Claude Code's /skill-doctor flags the skills eating your context
The same recent CLI drops raise inline command output to 128K chars, add a fullscreen /diff panel, and let orgs hand MCP servers to every user.
Copy markdown
See which skills are silently eating your context
The new /skill-doctor command lists every skill currently loaded, flags the ones you never actually invoke, and shows the token cost of each. Unload the dead weight and you win that context window back for the task at hand.
Agents can now hold 128K chars of output inline
New bashOutputMaxChars and taskOutputMaxChars settings let Claude take up to 128,000 characters of command or subagent output directly in context before it spills to a file. Full test logs and build traces stay in front of the model instead of getting truncated or shunted to disk.
A fullscreen /diff for everything the agent touched
/diff opens a fullscreen panel of all uncommitted changes beside the conversation, so you review the whole diff in one view before you commit rather than scrolling it inline.
Point subagents at a prompt file, not a wall of text
The new --append-subagent-system-prompt-file flag loads large custom subagent instructions from a file. Keep long orchestration prompts in version control instead of pasting them into every run.
Orgs can push MCP servers and run truly headless
The managedMcpServers setting lets an organization hand HTTP/SSE MCP servers to every user, and --permission-prompts none makes headless sessions auto-deny prompts instead of hanging. Standardize the toolchain, then let CI agents run unattended.