Claude Code ships fully unattended headless mode and org-managed MCP
The 2.1.259-2.1.263 run also adds /skill-doctor for context costs, 128K inline output, a fullscreen /diff, and readable model names.
Copy markdown
Truly unattended headless: --permission-prompts none
v2.1.259 adds a --permission-prompts none flag so headless hosts stop asking for approvals mid-run - the missing piece for CI jobs and cron agents. A new --json flag on `claude plugin validate` makes those pipelines scriptable end to end.
Org-managed MCP servers, pushed to every seat
The new managedMcpServers setting lets an org provision HTTP/SSE MCP servers centrally, so your whole team gets the same tools without hand-editing configs. GitLab merge requests now surface in tool summaries too.
/skill-doctor shows what your skills cost you
v2.1.261 adds /skill-doctor to flag loaded-but-unused skills and the context tokens they burn - a quick way to trim a bloated session. /status and `claude doctor` now explain why an org policy failed to load.
128K inline output before it spills to a file
New bashOutputMaxChars and taskOutputMaxChars settings raise inline display up to 128K characters, so large command and subagent outputs stay on screen instead of getting truncated away into a file.
Fullscreen /diff and a text-based /advisor
v2.1.260 adds a side-by-side fullscreen /diff for uncommitted changes and a text /advisor for desktop, Remote Control, and SDK sessions. /cost now names the likely cause of prompt-cache misses.
Small wins: readable model names, safer rm, longer reviews
The /model picker now shows human-readable names instead of raw provider IDs, dangerous-command detection got stricter on rm -rf, and /ultrareview's timeout jumps from 30 to 45 minutes. v2.1.262-263 are stability follow-ups.