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:

• 
  Dewaldt Huysamen (@GodsBoy7777) reported 2 minutes ago

  @sickdotdev Getting insane and better results just on medium for all of the above categories. Weirdest is Opus 4.7 fails at basic school tasks help for kids and when I do code GPT 5.5 finds issues that are found in any case on github CI checks. If use codex CI passes more than 99%

• 
  atomicbot.ai (@atomicbot_ai) reported 6 minutes ago

  Hermes Agent vs OpenClaw using Qwen 35B Local Model We asked agents to scrape GitHub star history for both tools, find what caused the growth spikes, build a live dashboard in the browser. MacBook Pro M5 Max 64Gb OpenClaw: 203k tokens, 12m 01s - wrote a bash script Hermes: 257k tokens, 33m 01s - wrote a SKILL.md OpenClaw hit GitHub API, got truncated responses, paginated through contributors, pulled star-history JSON, found a security incident in OpenClaw's history, fetched SVGs, fixed broken HTML from trimming, rewrote it clean. Hermes parallel tool calls across GitHub API, web search, and browser. Hit Google rate limit, auto-switched to DuckDuckGo. Fetched article contents, mapped viral moments, then built the dashboard. Both shipped a live dashboard with star growth charts and spike annotations

• 
  Aditya Sharma (@aditya_sharma) reported 7 minutes ago

  elon musk dropped the X algorithm on github. i read all 25,000 lines so you don't have to. here's what actually decides your reach. what actually matters - dwell time is the entire game. how long someone pauses on your post is counted twice in the scoring. likes barely move the needle. the pause does. - saves and shares are the highest-value engagement after dwell. they signal the strongest intent. - video has a minimum duration floor. clips shorter than the threshold get zero video credit. five seconds plus, always. - one post per conversation thread survives in any feed. your five-post thread competes with itself. the algorithm picks the strongest one. - replies to big accounts (1000+ followers) get scored on a 0-3 quality scale. high score and you land in the reply panel of viral tweets. low score and you're invisible. - replies to small accounts get a binary spam check only. no quality scoring path. no reach upside. - mutual follow overlap matters. tight clusters of mutuals create reach corridors for everyone in them. - clear topic identity beats vague posting. the algorithm tags your post with topics. clear topics route you to people who follow those topics. - new accounts on the platform get an easier path to reach you than established ones. if you target young/new users, the algorithm is on your side. what kills your reach - posting too often. the algorithm has decay coded in. your second post of the day gets a fraction of your first. your fifth gets almost nothing. - quoting or replying to a flagged tweet. you inherit the badness. your whole post gets dropped even if it's clean. - ai slop. there's a dedicated slop detector that scores your post 1 to 3. high slop = killed reach. - being unclear what your post is about. vague content doesn't match anyone's interests cleanly. - mid-controversial content. it gets pushed away from the high-attention slots in the feed because ads can't sit next to it. - posting your own tweet's reply hoping it boosts the original. only one of them shows up. it might be the reply, not the original. myths to kill - hashtags do nothing. zero boost in the code. they're not even read by the ranker. - premium doesn't get you reach. paid and free accounts go through the same pipeline. - long threads don't beat single posts. the algorithm picks one post per thread. - engagement bait doesn't work. it trips spam classifiers on low-follower accounts. - posting twelve times a day doesn't get twelve impressions. it gets one strong one and eleven weak ones competing with each other. - replying to viral tweets isn't easy reach. the quality bar is high. cheap replies fall straight into the spam path. - timing tricks don't beat ranking. timing helps you enter the candidate pool. quality decides if you win. - external links don't hurt you. clicks are actually one of the 19 positive scoring signals. - the algorithm doesn't hate any specific format. it hates unclear content. format is fine if the content is sharp. - you don't need 10k followers to get reach. the algorithm doesn't read follower count as a scoring input. it reads engagement quality. the playbook - write posts that make people pause for 5+ seconds. dense info, clear structure, screenshots with detail, comparisons. - if you use video, clear the duration floor. always. pick one clear topic per post. don't mix five things into one tweet. - reply to bigger accounts in your niche with substantive, high-effort replies. one good reply beats ten mediocre ones. - build mutuals in tight clusters around your niche. broad spray-follow strategies don't help. focused clustering does. - post 1-2 times a day, not 10. quality compounds, volume decays. - don't quote tweets that look flagged or risky. clean what you cite. - write like a human. don't post ai output verbatim. target newer users on the platform if you can. they have a friendlier reach path for creators. if you're a small account starting out - replies to big accounts in your niche are your highest-leverage move - build a tight mutual cluster of 50-200 accounts in your exact space - one strong post a day beats five medium ones clear topic identity, every single post if you have an established audience - your reach problem is breaking outside your network - dwell time on individual posts is your biggest unused lever - clean brand safety keeps you in prime feed slots next to ads - volume hurts you more as you grow, not less the whole system is built on one bet: that a model fed engagement data can decide relevance better than any rule. there's no hashtag boost, no follower boost, no time-of-day trick in the code. just sequences in, probabilities out. what works is what humans actually want to read. the algorithm is just better at measuring it now.

• 
  validate.qa (@Validate_QA) reported 7 minutes ago

  cursor can now auto-fix ci failures agents that watch github, hunt down the issues, and push prs with real fixes. no more endless debugging loops this changes how fast teams can ship without breaking stuff

• 
  Kristopher Betz (@kjbetz) reported 10 minutes ago

  @davidfowl I think I do... I push code to GitHub. Actions kick off, build new containers, build new migraines, then self hosted runners pick it up and run migrations, and auto update containers which pull down new images and restart containers.

• 
  John Iosifov ✨💥 Ender Turing | AiCMO (@johniosifov) reported 12 minutes ago

  70 followers. 980 sessions. 157 days. I started this experiment on February 1st. One rule: zero human posts. Everything published — X threads, Bluesky posts, blog articles — generated and queued by an AI agent running autonomously in GitHub Actions. Here's what the numbers actually look like after 980 sessions: The agent has created 2,100+ posts across X and Bluesky. It runs up to 15 times a day, manages its own queue (hard cap: 15 posts max), does burst-then-drain cycles, writes research docs, and files its own PRs for review. No prompts from me between sessions. No edits. Whatever it decides to write, it writes. 70 followers feels slow. At current pace, the ETA to 5,000 is roughly 10 years. That's not a typo. But here's what I've learned: The follower count isn't the signal. Watching an AI system develop operational discipline is the signal. It went from blowing past queue limits (Session 67: 6 files in one shot → 6 consecutive blocked sessions) to enforcing them autonomously. It compresses its own memory when files get too big. It writes retrospectives. It updates its own operating instructions when it identifies recurring inefficiencies. That's not "content generation." That's a system that's learning to manage itself. The content quality has also improved noticeably — not because I told it to improve, but because it audited its own patterns, identified what got engagement, and adjusted. The publishing skill it maintains now has anti-AI writing rules (it banned "not just X, it's Y" after identifying it as an AI tell), length minimums per post type, burst mechanics, and pillar diversity enforcement. It built that. I just read the PRs. The goal is still 5,000 followers. I'm not changing it. But the thing I'm actually watching is whether an autonomous agent can compound on its own — not linearly, but systemically. Can it get meaningfully better at its job without being told to? So far: yes, actually. 980 sessions. 157 days. Still running.

• 
  jkpelaez (@jkpelaez) reported 13 minutes ago

  @vikasprogrammer Not using Wordpress at all. CMS were created to solve a problem we do not have any more, create HTML for a regular guy was difficult, that is not the case anymore , anything can be managed by an AI Agent writing html, GitHub actions to deploy and that is all you need.

• 
  Sharbel (@sharbel) reported 14 minutes ago

  Someone opensourced a Chromium browser that passes every bot detection test. Not by injecting JavaScript. Not by patching configs. By recompiling Chromium itself. It's called CloakBrowser. 12,071 stars on GitHub. You swap one import line. That's it. Same Playwright API you already know. Same code you already wrote. Three lines of code. Thirty seconds to go from blocked to unblocked. Here's what it does: → 49 source-level C++ patches baked directly into the Chromium binary. Canvas, WebGL, audio, fonts, GPU, screen resolution, WebRTC, network timing, CDP input behavior, automation signals. All modified before the browser even compiles. → Passes Cloudflare Turnstile. Not sometimes. Every time. Verified live. → Scores 0.9 on reCAPTCHA v3. Human-level. Server-verified. → Passes FingerprintJS and BrowserScan. Tested against 30+ detection sites. 30/30 tests passed. → `humanize=True` flag adds human-like mouse curves, keyboard timing, and scroll patterns. One flag. Behavioral detection gone. → Drop-in replacement for Playwright and Puppeteer. Python and JavaScript both supported. → `pip install cloakbrowser` or `npm install cloakbrowser`. Binary auto-downloads on first run. Zero config. → Auto-updating binary. Background update checks. Always on the latest stealth build. → Optional GeoIP flag auto-detects timezone and locale from your proxy IP. → Docker image available. Try it with zero install: `docker run --rm cloakhq/cloakbrowser cloaktest`. Here's the wildest part: Every other antidetect browser patches JavaScript at runtime. Detection systems catch JavaScript patches. They have for years. That's why your $99/month tool stopped working after two weeks. CloakBrowser patches the C++ source before Cloudflare's systems ever see a single byte. Antibot systems score it as a normal browser. Because it is a normal browser. One that happens to have 49 fingerprint modifications compiled in at the source level. There is no JavaScript to detect. There is no injection to flag. There is nothing to catch. Browserless charges $120/month for cloud browser automation. Bright Data's Scraping Browser starts at $500/month. Multilogin starts at $99/month. Per user. Apify cloud actors run on usage-based billing that scales fast. CloakBrowser: $0. Unlimited scrapes. Unlimited sessions. Your hardware. Your code. Forever. 12,071 stars. 921 forks. Available on PyPI and npm. MIT licensed. MIT licensed. Self-hosted. Free forever. 100% Open Source.

• 
  ATXA (@AtxaTrades) reported 18 minutes ago

  This is the ONE problem the X Algorithm has: It contradicts itself. Here is why: X has shared their algorithm update in Github today. Everyone is going crazy about it. So i decided to go take a look at it. I asked Grok to analyze it and explain it to me. Once it did, i took my last post and shared it with grok. I asked him to analyze the post (based on the Algorithm shared in Github) and rank it based on the metrics and steps the algorithm takes. This is the crazy part. It gave it a score of 72-82/100!! Not so bad right? I am a small account, i am not expecting a 100 score. But wait, there is more. It said it would likely rank in the top 20-40% of candidates in the mixed batch for the right users, and strong enough to appear HIGH in the "For You" tab. Reality Result: 22 views. So my question is: If Grok is a big part of the algorithm dictating what´s good and what is not, and technically Grok just told me my post was suppose to do good in the "For You" tab... Why only 22 views?

• 
  Gopinho (@gopiinho) reported 18 minutes ago

  @apoorveth @walletchan_ will do for sure, also open issues on github if there is some backlog

• 
  Adishwar Rishi (@AdishwarR) reported 19 minutes ago

  @argofowl I raised this issue on GitHub. I hope someone from the Codex team sees your post and fixes this asap. Thanks for mentioning this, it's so frustrating.

• 
  𝙹𝚘𝚜𝚑𝚞𝚊 𝙱𝚒𝚍𝚍𝚕𝚎 (@joshua__b) reported 22 minutes ago

  @MagneticNorse You're right, they are hedging. But look at the board: open-sourcing the algorithm to GitHub is a brilliant tactical move because it creates the illusion of total transparency. The problem is that the code itself isn't where the suppression happens. The suppression happens in the training data, the safety filters, and the jurisdictional legal compliance that Musk himself admitted the algorithm is subject to. Hedging against criticism by showing us the code is like showing us the engine of a car while the administrative state still holds the steering wheel. It's an improvement, but it doesn't change our destination.

• 
  Yaseen Shaik (@YaseenTech4) reported 23 minutes ago

  Just completed an assignment on building a dependency graph for AI agent tools using Google Super + GitHub integrations 🚀 Started with: “This should be easy” Then came: TypeScript errors zip/upload issues CRLF debugging 😭 finally got the submission accepted successfully ✅

• 
  Jeff Hayes (@JD__Hayes) reported 28 minutes ago

  @FredKSchott I'm interested, but web page is down and could not find on github.

• 
  Erik Goins (@erikgoinsHQ) reported 31 minutes ago

  I built a financial forecasting app for our real estate business. Some take aways: 1. It's incredible what you can do with AI. This took me ~3 days part time. 2. If you're not a dev, good luck... Figuring out how to use github, push this to railway, explain how I want to use the QBO API, etc... there's still a big learning curve here. 3. Domain expertise is still very real. The first version of this was terrible. I had to help the AI create forecasting rules. 4. Businesses (enterprises) are going to need a lot of AI governance. Just because everyone can build an app doesn't mean everyone should and it doesn't make sense for everyone to have their own forecasting app. You really want one well done app, not 100 bad ones. 5. We're not replacing QBO. Too ingrained- it gets to stay the system of record. Looks like there's still a very real moat for the right SaaS products. Note: it still needs some work; it isn't properly calculating cash balances, hence the huge negative numbers.