AI-assisted code scoring. cossell analyzes a team's source repository and produces qualitative scores and commentary that feed into event scoring.

cossell.py loads a git repository, splits and summarizes the code, and uses an LLM to produce:

• a code summary and commentary, and
• numeric scores (1–10) for security, code smells, and complexity.

It uses LangChain with Anthropic and/or OpenAI chat models, plus sentence-transformer embeddings and clustering (scikit-learn, Chroma) to handle larger repositories.

The container image is built on the AWS CodeBuild Amazon Linux 2 base and also bundles companion analysis tools used elsewhere in the scoring pipeline: Infracost, Snyk, SCC (code counting), and Unlighthouse.

# directly
python3 cossell.py <path-to-repo> [branch]      # branch defaults to "main"

It expects LLM credentials in the environment:

docker build -t stackref-cossell .
docker run --rm \
  -e ANTHROPIC_API_KEY=... -e OPENAI_API_KEY=... \
  -v "$PWD/some-team-repo":/work \
  stackref-cossell cossell.py /work main

In production this image ran inside the isolated stackref-analysis-codescans account (see infra/) and was invoked as part of the kickoff scoring flow.

• The LangChain/model library versions here are from the original project (requirements.txt); when adapting, consider updating to current LLM SDKs and the latest models.