Context Lab

The lab where agent context is designed, and the package that ships it to every host in the fleet.

No clone required. marketplace add fetches the whole repository for you, so all three commands are typed as-is:

claude plugin marketplace add luutuankiet/context-lab
claude plugin install context-lab@context-lab
~/.claude/plugins/marketplaces/context-lab/install.sh

The third command is the one nothing will remind you about. The first two give you the skills; the plugin model has no post-install step of any kind, so the config tier — statusline, hooks, settings, the user-memory import — is applied only by running the installer out of the fetched clone. There is no credential to put on a machine.

The installer is a one-time bootstrap, not part of the update loop. To keep a host current you run one command, and it refreshes skills and user memory together — the memory import points into the directory this re-fetches:

claude plugin marketplace update context-lab

Re-run the installer only when the installer's own scope changes — a new owned settings.json key, a new hook, a new shell export. Editing memory or a skill never needs it:

~/.claude/plugins/marketplaces/context-lab/install.sh
~/.claude/plugins/marketplaces/context-lab/install.sh --check   # verify, mutate nothing

⚠️ marketplace update deletes the directory and re-clones rather than pulling. The path is stable, so the import survives — but an update that dies partway (no network, no git auth) leaves a host silently running with no memory at all, because an unresolved @-import produces no error. --check warns when the target is missing; it is the only thing that will tell you.

Developing on it instead? Clone it anywhere you like and run that clone's install.sh. Nothing follows the clone. Hooks, skills and the statusline contributor reach a host as plugin content addressed by ${CLAUDE_PLUGIN_ROOT}, so what runs is always the commit the marketplace published — never your working copy. Editing a file here changes nothing on any host until you push and re-run marketplace update (ADR 0014).

What is in here

tree	role	how it reaches a host
claude/	the payload — settings keys, the memory import, the hook script, the statusline dispatcher	mostly as plugin content; only the dispatcher is written to ~/.claude/
skills/	the distributed skills collection — only stable/ ships	as plugin content, enforced by "skills": ["./skills/stable"]
statusline.d/	this repo's statusline contributors — one executable per segment	discovered every render; no registration
hooks/hooks.json	the plugin's hook bindings	discovered by filename; no manifest key names it
install.sh	the distributor — registers plugins, merges owned settings keys, appends the memory import and the shell block	you, once per host
.claude-plugin/	the marketplace and plugin manifests	Claude Code, on install
scripts/, tests/	repo tooling and the dispatcher's own suite	nobody — never leaves this repo
docs/	the lab's own writing	humans, and agents on demand
AGENTS.md, CLAUDE.md, .claude/	this repo's own harness config	agents working on Context Lab

The marketplace publishes more than this repo

marketplace add gives you a catalogue of seven plugins, not one. context-lab is this repository; the other six are separate public repositories the manifest points at, so each is installed and updated on its own:

claude plugin install <name>@context-lab

plugin	source
context-lab	this repository
write-pr	luutuankiet/write-pr
learn-with-feedback-loop	luutuankiet/learn-with-feedback-loop
dbtcx	luutuankiet/dbtcx
looker-mcp-shim	luutuankiet/looker-mcp-shim
slides-mcp	luutuankiet/slides-mcp
skills-utils	luutuankiet/skills-utils

Installing one does not install the rest. install.sh registers the whole set from a list it holds itself, which is why re-running it is how a host picks up a newly published plugin — and why adding one to the marketplace manifest is only half the change.

The one structural requirement

Two trees must stay unambiguous: what Context Lab uses to maintain itself versus what it distributes. Otherwise the lab's own scaffolding leaks into every repo it touches.

The cut is the leading dot — .claude/ configures agents working on this repo and ships nowhere; claude/ is the payload. ⚠️ One character. Read the path twice before editing either. Why it was cut there: ADR 0001.

The naming rule

Context Lab names the lab. What it emits into any other repo carries no brand at all — a canonical shape any agent reads as "this repo is laid out well," not a methodology to learn first. If you find the string "Context Lab" inside a file destined for another repo, that file is wrong.

Documentation

Start at AGENTS.md — it is what an agent loads, and it is short enough to read in a minute.

Everything else is indexed in docs/README.md: how the installer works, the traps that have already cost somebody an afternoon, the reference detail, and the decisions behind the shape.