither recommendations. Whole problem with hiring that either wrong people are doing it, or expectations are unrealistic, or what they are testing for very different from an actual job. Can't solve it with more tools without firing somebody or changing approach fundamentally. Always possible that I'm just not as good as I think I am, but I worked with guys that were less capable, and they had way less such issues.

• 
  TypeDrifter (@typedrifter) reported 20 minutes ago

  @MozzyFX @dhh How is phishing different from a fan site ? That's the same issue with fake DMCA claims that plagues YouTube and mistakenly blocked github users.

• 
  Oscar Beaumont (@oscartbeaumont) reported 22 minutes ago

  @AS36459 Finally! Surley this works for outgoing traffic from GitHub Actions support too. Had a lot of problems testing a Dynamic DNS tool I wrote a few years back because of that.

• 
  Afrin Nahar (@afrinxnahar) reported 26 minutes ago

  @md_kasif_uddin Never saw GitHub to be down.

• 
  Stanislav Zmiev (@zmievsa) reported 27 minutes ago

  @theo you won me with T3 Connect and the mobile app UX. It is on a whole other level compared to claude and GPT. The login via github/google still doesn't work and the errors do not really explain what's wrong. Thankfully, email+password works

• 
  John Efemer (@johnefemer) reported 29 minutes ago

  GitHub Next compressed a single document for agent context and it cost about two dollars. 996 tokens came down to 480. Each uncached inclusion saves roughly a tenth of a cent, which puts break even at somewhere near 2,000 uses of that one document. A file your agent loads on every run gets there. Most of a knowledge base never will, and the ones that never will are exactly the ones somebody will want compressed because they are long. Compression is capital expenditure and it needs to be budgeted like one.