Aegis

A deterministic validator for AI-generated code. Apache 2.0.

What it does

aegis check ./path runs a 24-layer pipeline against a directory of code and reports whether the validation passes. Layers cover AST parsing, import resolution, cross-file consistency, package install, type check, test execution, design-brief fidelity, and feature coverage. 22 layers are deterministic (no model call); 1 is LLM-judge; 1 is hybrid (LLM verdict with a deterministic override).

Install

Install from source (a published pip install aegis-validator package is coming soon):

git clone https://github.com/andraste-labs/aegis.git
cd aegis
pip install -e .

The Anthropic SDK is an optional extra for the LLM-using layers:

pip install -e ".[anthropic]"

Quick start

aegis check ./my-code                          # deterministic + LLM
aegis check ./my-code --no-llm                 # deterministic only
aegis check ./my-code --brief brief.json       # include design / feature layers
aegis check ./my-code --json report.json       # machine-readable report
aegis check ./my-code --exit-on-fail           # exit 1 on FAIL

The LLM-using layers (design_fidelity, feature_coverage) skip unless an ANTHROPIC_API_KEY is set and a brief.json is supplied.

Layers

See docs/LAYER_INDEX.md for the full list of layers, their kinds (deterministic / hybrid / llm_judge), and the file under aegis/checks/ that implements each one.

Benchmark

aegis-bench/ contains a cohort of reproducible cases. Each case has a brief.json, an input/ directory, an expected.json describing the validator output, and a short technical README.

python -m aegis_cli check aegis-bench/cohort/<case>/input \
    --brief aegis-bench/cohort/<case>/brief.json \
    --no-llm

See aegis-bench/METHODOLOGY.md for case structure and reproducibility rules.

Contributing

See CONTRIBUTING.md.

License

Apache License 2.0.

Maintainer

Aegis is maintained by Andraste Labs. Contact: github@andrastelabs.com.