Claude Code v2.1.261
v2.1.261
Claude Code
v2.1.261
A skill auditor, bigger output limits, and a prompt input that stops dropping characters.
Claude Code v2.1.261 shipped on September fourth. It adds a command that audits your loaded skills, two settings that let more command output through, and a flag for long subagent prompts. The rest is repair work, and the prompt you type into got most of it.
01 — What's new
v2.1.261
>What's new
The additions.
/skill-doctor
v2.1.261
>Find out what your loaded skills cost you
> /skill-doctor- Shows
- which loaded skills went unused
- And
- what each one costs in context
- So
- you can prune them
Every skill you load sits in the context window whether Claude reaches for it or not. The new /skill-doctor command reports on that. It names the loaded skills that went unused and what each one costs you in context, so you can prune them. Run it in a session that has been going a while.
Output settings and the subagent prompt file
v2.1.261
>More command output inline, long subagent prompts from a file
{
"bashOutputMaxChars": 128000,
"taskOutputMaxChars": 128000
}128K characters is the ceiling for both.
claude --append-subagent-system-prompt-file ./subagent.mdWhen a command prints more than Claude can take inline, the output goes to a file instead. Two new settings raise that ceiling: bashOutputMaxChars for shell commands, taskOutputMaxChars for background tasks. Either one goes up to 128K characters. And a subagent system prompt too long for the command line can now come from a file.
02 — The prompt input
v2.1.261
>The prompt input
The prompt you type into.
Character order and word keys
v2.1.261
>Characters stopped scrambling, and word keys match Bash
- Out of order
- fast typing and key repeat scrambled or dropped characters
- Word keys
- Ctrl+W, Alt+F and Alt+D behave the way Bash does
- keybindingFlavor
- no longer has any effect
If you type fast, or hold a key down and let it repeat, characters were landing in the prompt out of order, and some were dropped. That is fixed in the input layer, so it holds wherever you type. The word editing keys now match Bash. Ctrl+W deletes back to whitespace, and punctuation separates words. keybindingFlavor no longer has any effect.
03 — Stops, resumes, Remote Control
v2.1.261
>Stops, resumes,
Remote Control
Stopping, resuming, and Remote Control.
Stop and Remote Control state
v2.1.261
>Stop lands, and Remote Control stops reporting stale state
- Stop
- ignored just after the first prompt in SDK and cloud sessions
- Permission mode
- stale when a phone or browser attached
- SendMessage
- said delivered for a machine that was offline
A Stop sent just after the first prompt was ignored in SDK and cloud sessions, and the run finished anyway. It stops now. Remote Control showed a stale permission mode when a phone or browser attached, and left the spinner running after a turn ended. And a SendMessage to an offline machine read back as delivered. It now says delivery is queued until that machine reconnects.
Resuming changed the request
v2.1.261
>A resumed session lost part of its context
Hook output and other context around parallel tool calls went missing on resume, which changed the resumed request.
claude -p --resume ./transcript.jsonl
# malformed session id -> now resumes under a fresh oneResuming a session was losing hook output and other context around parallel tool calls, so the conversation you came back to was not the one you left. And resuming from a transcript file would adopt a malformed session ID recorded inside it. It now resumes under a fresh one.
04 — Guardrails
v2.1.261
>Guardrails
What now stops and asks you first.
The delete prompt and packed diagram links
v2.1.261
>A wrapped delete and a packed diagram link both stop for approval
rm -rf "$@"
sh -c "rm -rf $TARGET"Both shapes now raise the safety prompt.
A link that packs your content into a public diagram renderer's URL counts as an upload: not auto-approved unless you asked for it.
The safety prompt for a dangerous delete now catches two more shapes: a recursive force delete on positional parameters, and the same command inside a double quoted shell script. Those are the shapes a wrapper script hides a delete in. Auto mode also treats a link that packs your content into a public diagram renderer as an upload to that site, so it is not approved for you.
05 — The VS Code extension
v2.1.261
>The VS Code
extension
The VS Code extension.
The VS Code extension
v2.1.261
>The VS Code extension edits MCP servers and shows what is open
- MCP servers
- an Add server form, and a Remove action
- Session list
- a hollow ring for a session open in a terminal or another window
- Model picker
- one flat list, without models your organization disabled
In the VS Code extension, the MCP servers dialog got an Add server form and a Remove action, so a server goes in without leaving the editor. A session already open in a terminal or another window now carries a hollow ring instead of looking closed. And the model picker is one flat list, with the models your organization disabled dropped from it.
Read the full entry
v2.1.261
Claude Code
v2.1.261
github.com/anthropics/claude-code
The full entry sits at the top of CHANGELOG.md in the repository, and on the release page for this tag. Claude Code v2.1.261.
