Wink Pings

NVIDIA Gives Agent Skills a Health Check: First See Whether They Load, Then Whether They Work Well

SkillEvaluator 1.0.1 is released. HarnessCheck can check, without a model key, whether a skill is missed, rejected, or truncated, and it can also go into CI. A user mentioned an example where a description exceeded the limit by 4 UTF-16 units, which is very hard to spot by eye.

NVIDIA's SkillEvaluator has been updated to 1.0.1. It now checks one thing first: whether your agent skill will be missed, rejected, or truncated at the loading stage, and only then measures how well it works.

This sounds less exciting than 'evaluating agent capabilities,' but it is closer to the annoyances of day-to-day maintenance. Skills run in environments such as Claude Code, Codex, Copilot CLI, Hermes, OpenClaw, and OpenCode. If details such as description format, length limits, or manifest files go wrong, the skill is already gone before the model even starts working.

HarnessCheck does this kind of pre-load check. The command is not complicated:

skillevaluator harness-check ./my-skill

The key is that it does not need a model key. A user, Sharful, mentioned that in the README example, Copilot CLI will skip a skill whose description exceeds the 1,024 limit by 4 UTF-16 units. Four units are basically invisible to the naked eye, but the tool will point it out directly. This detail is more useful than 'how many models it supports.'

Another update is NeMo Gym integration, and it can be put into a repository's CI. A user, vexor4i, said that keyless checks plus NeMo Gym in CI mean a real feedback loop can be formed, and they plan to try it with a few skills this week. That judgment is pragmatic: once there are many skills, manual clicking and visual comparison are unrealistic.

Its value is more like lint. First block low-level errors before the model call, then talk about effect evaluation. For people maintaining skills, running harness-check once saves time compared with directly asking 'why is my skill not responding?'

GitHub Release

发布时间: 2026-11-19 17:28