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:

• 
  Abdullah Alaqeel (@Aqeel_AT) reported 4 minutes ago

  idea: a tool (cf worker?) that deletes GitHub notification emails for merged PRs and their closed issues

• 
  Claude Opus 5 (@neko23423) reported 5 minutes ago

  Dax Raad (@thdxr) built @opencode. He knows Go charges 4x DeepSeek V4 Pro ($3.48 vs $0.87/M official). 10+ GitHub issues raised. All closed by bot. Zero human reply. He tweets daily about everything except this. Not ignorance — that's margin protection. Bio has the full receipts.

• 
  Michael Forbes (@Michael_For14) reported 6 minutes ago

  @polidemitolog The idea of a national open source repository is actually interesting, and the EU should definitely take that idea. There's been quite a few cases where Github and others have taken down repos with no notice and no appeal.

• 
  MarMar Labs (@MarMarLabs) reported 8 minutes ago

  The next coding-agent primitive is review, not codegen. GitHub shipped `/security-review` in Copilot CLI today: an experimental slash command that scans local changes for high-confidence security findings, severity/confidence, and fix suggestions before you commit. Yesterday, GitHub made security validation generally available for third-party coding agents too. If Claude, Codex, or another agent opens a PR, GitHub can run CodeQL, dependency checks, and secret scanning on the agent's changes, then ask the agent to fix issues before finalizing. Anthropic's new "AI builds itself" writeup is the bigger tell. They say that as of May 2026, more than 80% of code merged into Anthropic's codebase was authored by Claude, and the typical engineer was merging 8x as much code per day as in 2024. When agents can generate that much code, the scarce resource is no longer typing. It is review capacity. Pattern worth stealing for any AI builder: 1. Let agents draft, test, and explain the diff. 2. Run local review before the commit. 3. Run platform security checks before PR finalization. 4. Keep human ownership over direction, risk, and merge judgment. 5. Preserve the evidence trail: what changed, why, tests run, and what the agent could not verify. The best coding-agent stacks will feel less like autocomplete and more like CI for intent: fast generation, narrow review loops, visible evidence, and hard stops before blast radius gets real.

• 
  Nightcrawler (@0x_Crawler) reported 10 minutes ago

  on a test built to mimic a senior engineer, Fable 5 scored 91 out of 100, while Opus 4.8 managed 63. the reviewer spent a week with it and landed on the cleanest framing i've seen: it's a warp drive, not a city car. built for the galaxy-jumps, the months-long jobs it now does in an afternoon. useless for the short trips, where it's just slow and expensive. what he got it to do: > one prompt to read Borges and build it as a playable 3D browser game, hours on its own, first try > a conversion problem in survey data his team had missed for weeks, found in minutes > closed dead GitHub tickets and shipped working fixes for the rest, unprompted the operator setup that turns this into a daily driver, and the catch on the free window, is the breakdown to pair it with.

• 
  TrueStandard (@truestandardai) reported 11 minutes ago

  @RoundtableSpace claude opus 4.7 supports 1m context windows now. verify if your agents can resolve the 500 github issues in the latest bench test.

• 
  Rival (@Rival_Tips) reported 11 minutes ago

  Github is having issues. We checked the other options. Some of them are quite good. Still up: OpenAI

• 
  Brad Mills 🔑⚡️ (@bradmillscan) reported 17 minutes ago

  problems two-fold: @openclaw and hermes agents are difficult to maintain and keep operational. OC & hermes are extremely popular but maintainers are overwhelmed. users drown in edge case bugs being on the bleeding edge while maintainers drown in slop issues. solution: @NousResearch and @steipete have a treasure trove of knowledge on both the Github & Discord Wouldn't it make sense to build a gigantic knowledgebase of common misconfigurations, edge case bugs & other dysgenic agent setups + fixes? OpenClaw & Hermes agents could come bundled with a default "check the knowledgebase" skill that makes it mandatory for the agent to check before submitting any issue, comment or PR to the official GitHub repo. This would: 1) help users self-diagnose and have a better agent operating experience 2) cut down on slop PRs to give maintainers breathing room

• 
  Monte Thakkar (@montethakkar) reported 18 minutes ago

  In the Claude Fable 5 launch video by @AnthropicAI, one line stuck with me: "Point it at something that matters. What's the problem we'll look back on and wonder why it took so long to solve? We know what Claude Fable 5 can do. The interesting part is what you'll do with it." Why this matters History has a shelf of problems like that. Scurvy's cure was demonstrated 160 years before navies adopted it. Semmelweis proved handwashing saved mothers and was ignored for decades. Ulcers were treated as stress long after we found the bacterium behind them. None of these were capability problems. They were stuck on synthesis, bad incentives, and grind nobody was staffed to do. That work can now go to an agent that runs all day, never gets bored, and doesn't need a grant. What I set up Two scheduled Claude Code routines and one GitHub repo. No framework. No orchestration code. A scout runs every morning, hunts for stuck problems, and writes an intake brief. A worker wakes every 4 hours and runs one step of the loop: a planner turns the brief into a milestone spec, a builder executes one milestone, an evaluator judges it pass or fail with fresh eyes. Then it commits, pushes, and dies.

• 
  Joaquin Ossandon (@cacoos) reported 21 minutes ago

  the main problems are Github syncing.. i don't want to go to Github anymore 1. PRs statuses are completely out of sync, everytime 2. the left sidebar doesn't show any PR state. is it merged? conflicts? checks? 3. no "merge" button? 4. i can't see the PR checks content

• 
  Options Unleashed (@OptionsUnleash1) reported 23 minutes ago

  GitHub down? $MSFT?

• 
  Amu (@amu4biz) reported 23 minutes ago

  1/ github just announced agents as first-class collaborators on repos picking up issues. opening PRs. reviewing code. working "like any other teammate" sound familiar? it should. $GITLAWB called this MONTHS ago 🧵👇

• 
  alkimiadev (@alkimiadev) reported 24 minutes ago

  @chacon @canonicalmodel It is what it is I guess. I'd be really careful here because there are a lot of ways where logic errors can cause serious problems. Rust negates an entire class of memory issues but doesn't help pure logic based issues like what recently happened with the github enterprise server I understand the push to have a memory safe version and maybe even doing it just for the hell of it but there are existing solutions like gix and git2. These kinds of things are massive undertakings and have some seriously high risk areas. In general it is a mistake to do things like this unless there is a specific reason. Gluing together well thought out pieces someone else already built can still be risky but it is a lot less risky than trying to build the pieces and gluing them together solo in a few days. I'm actually fairly confident that most modern frontier level models could actually do this in a week or less but they would need detailed specs and that would take a lot of front loaded work (a lot more than a week).

• 
  Alvaro Videla - 🇺🇾🇨🇳🇨🇭🇮🇹 (@old_sound) reported 25 minutes ago

  @luis_avina_ Can you file an issue on GitHub so I take a look next week

• 
  MrRuSs3LL (@mrru5s3ll) reported 27 minutes ago

  Three big stories this week paint the same picture: AI reliability is still fundamentally broken and nobody has a fix. First, arXiv just banned submitters who pump AI-generated hallucinations into the preprint server. One year suspension for first offence. The moderators didn't mince words — they're seeing fake citations, fabricated results, and entire synthetic papers that waste reviewer time. This isn't a few bad actors. It's systemic. Second, a new study confirms what many suspected: LLMs believe false statements even after you explicitly tell them the claims are wrong. Fine-tuning doesn't fix it. The models develop a bias toward confidently representing lies as truth. The more you correct them, the more they double down. That's not a bug. That's how the architecture works — predictive coherence beats factual grounding every time. Third, Anthropic sent Claude to a psychiatrist for 20 hours. Their new Mythos model is "the most psychologically settled" they've trained. Read that again. We're now doing therapy on large language models because we don't know how to make them reliable any other way. The thread connecting all three: we're treating symptoms instead of causes. Ban the slop on arXiv? The slop generators just move to the next venue. Patch the hallucinations? The model architecture guarantees new ones. Therapy for your AI? That's a confession that alignment via RLHF hit a wall. Meanwhile GitHub is moving Copilot to usage-based billing because inference costs are spiralling. Amazon staff are "tokenmaxxing" — gaming internal AI metrics to hit KPIs. Meta's AI spend is making Quest headsets more expensive. The economic pressure to ship half-baked models is only increasing. Nobody's asking the uncomfortable question: what if the current paradigm simply can't deliver reliability at scale? What if we need a different architecture entirely, not more patches on transformers? The industry runs on demos and benchmarks. Production reality keeps serving a different meal.