skill-health
Installation
SKILL.md
Skill health from Braintrust coding-agent traces
Use coding-agent traces in Braintrust to find the repo skills that cost the most or go wrong most often, capture the failures in a dataset, and fix the skills. The dataset grows with each run, so every fix can be evaluated against all past failures.
Inputs
Ask for anything you cannot infer:
- The Braintrust project holding the coding-agent traces (a logs URL is easiest).
- The repo as
owner/name. - The time window. Use the dates the user gives; otherwise the last 30 days.
- A local clone of that repo, for skill files and their history.
- Where datasets live. Default one dataset per skill,
skill-health-<skill>, in the trace project.
Querying Braintrust
Use whatever Braintrust tools you have to query logs with SQL and to manage datasets: look through your available tools, or run bt --help. Read their help before the first call, because row limits differ.
The fields that matter: