Skip to content

Claude Code

Claude Code 2.1.292 gives sub-agents their own effort dial

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Claude Code versions and changes →

Sub-agents in Claude Code 2.1.292 can run at their own effort level: the Agent tool now takes an effort parameter, and Claude runs the sub-agent at the level you ask for, according to the October 6 release notes on GitHub. If you run local MCP servers, check the other change before you update. Stdio connections now negotiate MCP protocol version 2026-07-28 by default on every install, Bedrock, Vertex and Foundry included, and MCP_PROTOCOL_NEGOTIATION=legacy switches that off.

What to change, and when

The effort parameter is a setting on the Agent tool call. A main session that hands work to several sub-agents no longer has to run all of them at one effort level.

Plugin installs from a new source take one step instead of two. claude plugin install --marketplace <source> adds the marketplace if it is missing, runs the same policy checks as claude plugin marketplace add, then installs the plugin. On a first run, marketplace add and install also wait for an organization’s managed settings to load before they run.

If you hit overloaded (529) errors, CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS sets a longer base delay for the retry backoff. Teams behind a proxy get a fix too: with HTTPS_PROXY set, Claude Code had been ignoring NO_PROXY for its own API requests (sign-in, policy, feedback, artifacts).

Scripts built on claude -p or the Agent SDK behave differently in two ways. A background command used to be stopped 5 seconds after the final result, and one-shot -p runs dropped a scheduled wakeup. Both are now waited for, so a script can run longer than before. The first turn also no longer waits for HTTP and SSE MCP servers to answer resources/list. And an MCP tool whose name runs past 128 characters no longer makes every request fail: the tool is left out, and an MCP error names it.

The MCP default

The version string itself carries meaning. MCP versions are dates in YYYY-MM-DD form, marking the last time the protocol made backwards-incompatible changes, and 2026-07-28 is the current one, per the protocol’s versioning docs. Under that version, every request declares its protocol version in its _meta field, and the server accepts or rejects each request on its own.

Claude Code adds a fallback for older servers. A local server that ignores the newer protocol check costs one slow connect. After that, Claude Code remembers it for 7 days and connects to it the older way, without the wait. A team that would rather not wait for the fallback at all has the legacy opt-out above.

Security fixes

The release lists one item under the Security label. PreToolUse hook approvals and auto mode were skipping the permission prompt for file reads from network (UNC) paths, the \\server\share form Windows uses for shared folders. Several other fixes close holes in the sandbox and in permissions:

  • Sandboxed commands could read the staged copies of /ultrareview uploads under ~/.claude/seed-admin.
  • On macOS and Windows, reading a notebook or PDF could return a file outside what was approved if a link was swapped in mid-read.
  • A tampered on-disk cache of server-managed settings could switch off or unseat the built-in policy plugin while the settings fetch was failing.
  • rm -rf on the 8.3 short name, or another alternate Windows spelling of the home folder or a drive, was not treated as removing it.
  • Sub-agent definitions with permissionMode: auto entered auto mode even when it was unavailable: disabled by settings, tripped by the circuit breaker, or not supported by the model.
  • <system-reminder> tags in a hook’s output are now escaped before they reach Claude.

For mod authors

Mods, plugins whose JavaScript or TypeScript functions run inside Claude Code on events such as a tool call or a submitted prompt, arrived in 2.1.287 on October 1. Three additions here are for them. A new prompt.autocomplete event lets a mod add its own rows to the prompt box’s autocomplete list. $.model.complete gains prompt caching: prompt and system take blocks of text, and cache: true on a block caches the request up to that block. And the agent.spawn hook now sees workflow agents, with their run and index, so a mod can refuse them.

Most of the mod fixes sit in the chain that handlers form. As Digitalapplied describes it, a hook gets a function called next: calling it passes the event on to the next mod and then to Claude Code’s own behaviour, and returning without it stops the chain. Some of the fixed cases let a call through that should have been stopped. A hook that called next(e) and then failed while the turn was interrupted let the call through, and it is now rejected. Tool calls made while the plugin hooks worker restarted were answered without the plugins’ permission hooks. A tool.check hook answering allow could run a tool that needs your answer, such as a question or a plan approval, without showing its dialog.

Organization guards get two more fixes. Version 2.1.290 had already unloaded a user-installed mod that could make an organization’s guard skip its check. Now a guard hook with a .catch is asked about calls another mod’s hook makes beneath the guard’s own $ call, instead of being skipped silently. And when an organization’s plugin returns a $ name that a user-installed mod had added, the user’s mod is unloaded, not the organization’s plugin.

claude plugin test is stricter too. A failed expect inside a hook the test registered, or a stub answer the engine refuses, now fails the test. Before, those tests passed silently.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.