Skip to content

Claude Code

Claude Code 2.1.233 splits proxy spend by user

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: Claude Code versions and changes →

Claude Code 2.1.233 can now pass each signed-in developer’s identity from the Claude apps gateway to a proxy behind it. That proxy can then split spend by person, where before it saw one figure for the whole team. The setting is called forward_user_identity. It is opt-in, applies to Anthropic upstreams and sends the identity as request headers, according to the August 14 release notes on Github.

The release has about twenty entries. Most are fixes, but two need attention: a Windows credential-leak fix, and the removal of the todo and task tools on the newest models. The second one changes how sessions behave unless you turn the tools back on.

Who spent what behind the gateway

The AWS documentation describes the apps gateway as a self-hosted service between Claude Code clients and the model provider. Developers sign in through the corporate identity provider and never hold API keys or cloud credentials. The gateway holds the upstream credential. It ships inside the same claude binary and starts with claude gateway --config gateway.yaml. Because the gateway holds the only upstream credential, a proxy behind it would most likely see every request as coming from the gateway itself. The new headers give the proxy the person behind each request.

The gateway already tracks usage on its own side. AWS’s sample README describes per-user cost tracking as OTLP telemetry stamped with user identity, sent to Datadog, Splunk, CloudWatch or any OTLP backend. The same README lists daily, weekly or monthly spend caps per user, group or org, and the gateway blocks requests once a cap is reached. forward_user_identity puts the identity on the request itself. It looks aimed at companies whose cost accounting runs in a proxy rather than in a telemetry backend.

The release also fixes gateway error forwarding. When Vertex, Foundry or Claude Platform on AWS upstreams return 400 or 413 errors, the client now gets the upstream’s own message. The same change fixes a bug with auto-compact on the apps gateway.

Linux limits and the Windows rollback

On Linux, Bash tool commands can now run inside a memory cgroup, the kernel mechanism that caps the resources of a group of processes. You turn it on with CLAUDE_CODE_TOOL_MEMORY_LIMIT. Anthropic’s stated aim is that a runaway build can’t stall the session. A separate fix covers idle Linux sessions with sandboxing enabled, which sometimes kept one CPU core at 100%.

On Windows, paths written with the NT \??\ device prefix got past UNC path validation. UNC paths point at network shares, and opening one can make Windows send NTLM credentials to the remote host. The release notes say the fix closes an NTLM credential-leak vector.

The rest of the Windows work cleans up after 2.1.232. That release made auto mode stop repeatedly for manual approval on ordinary commands like cd <dir> && <command> > file. This is now fixed. Two of its Bash permission changes, covering Cygwin-style symlinks and input redirections (< file), have already been reverted, one release after they shipped. Anthropic says a narrower version will return in a later release.

MCP, hooks, skills and the rest

  • The --worktree flag and the claude agents view accept GitLab merge request URLs, and MRs display as !N.
  • The GitHub app setup tip no longer appears in repositories whose origin remote is on gitlab.com or bitbucket.org. The enterprise marketplace tip now covers non-GitHub internal git hosts.
  • CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS sets how long WebFetch caches URLs within a session. The default is still 15 minutes.
  • MCP v2 connections no longer keep reopening the subscriptions/listen stream against servers that cut long-held streams on a fixed timeout, such as serverless hosts.
  • Notification hooks now fire for permission prompts under Claude Desktop and VS Code.
  • Cloud sessions are no longer occasionally marked as lost when the environment shuts down while Claude waits on a permission prompt.
  • Bundled skill aliases like /checkup and /review no longer report “Unknown command” in -p mode or with plugins or MCP loaded, when a user or project skill shadows the bundled one.
  • Argument values in skill and command substitution are no longer re-expanded as template markers.
  • claude self-hosted-runner starts faster. It creates the session branch without rewriting the working tree, and two server round trips no longer block the agent’s launch.
  • claude plugin validate checks a bare .claude/skills directory and reports SKILL.md files whose frontmatter fails to parse.
  • In screen reader mode, the /effort selector renders as a numbered list with a typed-number prompt, and hint and dialog text is no longer clipped.
  • Print mode writes a [claude-code:unrecognized_model] line to stderr when a request goes out for a model ID Claude Code doesn’t recognize. Mapping that ID with modelOverrides silences it.

Where did the todo tools go?

TaskCreate, TaskGet, TaskUpdate, TaskList and TodoWrite are unavailable on Opus 4.8, Sonnet 5, Fable 5, Mythos 5 and newer models. On those models, a session will stop keeping a task list through these tools. Any hooks or permission rules that match these tool names will probably go quiet as well. Older models keep the tools.

To restore them on the newer models, set CLAUDE_CODE_ENABLE_TODO_TOOLS=1.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.