Claude Code v2.1.290 Simplifies Managed Agent Onboarding and Upgrades Plugin Hooks

Managed Agent Onboarding and CLI Usability#

This release introduces command-line support for setting up the Managed Agents pattern. Developers can now run /claude-api managed-agents-onboard <url> to configure files via ant apply, or deploy Console quickstarts (such as deep-researcher) with the ant CLI using /claude-api managed-agents-onboard <quickstart-name>.

Additionally, session management is more flexible: the commands claude attach <name> and claude logs <name> now accept a partial session name instead of requiring the full session ID. A new “Deny” button has also been added to the Claude apps gateway sign-in page to instantly terminate pending terminal sign-in requests.

Enhanced Plugin and Mod Hooks#

Building on earlier plugin capability expansions, v2.1.290 refines how custom mods interact with the tool execution lifecycle:

• Tool Tracking: The result of a mod’s turn.step hook now includes serverToolUses, detailing the specific tool calls executed by the advisor (including ID, name, input, start, and end parameters).
• Subagent Permissions: The tool.check plugin hook event now receives an agentId, allowing mods to differentiate permission checks originating from a subagent versus the main session.
• Enterprise Guardrails: The tool.check hook can now read a ceiling property, which specifies the explicit approval level an organization requires for a given tool.
• Theme Typings: The plugin hook typings now feature ThemeKey and Color types, enabling editors to resolve theme colors for mod drawings.
• Validation: The claude plugin validate command now outputs whether registered gating hooks include a .catch block, visible under the gatingHooks field when using the --json flag.

Reliability and Security Fixes#

Version 2.1.290 resolves several stability regressions, particularly in large codebases and multi-turn conversations:

• Long Session Failures: Fixed an issue where long sessions containing hundreds of images became unprocessable by the model.
• Content Filter Retries: If an output content filter stops Claude mid-thought, the system now retries the request once rather than immediately failing the turn.
• Prompt Caching: Resumed subagents and teammates no longer lose their prior thinking state or prompt cache when a message is received mid-run.
• WebFetch Limits: The WebFetch tool no longer silently truncates web pages past 100,000 characters. It now alerts the user to the unread volume and accepts an offset parameter to continue reading.
• Workspace Isolation: Fixed a security bypass where symlinked project configuration files (like CLAUDE.md, AGENTS.md, or custom rules) pointing outside the working directories could load despite permissions.blockReadsOutsideWorkingDirectories or active Read deny rules.
• Plan Mode Guardrails: Plan mode will no longer allow the auto-mode classifier to approve non-read-only connector tools that carry a server-pushed ask policy.