Claude Code 2.1.233 pulls the todo tools from Opus 4.8 and Sonnet 5
Newest models now plan without a checklist tool; one env var brings it back. Also: GitLab worktrees, per-user spend headers, and a Windows security fix.
Copy markdown
Todo tools are gone on the newest models
Claude Code 2.1.233 removes the built-in TaskCreate/Update/List and TodoWrite tools from Opus 4.8, Sonnet 5, Fable 5, Mythos 5 and newer — those models now plan inline instead of driving a checklist. If your hooks or workflow lean on the todo list, set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to bring it back.
GitLab merge requests get worktrees
The --worktree flag and the claude agents view now accept GitLab merge-request URLs, showing MRs as !N — the same MR-to-worktree flow GitHub users already had. GitLab-hosted teams can now spin a review branch straight from a merge request.
Attribute Claude spend per user
A new opt-in forward_user_identity apps-gateway setting forwards the signed-in user's identity as signed headers, so a proxy behind the gateway can bill or track Claude spend per person. Handy for teams routing Claude Code through one shared gateway.
Tune WebFetch caching, cap runaway builds
CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS now lets you change the WebFetch URL cache window (still 15 minutes by default), and an opt-in Linux memory cgroup (CLAUDE_CODE_TOOL_MEMORY_LIMIT) stops a runaway build from stalling the whole session.
A Windows credential leak, patched
The release closes a hole where Windows paths using the NT \??\ device prefix bypassed UNC validation — an NTLM credential-leak vector. It also stops idle sandboxed Linux sessions from pinning a CPU core at 100% and repairs 2.1.232's auto-mode regressions on Windows.
Elsewhere: OpenAI's Ultrafast mode hits 750 tok/s
OpenAI previewed an Ultrafast API tier that runs GPT-5.6 Sol up to 14x faster — around 750 output tokens/second — on Cerebras hardware. It's a gated, limited preview for now (incident response, finance, support, commerce), not general availability.