Budget-aware skill packing takes deliverability from 9% to 100% for just 1.14 points of quality

Notes on SkillSelect-Serve: QoS-Aware Budgeted Skill Service Recommendation for LLM Agents (arXiv:2607.00011) — Jingyuan Zheng, Dongjing Wang, Xin Zhang, Butian Huang, Haiping Zhang, Dongjin Yu, Shuguang Deng · May 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

SkillSelect-Serve treats a reusable agent skill as a Skill Service — a structured record of what it does, its input/output shape, which tools it needs, how many tokens it costs to load, and a risk label — rather than a passage to rank by text similarity. A five-stage pipeline builds these profiles from raw skill documents, then converts each incoming task into a structured requirement object without naming a candidate skill. Standard BM25/dense/cross-encoder retrieval pulls a high-recall candidate pool, and a task-conditioned suitability estimator scores each one. The final stage, constrained projection, is a greedy skip-and-continue packer: it walks the ranked list and admits a skill into the bundle only if the whole package still fits the token budget, stays under an aggregated risk ceiling, and draws only on tools the deployment actually has. Infeasible candidates get skipped, not stalled on — their budget rolls forward to the next-best option.

Against a registry of 35,353 skills, plain unconstrained top-5 retrieval builds a bundle that fits a realistic 4,000-token context in just 9.1% of tasks — the rest silently overflow. Constrained projection fixes that: 100% deliverability, at a cost of only 1.14 points of hit rate off the unconstrained ceiling, beating retrieve-and-rerank, budget truncation, and diversity-based (MMR) packing under the same token limit. The same mechanism roughly halves delivered risk exposure (53% at matched hit rate, up to 60% under a stricter risk setting) and drives tool-availability violations — 44–81% under tool-agnostic recommendation across five deployment environments — to zero. At a fixed three-skill budget, hit rate climbs from 0.8864 to 0.9091 over naive Top-3 selection. The authors flag one catch themselves: the offline hit-rate gains didn't fully carry over to execution. A small agent's task pass rate rose only from 0.3714 to 0.4000 despite the retrieval-quality improvement — a gap between recommendation quality and runtime utility.