Thanks for wanting to improve Codette.
This project moves fastest when contributions are reproducible, scoped, and easy to verify.
- Read README.md
- Read docs/proof.md
- If your change touches memory, routing, safety, or web research, also read:
git clone https://github.com/Raiff1982/Codette-Reasoning.git
cd Codette-Reasoning
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtModel setup:
Run this before pushing any change that touches reasoning_forge/ or inference/:
make cocoon-smoke27 checks covering schema, integrity scoring, echo detection, subsystem contracts, quarantine routing, and the v3 regression alarm. Exits 1 on any failure. See docs/cocoons_quickstart.md for what each section tests.
make dev # starts Forge locally with COCOON_STORE=./dev_cocoons/
make inspect-latest # human-readable summary of the most recent cocoon
make list-cocoons # last 20 cocoons: id, execution_path, integrity, echo risk
make health # avg integrity score, echo distribution, fallback alarm countpython3 -m unittest tests.test_event_embedded_value
python3 -m unittest tests.test_codette_runtime_benchmark
make test-cocoon # cocoon audit tests onlypython scripts/run_all_benchmarks.py
python scripts/run_all_benchmarks.py --include-runtimepython demo/run_local_api_demo.pyInstall the included hook to have cocoon-smoke run automatically when you push changes to forge or cocoon files:
cp .githooks/pre-push .git/hooks/pre-push
chmod +x .git/hooks/pre-pushThe hook only fires when the push touches reasoning_forge/, inference/codette_forge_bridge.py, or scripts/cocoon_smoke.py. To bypass in an emergency: git push --no-verify.
Good contribution targets:
- bug fixes
- benchmark improvements
- test coverage
- adapter/routing improvements
- memory and continuity improvements
- web research safety hardening
- documentation and reproducibility fixes
Please include:
- what changed
- why it changed
- how you verified it
- any benchmark/test output that supports the change
- any user-facing behavior change
If your change modifies public claims, update the evidence links too.
- Do not commit large model files or private cocoons.
- Treat cocoon memory, session data, and transcripts as potentially sensitive.
- If you add example transcripts, scrub personal data first.
If you add or retrain an adapter, document:
- what the adapter is for
- how it routes
- how it is tested
- how it changes behavior versus base generation
Guide: docs/adapter_workflow.md