lisa-setup-workstation

name	lisa-setup-workstation
description	Prepare a machine — a fresh…
allowed-tools	["Bash","Read","Skill"]

Setup Workstation: $ARGUMENTS

Get a machine ready to run agents. No checkout required.

Why this exists separately

lisa-setup-local-env reads remoteEnv.tools — a project's declared toolchain — so it needs a checkout to exist before it can answer anything. That leaves a gap at the start: a bare laptop or an empty container has no repository yet, and therefore nothing to ask.

This skill fills that gap and stops there. It describes a machine, never a project:

workstation  ->  git clone  ->  lisa-detect-tooling  ->  lisa-setup-local-env

Project-scoped tools stay in remoteEnv.tools where they belong. Nothing here reads a repository, and nothing here duplicates that manifest.

Usage

Start here — this is the entry point that presumes nothing:

npx -y @codyswann/lisa@latest workstation             # report what is present and missing
npx -y @codyswann/lisa@latest workstation --install   # install what is missing

That form needs only Node. No checkout, no agent, no Lisa install — npx resolves the package into its own cache.

Why it matters: this skill exists to prepare a machine with no coding agent and no checkout. But a skill can only be invoked by typing a slash command into an agent, and Lisa ships as a devDependency, so it does not exist until a repo is cloned and its dependencies installed. Both are exactly the state this is meant to create — so as a skill alone, the bootstrap was unreachable in the one scenario it was designed for. The CLI is the way in; the skill is the surface for people who already have an agent.

Every flag is the same either way:

lisa workstation --install --agents=claude,codex   # only these agents
lisa workstation --provider=bitwarden              # credential manager (asked if a TTY)
lisa workstation --json                            # machine-readable plan
lisa workstation --print-dockerfile                # an image that runs this same script

Inside an agent, /lisa:setup:workstation runs the identical script; and the script itself is directly runnable as node scripts/cli.mjs ….

Nothing is installed or written without --install. Exit status is non-zero when a required tool is absent or an install fails, so this is usable as a gate.

Ephemerality is not a mode

There is no Docker branch anywhere in this skill. A container gets a throwaway $HOME and dies with it; a laptop keeps its own. Same code path either way — so the rarely-run one cannot rot, which is the failure mode a second install path eventually always has.

--print-dockerfile emits an image rather than shipping a committed one, so the image cannot drift from the catalogue it provisions. It deliberately bakes in no repository: clone inside the container, then run lisa-detect-tooling and lisa-setup-local-env for that project's own tools.

node scripts/cli.mjs --print-dockerfile --agents=claude,codex --provider=bitwarden > Dockerfile
docker build -t lisa-workstation . && docker run --rm -it lisa-workstation

Already-installed wins, however it got there

A tool present by any means is reported and left alone. That dissolves the "Homebrew or ~/.local/bin" question entirely: we never contend for ownership of a tool someone else's package manager installed, and never shadow it with an earlier PATH entry that would silently win over their upgrades. The report names where each tool came from so the operator can see why it was skipped.

Running twice installs nothing the second time, and says so rather than going quiet — silence reads as "did nothing because it broke".

Install method is a per-tool property, not a global policy

The obvious design — put every binary in ~/.local/bin so one code path serves a laptop and a container alike — is wrong, and the agents show why:

claude        -> ~/.local/share/claude/versions/2.1.221
cursor-agent  -> ~/.local/share/cursor-agent/versions/2026.07.16-899851b
codex         -> ~/.codex/packages/standalone/current/bin/codex

Each vendor manages its own version directory, and ~/.local/bin holds only a symlink. Writing a raw binary there breaks the vendor's self-updater, which expects to swap a link target and would instead find a real file in its way.

So each catalogue entry declares its own method, and the honest cost is recorded with it:

kind	pinned	checksummed	notes
release-zip / release-tar	yes	yes	reuses the remote-env installer
npm-global	no	no	vendor registry
vendor-script	no	no	pipes a fetched script to a shell
manual	n/a	n/a	no headless installer published
required	n/a	n/a	expected from the OS or base image

vendor-script is a real weakening of the guarantee assertPinned exists to provide, so the report names every unchecksummed install rather than blending them in. Installing them trusts the vendor's script at fetch time. The alternative — repackaging vendor binaries ourselves — would break the self-updaters that own those directories.

The credential manager is asked, never assumed

lisa-secrets-access already treats the provider as an axis: Bitwarden, 1Password, Doppler, Vault, AWS, or plain environment variables. A bootstrap that installed bws unconditionally would contradict that — pushing every workstation onto one vendor and quietly making the others second-class.

So the provider is a question, none is a first-class answer, and only the selected one is installed. Installing all of them would leave four unused credential CLIs on the machine, each an extra thing to keep patched for no benefit.

Headless is the primary mode, so with no TTY and no --provider, the answer is none — not a prompt that would hang a container build forever. A misspelled provider stops the run instead of falling back, because quietly provisioning the wrong credential manager is worse than stopping.

What it will not do

• Read a repository. That is the next step, not this one.
• Install a credential manager you did not pick.
• Reinstall over a tool that is already there, whatever installed it.
• Install anything without --install.
• Install a tool whose vendor publishes no headless installer. Such an entry is reported as manual with a pointer, because "not installed" and "cannot be installed by this tool" are different answers to the operator. No agent is currently in that state — Antigravity was listed there in error and does publish a bootstrapper.