Claude Code 2.1.261 adds /skill-doctor and 128K tool-output caps
Sept 4's build also lets subagents load prompt files from disk, flags org-policy failures in /status, and fixes out-of-order keystrokes and ignored Stops.
Copy markdown
/skill-doctor finds the skills eating your context
The new `/skill-doctor` command scans your installed skills, flags the ones you never actually invoke, and shows the token cost each one adds to every prompt. It's a fast way to reclaim context window before a long agentic run.
Tool output stops truncating at the worst moment
New `bashOutputMaxChars` and `taskOutputMaxChars` settings raise captured output to 128K characters. A noisy test run or a subagent's full report now survives intact instead of getting cut off mid-result.
Subagents can load their prompt from a file
`--append-subagent-system-prompt-file` points a subagent at a large system prompt on disk instead of cramming it inline. Handy for reusable review or verify agents whose instructions have grown long.
/status now explains a failed org policy
A new "Organization policy" line in `/status` and `claude doctor` surfaces when a managed policy failed to load. It's the first place to look if managed MCP servers or permission rules are silently not applying.
Three fixes daily drivers will feel
Fast typing and pasting no longer lands characters out of order; Remote Control sessions stop showing stale permission modes; and SDK/cloud sessions no longer ignore a Stop sent right after the first prompt.
Elsewhere: Codex CLI wires up GPT-6 Astra
OpenAI's Codex CLI reached v0.153.x, adding GPT-6 Astra API configuration (plus a Fast tier), Vim undo/redo, and remote plugin-marketplace management. If you've switched to Astra, your Codex config now speaks it natively.