Development

Use Python 3.11 or newer. Keep runtime dependencies minimal and all failure paths abstaining. Run:

python -m unittest discover -s tests -t . -v
python benchmarks/run.py

Every schema change needs invalid/missing/ambiguous-input tests. Every routing change needs tests proving required-review and shadow behavior are preserved. Never add a success-shaped fallback for model or transport errors. Do not log raw context, provider error bodies or secrets.

Changes to questions.py change the question hash and must invalidate deployment approval. Changes to model, policy, scope or thresholds need fresh independent calibration/holdout evaluation. Benchmarks must clearly distinguish scripted fixtures, live model runs and real native integration results.

For native work, follow the pinned Codex repository's AGENTS.md, use its just workflows, and keep the core adapter narrow. Most policy/transport logic stays outside codex-core. Do not change sandbox environment variables or bypass managed reviews. Do not describe source-only validation as a passing native build.