Claude Code 2.1.245 fixes a startup crash on Arch, CachyOS, Fedora

The hotfix rides with 2.1.246's editable Auto mode rules and MCP error fixes, while Thomson Reuters ships a frontier model fine-tuned from open weights.

Nowline Aug 26 7:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • Update to 2.1.245 if the CLI won't launch

    2.1.245 fixes a startup crash on Linux distros shipping glibc 2.44 — Arch, CachyOS and Fedora Rawhide — where Claude Code died before it even opened. On a rolling-release distro? Upgrade now; it's a hotfix stacked on the big 2.1.243 feature drop.

  • Auto mode gets an editable rules tab

    2.1.246 adds an Auto mode tab to /permissions so you can finally read and edit the classifier rules that decide what runs without asking you — real visibility into what auto-approves. The end-of-turn line now also stamps the clock time it finished (e.g. 'done 6:05 PM'), handy for long unattended runs.

  • MCP and background sessions fail louder

    Interrupted MCP tool calls now surface an explicit error instead of a silent 'completed with no output', empty-schema MCP args stop getting mangled into JSON strings, and background sessions no longer die after 45s when the start directory is gone or the machine slept. Fewer mystery failures mid-agent-run.

  • Thomson Reuters ships a frontier model off open weights

    Thomson Reuters launched 'Thomson', built by mid- and post-training an open-source base on Westlaw and Reuters data across a ~$40M program, and says it's on par with frontier models on legal tasks. A small open-weight version is on Hugging Face (academic, non-commercial); a developer API is 'still very early'. The signal for builders: a domain-beating model from open weights, not a nine-figure pretrain.