Module llm 

submilli:llm — gated model calls from inside a Submilli program.

Rust host functions registered directly under the package name. Each op runs check_security before anything leaves the process; the one gated capability is llm.call, cataloged in crate::stdlib::capabilities — keep it in sync when adding or removing a gate (see AGENTS.md).

Dispatch is the embedder’s: StoreData::llm_provider holds the provider. A runtime with none configured refuses every op rather than inventing a completion, so a program never silently reasons over text no model produced — the rule submilli:session follows for its store.

One capability, per-model filtering. call, batch, and models() are the same grant, with prompt_count in the filter context rather than separate names: enumerating the operator’s models is not a distinct risk class from calling one. models() filters each candidate through the same model filter that gates calling, so a listing never offers a model the caller would be denied at call time — the session.list / session.read shape.

No prompt or completion text crosses this boundary in metadata. Not in the filter context, not in an error, not in a log. The context carries model and prompt_count — the numbers, never the payload.

Re-exports

pub use declaration::package_declaration;

Modules

declaration

Type surface of submilli:llm.

Constants

CAPABILITY

The single capability gating every op in this package. call and batch are the same risk and the same grant as enumeration, discriminated by the filter context rather than by three capability names.

MODULE_NAME

OPS

The ops a request is recorded under, call and batch. The provider is not told which one a request came from.

Functions

install

request_digest

The digest the call log records for a request of this op.