Claude Code v2.1.292 Adds Sub-Agent Effort Controls, Mod Caching, and Critical Security Patches

What changed#

Anthropic released Claude Code v2.1.292, introducing granular execution settings for sub-agents, new extension points for mods, performance enhancements, and security fixes.

Key changes in this release include:

• Sub-Agent Effort Controls: The Agent tool now accepts an effort parameter, allowing developers to set the precise reasoning or compute effort level when spawning sub-agents.
• Mod Hooks and Prompt Caching: Extension developers can hook into the new prompt.autocomplete event to insert custom completion options into the prompt input box. Inside $.model.complete, mods can now enable prompt caching by passing text blocks to prompt and system with cache: true. Additionally, the agent.spawn hook now passes workflow agents alongside their run and index details, allowing mods to block specific workflow executions.
• Plugin Installation Options: The CLI command claude plugin install now supports a --marketplace <source> flag, which adds the specified marketplace under standard policy checks prior to installing the plugin.
• Custom Backoff Delays: Operators can set the CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS environment variable to increase the base retry delay when handling HTTP 529 overloaded errors.

Security and path handling fixes:

• Network Path Bypass: Fixed a vulnerability where UNC network paths bypassed file-read permission prompts under PreToolUse hook approvals and auto mode.
• Sandboxing and Path Isolation: Commands running inside managed sandboxes can no longer read staged /ultrareview files stored in ~/.claude/seed-admin. macOS and Windows file reads for notebooks and PDFs now prevent mid-read symlink replacements from accessing unapproved files outside the target directory.
• Credential Injection & Policy Protections: Fixed managed read-deny sandbox paths so mid-session updates properly revoke project grants and halt file-based credential injection. On-disk settings caches that have been tampered with can no longer disable the built-in policy plugin if settings fail to fetch.
• Windows Short Name Handling: Deletion operations like rm -rf targeting Windows 8.3 short names or alternate path spellings of drive/home folders are now correctly identified and handled as home directory deletions.

Tooling, session, and API fixes:

• Proxy Rules: Fixed an issue where setting HTTPS_PROXY caused internal API calls for sign-in, policy, feedback, and artifacts to ignore NO_PROXY settings.
• MCP Tool Length: Tools provided via Model Context Protocol (MCP) with names longer than 128 characters are now omitted individually with a descriptive error message rather than causing the entire request to fail.
• File Handling: @-mentioned text files larger than 256KB are no longer silently dropped; Claude is now informed of the file size and directed to read the file in segments. Passing non-contiguous page lists (such as “6,9,15”) to PDF page reads now returns an error instructing the user to request single pages or continuous ranges.
• Session State: Restored plan mode state when resuming via claude --resume or /resume. Fixed issue where scheduled tasks created after /resume, /branch, or /clear failed to execute, and restored background /loop wakeups after process restarts.

Why it matters#

Adding an explicit effort parameter to sub-agents gives developers better control over API spend and execution time across complex, multi-agent workflows. Similarly, prompt caching support inside $.model.complete cuts latency and token costs for custom mods that reuse large system prompts.

On the security front, patching UNC path read bypasses, mid-read symlink swaps, and Windows 8.3 path handling hardens the agent execution sandbox against unexpected file system access. Fixing the NO_PROXY behavior ensures enterprise environments routing traffic through corporate proxies keep sign-in, policy checks, artifact fetching, and user feedback on internal routes when configured. These improvements build directly on earlier platform updates covering command safeguards and MCP updates as well as terminal layout and diagnostic controls.

What to do next#

Developers using Claude Code in automated environments or building custom mods should update to the latest build to apply these security patches and workflow controls:

1. Update Claude Code to v2.1.292 using your standard installation manager.
2. In sub-agent calls, pass the effort parameter to limit or expand compute depth based on task complexity.
3. Update custom mod calls using $.model.complete to set cache: true on static prompt and system blocks to reduce latency.
4. If using proxies, verify that your NO_PROXY configuration correctly accounts for sign-in, policy, feedback, and artifact endpoints.
5. Adjust any automated scripts relying on PDF reads to query single pages or continuous ranges rather than comma-separated lists.

Limitations & Caveats: MCP tools with names longer than 128 characters will now be excluded from the model’s available tools; tool developers must shorten their identifiers to make them accessible. Additionally, passing non-contiguous page lists to the Read tool for PDFs now raises an explicit error rather than silently returning only the first page in the list.

Sources#

• Claude Code v2.1.292 Release Notes