Claude Code 2.1.233 adds native GitLab merge request support
The Aug 14 release also disables the built-in todo tools on newer models, adds Bash memory limits on Linux, and patches a Windows credential-leak path.
Copy markdown
GitLab merge requests, first-class
The `--worktree` flag and the `claude agents` view now take GitLab MR URLs and render them as `!N`, the way GitHub PRs show as `#N`. If your team ships on GitLab, Claude Code's agent and worktree flow finally speaks your VCS end to end.
The todo tools are off on newer models
TaskCreate/Get/Update/List and TodoWrite are now disabled on newer models, which track tasks natively. If a hook, script, or MCP flow leans on them, set `CLAUDE_CODE_ENABLE_TODO_TOOLS=1` to bring them back.
Cap runaway Bash memory on Linux
A new `CLAUDE_CODE_TOOL_MEMORY_LIMIT` adds cgroup memory limits to Bash tools on Linux — handy for sandboxed CI. It lands with a fix for idle sandboxed sessions that were pegging a full core at 100% CPU.
A Windows credential-leak patch
The update closes an NTLM credential-leak vector via crafted device-prefix paths on Windows, on top of the PowerShell and symlink permission-bypass fixes from 2.1.232. Update now if you run Claude Code on Windows.
Per-user spend tracking and cached WebFetch
An opt-in `forward_user_identity` gateway setting attributes spend per user for teams on the apps gateway, and WebFetch now caches results for 15 minutes by default — tune it with `CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS`.
Elsewhere: OpenAI previews a 750 tok/s GPT-5.6 Sol
OpenAI opened a gated preview of an Ultrafast API tier for GPT-5.6 Sol that hits up to 750 output tokens/sec — roughly 14x Standard — on Cerebras hardware. No pricing or GA date yet and access is limited, but it points at real-time agent latencies.