Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-09-05 · high
Version 2.1.261 also raises output limits to 128K characters and fixes a bug that lost hook output on resumed sessions.
Why now: GitHub posted the release notes on September 4, dating the fixes and the new command precisely.
Claude Code 2.1.261 adds /skill-doctor, a command that lists loaded skills alongside how much context each one burns and whether it's being used. Loaded skills eat into the context window whether or not they get used in a session, and until now there was no way to see which ones were dead weight without guessing.
The same release raises two output ceilings: bashOutputMaxChars and taskOutputMaxChars. Inline command output and background task output can now run up to 128,000 characters before Claude Code spills the rest to a file. Anyone piping verbose build logs or long-running agent output straight into a session, instead of tailing a file afterward, gets far more room before that kicks in.
A third addition, --append-subagent-system-prompt-file, lets subagent prompts too long for a command-line argument load from a file instead.
Three bug fixes round out the 2.1.261 release notes. Resuming a session no longer drops hook output, and no longer loses context around parallel tool calls that had changed what the resumed request saw. A background agent that couldn't resume was retrying its wake-up in a tight loop and pegging the CPU. SDK and cloud sessions were also ignoring a Stop signal sent right after the first prompt, so a fast cancel didn't always cancel.
Each link below shares sources, entities, or timing with this story.
Fork skills chained to their own output now run before that output exists, unless they opt out.
Released September 4. The same build adds bashOutputMaxChars and taskOutputMaxChars, raising inline command and background-task output to 128K characters before spilling to a file, which saves a round trip on every long build log. It also adds --append-subagent-system-prompt-f...
Two of the six hid commands from the approval dialog itself, using tab padding, invisible Unicode, and split Bash strings.
The open-source tool now sits at 98.9k GitHub stars and pairs a skill with a proxy that compresses code, logs and JSON before they hit the model.
Boundary-Bench ran 12 agent harnesses through real firewall and filesystem locks, and costs climbed as much as 167 percent as those restrictions tightened.
The release also adds a command that migrates Python projects off the old anthropic 0.x SDK and stops WebFetch from serving stale content all session.
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.