claude-sandbox-doctor and claude-windows-doctor: nineteen tools, hub 1.8.0

[nrzz]

Two more tools join the toolkit, and the hub is at 1.8.0 with nineteen tools: claude-sandbox-doctor answers whether Claude Code's Bash sandbox can work on your machine and why a command ran inside it or outside it, and claude-windows-doctor answers which shell the Bash tool gets on Windows, and why.

claude-sandbox-doctor v1.0.0

202 open issues in anthropics/claude-code have "sandbox" in the title, and neither /sandbox nor /doctor says why a command ran outside the sandbox or was blocked.

• check tests whether the sandbox can start here: bubblewrap, socat and user namespaces on Linux and WSL2, Seatbelt on macOS; on Windows it says there is no Bash sandbox. With --live it also runs the real Claude Code in a throwaway config against a fake server on 127.0.0.1, so no request reaches Anthropic and no model is used, to see a command inside the sandbox.
• show lists the merged sandbox settings with the file each came from.
• why "<command>" says whether the command runs inside the sandbox, outside it (which excludedCommands entry took it out, or why an entry did not match), or not at all.
• lint finds the settings that never take effect, such as an allowedDomains entry written with a scheme, which is compared with the host name as a whole and never equals one.

What the real binary showed in CI, on six Claude Code versions (2.1.289 to 2.1.302) on Linux and macOS: 221 of 221 calls ran where why says; with bubblewrap missing, failIfUnavailable: false gives a "Sandbox disabled" warning and true refuses to start; in bypassPermissions mode the sandbox's network proxy lets every host through, so a lint rule says so. 218 tests.

claude-windows-doctor v1.0.0

884 open issues have "windows" in the title, and anthropics/claude-code#7490 (148 reactions) asks to choose the shell the Bash tool uses.

• check says which tools the model will get (Bash through Git Bash, or PowerShell) and why, lists every claude on the machine and which one runs first, and finds the setup problems behind common Windows issues: long paths, core.autocrlf, the PowerShell execution policy, variables set for one process but not for the user. On Linux and macOS it says there is nothing to check.
• fix --git-bash sets CLAUDE_CODE_GIT_BASH_PATH in your user environment when Claude Code cannot find Git Bash, and undo takes it out.
• explain <finding> says what a finding means, what Claude Code does in that case, and the fix.

What the real binary showed (2.1.295 and 2.1.299 on Windows 11, 2.1.301 on windows-latest; 129 cases): Claude Code looks for Git Bash in CLAUDE_CODE_GIT_BASH_PATH, then in C:\Program Files\Git and its (x86) twin, then two folders up from the first line of where git. So a Git installed elsewhere is found when its cmd\git.exe comes first on PATH, and missed when mingw64\bin comes first, as it does in a Git Bash window. A path in the variable that does not exist is ignored, and the variable is read from user settings but not from project or local settings. 197 tests.

Toolkit 1.8.0

• Both tools are on the setup page (npx -y github:nrzz/claude-code-toolkit) as on-demand tools whose buttons only read, and both plugins are in the marketplace: /plugin install sandbox-doctor@claude-code-toolkit and /plugin install windows-doctor@claude-code-toolkit. Each skill is user-only: about 440 and 540 tokens when you run it, and nothing in the model's request otherwise.
• The end-to-end test installs all sixteen plugins from GitHub on Windows, macOS and Linux and runs both new skills from the installed copies.

Questions about installing go in the install-help thread, and ideas for the next tools in this one. Both tools are mine; each repository has a good first issue.