Skip to content

Nested RLM recursion: subcalls are plain LLM calls, no rlm_query equivalent #2

Description

@shanev

Problem

SubcallClient exposes flat llm_query and llm_batch operations. A depth bound limits subcall-chain authorization, but a subcall is still a plain model response rather than a nested Droste loop with its own REPL state.

Proposal

Add an optional rlm_query(prompt, context) capability that runs a child loop using a strict allocation from the parent budget and returns the child's structured result. The child must preserve parent/child identity, typed failures, usage, and evidence without creating a second dispatch or accounting path.

Current flat subcalls remain valid. Any quality benefit from deeper recursion requires version-matched artifacts under #81; it must not be inferred from an external result or an average score.

Metadata

Metadata

Assignees

No one assigned

    Labels

    P3Valuable, not urgent: research, DX, single-consumer featurescross-repoRequires coordinated changes across tensor-systems repositories

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions