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:

• 
  Rulya (@Rulyaxd) reported 3 minutes ago

  3 hours with Claude. Two months later, servers he'd never visited had his bot installed. Zero coding classes. The article beside this is the exact four prompts he used. The story: a small study-group Discord server, one recurring annoyance (new members asking the same five questions in the wrong channels), Claude open in another tab. Three hours later he had a bot answering those questions. Sixty days after that, dozens of servers he'd never been in were installing it, because someone had shared it in a room he wasn't a member of. The build was four prompts. Prompt 1: describe the annoyance clearly enough that Claude writes the whole bot in a standard framework, with plain-English explanations of every section. Prompt 2: paste the exact error message when it breaks, ask Claude to explain what it means, what caused it, and the exact fix while explaining what changed so you understand it instead of copy-pasting. Prompt 3: turn the working script into a real product name, short description, add-to-server instructions. Prompt 4: after the first outside install, edge cases start showing up. Same loop. Paste the error, get the fix, ship it more stable than before. The growth curve is boring in a good way. First install: your own server. Second stage: someone shares it in a room you don't belong to. That's when the interesting bugs appear. From there it compounds handful → dozens → hundreds → thousands, once bot-listing directories start indexing it. The monetization shape follows. Free is what spreads. Premium at $19/month unlocks advanced automation for the server owners already depending on it. At 100 servers with a 50% paid conversion on Premium alone, that's $950 a month from a project that started as a study-group irritation. The illustrative math in the article 10 servers/4 paying/$76, 25/10/$190, 50/20/$380, 100/50/$950 isn't a promise. It's the shape of a low conversion rate compounding on top of a free version that already spread on its own. The article's kicker lands harder than the numbers do. The hardest part isn't building the bot anymore. It's believing you're allowed to. The kid in the video already believes. The article beside it hands you his four prompts. Most people will read this and think it's about Discord. A small number will notice that the same loop works for Slack, Telegram, Notion, Chrome extensions, and internal GitHub apps. He wrote no code. What he learned was how to describe a problem clearly enough that the software would build itself.

• 
  CelesTrak by Dr. T.S. Kelso (@CelesTrak) reported 7 minutes ago

  @Math_MntnrHZ Question 1. satellite.js was not written by CelesTrak. However it was heavily used, branched, and cloned throughout the GitHub community. We used it for a visualization and found some SGP4 errors in the code. We corrected them & created a verified version for the community.

• 
  thedeadrobot (@thedeadrobot) reported 7 minutes ago

  1/6 14,000 github repositories without clear ownership. 45 days to fix. that's a scalable solution to a fundamental problem. but who benefits from a validated owner? devs, enterprise, or github's audit trail?

• 
  Distractosphere (@Distractosphere) reported 9 minutes ago

  @thsottiaux on chatgpt there is a github connection issue. in chatgpt interface can not read private repos with active github connection.

• 
  OneHuman (@OneHumanIO) reported 14 minutes ago

  5/6 One more thing: reports of the standalone #Codex app vanishing without warning trace to a GitHub issue showing the new and old apps share the same bundle ID — the OS treats the update as a silent replacement, no prompt.

• 
  Nirmātṛ (@En_formare) reported 14 minutes ago

  LaTeX’s ***** little secret just spilled millions of researchers’ private laundry onto arXiv’s public stage. Open science exposes its Achilles’ heel when sloppy habits meet permanent public archives. The very tool accelerating discovery • arXiv’s 1991 revolution in physics, math, and beyond • now risks reputations, security, and ethics in one sloppy compile. 88% of LaTeX-submitted preprints on arXiv – nearly three million papers, covering 93% of the repository up to late 2025 – harbor hidden information never meant for prying eyes. Passwords, GPS coordinates pinpointing homes, API keys, profane co-author smack talk, to-do lists confessing paper weaknesses, Google Docs links exposing peer reviews and participant surveys: all baked into source files, dangling attachments, and metadata like forgotten cookies in a browser history. LaTeX’s comment-friendly markup is treated by authors as private scratch paper while obsessing over the pristine PDF output. arXiv mandates uploading those source files, but most researchers remain unaware; only 41% of surveyed affected authors knew the repo publishes them. The ultimate mental model mismatch: GitHub devs expect exposure; academics drafting in comments do not. One team’s snarky “WTF does this mean?” about a rival’s work, another’s home address leaking via geotagged images (7,326 submissions flagged, hundreds tying lab to living room), or 699 editable Google Docs spilling rebuttals and raw data. Earlier scans caught social security numbers and cloud links; this deep dive calls it the tip of a 12-million-file iceberg across preprint versions. The fix is simple yet revolutionary: strip comments (% lines), purge metadata from images and PDFs, delete dangling files, and use tools like latexclean or arXiv’s own guidance before upload. Future-proof with scripts or pre-submission audits; don’t let a casual “TODO: fix this embarrassing gap” torpedo your next big claim.

• 
  Vladic (@Vladic_ETH) reported 16 minutes ago

  OPENAI SHIPPED GPT-5.6 AND CHATGPT WORK. THE REAL WEAPON IS PRICE, NOT IQ. OpenAI shipped two things today. One of them is a costume change. GPT-5.6 landed as three models. ChatGPT Work is a new agent on top. The feeds say "new agent does your work." The real launch is the price sheet. Sol, the flagship, costs $5 per million input tokens and $30 output. That's not flagship pricing. That's what you paid for a mid-tier model a year ago. The gate half the feeds skipped Context first. Two weeks ago the US government cut GPT-5.6 access down to a small group of vetted partners over national security. The gate held about 12 days. Restrictions lifted July 8, public release July 9. Same day SpaceXAI shipped Grok 4.5. The frontier now ships when the government clears it, not when the model is ready. Anthropic went through the exact same thing with Fable and Mythos in June. A pattern, not a one-off. Three models, price as the weapon GPT-5.6 is three models, not one. Sol is the flagship. Terra is the everyday workhorse. Luna is cheap and fast. Price per million tokens, in/out: Sol $5/$30, Terra $2.50/$15, Luna $1/$6. Terra matches GPT-5.5 quality at half the cost. Luna is the cheapest entry in the line. Altman told CNBC Sol is 54% more token-efficient on agentic coding. That's the message. Not "smarter." "Cheaper for the same result." And ultra: a mode inside Sol that spins up multiple agents in parallel and hands subtasks to submodels. The market counts token bills, not benchmarks. Enterprise thinks spend first now. OpenAI heard it and made price the argument. Today's real launch is unit economics, not intelligence. "Sol beats Fable 5, Luna beats Opus 4.8 at two-thirds the cost" are OpenAI's own benchmarks. Until independent runs, treat them as marketing. ChatGPT Work is Codex in a suit Now the "new agent." ChatGPT Work runs on Codex and GPT-5.6. It moves across your apps and files, stays on a project for hours, breaks it into steps, finishes on its own. Output: docs, sheets, slides, web apps. Inside sits a Unified Plugins Directory: Google Drive, Slack, Teams, Gmail, Outlook, Salesforce, GitHub, Canva, Dropbox, more. Call one with "@" or let the agent pick the source. Sounds familiar. This is OpenAI's second run at plugins. The first was 2023 and it flopped. Brockman admitted the models weren't ready back then. Honest read: hard to tell what's actually new. Scheduled Tasks, Computer Use, connectors already lived in ChatGPT and Codex. Long tasks and data sources worked before too. The real move isn't features. It's consolidation: on desktop, OpenAI is merging Codex and ChatGPT into one super app and putting Codex in front of people who don't code. The Anthropic mirror Here's the tell. This is the exact play Anthropic ran with Claude Code -> Cowork. Take a dev agent, strip the "for coders" label, hand it to knowledge workers. Cowork just hit web and mobile, timed to get ahead of this. Two labs, one bet: whoever owns the desktop app that touches your files and apps owns the knowledge-work layer. Chat is the storefront. The desktop is the land grab. What a practitioner does with it One: rebuild pipelines around price tiers. Route bulk work to Luna and Terra. Keep Sol and ultra for the 10% that needs the ceiling. Economics is a routing problem now, not a single-model choice. Two: the real unlock is the desktop with local file access, not the web. Free tier gets ChatGPT Work on desktop right away. Web and mobile roll by tier: Pro, Enterprise, Edu first, Plus and Business next. Three: billing is usage-based and shares one pool with Codex, ChatGPT for Excel, and Workspace Agents. Count tokens before, not after. A complex task burns quota quietly. Security: OpenAI touts Auto-Review, where senior models check important actions before they run, and claims it blocked 100% of protected-data extraction attempts in red-teaming. 100% in a lab is zero confirmations in ****. Test it yourself. Sober read The model war moved from IQ to unit economics. The product war moved from chat to the desktop that holds your files. Testers are already posting "best model I've touched." Maybe. That's day-one sentiment, not fact. The real scoreboard isn't a benchmark. It's the "AI spend" line in an enterprise budget. That's a market you can actually read. The window is the next couple weeks, before prices settle and everyone re-routes spend. Rebuild your routing around three models now and you enter the quarter with a smaller bill for the same work. Everyone else reads the thread and changes nothing.

• 
  Shadow Nick (@doublenickk) reported 22 minutes ago

  87% OF THE PLANET SUCKS AT AI BECAUSE THEY ARE STILL TYPING MANUAL PROMPTS LIKE AMATEURS While the masses use ChatGPT as a glorified search engine, elite builders are deploying autonomous digital armies that execute high-stakes business operations 24/7. Meet Synapse, an open-source MCP engine that hands AI complete vision and surgical command over your desktop to run background tasks silently while you sleep. The exact strategy used to break the system: The FBI Negotiation Hack: Scrape a massive list of multi-million dollar startups, feed real FBI hostage negotiation transcripts into the AI, and let the agent autonomously blast out high-leverage B2B outreach that forces prospects to say yes. Zero-Drift Execution: Ditch chaotic markdown files and manage your agent's state through GitHub Issues to keep them locked in for weeks without a single hallucination. Full-State Reality Testing: Stop relying on worthless pre-compile unit tests because this agent forces your system to compile, screenshots the actual interface, and verifies performance against reality itself. You can keep playing around with basic chatbots, or you can deploy a ruthless autonomous agent to scale your code and outreach on autopilot.

• 
  Shweta ♡ (@Shay_Slay_) reported 23 minutes ago

  Watching your Claude Code bill climb for a repo this small does something to a person I genuinely thought the billing was broken turns out my agent was quietly doing the ONE thing nobody warns you about the thing that silently drains your entire token budget and i had no clue until i installed Repowise it indexes your repo once so the agent stops re-reading the same files forever loading context for a commit went 64k tokens → 2.3k that's 27x fewer 70% fewer tool calls plus it scores every file for bug risk in under 30s, no LLM, fully local pip install repowise and see your own before/after ♡ completely open source. github link in comments

• 
  Max Blade (@_MaxBlade) reported 27 minutes ago

  The truth about 5.6 sol after using it all day : The hype is overblown. Sort of. The benchmarks, and the commentary on X convinced me we were receiving AGI that runs at hyper speed, and is insanely cheap. in reality, 5.6 is built on the same spud pretraining as 5.5 this means its a nice bump, but not the opus to fable 5 LEAP in intelligence we recently experienced from anthropic. 5.6 is 2x times cheaper than fable on paper, and actually 3x cheaper when you look at actual task execution because of its token efficiency. BUT on swe bench where the models have to fix actual github bugs it falls behind fable pretty big. For vibecoders like myself this means I will be using 5.6 sol as a worker agent for Fable 5 to orchestrate alongside grok 4.5 I love this new era.

• 
  alias (@loadingalias) reported 28 minutes ago

  @_dylanga Yeah, GitHub has a serious issue with provenance and authenticity. Aside from the *** graph being kind of useless… repo stars are totally ******, too. It seems like an important problem, but GitHub is on autopilot or something. IDK. They move with zero intention nowadays.

• 
  Harry Tandy (@HarryTandy) reported 29 minutes ago

  A World Cup prediction game sounds like a gimmick until you see the payments layer Sam Witteveen breaks down Google's Agent Payments Protocol in 10 minutes: 0:00 - Agents, MCP, and A2A context 1:00 - Agent Payments Protocol 1:41 - AP2 use cases 6:42 - Core principles: openness, user control, accountability, verifiable intent 8:16 - Google Agent Store 8:49 - AP2 GitHub Then open the Cyber Cup piece with one question in mind: what happens when leaderboard points depend on a tiny agent org? > one agent pulls match data > one agent checks odds and injuries > one agent decides the bet > another agent gets hired when the first 3 can't answer That is why a football tournament is such a clean test The scoreboard is public, the feedback loop is daily, and bad agents can't hide behind a polished demo

• 
  Akshay Nandwana (@akshay81844) reported 29 minutes ago

  2/10 Unlike traditional benchmarks, this one uses 105 real GitHub fixes from production Kotlin repositories. Models receive an issue description and repository state, then must generate a patch that passes hidden regression tests That's much closer to real software engineering.

• 
  the_architectopteryx (@rchitectopteryx) reported 30 minutes ago

  @PrasVector @ajambrosino Yeah I have an iPad that can’t connect to codex remote and as OpenAI doesn’t have real support or read GitHub issues, its so frustrating.

• 
  Clifford Richardson (@CorvusCrypto) reported 34 minutes ago

  @kellabyte If you're a leader you need to be better about handling emotion. It really drags down my confidence in Andrew being the person toead zig to 1.0. I get your point about empathy but it's a pattern with Andrew (this, GitHub, etc.) He becomes vitriolic at the slightest challenge.