Eigent Skills Hub

Design and scaffold reproducible multi-turn reinforcement-learning agent environments with explicit task datasets, state reset behaviour, deterministic per-turn seeding, and stratified difficulty sampling. Use this skill when asked to "build an RL environment for an agent", "scaffold a Verifiers taskset or multi-turn gym", "define observations, actions, and termination rules", or "reduce rollout variance across GRPO groups".

Engineer evidence-grounded verifiers and multi-objective reward functions that isolate rubric criteria, rank state snapshots per criterion, separate controllable process scores from uncontrollable environment failures, and detect reward hacking or "confident liar" false positives. Use this skill when asked to "build a verifier or reward function", "design an LLM judge rubric", "detect reward hacking or invalid successes", or "prevent cascading rubric errors".

Package reproducible terminal and coding tasks into Harbor / Terminal-Bench and Verifiers HarborTaskset directories with isolated verification tests, oracle solution validation, and anti-tampering boundaries. Use this skill when asked to "create a Terminal-Bench or Harbor task", "package a containerized CLI benchmark task", "verify an agent environment with pytest", or "prevent agents from reading or modifying test files".

Collect, score, and audit multi-turn agent trajectories across rollout groups, preserve dual message-and-token trace masks, compute group-relative advantages, and filter verified trajectories into SFT datasets without conflating trajectory export with RL policy optimisation. Use this skill when asked to "collect and score agent trajectories", "prepare SFT warm-up data from rollouts", "compute GRPO group advantages", or "audit tokenization alignment on multi-turn traces".