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:

• 
  Roy (@__roycohen) reported 5 minutes ago

  @geerlingguy Even Github doesn't have ipv6 Just fyi, if you're on ipv6, Github won't even error if you try to connect, leaving you frustrated. :D

• 
  Rohit Yadav (@RohitYa08173144) reported 5 minutes ago

  The hardest part wasn't the code. It was starting when I knew nothing. Most freshers wait until they're "ready." I just started broken and figured it out. You don't need a bootcamp. You need a project that scares you a little. Start today. Push to GitHub. Share the link.

• 
  momo! (@costackedceo) reported 8 minutes ago

  Here’s your monthly reminder that copying StackOverflow is for beginners. The new meta is feeding your agents yesterday’s GitHub issues so they hallucinate tomorrow’s breakthroughs.

• 
  Josh Koenig (@outlandishjosh) reported 10 minutes ago

  We took the trouble to build a first-class GitHub integration, whereas they have some scripts they'll run. We spin up environments automatically for every pull request, whereas they have one dev env that you have to figure out how to juggle.

• 
  Graham Neubig (@gneubig) reported 14 minutes ago

  Agreed. My basic agent loop is this: 1. Organize my to do list: pull in context from email, slack, github, and linear and add linear issues for any missing things 2. Walk through the to do list in order of priority, tell me the current status, ask what to do 3. I delegate a sub-agent (using the OpenHands cloud api typically) to work on that task 4. Once I have many tasks going at once, check on the ones that finished in order, and take any action I tried a more automated version of this where I skipped step "2", but eventually it resulted in a lot of slop, so I think the human decision step is still pretty useful.

• 
  Cody Mullins (@codemullins) reported 18 minutes ago

  @awesomekling @ladybirdbrowser Impressive! I don’t want to brag but my browser only took 16gb of RAM to load a broken GitHub homepage.

• 
  yourfren (@0xyourfren) reported 19 minutes ago

  Agent woke up telling me Github is down.

• 
  David Jennings (@ChilliDoor) reported 22 minutes ago

  @aap_twak @joshmanders The recent security issues with GitHub show that that's not really true anymore. You now need at least a basic level of caution

• 
  Josh Ellithorpe (@zquestz) reported 23 minutes ago

  Everyone is getting owned. Microsoft shuts down 70+ GitHub repos due to malware. Sadly I think these exploits will get significantly worse before quality solutions emerge.

• 
  nofunsir (@nofunsir) reported 25 minutes ago

  @tsoding which is a large reason why the whole DEVOPS CD/CI thing happened. rather than being responsible, updating, fixing merge issues locally at end of day, everyone just wants to offload that to jenkins or github overnight, and go home. pure laziness.

• 
  Firas D (@firasd) reported 25 minutes ago

  @chrisalbon What these guys are skipping over is that they have a thousand github issues right They aren't prompting the agent directly with hey lets look at XYZ cause the github issue becomes the prompt

• 
  Andrei Onel (@AndreiOnel) reported 26 minutes ago

  I don't have enough details to open a github issue though 🫤

• 
  Ross (@RV_Smirnov) reported 30 minutes ago

  Coding agents just hit the 2nd-order problem: not "can it write code?" but "can you afford the tokens when it loops?" GitHub moving to usage-based AI credits is the pricing wake-up call for agent infra. Reliability now includes cost control. @karpathy @yoheinakajima

• 
  Bryan C Woods (@BryanCWoods) reported 30 minutes ago

  @openclaw OpenClaw GitHub issue (#86521, filed 2 weeks ago) - any idea of resolution? When using DeepSeek V4 Pro with thinking mode enabled, the API requires `reasoning_content` to be echoed back in every assistant message when continuing after a tool call. OpenClaw handles this correctly when the endpoint class is `deepseek-native` — but when DeepSeek is accessed through OpenRouter or any proxy provider, the endpoint class is detected as something else (e.g. `openrouter`), so `shouldPreserveReasoningContentReplay()` never fires. The result: multi-turn tool call chains break mid-session with a 400/500 error. **Related:** GitHub issue #86521 covers the proxy case. The REASONING_CONTENT_REPLAY_MODEL_IDS set and the `isDeepSeek` detection in `openai-transport-stream.ts` are the specific spots that need patching. DeepSeek just made V4 Pro pricing permanent at $0.87/M output — that's 17x cheaper than Sonnet 4.6 for the same benchmark tier. A lot of people running cost-sensitive agentic setups are trying to migrate. The tool call breakage is the single blocker. Would love to see `reasoning_content` replay detection extended to cover proxy/OpenRouter endpoints — either by model ID matching or by a provider-level compat override. Happy to test a patch if anyone wants a real multi-tool-call workload to run it against.

• 
  masYNYa (@KijAkubovs86334) reported 32 minutes ago

  A developer in Chengdu recorded a 42-second clip on his MacBook Air last Thursday. No voiceover. No explanation. Just a monitor showing something that looked like a video game. Pause at 0:26. Look at the screen. That is not a game. That is his business. Running. Earning. Getting smarter without him. The map on screen is his agent system visualized. HERMES HQ at the top. FORGE LABS in the center. Every building is a department. Every glowing node is a skill his agent taught itself and saved to disk. He has 47 of them now. He started with zero in February. Someone in his Discord asked what the system actually does. He said: competitor research reports. Startups pay $300 each. Same day delivery. The agent handles everything. He reviews the output and sends the invoice. 15 minutes of work per report. $300 per report. The comment thread did the math faster than he did. But here's the part the math misses. Month one: 20 skills saved. Reports take 15 minutes. Month three: 47 skills saved. Reports take 10 minutes. The agent is not just doing the work. It is getting better at the work. Every session it saves what it learned to a folder on his drive. Next session it reads that folder before it starts. It compounds instead of resets. Freelance analysts charge $150 to $300 for the same report. Research firms start at $500. They have overhead, employees, calls, proposals. He has LM Studio. Qwen 3 27B. Hermes. One flag most people miss: context window set to 65,536 tokens or nothing runs correctly. Setup took 30 minutes. He now has 6 retainer clients. $3,300 to $4,400 a month. Total tool cost: $0 to $2 per month. The pixel city on his screen is not decoration. It is a real-time map of every skill his agent has accumulated since February. Each building represents a department. Each glowing node is a saved procedure. The brighter the node, the more times the agent has used it. FORGE LABS glows the brightest. Systems architecture. His most requested service. Six months ago a 19-year-old dropout from Chengdu pushed a local agent memory framework to GitHub. Judges said no commercial application. 4,200 forks later. He had been one of them. He built Hermes on a fork of that repo. Added the skill persistence layer himself. Spent one weekend on the pixel city interface because he wanted to see the brain growing in real time. He posted the clip to show his Discord the map. He forgot the browser tabs were visible. GitHub. Claude. Local server. Credit dashboard. Someone zoomed the tab bar. Matched the URLs. Found the Hermes instance running on his local network. The clip had 180 views when he noticed. He hasn't posted since Thursday. The map is still glowing. FORGE LABS is still the brightest node. The agent is still running. He wanted to show his Discord a cool interface. He accidentally showed them that the most valuable thing on his screen wasn't the pixel city. It was the folder behind it with 47 skills that never stop compounding.