simonlqy/skillgate

---

Agent skills are instruction files an agent reads on demand: a name, a one-line description, a body of procedure. With thousands in a library, which skill to read is a decision the policy makes mid-episode — from names and descriptions alone — and outcome-rewarded RL cannot teach it. The tokens that name the chosen skill carry a median 0.14% of their trajectory’s loss weight, and two in five receive a negative advantage because execution afterwards failed. We call this selector credit starvation. SkillGate removes it by construction: one GRPO update, two disjoint credit channels — outcome credit reaches only execution tokens, while an action-local advantage reaches exactly the skill-naming tokens, positive only when the trajectory’s single read is the correct skill.

🎬 Demo

The problem and the fix, in 30 seconds

Four scenes, no mock data. A held-out task where the whole decision is 8 tokens inside a read call; the failing run in which those tokens are 11 of 30,487 — 0.036% of the loss, counted with the released tokenizer over the archived trajectory; the credit partition SkillGate applies instead; and what that one change buys on the same task pair.

Case study: one SkillGate episode, end to end

An archived run on sb_ns/lake-warming-attribution (verifier-checked data analysis). The policy starts working, reaches for the library mid-episode, picks the oracle out of six lake/trend lookalikes in one read, and lands the two traps the skill documents (Mann-Kendall instead of linregress; the category label Heat, not the individual driver). Score 1.0 in 14 tool calls — on the same slate, the outcome-only baseline read four skills, none of them the oracle, and failed.

Overview

The problem — one broadcast advantage updates both the few tokens that chose