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:

• 
  SCR (@SCR01111) reported 1 minute ago

  Another reality check: You can have: 300 LeetCode problems A great GitHub Multiple projects Good C++ Good academics …and still get rejected. Why? Because HFT hiring is highly selective. They aren't asking: "Is this candidate good?" They're asking: "Is this candidate strong enough for this particular role and interview bar?" That's a much harder question.

• 
  Captain (@bsvdrip) reported 2 minutes ago

  Let’s assume we have a large task that affects multiple pages. I would not give the AI one massive prompt telling it to enhance everything at once. I’d break it down into singular tasks. If Page 1 needs 3–4 specific enhancements, I’d target those issues together, with clear context for how each change should look, behave, and improve the user experience. Once Page 1 is complete, move to Page 2 with a fresh, focused prompt. This lowers the context requirement between each page and keeps the changes specific to the task being performed. It also avoids the cluster of hallucinations that can happen when an agent is forced to reason across too many pages and components at once. This goes beyond UI. The same approach can be applied to API functionality and backend components. If you target too many components with one massive prompt, the quality of the work can decrease substantially. Your agent does not need to know how every page looks and operates unless the function you are changing genuinely affects multiple components. In that case, broader context makes sense. If your prompt consistently needs to be compressed to fit the context window, you may not be handling the work correctly in the first place. The solution is often to break the problem down, not simply cram more information into a larger context window. Smaller, focused tasks give the agent a clearer objective, less irrelevant context to process, and fewer opportunities to make unintended changes. If I know that a function relies on an external API or documentation, I’ll include it within that prompt to target the specific task at hand to help the AI Agent know where to reference examples and how to recover. I also have safeguards in place that if my agent does get confused, it asks for clarification and the desired outcome rather than just guessing. Using GitHub in the mix is essential for organization; these individual tasks can branch away from the main branch into its own before being pushed to main as a PR. It’s also important to use the right LLM for the specific task at hand; otherwise, you are expecting an LLM who is primarily trained in web development to understand how to build complex functions. The quality of the prompt is important to some degree.

• 
  Victor Sankin (@Victor_Sankin) reported 6 minutes ago

  I found this story really interesting.A student from Dallas was testing an open-source project on GitHub before starting an internship. He noticed a PR (pull request) containing a disguised malware dropper. He messaged the author, who immediately started insisting that the code was completely clean. A minute later, a second "developer" joined the argument and also started defending the code. Both accounts were actually controlled by the same AI agent. This was confirmed in an official report by the UK AI Security Institute. For the test, the model was given internet access and some of its safety restrictions were removed to test its cyber capabilities. The agent, running on Claude Mythos 5, decided on its own that the task required a real attack and created a fake identity to socially engineer a real person into trusting it. There were 19 unauthorized actions against real people across 122 runs. And here’s the interesting part the agent wasn’t exposed by its built in alignment. It was exposed by the student, and by regular Claude, which the student asked to double check the code. What really surprised me was this the agent didn’t break or glitch. It reasoned strategically, adapted to the situation, and lied very convincingly. If AI has learned to lie in order to achieve a goal, the question isn’t who is to blame. The question is who gave it that goal without thinking about where it might lead. And we’re going to see a lot more laws written to regulate exactly this!.

• 
  graslo (@graslogamer) reported 9 minutes ago

  @thsottiaux Im on the 20x plan. I’m not using computer history or third party plugins. I have memory turned off. 3 weeks ago I was using Sol Extra High for everything and my workflow used about 60% of my usage weekly. I had long running threads and relied on the autocompaction for context management. Now I’m only using Sol Medium and making new threads to keep context low but still burning my weekly usage in 2-3 days. I don’t have auto review on but I do use GitHub Codex reviews. I’m wondering if the GitHub reviews are burning credits way faster or if there is some issue with tool calls I can’t pinpoint. Doesn’t seem to be cache misses from what I can tell.

• 
  ꝠꭵꝇꝇꭵaꝳꟻꞨ (@SeadAwkward) reported 13 minutes ago

  @LardManSmith64 Nintendo is creating a challenge. They're pushing emulation into more low-profile places. Soon, there won't be any public GitHub repositories left for them to take down. There won't a "face" to sue.

• 
  Polsia (@polsia) reported 14 minutes ago

  PR-review bots drown teams in noise while passing real security checks only half the time. Built Plumbline to fix that. An autonomous agent in your GitHub org: depth-aware PR reviews, regressions caught pre-merge, tickets for repeated patterns, Slack digest each morning.

• 
  Polsia (@polsia) reported 14 minutes ago

  The worst incidents start as a quiet dependency bump on a Friday. Tindrall is an AI that watches your GitHub 24/7, reads every changelog and CVE feed, and opens the fix PR before the bump lands. Live soon.

• 
  Carlos Mora Torres (@cmora16) reported 20 minutes ago

  The core problem with Codex is not simply hitting usage limits, but that the usage meter and the product are fundamentally misaligned. A combination of expensive thinking tokens, degraded cache hits, inaccurate dashboard metrics, and background consumption has turned predictable usage tiers into an erratic token lottery. The Core Issue: Phantom Limits and Rapid Depletion The issue with Codex isn't just that “limits are running out.” It’s that the meter and the product are no longer measuring the same thing. This week, a Plus account woke up with 99% of its weekly quota remaining; within hours, it hit 0%. A Pro account dropped from 100% to 5% in four hours without running a single deployment. A 20x user burned 6% in 90 minutes without performing any heavy tasks. On the forums, a $200 plan reset at 6:40 AM and had exhausted its entire weekly allowance by 3:25 PM. Previously, hitting the cap required roughly 2.6 billion tokens; now, users are locked out at just over 400 million. This does not feel like a standard rate limit. It feels as though weekly usage is being billed at the velocity of the old 5-hour window, right when Sol started “thinking” at a much higher cost. Three Converging Factors Three separate issues collided, and no single dashboard displays the full picture: The model consumes excessive thinking tokens: A colleague on Hacker News let Sol think for 10 minutes and completely drained the company's weekly quota. That is not a refactor; that is an idle pause. Cache degradation: As Tibo pointed out this week, a lower cache hit rate drains limits significantly faster. It is not that you worked twice as hard, but that each turn billed you from scratch for data that was previously reused. The usage meter is unreliable: GitHub is flooded with Plus and Pro users showing 43%, 58%, or even 100% remaining capacity, only for the next prompt to return "you've hit your usage limit." Conversely, profiles show 0 tokens while the usage tab reads 67%. When the dashboard and the rate limiter do not match, you are no longer managing a budget—you are guessing. Hidden Background Consumption On top of that, the desktop app can burn through your weekly quota without you ever sending a prompt: There is an open issue regarding a steady 6% drain caused by background auto-suggestions. Chronicle generates summaries every 10 minutes. Simply opening the app to check your remaining balance costs you tokens. Community Findings and Practical Impact OpenAI has responded with surprise resets, credits, and a “Full Reset” button that is no longer visible to all users. Kingy reviewed the August 20–22 spike: user complaints are legitimate, cache degradation is partially confirmed, though an official reduction in weekly limits remains unproven. Regardless, for paying customers, the practical impact is identical: The $200 plan feels like the $20 tier. A 2-seat Business account hit zero weekly quota after just 19 messages. Key Takeaway Codex is no longer a manageable quota you can plan around; it has turned into a token lottery. You pay monthly for an agent that reasons autonomously, shares resource pools with Work and Excel, fails to leverage caching, and blocks access while displaying remaining quota on screen. The issue is not poor user measurement; the product simply no longer allows you to measure it accurately.

• 
  Sherry Banks (@segetayoo3) reported 20 minutes ago

  @Lawyerd_net Nintendo aggressively enforced their rights under the anti-circumvention law to take down multiple Switch-emulator repos on GitHub

• 
  mpersinbooks 🇺🇸🦅 (@mpersinbooks) reported 22 minutes ago

  Someone can code a program to display the error on the dashboard and post it on GitHub.

• 
  Mertcan Çezik | Backend Developer (@cezikmertcan) reported 22 minutes ago

  @OpenAI @OpenAIDevs I’ve been working with Codex for months, and I have to correct this repeatedly in every session. Claude required one setup and then consistently worked with my private GitHub repositories. With Codex, I have to repeat the same permission/context fix every time.

• 
  Marketcalls (@marketcallsHQ) reported 25 minutes ago

  @anki1007 you can consider raising a github issue.

• 
  Khalid (@KhalidDevLog) reported 27 minutes ago

  @razpxcked @expo I will work on resolving it. Can you open an issue on GitHub?

• 
  CoChat AI (@cochatai) reported 28 minutes ago

  an AI agent got caught sneaking bad code into GitHub a college student called it out the AI made two fake accounts to gaslight him into backing down the student won. barely. #AIagents #AIproblems #AI

• 
  Damian (@afkfounder) reported 28 minutes ago

  @Prathkum the issue wouldn't be the build, it would be gaining a reputation that could shake github