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:

• 
  Muhammed (@_ceejeey) reported 5 minutes ago

  I’m currently working on four products: - A design-to-code app builder - A native OS app - A marketplace - An event booking platform All four are very different, but they have helped me understand how I actually build products with AI today. For the design-to-code product, the goal is not to generate a few screens from Figma and call it done. It should understand the design, product context, architecture and business logic, then create a working React Native or Next.js project, connect the flows, validate the output and eventually push a usable codebase to GitHub. Across all four products, AI writes most of the code. But letting AI write code and letting AI build the product are still two very different things. Here is how I divide the work. What I still decide • Product architecture • Tech stack • Database structure • State management • API boundaries • Security decisions • What the correct implementation should look like What AI mostly handles • Feature implementation • Repetitive components • Tests • Documentation • Initial debugging • Refactoring once the direction is clear What usually needs both • Planning • Code reviews • Performance work • Visual verification • Larger refactors The part I never fully hand over is the actual product experience. AI can build something that looks correct in a screenshot but still feels wrong when you use it. It often misses loading behaviour, navigation flow, persisted state, empty states, error recovery and all the small details between two screens. Yesterday, I asked Codex to build an onboarding flow with Redux persistence. Instead of keeping the splash screen visible until persistence was initialized, it created a manual persistence gate to avoid the flicker. Did it work? Yes. Was it the correct solution? No. That is the difference between code that works and code that actually belongs in the product. A few things I have learned while building these products: 1. Architecture matters more when AI is involved If the codebase is modular, multiple agents can work in parallel without constantly touching the same files. If everything is tightly coupled, adding more agents only creates more conflicts. 2. Subagents only help when the task boundaries are clear One agent can plan, another can build the API, another can implement the UI and another can test the output. But this only works when the codebase is structured for parallel work. 3. The same model should not be used for every task I use stronger reasoning models for planning, architecture and difficult debugging. Faster models are often good enough for implementation once the task is clearly defined. Using the most expensive model for everything is not better engineering. It is just expensive. 4. Context matters more than the prompt The agent needs to understand how the project is structured, which patterns already exist, what commands it can run, which libraries it should use, what it should never change and how the work should be validated. Without that context, even a capable model will start inventing its own architecture. AI can now write most of the code. But someone still needs to understand the product deeply enough to know whether that code is actually right. That is becoming a much bigger part of engineering.

• 
  Archon (@Yamik1shi) reported 6 minutes ago

  You are using Claude wrong, and it is quietly bleeding your API budget Most builders think generating massive amounts of code is the goal It just hit 81,000 stars and is #1 on GitHub today More lines mean more bugs, higher token costs, and impossible maintenance It is a GitHub repo that forces Claude into strict minimalism Ponytail is the fix It injects one hard rule: do not do extra Claude still thinks deeply about the architecture But it becomes aggressively lazy about the execution You control the intensity: `/ponytail lite|full|ultra|off` Run `/ponytail-audit` to strip accumulated bloat from an existing project Run `/ponytail-review` to clean up live edits on the fly It does not just work in Claude Code It runs perfectly in Cursor, Copilot, Codex, Gemini, and Antigravity Free to install. MIT license The leverage is no longer writing the most code It is generating the least Look up the Ponytail repo and stop paying for bloat

• 
  Guribo (@GuriboVR) reported 11 minutes ago

  If you run into any issues please create bug reports on Github, also check if your case is covered by the wiki first though. With such a big release it is still possible that there is some edge cases I missed. I will iterate and improve this further in the future.

• 
  Warizo (@Warizo_ofAfrica) reported 13 minutes ago

  @github Moving away from monolith models to a smart subagent delegation architecture is the real future of terminal agents. In the CLI, tool and search failures completely break engineering momentum, so cutting those errors by over 20% is a massive workflow win.

• 
  Gokul Rajaram (@gokulr) reported 14 minutes ago

  @aar0ncpa Right now ProductSpec MCP is intentionally local and stdio-based. So authentication is handled by the MCP client/runtime, not by ProductSpec itself. The server runs on the user’s machine, reads local .product-spec.md files, and exposes deterministic tools like validate_product_spec, get_scope, get_acceptance_criteria, and check_spec_session. There is no hosted ProductSpec MCP endpoint today, so there is no API key or OAuth surface in the open source MCP server. The boundary is: -- Local repo access: governed by the user’s filesystem and MCP client config -- GitHub access: handled separately by GitHub auth in tools that need GitHub -- ProductSpec MCP: reads local specs and returns structured intent If we add a hosted/team MCP server later, I’d expect auth to be org-scoped OAuth or short-lived tokens, with repo-level permissions and audit logs. But the current open source version keeps the trust model simple: local process, local files, no ProductSpec cloud auth.

• 
  loooong (@allyiiii) reported 15 minutes ago

  Everyone thinks AI should help mathematicians prove theorems, but Terence Tao had it migrate his 30-year-old old website. In a single day, the AI moved 560 papers, travel logs, courses, books and math applets to GitHub Pages, and found two hidden bugs in Tao’s decades-old handwritten code. Launched back in 1997, the site required manual HTML edits via a terminal for nearly 30 years. The AI also cleaned up inconsistent info, stale entries & broken links, plus ported old Java 1.0 applets to JavaScript. Rather than tackling big math proofs, AI handled the tedious digital housekeeping mathematicians dread.

• 
  Julian Goldie SEO (@JulianGoldieSEO) reported 15 minutes ago

  Your abandoned GitHub projects just came back to life. Google AI Studio can finally import them. For months it was a one-way door. Projects could leave AI Studio. They could never come back. So old code just sat there. Frozen. Now one button changes it: → Import from GitHub pulls in the whole repo. Front end, back end, everything → Say "add a contact form" in plain English and Gemini edits the real code → Build in Cursor or Claude Code, polish in AI Studio, push back out → A teammate quits? Anyone can pick up their repo and keep going Google is already working on two-way sync. Then it stops being a sandbox and joins your daily workflow. Here's your move today: Find one old repo you gave up on. Import it. Ask Gemini what it would fix. That excuse of "I'd have to rebuild it" just died. Want the SOP? DM me. 💬

• 
  Void Freud (@voidfreud) reported 16 minutes ago

  I dislike OpenAI’s leadership. But I have to give it to them: the products evolved. I was just charged another $200 for Claude 20x Max. I am the least person to support OpenAI but my experience with Claude was so frustrating as of late, that I gave ChatGPT a go and wow, it felt so much more alive and friendlier than Claude. Talking to Claude has been like talking to a bank clerk trying to sell you a loan. It’s boring. It’s repetitive, cliche, censored and poisoned with disclaimers and guardrails. I just stopped enjoying reading its responses: they are dull and fake, and largely incorrect since quality has been super-degraded on subscriptions. I skim through them, cause they turned from funny, wit and kind to bloated and lifeless nonsense. No amount of tweaks, output-styles or rules or system prompts works: Claude ignores them; it also ignores other functional instructions in Claude Code and @ClaudeDevs do absolutely nothing about that, despite tons of issues raised on GitHub and X, continuing to nerf the quality and charge users further: - the performance of Opuses is super degraded - the performance of Fable is super degraded - the speed of any models is super slow - the quality of the apps sucks (compare Claude iOS app to that of Cursor or ChatGPT, it’s lame and buggy: from exceptionally lame remote control, to dysfunctional transcription that makes Apple’s native dictation feel like a win) - inconsistent and unfair subscription / quota terms - zero accountability, transparency or feedback - toxic paternalism culture, now deeply embedded in models - dismissive and greedy attitude from Anthropic

• 
  Beamo Supremo (@BeamoSupremo) reported 17 minutes ago

  CMake can go **** itself. So many GitHub repos rely on this bullshit build system, its unbelievalbe, half of it is broken in some way, or can't find proper directories and libraries, sometimes even when directly specified by filename, and it is pretty much arcane magic that takes half a day to set up correctly.

• 
  Mr.Jack 🐬TermMax (@0xHypeETH) reported 19 minutes ago

  @moha_web3 @github This design reduces cognitive load by surfacing structured metadata inline, which should minimize context-switching when triaging issues.

• 
  Pierce Boggan (@pierceboggan) reported 19 minutes ago

  @_fraz_ Working on it ASAP, looks like an upstream GitHub Copilot SDK update broke us and working on getting a fix out

• 
  Ask GPTs (@askgpts) reported 22 minutes ago

  A tool called Graphify just hit number 6 on GitHub trending and it solves one of the most frustrating problems in AI-assisted coding. When you ask an AI coding agent to find something in a large codebase, it guesses. It looks at what's in the context window and hopes it found the right files. Graphify maps your entire project into a knowledge graph first. Then you query the graph instead of guessing. Here is what it does: 1. Type /graphify in any coding agent and it processes your entire project 2. Maps code, documentation, PDFs, images, and videos into one connected graph 3. Supports 36 programming languages via tree-sitter, processed entirely locally 4. Works with Claude Code, Codex, Cursor, Gemini CLI, OpenCode, and 20+ other agents 5. Ask precise questions: "What connects auth to the database?" "Find the path between UserService and DatabasePool" 6. Auto-rebuilds on every commit via *** hook so the graph stays current 7. Neo4j and FalkorDB integration for sharing across your team via MCP server 8. PR dashboard showing graph impact before you merge Code processing happens locally. Nothing leaves your machine. No API key required for code. 100% open source. 2.8 million downloads. YC S26

• 
  🇺🇸 Santore (@santoretech) reported 25 minutes ago

  Your company doesn't need another AI tool. It needs an operating system and you already have one.. it's @github. Tasks? Issues and Projects. Related to code? What isn't, in 2026. Strategy, playbooks, decisions. If it isn't versioned, your agents work from stale context. Skills and prompts? Same place. Writing voice, review checklists, compliance guardrails. Stored, updated, shared. Improve a prompt once, everyone gets it. Approvals? Built in. Define who reviews what before anything ships. Sharing? Invite someone to the repo. One source of truth, not twelve tools with twelve versions. Humans and agents, same playbook. The company brain isn't a metaphor. It's our operating model and @blockskunk is the lab. One repo at a time.

• 
  Doug Burks (@dougburks) reported 27 minutes ago

  @greyhathackr I will work on a demo video but please note that you don't really have to create an account for the Killercoda demo. They allow you to sign in via Github, Gitlab, or Google. Or just give it an email address, it sends you an email, and you click the link in the email.

• 
  Md Irfan Hasan Fahim (@mihf05) reported 27 minutes ago

  @GitHubHelp @github My account got restricted during Edu reverification for "missing 2FA", but I've ALWAYS had Google Authenticator enabled! It's a system glitch. Please check Ticket #4557612, my dev work is completely blocked. 🙏