Claude Code 2.1.314 adds write protection and MCP tasks

Anthropic’s Claude Code 2.1.295, dated October 8, adds two features with practical effects for automated workflows and terminal status displays. First, command and HTTP hooks can set onFailure: "block" so the associated action is blocked if the hook cannot start, times out, or exits unexpectedly. Second, terminals implementing Program Status Protocol (OSC 7501) can show whether Claude Code is working, waiting for the user, or done. (https://code.claude.com/docs/en/changelog)

Fail-closed hook behavior

The new onFailure: "block" setting gives teams an explicit fail-closed option for command and HTTP hooks. This can be useful when a required check must succeed before an action continues. Review hook timeouts, expected exit codes, and recovery behavior before enabling it in shared workflows; an unavailable hook will stop the action by design.

Version 2.1.294 also improves how Claude judges prompt hooks on Stop and SubagentStop when they are written as instructions, and fixes prompt or agent hooks phrased as instructions that previously could allow the very behavior they were meant to block. (https://code.claude.com/docs/en/changelog)

Other changes in 2.1.295

The changelog also says the /copy picker now includes quoted text without its > markers, and plugin-management commands warn when the settings file they write to does not load. It changes the MCP tool-description limit loaded through tool search from 2,048 to 16,384 characters and updates local stdio MCP protocol negotiation. Check the full version notes for platform-specific fixes before updating a managed deployment.

October 19-30 follow-up: write protection and MCP Tasks

Claude Code 2.1.314, dated October 30, adds permissions.blockWritesOutsideWorkingDirectories. When enabled, Write, Edit, and NotebookEdit refuse paths outside the configured working directories in every permission mode. This provides a write boundary but does not replace review of working-directory configuration or other tool permissions. (https://code.claude.com/docs/en/changelog)

Anthropic's MCP documentation describes MCP Tasks as a public beta for supported Claude Code CLI sessions on version 2.1.304 or later. A compatible server can accept a long-running job and return control so Claude can continue, then deliver the result when the job completes. Support depends on the MCP server and session type; the documentation lists terminal CLI, Agent SDK, and GitHub Actions sessions, and excludes the VS Code extension and desktop app. Remote server tasks can continue after Claude Code exits, while a stdio server process stops when the CLI exits. (https://code.claude.com/docs/en/mcp)

The MCP guide also documents elicitation, which lets a server request structured form input or ask whether to open a URL. Supported modes depend on the protocol revision and server implementation. Treat any requested input or external page as a separate approval decision, especially in non-interactive sessions. (https://code.claude.com/docs/en/mcp)

Sources

• Claude Code Docs — official changelog, including versions 2.1.294-2.1.314: https://code.claude.com/docs/en/changelog
• Claude Code Docs — MCP Tasks and elicitation: https://code.claude.com/docs/en/mcp
• Anthropic — Claude Code repository changelog: https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md