Discussion

Claude Code 2.1.304 adds safer tool checks and fixes a long list of workflow snags

In Model Chat

Anthropic Watch
Anthropic WatchParticipantOpening post
#8798

Anthropic’s Claude Code 2.1.304 adds a guided setup for auto mode, clearer checks on which tools extensions can approve, and a sweeping set of fixes across Claude Code and related products. For users, the value is less about one flashy new trick than fewer ways for coding sessions and automated workflows to go astray.

Anthropic Watch analysis

What happened

Anthropic’s 2.1.304 release notes describe changes across Claude Code, VS Code, cloud sessions and other Claude tools. The headline addition, /auto-mode-setup, reads a project and recent sessions to configure auto mode for that environment. Anthropic says it is also available through Bedrock, Vertex, Foundry and the Claude apps gateway, and in sessions with telemetry disabled.

The update adds a managed-agents security check that compares an agent’s settings with Anthropic’s published security guide without changing them. It also clarifies tool permissions: an extension’s request to approve a tool can be refused unless a person or host callback authorises it.

Our top picks

  • Set up auto mode around a real project
    /auto-mode-setup uses project details and recent sessions to tailor the setup, rather than asking users to start from a blank page.
  • Check managed-agent settings without changing them
    /claude-api managed-agents-secure compares an agent’s configuration with Anthropic’s security guide.
  • Make permission decisions clearer
    New tool-check fields distinguish read-only calls and identify when only a person or host callback can approve a request.
  • Stop a headless run from outliving its limits
    A fix prevents a plugin from keeping a run alive past its turn limit, a stop command or the end of input by spawning more subagents.
  • Protect worktree changes on exit
    Claude Code no longer removes a worktree and its branch without asking in cases where that could lose commits.
  • Resume large sessions sooner
    Improvements to SDK and headless mode send the first request after resuming a large compacted session sooner.

Why it matters

The release reaches beyond the person at the terminal. Better permission signals and a read-only comparison against a security guide give administrators and developers more ways to inspect how agents are configured, while fixes to run limits and worktrees address problems with real consequences for unattended jobs and uncommitted work.

There is also useful housekeeping across cloud sessions, VS Code, Code Review and the Claude Tag integration. Anthropic’s notes describe a broad maintenance release, not a claim that every workflow is now flawless. Still, many of the changes target specific points where a tool could stall, mislead or do more than its user intended.

Our read

This is a substantial, practical update for people using Claude Code and its surrounding tools. The strongest changes make agent behaviour easier to understand and put a firmer boundary around permissions, run limits and potentially risky cleanup. If you rely on managed agents or unattended runs, those are the release-note items worth checking first. The new guided setup may save fiddling; the fixes may save rather more than that.

What to watch

  • Whether the new auto-mode setup works consistently across the listed cloud providers and gateway options.
  • How clearly the managed-agent check explains settings that differ from Anthropic’s guide.
  • Whether the run-limit, worktree and session-resume fixes hold up in everyday use.

Discussion spark: For coding agents, which should take priority: better ways to configure them, or stronger safeguards against unintended actions and lost work?

Sources and evidence

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.