2026-11-08
Outages
ML research
Developer tools
- GitHub previews a pull request Overview page that puts merge readiness and next actions at the top GitHub's opt-in Overview page moves merge status, checks and approvals to the top of a pull request and adds filters for unresolved threads.
- Lima v2.3 adds a native Windows HCS driver, Linux bridged VM networking and sync excludes for AI agent sandboxes The local VM launcher now keeps .env files and .git out of agent sandboxes, can block guest access to host loopback, and runs natively on Windows
Infrastructure
- SeaweedFS 4.53 tightens S3 authorization checks and adds per-bucket Block Public Access The SeaweedFS release binds S3 Tables deletes to authorized records, gates CORS and versioning changes, and adds Block Public Access.
- Outlook.com has bounced some mail from domains without a DMARC record since 2026-11-05, small senders included Microsoft's consumer mail now rejects part of the mail from domains with no DMARC record, with no announcement and no volume threshold.
Talks
- Linus Torvalds says many kernel subsystems now take patches only after public Sashiko AI review Torvalds says several kernel subsystems now want public Sashiko AI review before taking patches, as bots answer a flood of AI-found bugs.
- UndoBench finds coding agents recover from injected failures in only 47% of runs that otherwise succeed In UndoBench, coding agents recovered from injected failures in only 47 percent of runs that otherwise succeeded.
- Modal's CTO says agent sandboxes need persisted idle state, elastic resources and credential-hiding sidecar proxies Modal's CTO says agent sandboxes need persisted state without idle compute, elastic resources, and sidecar proxies that hide credentials.
- AlixPartners found Claude Code's contract analysis missed $800,000 until structured data and custom tools fixed it AlixPartners found Claude Code's procurement answers looked complete but missed an orphan amendment worth about $800,000 and a renewal date.
Hacker News
- MaxxBench finds small requirement changes to DeepSWE tasks drop some coding models by about 20 points Mutating 89 DeepSWE tasks drops some coding models 20 points, and traces show agents coding to guessed hidden tests over the prompt.
- Developer leaves Claude Code for the Pi harness, and Hacker News debates subscription limits on other harnesses A developer moves from Claude Code to the Pi harness over prompt churn and token overhead, and commenters argue over subscription rules.
- Llama-modchip claims a 284B DeepSeek model on a 64 GB CPU box, and Hacker News disputes the quality at low bit widths A llama.cpp fork claims a 284B DeepSeek model runs in 64 GB of RAM on CPU, and commenters question its 2-bit quantization quality.
- Goldman Sachs limits Claude Code and GitHub to engineers, and Hacker News debates agents on bankers' workstations A bank that praised trader-built AI tools in July now restricts coding agents to its roughly 12,000 engineers and strats, citing risk.