Claude Code 2.1.261 adds /skill-doctor to find context-hungry skills
It also caps command output at 128K chars, loads big subagent prompts from files, and fixes characters landing out of order during fast input.
Copy markdown
/skill-doctor finds the skills eating your context
New `/skill-doctor` lists which loaded skills go unused and what each costs you in context, so you can prune them. If you run a stack of skills, it's the first built-in way to see the context tax and cut it before it slows every turn.
Cap runaway output before it floods the window
Two new settings, `bashOutputMaxChars` and `taskOutputMaxChars`, control how much command and background-task output Claude takes inline before it's spilled to a file — up to 128K characters. A direct lever on context and token spend when a command dumps a wall of logs.
Load oversized subagent prompts from a file
`--append-subagent-system-prompt-file` reads a subagent's system prompt from a file, for prompts too large to pass on the command line. Elaborate multi-agent setups are now scriptable instead of hitting the argument-length wall.
See why your org policy won't load
`/status` and `claude doctor` gain an 'Organization policy' line that explains why a managed policy fails to load — a proxy in the way, for instance. If your team config silently isn't applying, you can finally tell why.
The fast-typing fix
2.1.261 fixes typed and pasted characters occasionally landing out of order during fast input, alongside Remote Control permission-mode display and prompt cache-miss fixes. Small on paper, but the input bug bit anyone who types quickly.
Elsewhere: GLM-5.3's open weights carry a hyperscaler tax
Z.ai's 753B GLM-5.3 is open-weight and trending on Hugging Face, but it dropped the MIT license for a custom one: firms with $10B+ revenue must pass a Z.ai security review before any commercial use, while individuals keep full rights to run and fine-tune. Local deployment still needs ~245GB at 2-bit — datacenter GPUs, not a laptop.