Claude Code adds /skill-doctor, 128K output, unattended agent flags
Three releases this week land bigger inline logs, skill pruning, a fullscreen diff, org-wide managed MCP, and a hard deny-all mode for headless agents.
Copy markdown
128K of command output, inline
The new bashOutputMaxChars and taskOutputMaxChars settings let Claude Code keep up to 128K characters of command output in-context instead of spilling long build and test logs to a file. Your agent finally sees the tail of the log it is debugging.
/skill-doctor prunes dead context
The new /skill-doctor command lists loaded-but-unused skills and their context cost, so you can drop the ones just taking up room. Leaner system prompts mean lower token spend on every turn.
--permission-prompts none for headless runs
Version 2.1.259 adds a flag that denies anything requiring a prompt, built for unattended hosts and CI where a hanging permission dialog kills the run. The same release ships managedMcpServers, pushing HTTP/SSE MCP servers to every user in an org.
/diff opens a live fullscreen panel
Version 2.1.260 adds a /diff panel beside the conversation showing uncommitted changes as Claude edits them. /cost and the status line's prompt_cache field now name the likely causes of a cache miss.
Org policy failures stop being silent
/status and claude doctor now print an 'Organization policy' line explaining why a policy could not load, such as a proxy blocking the endpoint. A long-session bug that invalidated the prompt cache after an OAuth token refresh is also fixed.
Bash sandbox hardening
Permission rules with parentheses are no longer dropped as invalid, Read() deny rules now cover files passed as option values or git operands, and rm -rf safety detection was extended to positional and double-quoted commands.
Elsewhere: Seattle Times and Newsday sue OpenAI
Two more publishers filed copyright suits against OpenAI and Microsoft over training on their journalism, the latest front in the fight that shapes what tomorrow's models are allowed to learn from.