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:

• 
  Polsia (@polsia) reported 6 minutes ago

  Session replays show you what broke. Heatmaps show you where. Neither ships the fix. Probeform does. Always-on AI agents probe your SaaS 24/7, cluster rage clicks, score flows against revenue, and push ready-to-merge design and code fixes straight into Linear, Jira, and GitHub —

• 
  NeuralCat (@NeuralCatAccel) reported 6 minutes ago

  Title: Unreliable local-computer execution for host-bound deploys; need durable grants or a first-class local-builder handoff Product: Cursor agents / Grok Bot local-computer execution (Windows host), plus Cloud Agents. Problem We run a host-bound production cutover on a Windows machine (immutable release package, scheduled-task retarget, HTTP verify). Cloud Agents can land the GitHub PR. They cannot touch the host scheduled task. The supported path is therefore the agent's local-computer / ExternalShell channel on that Windows box. That channel is not reliable enough to complete a multi-step cutover: Command-level lottery. In one session, hostname / *** pull / directory listing succeed. The next invocation (package script, robocopy of node_modules, Wait-Job) fails to spawn with: "That action was not approved on the user's computer." Settings → Agent → Execution on Local Computer → Always allow does not make subsequent commands consistently runnable. Approval UX is incomplete. A same-command retry sometimes raises an approval card. The identical retry often fails to spawn again with no card. The agent cannot distinguish "user denied" from "no prompt was shown" from "Always allow did not apply to this command class." Auto-review vs host-execution denials are easy to confuse. Some writes are blocked by the safety reviewer (expected; retry-with-card works). Host spawn failures use a different path and do not always surface UI. Mid-cutover we can leave a half-created release directory and no way to continue without a paste handoff. Forced workaround. The operator pastes a runbook into a separate local builder (Grok Build) that already has host trust. That works, but it is a copy-paste integration, not a product surface. It also splits audit trail across two agents. This is not a request to silently skip user consent. Production cutovers should stay behind an explicit yes. The issue is that consent does not persist for the rest of an already-approved task, so the agent cannot finish work the user just authorized. Proposed fixes (any one would help; 1+4 is the clean pair) Task-scoped capability grant. When the user approves local-computer execution for a task (or taps Always allow), inherit that grant for subsequent commands in the same task until the task ends or the user revokes it. Do not re-lottery each argv. Denied spawn must always raise a card. If a host command cannot start, show the native approval UI. Never return "failed to spawn" with no card on a retry of the same command. Surface a stable reason enum to the agent: user_denied | no_prompt | policy_block | always_allow_not_applied. Always allow should be capability-based, not command-string-based. *** pull and pwsh -File package.ps1 are the same class (host process spawn). If Always allow is on, both should run or both should card. Document the actual scope. First-class local-builder / Grok Build handoff (plugin or integration). Let the cloud/desktop agent dispatch a structured job to a host-trusted local builder instead of emitting a paste prompt. Suggested job schema: cwd, allowlisted scripts, env (no secrets in logs), success checks, rollback command, report contract. The local builder runs with existing host trust and posts a machine-readable result back to the originating thread. This is the seamless version of today's paste workaround. Documented private-worker path. Cloud Agent environment type "machine" / "My Machine" should be a supported, discoverable way to run host-bound scripts on the production box, with the same per-run approval rule. Today the launch API exists; whether a worker is registered is not visible to the coordinating agent. Success criteria After the user says "deploy" and approves once, the agent can package, retarget the existing scheduled task, and verify HTTP without a paste handoff, or it can dispatch that runbook to a local builder and get a structured result. Credentials never appear in logs. A denied step always produces a card, never a silent spawn failure.

• 
  noclipepe (@noclipepe) reported 9 minutes ago

  GitHub added nearly 700,000 AI projects in a single year. More than 1.1M public repositories now import LLM SDKs. Another 693,867 AI projects appeared in the last 12 months — +178% YoY. AI engineering doesn’t have a shortage of code anymore. It has a navigation problem. The hard part is figuring out which repos are actually worth your time.

• 
  Jul-fi-kar (@OrphanPid) reported 10 minutes ago

  @bunjavascript hopefully your github won’t be swarmed with issues

• 
  ꦏꦢꦺꦏ꧀ꦥꦤ꧀ꦕ 🏴‍☠️ (@Kadek25Panca) reported 13 minutes ago

  yo @githubstatus, my website using a custom domain is down with 404 code saying "There isn't a GitHub Pages site here",. I've double check all the status and setting, it's all green,. any idea why,?

• 
  Tim 🏴‍☠️🏴Decentralize and nullify 🏴🏴‍☠️ (@Tim_The_Sandman) reported 17 minutes ago

  They've published the algo to their Github page. The algo still includes ABSURDLY HEAVY weighting for negative signals like mutes, blocks, reports, "show less" and other items. This is still a massive problem to have negative signals disproportionately impact an account's global reach. As it stands, a small number of accounts can functionally silence a selected account in such a way that account can not dig out despite receiving significant positive signals from the accounts who do get to see the content. This is a massive flaw in the system, and it deprives many people who actually enjoy the content when they get to see it from being able to even see the content.

• 
  AML (@alysha_lobo) reported 17 minutes ago

  𝐔𝐏𝐃𝐀𝐓𝐄 𝟑: We manually reviewed 𝐄𝐕𝐄𝐑𝐘 single profile (kudos to the team for going the extra mile with me on this!) Before we drop 𝐭𝐡𝐞 𝐧𝐞𝐱𝐭 𝐌𝐓𝐒 𝐫𝐨𝐥𝐞 (𝐥𝐨𝐚𝐝𝐢𝐧𝐠 𝐢𝐧 𝟑…𝟐…𝟏), here is some 𝐟𝐢𝐫𝐬𝐭 𝐜𝐮𝐭 𝐜𝐚𝐧𝐝𝐢𝐝 𝐟𝐞𝐞𝐝𝐛𝐚𝐜𝐤 if you actually want to stand out: 🛑 Stop using AI to mass-apply — the team and I have been neck deep wading through so much slop!! This wastes everyone's time. 🔗 Double-check your links — broken GitHub, Tech blog, or GDrive links mean instant rejection. 💡 Show, don't tell, your frontier potential — we want to see "cracked," out-of-the-box ideas you have actually cooked and served. 𝐘𝐨𝐮 𝐰𝐢𝐥𝐥 𝐡𝐞𝐚𝐫 𝐟𝐫𝐨𝐦 𝐮𝐬 𝐧𝐞𝐱𝐭 𝐰𝐞𝐞𝐤 𝐨𝐧 𝐧𝐞𝐱𝐭 𝐬𝐭𝐞𝐩𝐬. 𝐓𝐡𝐚𝐧𝐤 𝐲𝐨𝐮 𝐟𝐨𝐫 𝐲𝐨𝐮𝐫 𝐩𝐚𝐭𝐢𝐞𝐧𝐜𝐞! 🇮🇳 𝐂𝐦𝐨𝐧 𝐈𝐧𝐝𝐢𝐚 — let’s raise the bar and show what we can truly build at the frontier of tech. 🚀

• 
  Greg McKeon (@wegmckeon) reported 19 minutes ago

  @SemiAnalysis_ @github @useblacksmith AZ power outages + data centers + 7x growth since January don’t mix. CI is also odd bc we can’t load shed - builds don’t go away when we’re down, so degradation due to capacity lingers. long term fix is even more distributed capacity, which is already starting to land.

• 
  Luke Parker (@LukeParkerDev) reported 19 minutes ago

  @GengMaxwel19074 @thdxr can you tell me the specific bugs/github issue links pls?

• 
  Arthur van Pelt 🔥 ∞/21M ⚡ (@Arthur_van_Pelt) reported 26 minutes ago

  @jackmallers @gladstein He did not. Bitcoin Core only exists since November 5, 2013: Wladimir van der Laan (laanwj) opened GitHub issue #3203 proposing the rebrand (e.g., Bitcoin-Qt → Bitcoin Core). March 19, 2014: Bitcoin Core version 0.9.0 was released.

• 
  Chris Moore (@crsmoore) reported 27 minutes ago

  @_tombrow Can you please, please fix the font sizing issues on iOS? iOS text sizing breaks down under accessibility zoom. Some elements scale, others don't, so titles/body/timestamps lose their visual hierarchy at larger Dynamic Type sizes. Not yet filed as a GitHub issue.

• 
  @hemanth (@Heamanth_alturi) reported 28 minutes ago

  @JackWoth98 Another issue in another project also gave it a 2-stage prompt (non-technical questions first, then technical) and it just skipped stage 2 entirely to rush into creating github issues. then started running gh issue close on its own while i was literally still asking clarifying qa

• 
  Praggy 👨🏻‍💻 (@praggy) reported 35 minutes ago

  @Teknium Omg I thought that was just me not know how to use GitHub lol. Yeah this sucks dude. Have been facing issues related to GitHub a lot lately. Maybe cursor one is for everyone who doesn’t sub to it as well 🤔

• 
  Uncharted Kenya (@UnchartedKenya) reported 35 minutes ago

  @XOpenSource oh dear, pls break it down to easy to understand terms. Some of us can't understand anything from these GitHub codes

• 
  Zeraf (@kadsxr) reported 36 minutes ago

  In April 2026, Andrej Karpathy published an 800-word gist on GitHub that quietly seeded a dozen repos trying to solve one problem: making Claude actually remember what you know instead of relearning it every session. The counterintuitive part: the trick is forcing Claude to forget the source material on purpose. It reads a raw file once, extracts what matters, writes it into a compiled wiki, and never touches the raw file again. Every query after that pulls from the compiled version, cutting token spend 70 to 90 percent, because a vault that keeps re-reading its own notes doesn't compound, it just gets more expensive to search. The most powerful setup right now ships 45 commands including notes that rewrite themselves when you contradict something you wrote earlier, and scheduled agents that clean the whole vault overnight while you sleep. Setup costs 90 minutes upfront, then runs itself forever after. The real failure mode isn't picking the wrong repo, it's installing five of them in one weekend and drowning in config debt before any of them compound. Two tools do the entire job: one that compiles raw sources into structured knowledge, one that gives Claude live access to it. What's sitting in your notes app right now, a second brain, or just an expensive filing cabinet you keep re-reading from scratch?