Reproducible dev environments: one-command setup for AI coding agents
This check looks for files that set up the project with one command: a Makefile, a devcontainer, a Dockerfile or compose file, Nix files, a justfile, a Taskfile, tox.ini, noxfile.py, the Maven or Gradle wrapper, or bin/setup.
How many repos pass
Of the 371 repos on the leaderboard, 79 pass in full (21%), 192 pass in part (52%) and 100 fail (27%).
Why agents need it
Agents that run in their own machine have to get the project working there first. Devin starts from a virtual machine with the repo, its tools and its dependencies already set up, per Cognition's environment docs, so its profile weighs a runnable environment above CI. A setup that is written down as a file is one an agent can run.
How we score it
- Full pass: two or more of these files.
- 70%: exactly one.
- 60%: none, but
package.jsondefines three or more scripts. - Fail: none of the above. We do not check that the setup works.
The methodology has the formula that turns checks into a score.
How to pass
Add a Makefile or justfile with setup, test and lint targets, and a devcontainer or Dockerfile that installs the toolchain. Two of these files earn the full pass.
Which agents count it most
- Devin13% of its score
- Claude Code12% of its score
- OpenHands10% of its score
- Kimi CLI10% of its score
- GPT-5 Codex9% of its score
- Pi8% of its score
- Gemini CLI8% of its score
- Aider7% of its score
- Cursor6% of its score
Pass rate by language
Top-scoring repos that pass
- Rust · 33.1k ★
- gitlab-org/cli92.7— · 1.1k ★
- apache/superset92.2Python · 75.1k ★
- C++ · 131k ★
- Rust · 6.7k ★
- Python · 46k ★
- Clojure · 49.6k ★
- milvus-io/milvus89.6Go · 46.4k ★
- pingcap/tidb89.6Go · 40.6k ★
- Go · 16.9k ★
- Python · 9.9k ★
- cli/cli88.2Go · 46.7k ★
Popular repos that fail it
- TypeScript · 190.9k ★
- Python · 187.5k ★
- flutter/flutter53.5Dart · 180k ★
- Python · 165k ★
- golang/go31.2Go · 139.2k ★
- Python · 137.2k ★
- rust-lang/rust71.4Rust · 119.8k ★
- C++ · 118.3k ★