# CmdBench

> CmdBench: AI coding agent benchmark by Jaanus Lepik. Made by CmdBrief. Agents run the same frozen task under the same deadline and are verified by the same hidden checks; every run is published with its checks, transcript and source. No blended overall score.

CmdBench is not related to the cmdbench Python package or to CMDBench, the Megagon Labs data-discovery benchmark.

## Method
- [Protocol v1](https://cmdbench.com/protocol/v1/): how runs are frozen, ordered, verified and published
- [Methodology](https://cmdbench.com/methodology/): how agents are tested and how results are summarised
- [Corrections](https://cmdbench.com/corrections/): every correction and regrade

## Results
- [Leaderboard](https://cmdbench.com/leaderboard/): per-episode tables, no overall rank

## Episodes
- None published yet.

## Benchmarks
- [Lava lamp v1](https://cmdbench.com/benchmarks/lava-lamp/v1/): data https://cmdbench.com/benchmarks/lava-lamp/v1/benchmark.json