← Back to the briefs
Coding

Anthropic tightens path boundaries and subagent permissions in Claude Code v2.1.327

Anthropic has released Claude Code v2.1.327, adding command-line stdin input for MCP configurations, enforcing stricter working-directory checks on Bash commands, and aligning resumed subagents with session permission rules.

Tighter boundaries for tools and paths

Anthropic updated Claude Code to version 2.1.327 with stricter constraints on filesystem access and tool execution. The release alters what qualifies as inside a working directory and instructs path checks to reject select symbolic links. Configurations using permissions.blockReadsOutsideWorkingDirectories now deny Bash commands that target paths covered by Read or Edit deny rules, rather than asking the operator for approval. The same setting also prompts on more Bash commands in auto and bypass permissions modes, denying several in auto mode.

Auto mode now disregards wide package allow rules, including package runners like pnpm dlx, deferring those executions to the safety check. To keep credentials off the command line, users can pass Model Context Protocol server configuration JSON through standard input when adding servers.

Subagent isolation and worktree handling

Anthropic addressed several multi-agent coordination behaviors. A subagent resumed by a sibling message must now obey session permission settings, prompting for approval or failing outright if no operator is present to respond. In-process agent teammates assigned to a different model from the lead agent now calculate compaction using their own model and context limit rather than those of the lead.

Worktree management also received safeguards. ExitWorktree now asks before removing a worktree that holds unsaved work, refusing removal in headless or unattended environments unless explicit allow rules apply. Anthropic also corrected a fault where stopping the server in worktree mode deleted git-ignored files like .env. Diagnostic reporting expanded as well, with the status view listing broken agents, skills, and ignored permission settings.

Our assessment

Teams running autonomous or headless pipelines should verify that their subagent tasks and Bash scripts do not rely on implicit path traversal or package runner wildcard rules. If automation relies on unattended subagent handoffs, verify whether sibling agent wake-ups stall on permission prompts that lack an active responder.

Source: Anthropic original publication.

Original source

Read the original material

Anthropic (opens in a new tab)

The article date belongs to Agentic Horizon. The source publication date is listed separately. Vendor claims remain attributed; we have not independently tested the reported capability.

Search the publication

Find a story by title, topic, or keyword.

Press Escape to close

Ask HorizonYour site guide
Your guide to the publication

Ask Horizon.

Ask a question about the publication.

Privacy