the install-introspection language

lmi

Ask the running deployment what it is and what it can actually do. Safe mode, needs the host's install bridge — so POST /api/lmx/run or Studio's LMX Playground, not the CLI.

`lmi` — the install-introspection language

Ask the running deployment what it is and what it can actually do. Safe mode, needs the host's install bridge — so POST /api/lmx/run or Studio's LMX Playground, not the CLI.

INSTALL
KINDS

LIST apps LIMIT 5
FIND "invoice"

CAN ai.text.generate
PROVIDERS FOR email.send

Why this exists

Leumas is a product other people run. "Leumas can do X" and "you can do X" are different claims, and the second is the one that matters when answering someone about their own deployment.

A given install has a role, an edition, a set of published apps and tools, its own index sources, a set of connected providers, and a per-user entitlement map. None of that is knowable from the source tree — the repo says what Leumas can ship, not what this box has. lmi asks the instance.

Like lma, it holds no vocabulary. LIST <kind> accepts whatever kinds the install indexes, because the bridge takes the same { <kind>: async () => Array } source map createIndexRouter takes — in leumas-api literally the same hoisted object. Add a source there and the language gains a kind.

[critical] The rule: never answer "yes" by default

Every answer here is three-state:

	Means
`yes`	this install said so
`no`	this install said so
`unknown`	nothing was wired to answer, so nothing was said

An unknown reported as yes produces a confident, wrong claim about somebody else's deployment — exactly the failure this language exists to stop. So:

CAN with no entitlement engine wired prints unknown — no entitlement engine is wired here,

and its allowed field is null, never true.

LIST on a kind this install does not index throws, naming the kinds that do exist. Returning

[] would read as "you have none of those", which is a different and misleading answer.

A source that fails is reported as broken (count: null plus the error), not as empty.

And: entitled ≠ fulfillable

CAN and PROVIDERS FOR answer different questions, and a thing only works here when both are satisfied. A user can be perfectly entitled to email.send on an install where nothing is connected to send it. PROVIDERS FOR says so in as many words:

--- providers for sms.send (0) ---
  nothing here fulfils it — CAN may still say yes, and it would still fail

Statements

# starts a comment.

Statement	Answers
`INSTALL`	role, edition, plane flags, which kinds are indexed, who you are
`KINDS`	every indexed kind with a live count
`LIST <kind> [WHERE …] [LIMIT n] [-> var]`	that kind's rows
`FIND "query"`	search every kind at once
`CAN <capability> [-> var]`	is THIS user entitled, and up to what limit
`PROVIDERS FOR <capability> [-> var]`	which connected provider would actually serve it
`ORPHANS`	capabilities nothing here fulfils — a worklist
`SET` · `PRINT` · `WHEN … THEN` · `RETURN`	the usual

WHERE takes one comparison against a row field: LIST tools WHERE category = "media" LIMIT 20.

Every question and its answer is kept in answers on the result — a report about an install should be quotable, and that is the quotable part.

What it deliberately cannot do

Change anything. No method on the bridge installs, enables, entitles or connects. It is a

read-only surface by construction, not by check.

Default a missing fact. A describeInstall that omits edition yields no edition field —

"this install did not say" is not "community".

Run from the CLI. No install bridge there, and answering from nothing is the one thing this

language must never do.

Running it

curl -s -X POST localhost:3000/api/lmx/run \
  -H 'content-type: application/json' \
  -d '{"mode":"lmi","source":"INSTALL\nKINDS"}'

Bundled examples: shared/engines/lmx/src/scripts/demo.lmi, shared/engines/lmx/src/scripts/audit-install.lmi — the latter is the report to run before telling a user what they can do on their install.

Tests: shared/engines/lmx/src/test/lmi.test.js. Guard: pnpm smoke:lmx.

lmi

lmi — the install-introspection language