Lab

Lab: dotclaude-lab

The plugin dotclaude-lab holds tools for plugin authors. It has four bins and one skill. It registers no hook and adds no tool. It works without the core plugin, and the repository justfile calls its bins. Nothing in it changes your own Claude Code config.

At a glance

Item	Value
Install	/plugin install dotclaude-lab@dotclaude, with Node.js 22.18 or later
Skills	eval-compare, user-invoked only
Bins	dotclaude-sandbox, dotclaude-capture, dotclaude-schema, and dotclaude-eval-compare, in bin/
Hooks	None
Options	None. The bins take flags
Limits	dotclaude-eval-compare adds --max-cost-usd 5 for a run that sets no bound
Problems	See Troubleshooting

Bins

dotclaude-sandbox

Runs Claude Code in a config directory apart from your own, so a test cannot change your settings, sessions, or plugins.

dotclaude-sandbox [--dir DIR] [claude arguments...]
dotclaude-sandbox [--dir DIR] --clean
dotclaude-sandbox [--dir DIR] --check-frontmatter ...

• --dir DIR sets the sandbox directory. The default is dotclaude-sandbox in the temp folder. The directory holds config/ (CLAUDE_CONFIG_DIR), project/ (an empty git repository), and home/ (the HOME of claude).
• The bin loads no plugin by itself. Pass --plugin-dir DIR once for each plugin folder.
• --clean removes the sandbox.
• --check-frontmatter runs one headless session with a debug log, and exits 1 when the log names an unknown field.
• The login is CLAUDE_CODE_OAUTH_TOKEN, or your own login token, which goes to claude in its environment only.

See Sandbox for the full guide.

dotclaude-capture

A local logging proxy for the requests of Claude Code.

dotclaude-capture [--port 8787] [--out DIR] [--upstream URL] [--no-count] [--count FILE]

• --upstream URL sets the upstream. Without it, the upstream is ANTHROPIC_BASE_URL, then https://api.anthropic.com. The proxy stops when the upstream is its own address, because requests would loop.
• --out DIR sets the folder for the request bodies and the report. The default is dotclaude-capture in the temp folder.
• It never writes a request header, because the headers carry the credential.
• For the first request with tools, it sizes each part with the count_tokens endpoint. --no-count turns this off, and --count FILE sizes FILE instead.

dotclaude-schema

Writes the JSON Schema of settings.json from the installed Claude Code bundle.

dotclaude-schema --out DIR [--bundle FILE] [--version X.Y.Z]

--out is required. The bin reads the bundle and never changes it. See Settings schema.

dotclaude-eval-compare

Runs claude plugin eval on two plugin folders, one after the other.

dotclaude-eval-compare A B [claude plugin eval options...]

Each run scores a plugin arm against the same no-plugin baseline arm, so the WITH column of the two tables shows which folder scores better. Each run costs money, so the bin adds --max-cost-usd 5 when the options set no bound. The bin adds no --scaffold and no --allow-tools, because each one runs the scripts or the tools of a suite as you. Add them in the options only for suites that you trust. just eval-variant adds both, because the suite of this repository needs them. The results stay in evals/results/ of each plugin folder. See Evals.

Skill

eval-compare is user-invoked only (disable-model-invocation: true), so it has no listing line. It confirms the two folders with you, runs dotclaude-eval-compare, and compares the WITH columns case by case. The skill grants no tool, so Claude Code asks you before each run. When a run stops at the cost bound, Claude stops and tells you, and it does not raise the bound without you.

Troubleshooting

See Sandbox for the sandbox problems, and Troubleshooting for the list.

Symptom	Cause and fix
dotclaude-schema stops with an error that names a version	The script has no type table for that Claude Code version, and it writes no file.
dotclaude-eval-compare prints a usage message and exits 2	One of the two folders is missing, or it starts with -.

Related pages

• Sandbox
• Settings schema
• Evals
• Development