Claude Code adds a 1-hour prompt-cache setting for long sessions
The v2.1.243 release also curates the /model picker, breaks token use down per loop, adds keyless Console sign-in, and fixes a Linux startup crash.
Copy markdown
The setting, and who gets it
New promptCacheTtl and subagentPromptCacheTtl options let API-key, Bedrock and Vertex users hold a 1-hour prompt cache on the main thread while subagents stay at 5 minutes. Step away and return mid-session and you pay cheap cached-read rates instead of reprocessing the whole context.
A /model picker you can curate
The new modelPicker setting builds an ordered, labeled model list for /model — any id spelling, including Vertex and Bedrock — appended to or replacing the built-in lineup. Stop scrolling past models your team never touches.
/usage now itemizes your loops
/usage gained a Loops breakdown: per-loop run count, total tokens, tokens per run, and last-run time. A runaway or chatty /loop task is now easy to spot before it quietly drains your budget.
Keyless sign-in for locked-down orgs
/login adds 'Sign in with your Console account' next to the API-key path, so organizations that don't allow raw API keys can finally authenticate Claude Code.
Remote MCP reconnects on its own
In non-interactive (-p) and SDK sessions, dropped remote MCP connections that used to hang forever now reconnect automatically or report as failed — a real fix for CI and headless agents.
Arch and Fedora startup crash fixed
v2.1.245 fixes a startup crash on distros shipping glibc 2.44 — Arch Linux, CachyOS and Fedora Rawhide. If Claude Code stopped launching after a system update, upgrade to get it back.