Claude Code release notes: 2.1.219–2.1.220 bounds workspace growth

Claude Code 2.1.219 makes it easier to see when an agent's workspace grows. A new hook fires when a workspace root is added, the sandbox can deny network traffic outside an approved list, and nested subagents get new default boundaries. Version 2.1.220 follows as a maintenance-only release.

That is more useful than another model-performance list. When Claude can enter more directories and delegate further down an agent chain, the boundary needs to travel with the job.

Source: Claude Code changelog for 2.1.219 and 2.1.220.

Claude Code 2.1.219 makes new workspace roots traceable

A hook is a user-defined event handler that runs a command, HTTP request, or LLM prompt at a specific point in the Claude Code lifecycle. The new DirectoryAdded hook fires after /add-dir or the Agent SDK control request register_repo_root registers another working directory during a session.

The hook does not block the addition beforehand. It gives the team a clear place to log it, send a notification, or trigger a follow-up policy check. That matters when an agent starts in one bounded repository but later needs a shared component, document library, or separate test environment.

Sources: Claude Code 2.1.219 on GitHub. Anthropic's hooks reference.

A strict network list makes the sandbox boundary real

A sandbox is an isolated runtime that limits which files and network hosts a command can reach. In 2.1.219, sandbox.network.strictAllowlist can deny connections to hosts outside the allowlist without opening an exception prompt.

This suits teams that want Claude Code connected to real build tools, package registries, and internal APIs without giving every run unrestricted egress. Keep secrets in environment variables or a secret manager, use scoped keys, and record the hosts and workspace roots each run actually touched.

Sources: Claude Code 2.1.219 release notes. Anthropic's Claude Code sandbox guide.

Nested agents reopen with a visible operating limit

A subagent is a specialized Claude assistant with its own context and tool boundaries. Version 2.1.219 again allows nested subagents to a default depth of three. Teams that want to disable nesting can set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1.

Dynamic workflows now also default to the medium size guideline, which aims for fewer than 15 agents. The technical room has expanded since 2.1.217 disabled nesting by default. A team's operating limit should still match the number of results people can review and approve.

Version 2.1.220 documents no new user-facing features. Anthropic lists only bug fixes and reliability improvements, while the npm registry shows 2.1.220 as the current package. There is no reason to pad that point with guesses.

Sources: Anthropic's subagent guide. Claude Code 2.1.220 on GitHub. Current Claude Code package in the npm registry.

Try this prompt this week

Human step: Update through your approved distribution path, run claude --version, and confirm 2.1.220. Open the correct codebase and use /status to check which settings sources Claude Code loaded. Then run the prompt without changing any settings.

Sources: Claude Code changelog. Anthropic's settings and /status documentation.

Read CLAUDE.md and every loaded .claude/settings*.json file.
List registered workspace roots, sandbox network rules, DirectoryAdded hooks, and agent definitions.
Map where this session can add a directory or spawn nested subagents.
Compare that map with the actual task and flag anything you cannot verify.
Propose one boundary contract for roots, hosts, agent depth, and human review.
Make no changes. Finish with a short boundary receipt.

Good output should:

• Name the workspace roots and network hosts the actual task needs.
• Separate loaded rules from assumptions and proposed changes.
• Show who reviews a new workspace root or a result from a nested agent.

This is practical Tool Forge work: let the agent use real tools and directories, but record every expansion in the same run receipt as the result.