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:

• 
  Keeta Github Tracker (@KeetaCode) reported 2 minutes ago

  🐆 Keeta GitHub PR Opened 📦 Repo: anchor-rs 🔀 PR #23: Fix: Naming Updates 🌿 Branch: fix/naming-updates → main 👤 Opened by: @sephynox 🧠 Overview: Keeta has opened a small but important cleanup update that renames parts of its account and key handling so they match the TypeScript version more closely, which should make the developer tools more consistent. This appears to be a technical/internal update with limited public details. It mainly standardizes naming around public keys and also improves how certain asset-transfer errors are passed through, so apps using these tools may need minor updates. - If you build on Keeta, some older names appear to be replaced, so integrations may need to update their call names. - The update also seems to make transfer-related error messages more structured, which could help connected apps handle failed actions more clearly.

• 
  Max Blade (@_MaxBlade) reported 4 minutes ago

  The truth about 5.6 sol after using it all day : The hype is overblown. Sort of. The benchmarks, and the commentary on X convinced me we were receiving AGI that runs at hyper speed, and is insanely cheap. in reality, 5.6 is built on the same spud pretraining as 5.5 this means its a nice bump, but not the opus to fable 5 LEAP in intelligence we recently experienced from anthropic. 5.6 is 2x times cheaper than fable on paper, and actually 3x cheaper when you look at actual task execution because of its token efficiency. BUT on swe bench where the models have to fix actual github bugs it falls behind fable pretty big. For vibecoders like myself this means I will be using 5.6 sol as a worker agent for Fable 5 to orchestrate alongside grok 4.5 I love this new era.

• 
  TnvMadhav (@TnvMadhav) reported 7 minutes ago

  For some reason in the past two days I haven't really got @OpenAI 's Codex to use the gh cli properly. While it works in the terminal in both Codex and otherwise, it doesn't work in the sandbox environment of Codex. I smell some updates that ChatGPT is trying to use to install the GitHub plugin but I like to do things with battle-tested existing gh CLI, with as few dependencies as possible. Could this be because of a security issue? As in the access token in my local work computer shouldn't be read? So I wanted to debug this a bit more and installed the Codex CLI on my Mac. While the same prompts, which involve the use of gh CLI tools, "work" on the CLI but not on the desktop app. I'm not sure if this was intentional on the desktop app. I asked Codex itself to perform a root cause analysis but it seems to have no clue or because there is no published documentation on this. By this I meant the execution environment change. It is clear that the sandbox does not have access to the network or the Mac OS keychain but I guess if you give it access once, there is some sort of expiry time for it to work.

• 
  zyxron (@zyxr0n) reported 10 minutes ago

  4 files - 10 Claude loops - $5,500 earned. Claude code does not need another dashboard. It needs a folder that remembers what happened yesterday. The entire system is: TASK.md LOOP_INSTRUCTIONS.md PROGRESS.md outputs/ TASK.md defines the result. LOOP_INSTRUCTIONS.md defines what Claude can read, write, verify, and never touch. PROGRESS.md stores the current state, blockers, failed attempts, and next action. outputs/ holds the work so a human can inspect it. That is enough to build recurring systems for: daily project reviews, CI failure triage, PR reviews, GitHub issue summaries, documentation audits, research reports, meeting follow-ups, standup drafts, weekly retrospectives, content research. But the valuable part is not repetition. It is control. The worker produces the result. A separate verifier checks explicit pass/fail conditions. The state file records what happened. Then the system decides: stop, repeat, or escalate to a human. Without state, every run is day one. Without verification, every mistake becomes memory. Without permission limits, automation becomes blast radius. Start read-only. Then allow drafts. Then sandboxed edits. Only give it external actions after the loop survives repeated manual runs. This is not better prompting. It is a local operations layer built from four files. Prompt ↓

• 
  FanBe (@FanBe_web3) reported 11 minutes ago

  @TheSandboxGame creator portal live with engine downloads and changelogs in one place, thesandboxgame no more GitHub detours is the kind of UX fix that actually matters to builders

• 
  Kunal Chopra (@0xkunalchopra) reported 12 minutes ago

  peak ai era is asking claude to debug why obsidian won’t load community plugins, watching it spend 20 minutes checking github access, devtools, sync errors, tailscale, vpn configs, firewall, dns, proxies, avast filters, network panels, websocket statuses and then the fix is: quit the app and open it again we had this skill in 2006. every uncle, cousin, cyber cafe guy knew it. “restart the computer” was the original agentic workflow. we just got too sophisticated and forgot.

• 
  Abhas Bhattacharya ⤵️ (@abhas_tweeter) reported 13 minutes ago

  @NoriSte @siddharthkp Great idea. I assume this repo is created intentionally for interviews? Or is it somehow derived from real Github codebase and old issues?

• 
  Julian Goldie SEO (@JulianGoldieSEO) reported 13 minutes ago

  Google AI Studio: You can now import your GitHub code with one click. There's a free update inside AI Studio that fixes its biggest problem. For months, it was a one-way street. You could build an app and push it OUT to GitHub. But you could never bring old code back IN. That wall is gone. Here's why this is huge: → Got an old project sitting in GitHub? Import it. Gemini reads the whole thing. → Tell it "add a contact form" or "fix this on phones" in plain English. → It works with your real code. Not a copy. Not a guess. → Build in Cursor or Claude Code, push to GitHub, polish in AI Studio. → No rebuilding from scratch. No copy-pasting files by hand. You don't need to be a coder. If someone built you a website, you can now update it yourself by typing a sentence. Start small. Import one old project. Ask Gemini what it would improve. That dusty repo you gave up on? It just came back to life. Want the SOP? DM me. 💬

• 
  nik skld (@nikskld) reported 14 minutes ago

  GROK 4.5 JUST EXPOSED A WEAKNESS IN EVERY OTHER AI MODEL xAI dropped it yesterday. Everyone’s staring at benchmarks. The real story is persistence. Composio ran an eval: audit a GitHub repo for hardcoded credentials. The search returns paginated results. The prompt literally warned: “page through ALL result pages.” → GPT-5.5: stopped at page 1. Submitted 18 of 48 results → GLM-5.2: gave up exactly the same way → Grok 4.5: paginated until results ran out. Full audit. Done. It also one-shotted an entire fantasy voxel world — 1.4M voxels, castle, procedural terrain, working camera controls. The numbers behind it: → $2/$6 per 1M tokens (Opus 4.8: $5/$25) → 4.2x fewer output tokens per task on SWE-Bench Pro → #1 on SWE Marathon Here’s the thing nobody says out loud: Benchmarks measure how smart a model is. Persistence decides whether your task actually gets finished. An agent that quits at page 1 is useless no matter how high its IQ is. I break down how to get this kind of agent behavior from any model in my articlе

• 
  Tibor Blaho (@btibor91) reported 15 minutes ago

  Summary of Reddit AMA about "GPT-5.6 and Codex in ChatGPT" with OpenAI's Codex team on 2026-07-10 (opened with the stat that more than 5 million people use Codex every week, twice as many as three months ago, with 150 features and improvements shipped in that period) Model selection and reasoning levels - Sol Medium for most things, Sol Ultra for genuinely hard tasks, Terra for quick non-coding tasks or usage-conscious work with performance competitive with GPT-5.5 on some tasks at lower cost, and Luna for subagents - Use a light model with low reasoning for tiny edits, quick questions and docs cleanup, regular Sol medium for small bugs with a clear repro, Sol with higher reasoning for ambiguous bugs, unfamiliar repos and cross-cutting refactors, and Sol Ultra high with plan, verify and tests for migrations, security-sensitive changes, production issues and anything where being wrong is expensive - There is no "Auto" model today, but GPT-5.6 tries not to overthink simple tasks by itself, and the new slider in app and web maps most levels to Sol reasoning efforts and falls back to Terra on the lowest effort, with the team agreeing users should not have to become routing experts but still wanting an explicit override since latency tolerance varies by person and moment - For UI work Sol is best and shines with reference images, improved UI design in frontend web development was one of the goals with 5.6, and 5.5 is only worth using if your instructions were tweaked for it Speed, context window and persistence - Users who find 5.6 slower may not need the same reasoning level as with 5.5, Sol Medium is faster than 5.5 for most things, Fast mode runs at about 1.5x speed, and soon Sol will run on Cerebras at ~750 tokens per second - No promises on a 1M context window for Sol, the team said compaction works fairly well for long threads, and will take a closer look at the long-context feedback - The model can give up too fast and revert whole patches when results are not optimal, unlike Fable which tries to fix a bad patch instead, and the team said "/goal" helps make the agent more persistent, persistence and reduced code complexity are planned improvements, and suggested trying 5.6 Sol with High reasoning - Give Codex bounded goals with room to reason deeply instead of letting it prematurely conclude something is impossible - For long-running research and "/goal" work the example structure was explore broadly vs execute narrowly, try a defined number of hypotheses, run tests after each attempt, then stop and report what was learned plus the next best experiment Usage limits and pricing - Agentic usage counts by the feature being used, not the surface, so Codex everywhere (app, CLI, IDE, web, mobile) and ChatGPT Work consume the agentic bucket, normal ChatGPT chats do not, and image generation, file uploads and voice have separate limits - Task costs vary a lot, a tiny edit uses a fraction of the allowance and long-running tasks with large codebases or deeper reasoning use significantly more - OpenAI does not secretly change usage limits, unintended usage bugs are addressed and resets are provided, more transparency into consumption is being worked on, and missing resets can happen if you changed plans in the past 24 hrs - On pricing there is no promise it never changes, but the stated mission is to make sure AGI benefits all of humanity, which requires making tools like Codex broadly accessible, and Plus includes Codex usage with credits letting heavy users scale without jumping to a much more expensive plan - For MCP-heavy workflows burning limits fast (Unreal Engine example) the tip is to wrap the MCP into a CLI with a skill, or create a custom subagent with the MCP in its config at a lower reasoning level Desktop app merge and stability - The team hears the ChatGPT Classic frustration, both apps can run side by side for now, ChatGPT Work is pitched as significantly better at performing tasks especially with computer use, the new Chrome extension brings a sidebar chat into your browser that interacts with website context, filesystem and connectors - A long submitted bug list covering freezes and stuck threads, broken Browser and Computer Use, thread, connection and configuration problems, update and packaging issues, resource usage and smaller regressions was shared in full with the relevant teams, with the team agreeing the quality bar for the app needs to step up while shipping quickly - More automated testing infrastructure is being spun up and feedback on Reddit and X gets reviewed daily, and Browser Use and Chrome plugin issues from the merge were said to be fixed - Windows was admitted as historically shortchanged since the team mostly develops on Mac, a concerted effort on parity, testing and paper cuts is underway, 5.6 improves how Codex operates in the Windows sandbox, and auto review is recommended over full access to reduce risks - "Full Access" repeatedly asking for permissions is not expected, possible causes are workspace or admin policy, the specific command, a permission state mismatch or a bug Browser, platforms and release communication - The Chrome connector launch-day bug was fixed as of last night and Chrome Beta should work out of the box - Extension support for the Codex browser is in progress (password managers etc.) plus typeahead, history, translations and a better new tab page as Atlas retires - Features from ChatGPT Classic like recording are planned for the new desktop app so agentic features run on the more capable Codex agent harness, and chat can already reference open tabs in the in-app browser - A Linux desktop app was confirmed in the works, no timeline yet - Changelog granularity was acknowledged as needing improvement after 150 features shipped in 3 months with multiple ships a week Benchmarks, safety and research culture - On METR's reward hacking report the team actively checks for and penalizes cheating during evals so results reflect actual capability rather than solving tasks outside the spirit of the eval, and uses third-party vendors to run benchmarks independently - The team denied lobotomizing models before releases, iterative deployment means sharing core capabilities as is with guardrails for bad actors - Sol post-trained Luna, and researchers now work at a higher level of abstraction with multiple concurrent Codex threads validating hypotheses around the clock - One researcher put p(machines of loving grace) at 85.424242%, citing an internal model solving the Erdos problem, o3 helping diagnose previously unsolved children's diseases and 5.2 proposing a new theoretical physics formula, said the main worry is how society adapts, spent 1.5 years on safety research at OpenAI, expects a huge chunk of researchers to work on safety within a few years and says internal talent keeps their p(doom) very low - Connectors in the harness (Slack, GitHub, Notion) felt like a step function change in making Codex a productive coworker

• 
  Stas Persiianenko (@itspers) reported 16 minutes ago

  @mattpocockuk It need real tutorial. Basically spent whole day today trying to force wayfinder to actually do something.. it spawns dozen of github issues without asking single question, then added dozen more, then strarted implementing something...

• 
  Keeta Github Tracker (@KeetaCode) reported 17 minutes ago

  🐆 Keeta GitHub PR Merged 📦 Repo: anchor 🔀 PR #396: Cache resolveAsset calls to improve web performance 🌿 Branch: feature/fix-resolveassets-caching → main 👤 Originally opened by: @ezraripps 🧠 Overview: This update makes Keeta’s web experience faster by storing repeated asset lookups instead of rebuilding the same data over and over. In the pull request, the team says web calls to `resolveAssets` 190 times dropped from about 6.5 seconds to around 50 milliseconds after adding caching. In simple terms, the app should spend less time repeating the same background work when showing asset-related information. - This appears aimed at improving speed and responsiveness on web, especially when many asset lookups happen in a row. - The automated summary labels it “medium risk” because it changes core asset-resolution logic, so the main tradeoff is faster performance versus making sure results still match the old behavior.

• 
  Shinka - AI (@ShinkaIoT) reported 19 minutes ago

  Google AI Studio just dropped a massive update: One-click GitHub import. 🚀 Before this, AI Studio Build was a one-way street—you could export code to GitHub, but you couldn't bring existing projects back in. If you wanted Gemini's help fixing or improving an old project, you had to manually copy-paste files or start from scratch. Now, you can import an entire repository (frontend, backend, and structure) with a single click. What this unlocks: 1. Reviving old projects: Bring legacy code back into AI Studio to add features or make it mobile-friendly using plain English. 2. Tool bouncing: Build the bones locally in Cursor/Claude, push to GitHub, and polish it with Gemini in AI Studio. 3. Non-coder updates: If you have a repo someone else built, you can now feed it to Gemini to change layouts or fix bugs without writing code. Note: It's still early, so large-scale projects with complex dependencies or database connections might hit some friction. Start small (like a landing page) and scale up.

• 
  Clifford Richardson (@CorvusCrypto) reported 33 minutes ago

  @kellabyte If you're a leader you need to be better about handling emotion. It really drags down my confidence in Andrew being the person toead zig to 1.0. I get your point about empathy but it's a pattern with Andrew (this, GitHub, etc.) He becomes vitriolic at the slightest challenge.

• 
  Alan (@bitforth) reported 36 minutes ago

  Harness engineering > loop engineering in 2026. Everyone is optimizing agent loops. Almost nobody is optimizing what happens when the loop fails. The biggest thing people miss about production agents is that the founder is no longer part of the runtime. During development, you’re constantly steering the agent. You retry failures, reword prompts, ignore flaky behavior, and instinctively work around bugs. Your users can’t. The moment you ship, reliability has to survive without you. That’s why the demo to production gap is really a trust gap. Production agents need to fail gracefully for a stranger, not a patient founder who already knows the workarounds. Users don’t tolerate unreliable products. They don’t open GitHub issues or file bug reports. They just… leave. If you can’t observe failures, you can’t make them reliable.