pi

The control host is where you write code. It needs no GPU — only the address of a host that has one, and a key.

Install pi first

curl -fsSL https://pi.dev/install.sh | sh

The Windows and package-manager variants are on pi.dev. pi is a prereq, not a recommendation: make pi-install writes into ~/.pi/agent/ and needs that directory structure to exist.

Point it at lmstack

make pi-install
$EDITOR ~/.pi/agent/extensions/.env    # host URL + LiteLLM key
pi --list-models | grep lmstack

lmstack-h1  qwen2.5-coder-7b  16.4K  4.1K  no  no
lmstack-h2  qwen2.5-coder-7b  16.4K  4.1K  no  no

It merges, it does not replace

If you already use pi, you already have a packages list. make pi-install adds three entries to it and leaves everything else alone, including settings this repo knows nothing about. make pi-dump is symmetric: it captures back only the entries the repo owns, so your personal configuration does not end up in a public repository.

The script this was ported from copied settings.json but not the extensions it names. A fresh machine then got a pi that started cleanly with no providers registered and nothing useful in the logs. That is test T4.1.

Configuration

~/.pi/agent/extensions/.env, created mode 0600 on first install and never touched again — not by a re-install, not by dump.

LMSTACK_H1_URL=http://100.64.0.2:4000/v1
LMSTACK_H1_KEY=sk-...
LMSTACK_H2_URL=http://127.0.0.1:4000/v1
LMSTACK_H2_KEY=sk-...

Read the key from the host rather than from your notes:

ssh h1-nvidia 'grep ^LITELLM_MASTER_KEY ~/.lmstack/stack.env | cut -d= -f2'

LMSTACK_H2_URL defaults to loopback, which is right when h2-amd is this machine. LMSTACK_H1_URL defaults to http://h1-nvidia:4000/v1, which assumes an ~/.ssh/config alias or a Tailscale name.

An unconfigured provider still registers and then fails to connect. That is deliberate: a clear connection error beats a provider that quietly vanishes from --list-models.

The status line

The installed layout shows provider, model, branch, context, tokens, and time. It omits cost, because local inference has none and $0.00 on every turn is noise. Context is first among the numbers for a reason — a 16k window is the binding constraint when you serve your own model.

Advertised models must be real

Each extension lists the models its provider serves, and make validate (T0.10) fails if any of those ids is not a LiteLLM alias some host actually exposes.

That check is the reason to edit pi-config/extensions/ rather than ~/.pi/agent/extensions/ directly. Hand-edits to the installed copy are lost on the next sync, and they skip the check.

Switching hosts

Both providers serve qwen2.5-coder-7b. Moving from the laptop to the DGX is a provider change and nothing else — no prompt changes, no model name changes, no per-host branching in whatever you have built on top.

What lmstack turns off in pi

The settings.json this repo ships is intentionally lean — a fresh pi loaded with lmstack should start in the fewest tokens of context possible, so the local model has room for your work rather than for pi's own scaffolding.

Setting	Value	Why
defaultProvider	lmstack-h2	Pre-selects the lmstack provider so the picker never appears at startup. Change to lmstack-h1 if that is your primary.
defaultModel	qwen2.5-coder-7b	Same. Skips the model prompt.
enableInstallTelemetry	false	No install/update ping to pi.dev.
quietStartup	true	Hides the startup header.
warnings.anthropicExtraUsage	false	Not relevant when there is no Anthropic provider registered.

pi's cloud providers (Anthropic, OpenAI, Codex, GitHub Copilot, etc.) do not load unless you /login to them or set their API-key env vars, so lmstack does not need to disable them explicitly — they simply do not appear. If you want a pi that speaks only to lmstack, don't /login and don't export those keys.

Merge semantics still apply. If your existing settings.json already sets any of these, make pi-install leaves your value in place — same rule as with packages. Delete the key from your settings if you want the lmstack default back.