↓ Skip to main content
  1. Articles/

Claude Code v2.1.330: Enterprise Sign-In, 30-Day /loop Runs, and Safer rm and Worktree Handling

·1349 words·7 mins·
Florent Clairambault
Author
Florent Clairambault
CTO & software engineer — writing daily about spec-driven development and agentic coding

Version 2.1.330 is the newest entry in the official Claude Code changelog. I read the full bullet list from the raw file. I did not run it, and the changelog gives no release date. What follows is a reading of release notes, not a field test.

Last week’s releases were about bounding the agent. This one continues that theme and adds two things that matter to teams: organization-managed sign-in and much longer-lived scheduled work.

Enterprise sign-in
#

The first bullet: after claude auth login, Claude Code can now sign in to organizations managed in Anthropic’s Enterprise Admin Console. It applies the organization’s settings and uses the organization’s cloud provider or gateway.

For anyone rolling Claude Code out to a team, this is the interesting line. Today the usual path is distributing settings files, environment variables and provider credentials. Pulling configuration from the account the developer signs in with moves that to a central place. The changelog does not say which settings travel this way, so check your own policy keys before assuming parity with managed settings files.

Related changes in the same release:

  • CLAUDE_CODE_FULL_SIGN_IN_LINK=1 prints the full sign-in link when the short one errors or a script needs to read values from it. The default sign-in link is also now “much shorter”.
  • Startup under forceLoginOrgUUID no longer gets refused after a single temporary server error.
  • /install-github-app and /design-login now show an on-hold notice with an appeal link for accounts on hold, instead of a generic failure.

30 days instead of 7
#

Recurring scheduled tasks and /loop now run up to 30 days, up from 7, as long as their session keeps running, and they say when they end. Tasks created before this release still end at 7 days.

That caveat is easy to miss. If you have a loop watching a long migration or a flaky pipeline and you expect the new limit, it only applies to tasks created on this version. Recreate old ones.

Also note the condition: the session has to keep running. A 30-day loop is a long-lived process on a machine that must stay up, not a cloud job. For the cloud-hosted alternative, see Claude Code Routines.

Fewer ways to delete the wrong thing
#

Two changes in the same family:

  • More rm commands aimed at the root, the home directory, a top-level folder or the working folder now ask for approval. Where nobody can answer (headless runs), they are refused.
  • ExitWorktree and the worktree exit dialog no longer delete a worktree that another running Claude Code session was working in.

The second is a real hazard if you run parallel sessions in git worktrees, which is the standard pattern for agentic work. Two agents, one worktree, one of them tidies up: the other loses its working directory. It is fixed here, but if you rely on parallel worktree sessions, upgrade before your next long run.

Other data-integrity fixes in the same spirit:

  • Changes to user settings being lost when several Claude Code processes saved at the same time.
  • Task dependencies and metadata lost when several task updates ran concurrently.
  • Plugins disappearing from the installed list when several claude plugin commands ran at once.
  • Incomplete backups of ~/.claude.json displacing good ones when the disk was full.
  • claude -p exiting 0 with no message when output went to a full disk. It now prints an error and exits 1, though the changelog notes a full NFS export may go unnoticed.

The common theme is concurrency. These are the bugs you hit when you stop running one session and start running several.

MCP and permissions
#

  • A change to deniedMcpServers: name entries now hide only the named claude.ai connector, not others at the same URL. To block all of them, deny the exact URL. If you maintain a denylist, review it.
  • Fixed claude.ai connectors with different names being dropped when they share a server URL.
  • Fixed a subagent-declared MCP server being shut down when one subagent finished while others still used it.
  • Fixed turns failing with “JSON schema is invalid”, and session files growing by megabytes, when many MCP tools have an input schema the API refuses.
  • /mcp now groups plugin servers under “Plugin MCPs” and --mcp-config and built-in servers under “Other MCPs”.
  • Fixed text and images added when approving a permission prompt being dropped when the approved command then failed.

easel and Code Mode
#

The release adds easel, described as a built-in mod that draws charts when asked and under MCP metrics results. It is available on Pro, Max and, where available, Team and Enterprise, on macOS and Linux, and automatic charts can be turned off in /config. A separate fix covers the chart tool failing in Code Mode, and another handles an “MCP server blocked by enterprise policy: easel” warning on every start when an allowlist or denylist excludes it (the tool stays blocked).

Code Mode, which landed experimentally earlier this month, gets a long list of fixes: a pause when code read hundreds of files under path-scoped rules, ignoring an MCP tool’s request to end the turn in auto and bypass modes, not restoring a read file after Esc and /compact, and /context undercounting system tools. Claude is also now told how to look at images and PDFs in Code Mode. A subagent’s script making thousands of tool calls freezes the terminal for far less time. This reads like a feature being hardened in public. Keep it away from anything critical until the fix rate drops.

Self-hosted runners
#

A large share of the bullets target the self-hosted runner: --start-dir / SELF_HOSTED_RUNNER_START_DIR, two new /metrics for order-to-registered and order-to-first-session latency, concurrent spawn-runner hooks (4 by default, --hook-concurrency to change), and a stop-at-once exit code 108 for checkout hooks.

Two behavior changes deserve a log-alert review:

  • A 403, such as from an IP allowlist, was previously logged as a bad secret or token. Most such lines now lead with the server’s reason, so alerts matching the old wording need updating.
  • Bash commands skip the file diff when nothing needs it, so transcripts can lack it. Set bashEditDiffEnabled to true in managed settings to get every diff. If you audit agent edits from transcripts, you need this.

Smaller items worth knowing
#

  • Fixed a freeze of up to several minutes right after compacting a very long conversation, and WebFetch freezing a session for minutes on some pages.
  • Fixed /rewind sometimes restoring an older file version when queued prompts or stacked skills ran together.
  • Linux: Bash commands no longer wait seconds in slow repositories with large ignored folders such as node_modules.
  • Skills run with a model setting now leave subagents, forks, compaction and /recap on the session’s model.
  • Prompt hooks accept aliases like haiku or sonnet in their model field.
  • Unrecognized model names containing “sonnet” or “haiku” now use the shorter system prompt. CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 restores the full one.
  • 2.1.329’s removal of the automatic retry when max_tokens exceeds a model’s or gateway’s limit was reverted.
  • /ultrareview now stops and names both folders if the shell moved into another repository, and on Team and Enterprise it says how much monthly spend limit remains when that is less than a review usually costs.
  • Plus a block of VS Code fixes and additions, including /fork, /workflows and several Enter-key safety fixes.

Why this matters for spec-driven work
#

If your workflow is spec in, implementation out, you want the agent to run for a long time, in parallel, without supervision. Each item above is a constraint that shows up only at that scale: 30-day loops, shared worktrees, concurrent settings writes, headless rm, centrally managed org identity. The tool is being shaped for operators, not for someone watching every diff in an editor pane.

What I did not check
#

  • I read the changelog only. I did not run 2.1.330 or test any flag.
  • The changelog gives no release date.
  • Which organization settings apply through Enterprise Admin Console sign-in is not specified in the entry. Verify against the documentation.
  • easel and the self-hosted runner features are described only in one-line bullets.

Sources
#

AI Tools

Claude Code v2.1.309 Is a Sandbox and Permissions Hardening Release. Read It Like One.

Claude Code v2.1.309 ships roughly 60 changelog bullets, and the heart of it is closing gaps in the sandbox, Bash permission checks and worktree isolation guard. It also makes claude ultrareview require workspace trust and denies dangerous rm calls outright in bypass mode on Desktop and VS Code. Changelog-only analysis; not run.

Oct 23, 2026 · 6 min read

AI Tools

Claude Code's June Sprint: Four Releases, the Ultracode Rename, and Enterprise Version Locks

Between June 2 and June 4, Anthropic shipped four Claude Code releases — v2.1.160 through 2.1.163. The headline change: the dynamic-workflow trigger keyword is now ‘ultracode’, not ‘workflow’. Enterprise admins get version range enforcement. Hooks can now talk back. Here’s what actually matters.

Jun 5, 2026 · 6 min read

AI Tools

Claude Code v2.1.129: Bedrock Tiers, Smarter MCP, and a Gateway Reversal

Claude Code v2.1.129 shipped quietly on May 6. For most users it’s invisible. For Bedrock enterprise shops, MCP server operators, and anyone using third-party model gateways, it changes real behavior — and one change is a deliberate reversal of v2.1.126.

May 15, 2026 · 6 min read

AI Tools

Claude Code v2.1.329: Code Mode Gets Guardrails, /tasks Becomes a Control Panel, and /code-review Stops Guessing

Claude Code v2.1.329 is a hardening release: Code Mode gets memory limits and clearer failure semantics, /tasks now covers scheduled tasks, /loop wakeups and forks, and /code-review reads the actual pull request via gh pr diff or says it can’t. Based on the official changelog only; I have not run this version.

Nov 20, 2026 · 7 min read

AI Tools

Claude Code v2.1.328: Auto Mode Reaches Bedrock, Foundry and Google Cloud, and MCP Hooks Can Fail Closed

Claude Code v2.1.328 extends auto mode to Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry and the Claude apps gateway, adds “onFailure”: “block” for mcp_tool hooks, and narrows /code-review to bugs and CLAUDE.md violations. Based on the official changelog only; I have not run this version.

Nov 19, 2026 · 6 min read