Recursive large language model (LLM) agents tackle long‑horizon tasks by decomposing them, executing sub‑tasks, and receiving environmental feedback. After producing an atomic sub‑task, they usually call the same large model again to translate a familiar instruction into an executable action. We notice that high‑level planning is task‑specific, while leaf‑level action grounding recurs across tasks and can be validated against actions admitted by the environment. To address this, we introduce LeafSkill, an external reuse layer that retrieves semantically related grounding experiences, binds action templates to the current objects and admissible actions, and employs a lightweight 1.7 B selector to either choose the bound action or defer the unchanged sub‑task back to the original agent. Skills reside outside the model parameters, allowing a frozen selector to leverage post‑training additions. Under lexical and syntactic paraphrases, exact matching retrieves none of the reference actions, whereas semantic retrieval preserves 96.0 % and 97.0 % candidate coverage; the frozen selector adopts 97.34 % of unseen skill types. On an independent confirmation set of 200 Robotouille tasks, LeafSkill completes 137 tasks versus 121 for ReCAP (68.5 % vs. 60.5 %, exact McNemar p=.0259). Among the 106 tasks completed by both, LeafSkill reduces Qwen3.5‑27B token usage by 14.42 % and calls by 14.08 %. A separate held‑out repeat yields a consistent 9.0‑point success gain and 18.98 % token reduction. Reusable leaf grounding can therefore improve task completion while cutting host‑model inference cost without replacing high‑level planning.
Review