[P1] Implement adaptive recursion routing and candidate context-program selection #17

Closed
opened 2026-08-12 01:19:53 +00:00 by lost-rob0t · 0 comments
lost-rob0t commented 2026-08-12 01:19:53 +00:00 (Migrated from github.com)

Goal

Treat recursion as a costed routing decision rather than a static depth knob.

Scope

  • define routing choices among direct continuation, cheap submodel, recursive RLM, delegated subagent, and deterministic context operation;
  • add heuristics/policy hooks for expected utility versus cost;
  • record reason, estimated cost, actual cost, parent/child identity, and depth for every recursive/delegated call;
  • detect duplicate subproblems and no-progress cycles;
  • add bounded candidate-plan generation/selection hooks inspired by self-reflective program search;
  • keep depth >1 experimental behind explicit options.

Acceptance criteria

  • trivial retrieval can complete with no recursive call;
  • a decomposable long-context fixture can select depth 1;
  • duplicate/no-progress recursion is rejected or redirected;
  • routing decisions are traceable;
  • depth 0/1/2 can be benchmarked under identical budgets;
  • default production policy does not silently enable depth >1.

Dependencies

Research

See RLM-RESEARCH-008-adaptive-recursion.org.

## Goal Treat recursion as a costed routing decision rather than a static depth knob. ## Scope - define routing choices among direct continuation, cheap submodel, recursive RLM, delegated subagent, and deterministic context operation; - add heuristics/policy hooks for expected utility versus cost; - record reason, estimated cost, actual cost, parent/child identity, and depth for every recursive/delegated call; - detect duplicate subproblems and no-progress cycles; - add bounded candidate-plan generation/selection hooks inspired by self-reflective program search; - keep depth >1 experimental behind explicit options. ## Acceptance criteria - trivial retrieval can complete with no recursive call; - a decomposable long-context fixture can select depth 1; - duplicate/no-progress recursion is rejected or redirected; - routing decisions are traceable; - depth 0/1/2 can be benchmarked under identical budgets; - default production policy does not silently enable depth >1. ## Dependencies - #9 - #10 - #11 - parent #3 ## Research See `RLM-RESEARCH-008-adaptive-recursion.org`.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
nsaspy/prolog-rlm#17
No description provided.