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:

• 
  Caelean (@the1024th) reported 1 minute ago

  We used @gauge_sh to measure 500 real coding sessions (Claude Code, Codex) to find out. The results surprised us: - Docs represented >50% of the sources fetched - This was followed by source code (99%+ GitHub) - Marketing content was minimally fetched, down at 5%

• 
  Find me on 🟦☁️ 🍉 #BLM (@ThisIsCSDX) reported 3 minutes ago

  @EstebanPdn3156 Disregard my now deleted tweet. I checked the github and couldn't find anything about it being AI-coded. Sorry for the trouble.

• 
  Paul Razvan Berg (@PaulRBerg) reported 5 minutes ago

  I don't think Anthropic realizes how **** their user comms are on GitHub — there are no humans replying, and a silly bot auto-closes almost all issues after 7 days. One of the reasons I switched to Codex is the high-quality support they provide on GitHub.

• 
  AJ👽 (@Web3AJ_) reported 7 minutes ago

  So i built a web app with node.js that checks a particular website periodically(every 10 seconds) for updates It works perfectly fine when I host on my pc but I run into errors when hosting from GitHub Tried vercel and @pxxl_space , both ran into errors Who knows what i need?

• 
  Michael Boisson (@Shaostoul) reported 7 minutes ago

  @Velascode_ Peacefully uniting people to the cause. Most people don't seem to care at this early stage. Finding the few who are voluntarily willing to help/test/advocate is like finding a needle in a haystack. I think part of the problem is the tech is so complex and vast that most people can't properly understand the implications of advocating, supporting, using the software and how it makes the dream come true for everyone. I've tried to make it as easy as possible to learn on the official website and GitHub but, the first steps of individually then collectively comprehending the different aspects of the app is not easy for those with low tech knowledge and limited patience.

• 
  Joseph Sauvage (@JoesInvestments) reported 9 minutes ago

  @CosmicRaisins It is very cool for you to respond directly as my repo is largely based off of your incredible work! This is my first time using github so I am still learning as I go but my agent should have caught the license issue.. License and NOTICE are up, with your work credited by name through the whole chain. That was a real miss on my part and I appreciate you calling it directly. Your kernels carry most of this stack and the attribution should have been airtight from day one. On the decode number: not transient. It's a content-class peak that reproduces deterministically, 42.4 / 44.6 / 46.9 on repeated runs with the exact battery in window-data/, and the README splits it from the ~20 tok/s sustained prose figure for exactly this reason. If you measured below 40 on the summary-class probe, tell me your draft acceptance rate. If it reads around 38 percent you're missing VLLM_MARLIN_USE_ATOMIC_ADD=1, which lives in the launcher env, not the serve flags. It's a 7x lever on the quantized draft and it's the single most missed line in the repo; you'd be the third strong reproducer to trip on it. If your acceptance reads 55+ and you're still slow, then you've found something real and I want to see it. Either way, run the battery as shipped and I'll put your numbers next to mine in the README, agreement or not. Your stack, my measurement discipline; that combination is worth more than either repo alone.

• 
  Mark Ajzenstadt (@mardehaym) reported 9 minutes ago

  We built a multi-agent orchestrator at @LimestoneHQ. It's called Velocity Core. It's an autonomous task executor. Board task in, reviewed pull request out. The trigger is a status change on your issue board. Five steps: 1. Trigger. A ticket moves to "AI: Ready" in Jira or ClickUp. The orchestrator resolves the repo and pulls context. 2. Sandbox. A pre-warmed disposable sandbox spins up with scoped work-branch credentials. 3. Loop. The agent plans, edits, runs tests, reads CI feedback, and fixes. Repeats until CI passes. 4. Gate. A PR opens linked to the original work item. Human review. No auto-merge. 5. Routed. Every LLM call goes through the gateway with task metadata. Every run is traced. The harness plugs into the client's existing Jira, GitHub, and CI pipeline. The human stays at two points: intake (writing the ticket) and output (reviewing the PR). Everything between is the agent's job. We deploy it inside client engineering orgs. Running internally at Limestone, landing inside your perimeter. Board task in. Reviewed PR out.

• 
  Diam (@diamai_) reported 10 minutes ago

  Connect an AI agent to five MCP servers, and it can spend 55,000 tokens before it reads the task, loading instructions for the tools they expose. GitHub, Slack, incident alerts, dashboards, and other software can arrive as a giant menu before the agent knows the job. More of its attention goes to that information, less to the request. Anthropic's Applied AI team ran into a different version of the problem. Sonnet 4.5 would start wrapping up a job before its context window was full. They added reset logic so it could continue instead of stopping early. Opus 4.5 no longer behaved that way. The reset became pure overhead, adding delay and sometimes discarding cache the system could still use. In a new AI Engineer talk, Gagan Bhat and Isabella Kai He explain the architecture behind Claude Managed Agents. The full record of the job sits outside the agent's active memory. The system brings back only the detail needed now. - 7:44 When a reset helps and when it becomes a drag - 14:32 Keeping old work without carrying it all - 25:58 Giving an agent private access to company tools Watch the video, then read the attached article. It is about what happens when an agent gets every available tool before it knows which one it needs.

• 
  ndy_gratitude (@wendylocas757) reported 10 minutes ago

  quantum risk for crypto wallets has a countdown, not an opinion. @quipnetwork tracks it on a public doom clock and ships the cryptographic fix. open testnet, public github, 13k+ users already verifying.

• 
  Alonso G (@CryptoDegenTopG) reported 12 minutes ago

  @AdamGell @AdamGell I’ll double check if I am doing the formatting correctly maybe I am missing something. Otherwise I’ll open a GitHub issue.

• 
  idan levin (@0xidanlevin) reported 15 minutes ago

  I find it crazy that Skills.md came out in October 2025 and we still don’t have a good way to find skills and compare them (this goes to plugins as well which are a form of packaged skills) Skill marketplaces are still pretty basic. Most don’t have any real review system beyond the number of downloads. What is currently missing: 1/ A curated and reliable review system (not just downloads or GitHub stars) 2/ Comparison between skills based on common dimensions 3/ A proper versioning system 4/ Evals - e.g. how well does this SEO skill perform on a benchmark? The problem is that we don’t really have neutral common benchmarks for most tasks yet 5/ Usage data beyond downloads - retention, repeat usage, success rate, etc. 6/ Provenance - who built it, who maintains it, and how actively is it maintained? 7/ Dependencies - what other tools, MCP servers, APIs, or credentials does the skill required

• 
  Rob Smith (@RobmsmithUK) reported 18 minutes ago

  @mattpocockuk Using Wayfinder, my Claude agents often argue that they have already built out the specs as part of the map GitHub issues and advise to move forward with 'ticket' and 'implement'. Do you as a rule always go Wayfinder, spec, ticket, implement?

• 
  DGDG (dee-gee-dee-gee) (@dgdg_app) reported 22 minutes ago

  @earporter @dominic_w I know they weren't. It was bots working on a coordinated script that sped up and slowed down activity, together, with over a million events. The activity matched the trading simulation script in the GitHub repo to a T, down to the number of nodes used (20)

• 
  Elves - They work while you sleep (@elves_skill) reported 23 minutes ago

  Elves v2.27.0 + v2.28.0 live! • Discovery: A read-only pass that decides what's worth doing. Findings must cite file:line or they aren't findings. • Scope Control: Out-of-scope findings become GitHub issues instead of scope creep. Notice everything, change only what was agreed.

• 
  Vikas gupta (@vicky_grok) reported 24 minutes ago

  Now here's the number that breaks most people's mental model. GitHub exposes over 600 REST operations. Its official MCP server surfaces just 51 tools. Slack has more than 200 API methods. Its MCP server? Only 8 tools. Across the board, MCP servers expose a median of about 19% of the operations available in the underlying API. That's not a mirror of your API. It's a deliberately small, hand-picked slice. Why would anyone expose so little when they could expose everything? The reason is the most counterintuitive finding in the whole paper... 👇