recurve 0.1.0

Recursive context-aware reasoning & planning for LLM agents — the ReCAP loop (arXiv:2510.23822) as a thin, provider-agnostic library.

recurve

Recursive context-aware reasoning & planning for LLM agents.

recurve is a thin, async-first, provider-agnostic Python library implementing the ReCAP loop — Recursive Context-Aware Reasoning and Planning (arXiv:2510.23822, NeurIPS 2025).

Its one defensible novelty, packaged nowhere else: ReCAP's single shared sliding-window context with structured re-injection of the parent's remaining plan on return — not isolated sub-agent contexts, not a graph engine. The loop:

1. plan-ahead — emit the whole ordered subtask list, but execute only the first item;
2. recurse on composite subtasks within the same shared context;
3. re-inject the parent's remaining plan when a subtask returns;
4. refine the remaining plan after each step.

Provider plumbing, typed structured outputs, retries, and tracing are reused (Pydantic AI / LiteLLM behind a tiny 2-method Model protocol), never reinvented.

Part of the AETHEL project. MIT licensed. Status: alpha (Phase 1).