Claude Code: subagents run at the effort you ask for, plugins install from a marketplace in one command, and network-path reads ask first

5 days ago

What changed in Claude Code 2.1.291 and 2.1.292: the Agent tool takes an effort parameter so subagents run at the effort you ask for, claude plugin install adds a marketplace and installs from it in one command, reads from network paths ask for permission again, quitting no longer loses a session's last messages, scheduled tasks fire when they should, and local MCP servers negotiate the newer protocol by default. https://github.com/anthropics/claude-code/releases/tag/v2.1.291 https://github.com/anthropics/claude-code/releases/tag/v2.1.292 https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md

Ask

Ask about this presentation

Answers are generated from this presentation.

Chapters

  1. 0:00Subagent effort
  2. 0:09Plugin install from a marketplace
  3. 0:20Security fixes
  4. 0:35Sessions keep what you gave them
  5. 0:47Scheduled tasks fire
  6. 1:00Newer MCP protocol
  7. 1:14No more silent empties
  8. 1:26The full changelog
Show transcript

Subagent effort

Claude Code

Subagents now run at the effort you ask for

Agent tool
takes an effort parameter
Each subagent
runs at the level you asked for
Release notes, October 6 · anthropics/claude-code

The Agent tool in Claude Code gained an effort parameter. When Claude hands work to a subagent, you can name an effort level, and that subagent runs at it.

Plugin install from a marketplace

Claude Code

>One command adds a marketplace and installs a plugin from it

claude plugin install
--marketplace <source>
Policy
the same checks as adding the marketplace yourself
Release notes, October 6 · anthropics/claude-code

Claude plugin install now takes a marketplace option. Claude Code adds that marketplace if needed, under the usual policy checks, then installs the plugin from it.

Security fixes

Claude Code

>Security: reads from network paths ask for permission again

Network (UNC) paths
hook approvals and auto mode no longer skip the prompt
Sandboxed commands
can no longer read staged /ultrareview uploads
rm -rf on Windows
an alternate spelling of the home folder counts as removing it
Release notes, October 6 · anthropics/claude-code

In a security fix, hook approvals and auto mode no longer skip the prompt for reads from network paths. Sandboxed commands can't read staged /ultrareview uploads, and Windows now recognizes a delete of your home folder under another spelling.

Sessions keep what you gave them

Claude Code

>Quitting no longer loses a session’s last messages

Quitting
the last messages of a session are kept
Cloud sessions
answers to permission prompts are no longer dropped
Resume
plan mode comes back when you resume a session
Release notes, October 6 · anthropics/claude-code

Quitting Claude Code no longer loses a session's last messages. Cloud sessions stop dropping your answers to permission prompts, and resuming a session now brings plan mode back.

Scheduled tasks fire

Claude Code

>Scheduled tasks and wakeups now fire when they should

Saved tasks
fire when created after /resume, /branch or /clear
Background /loop
keeps going after the session’s process restarts
claude -p
waits for background commands and scheduled wakeups
Release notes, October 6 · anthropics/claude-code

Saved scheduled tasks created after /resume, /branch or /clear now fire. A background session's /loop survives a process restart, and one-shot claude -p runs wait for background commands and scheduled wakeups.

Newer MCP protocol

Claude Code

>Local MCP servers now negotiate the newer protocol by default

Every install
protocol 2026-07-28, including Bedrock, Vertex and Foundry
Slow servers
remembered for 7 days and connected the older way
Opt out
an environment variable keeps the legacy handshake
Release notes, October 6 · anthropics/claude-code

Local MCP servers now negotiate the July protocol version by default, even on Bedrock, Vertex and Foundry. Servers that ignore it connect the older way, and an environment variable opts out.

No more silent empties

Claude Code

>Failures that looked like empty results now say so

Grep and Glob
an unreadable path no longer reports “no matches”
PDF page lists
Read errors instead of returning only the first range
@-mentions over 256KB
Claude is told the size and to read in portions
Release notes, October 6 · anthropics/claude-code

When Grep or Glob can't read their path, Claude now retries or tells you, instead of reporting no matches. PDF page lists return an error, and large mentioned files stop vanishing.

The full changelog

Claude Code

Every change is in the Claude Code changelog

CHANGELOG.md in anthropics/claude-code

github.com/anthropics/claude-code
Release notes, October 6 · anthropics/claude-code

Long bulleted replies also render faster. The full list is in the Claude Code changelog on GitHub.