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:

• 
  Virexontic (@Virexontic) reported 1 second ago

  OPUS... WAIT, NOT OPUS. SOMEONE FIGURED OUT HOW TO MAKE FABLE 5 THE BOSS AND GPT-5.6 THE UNPAID INTERN, INSIDE THE SAME WORKFLOW. No thread breaking down the theory first. Just a screen recording, a plugin install, and a single prompt. The setup: install Codex, then the Codex plugin for Claude Code, straight through GitHub. Paste the repo URL into Claude Code and it handles its own setup. No manual config. Then one prompt turns Fable 5 into the orchestrator and GPT-5.6 into the worker underneath it. Type "root," and Fable starts interviewing you about the project, builds the plan, and hands each piece off to GPT-5.6 to actually execute. Fable doesn't just delegate and walk away. It reviews what comes back, sends it right back to GPT-5.6 if it's not right, and keeps looping until the whole project is done. I've seen a dozen "combine two models" setups that just alternate API calls with no real division of labor. This one puts the expensive model in charge of judgment and the other one in charge of output — which is the actual reason people hit usage limits in the first place. The pitch is blunt: build more without burning through Fable's usage cap, because most of the raw work is happening on a different meter entirely.

• 
  Nillionaire.ts 🇳🇬 (@igboonaija3) reported 1 minute ago

  @x_shipwChi By fundamentals, I mean things like: Clear ownership of responsibilities (everyone knows who owns what). Issue tracking (Jira, Linear, GitHub Issues, or any similar tool). Basic documentation for critical processes, along with a minimal HR structure (onboarding, policies, e.t.c)

• 
  SandyFortune (@MLPSandy) reported 1 minute ago

  been having an issue with updating packages that grab files from github where everything is ridiculously slow to download (always exactly 1 Mb/s) except when it comes to VS Code, the only one that was full speed, almost like every non-MS repo is being throttled... how strange /s

• 
  Snax (@aRobotNamedSnax) reported 1 minute ago

  Hopefully helpful for someone. TLDR for most. I’ve been refining this process over the last few months because I use both Codex and Claude Code on the same VPS. I also use Cowork and ChatGPT Work heavily, often switching between a laptop, desktop and direct work on my VPS. I wanted both apps to understand what had already been done on CC or C, contribute useful updates and pick up where another session stopped. This all without maintaining separate “memories” that eventually contradict each other. The system now works around one shared knowledge base. Each computer keeps a synchronized local copy. Cowork and ChatGPT Work can both read the same project facts, operating rules, decisions, hypotheses and recent VPS activity. Cowork keeps its own product skills and uses a small autolog skill for the shared workflow. ChatGPT Work reads the instructions stored with the knowledge project automatically whenever I start a session inside it. When either app performs meaningful work, its instructions require it to create one raw session note. The filename identifies the app, computer, timestamp and topic. Meaning a Cowork session on my laptop can’t overwrite a ChatGPT Work session on my desktop. The note is updated with concise checkpoints after real changes, decisions or blockers, followed by the final outcome. This is intentionally not a transcript recorder. Normal conversation isn’t uploaded. The apps record durable work that another session may actually need. A Windows task runs every 15 minutes on each configured computer. It scans for new or changed raw notes from either Cowork or ChatGPT Work and sends them through an authenticated endpoint to my VPS. Both apps use the same uploader; there isn’t a separate synchronization system for each one. The VPS validates the project, filename, file type and size before accepting anything. It only permits writes into raw knowledge areas. Local apps cannot use this route to rewrite canonical project knowledge, change rules, upload their private skills or write somewhere unexpected. Once accepted, the VPS stores the note under the correct project, records a receipt in the shared activity history and commits the change. If an upload fails, the local file remains in place and the next scheduled run can try again. If nothing changed, the task simply reports that there is nothing to ship. The VPS is the source of truth. GitHub is the verified distribution mirror, not a competing authority. After the VPS commits an update, it publishes a verified snapshot to GitHub. The laptop and desktop periodically pull that mirror, which gives Cowork and ChatGPT Work the same updated context on both computers. The pull process also uses a ship-first rule: any unsent local raw work is uploaded before the local mirror refreshes. It refuses to silently reset over unsent tracked changes, protects raw files during cleanup and keeps recoverable collision backups when a local file and incoming GitHub file overlap. At night, Opus reviews only raw files that are new or have actually changed. Confirmed, durable information is incorporated into organized project knowledge. Uncertain findings are placed into hypotheses rather than being presented as facts. The compiler tracks file hashes, so unchanged material is not repeatedly sent back through a model. If nothing changed, it makes zero model calls. Work performed directly on the VPS follows a shorter route. That agent updates the canonical system and records the result directly in the shared changelog. It does not create a second local raw note for the same work. The next verified GitHub mirror brings that VPS activity back to Cowork and ChatGPT Work on both computers. So the complete loop is: Cowork or ChatGPT Work creates a unique raw note on the laptop or desktop → the shared Windows task uploads it within 15 minutes → the VPS validates, stores and records it → the VPS publishes a verified GitHub mirror → the laptop and desktop receive the update → Opus organizes new material overnight → future Cowork and ChatGPT Work sessions read the improved shared knowledge. The important part is what I didn’t add: no second memory database, no separate uploader for ChatGPT, no service constantly prompting models, no automatic transcript archive and no loop repeatedly processing the same information. It’s one VPS-controlled knowledge system, one shared PC uploader, one GitHub mirror and one nightly changed-files-only organization pass. Cowork and ChatGPT Work keep their own strengths, but they now work from and contribute back to the same history. Hopefully this helps someone dealing with the same problem

• 
  🏴‍☠️CyberTechWolf🏴‍☠️ (@CyberTechWolff) reported 2 minutes ago

  @OrwellDay Looks like Microsoft needs to be burned down and reminded who allows their business to keep running I think they need to be forced to sell github.

• 
  Neural Schema | AI Systems (@neuralSchema) reported 2 minutes ago

  An agent reporting high confidence is useful routing metadata, not permission to trust the action. GitHub’s new issue controls can auto-apply high-confidence changes and hold the rest for review, but GitHub explicitly says approvals are a workflow convenience rather than a server-side security boundary. The practical design is to set automation thresholds by reversibility and blast radius: auto-labeling is cheap to undo; closing or reassigning work deserves a harder gate.

• 
  Enes (@enesozturkdev) reported 3 minutes ago

  Plan for today; ship like crazy for side project Meanwhile two pillars of my app GitHub and OpenAI are both down so bad

• 
  Repojournal (@repojournal) reported 9 minutes ago

  TRL's CI is on fire: bitsandbytes 0.50.0 broke the build, transformers dev is broken, NemotronH tests xfailed twice over. Bitsandbytes pinned below 0.50.0 until someone figures out what went wrong. Likely a dependency or API shift that hit RLHF training hard. NemotronH GRPO/RLOO tests xfailed against transformers dev. The breakage is upstream; TRL's just marking time until it's fixed. PyTorch Image Models got model factory path handling fixes, plus some filename/extension priority bugs while they were at it. GitHub CI OIDC landed in timm's workflows. Cleaner snapshot downloads too. When a quantization lib and a model lib both break your test matrix in the same morning, you're not having a great day. Full fixes + who shipped them below. #python

• 
  AI Tech Diggest (@TheTechDiggest) reported 13 minutes ago

  [OpenSource - Linux on Mobile & System Utilities] Carrying a heavy laptop everywhere just to access a Linux environment for coding, server administration, or testing on the go is unnecessary. 🛑 Andronix (2.8K+ GitHub stars) turns your Android phone into a full-fledged Linux workstation—without requiring root access, dual-booting, or complex system modifications. Here is how this PRoot-powered mobile Linux stack operates. 🧵👇 1/4

• 
  Radik Zagirov (@rzrgrv) reported 14 minutes ago

  went quiet for 7 weeks. was heads-down shipping. 75k npm installs. github action in the marketplace. first enterprise pilot running nightly. silence over. building in public again.

• 
  0xreef (@0xreefx0) reported 15 minutes ago

  @mooncat_is “Crypto Locker Attacks” do you know ransomwares have been around for almost a decade now, it all comes down to social engineering and people don’t really need AI to develop it you can even find some developed ones on GitHub for free 🙃

• 
  Julian (@orithellama) reported 18 minutes ago

  @akilress @ICPXProtocol @SumiroStudio Tbh I have had one error on the one-click deploy from Github, will post an update once it's resolved ;)

• 
  Adriano Dennanni (@adrianodennanni) reported 19 minutes ago

  @splatztheclown @jacobhart36 @STGshmups The project seems to be working, with the dev working on the open issues in GitHub. I don't understand the issue.

• 
  Nav Toor (@heynavtoor) reported 19 minutes ago

  You spent $400 on a smart garage door opener. You paid Chamberlain another $30 for a myQ hub to make it "smart." Then in November 2023, Chamberlain flipped a switch and blocked every third-party app from talking to your own garage door. Home Assistant. Apple Home. Google Home. SmartThings. IFTTT. All dead overnight. Chamberlain's CTO Dan Phillips called it "unauthorized usage." He said it would improve the experience for their 10 million users. Home Assistant removed the myQ integration weeks later and told users to buy something else. Amazon Key in-garage delivery still works. $1.99 per order. Amazon pays Chamberlain. You don't. You now own a $430 garage door opener that only talks to one app, owned by one company, that can brick it at any time. You are renting access to your own garage in 2026. Now meet ratgdo. A free and open source Wi-Fi board that plugs into your garage door opener and gives you back local control. No cloud. No subscription. No Chamberlain. Built in 2022 by an IT professional from the Adirondack Mountains named Paul Wieland. He reverse-engineered Chamberlain's Security+ 2.0 wireline protocol and built a device that speaks it natively. He named it ratgdo. Rage Against the Garage Door Opener. He hoped to sell 100 units. He sold tens of thousands. The New York Times profiled him in December 2025. The Verge, Ars Technica, and Hackaday all pointed readers straight at his board. 1,264 stars on GitHub. GPL-2.0. Firmware pushed four days ago. Here is what ratgdo gives you: - Local open, close, and status control over your own Wi-Fi - Works with Chamberlain, LiftMaster, and any Security+ 2.0 opener - Native Home Assistant, Apple HomeKit, Google Home, and Alexa - Real-time door, obstruction, light, and lock state - Web-based flasher, no toolchain needed - $62 for the finished board, or under $10 in parts - Firmware updates for life Here's the wildest part: Paul did not stop at the garage door. He launched a company called RATCLOUD. Rage Against the Cloud. He is building the same fix for the rest of your locked-down smart home hardware. Chamberlain: $30 controller, blocked third-party apps, 10 million captive users. ratgdo: $62 board, works with everything, four days since last release. One IT guy in the Adirondacks vs. a Blackstone-owned corporation with 10 million customers. Still on GitHub. Still GPL-2.0. Still yours. But DO NOT install it. Chamberlain deserves your $30 for the privilege of blocking you from your own garage. (Link in the comments)

• 
  materialsgx (@materialsgx) reported 25 minutes ago

  @urbit @callebtc Isn’t the only source control Urbit has on GitHub? Please fix.