Benchmarks agent code and test quality through four adversarial review rounds.
npx skills add https://github.com/greglas75/zuvo --skill agent-benchmark-greglas75
Please help me install this Agent Skill. Skill: agent-benchmark Source: https://github.com/greglas75/zuvo/tree/main/skills/agent-benchmark Command: npx skills add https://github.com/greglas75/zuvo --skill agent-benchmark-greglas75
High-intent search queries and answers about installing and using this skill.
Run the agent-benchmark skill, which has the agent write TypeScript code and tests itself, then applies multi-provider adversarial review between rounds. It scores the results against a rubric and writes a JSON report you can compare across models.
Run this benchmark once under each model; it detects the current model, tags the run with a model slug, and records scores, timing, and estimated API cost. Comparing the agent-benchmark.json files shows quality and cost differences.
Yes. Use --quick to skip adversarial rounds and only run code and test writing, or --no-tests to run only the code rounds. The --dry-run flag prints what would happen without executing.
It requires the adversarial-review.sh script (from the plugin repo or platform install paths) and shared include files for the benchmark corpus, scoring rubric, run logger, and retrospective. If any mandatory file is missing, the run stops.
R1 and R3 are the baseline artifacts; fixes are written as new R2 and R4 files. Editing the originals would destroy the before/after comparison that measures how much adversarial review improved the output.
agwacomQuery the Codex CLI for structured code review and adversarial patch testing.
domattioliQuery HuggingFace benchmark leaderboards and aggregate model performance data into a device-aware comparison.
torikhayesLink duplicate security-review findings so consolidated reports merge them into one entry.
woodfishhhhDiscover and download public datasets from Kaggle, UCI, HuggingFace, and Tianchi.
