Thanks for contributing to OpenAdapt. A few things keep the project healthy and keep the company able to steward it.
This repository's code is MIT-licensed, and your contribution goes in under the MIT License. Two things cover it.
-
Developer Certificate of Origin (required now). Add a
Signed-off-byline to every commit, certifying you wrote the change or that you have the right to submit it under the project license:git commit -s -m "fix: ..."This produces
Signed-off-by: Your Name <you@example.com>. The full text is at https://developercertificate.org. -
Contributor License Agreement (published, not yet enforced). The canonical text is
openadapt-flow/CLA.mdfor individuals, andopenadapt-flow/CCLA.mdfor companies whose employees contribute on company time. It gives MLDSAI Inc. an explicit copyright and patent license, which the MIT License alone doesn't provide, and it keeps the option of relicensing the combined work later.You don't agree to the CLA by opening a pull request. Nothing is implied. You agree when you sign, either through the automated CLA check once that check is turned on for this repository, or by email. Until the check is on, the DCO sign-off and the MIT License are what govern your contribution.
OpenAdapt is open-core. MLDSAI Inc. sells proprietary products built on this code, including a hosted control plane, and your contribution may end up in them. The MIT License already permits that. The CLA says it out loud so nobody is surprised.
- Use Conventional Commits for titles and commits (
feat:,fix:,docs:,chore:,refactor:,test:,ci:). - Keep PRs focused; separate mechanical changes from behavior changes.
- Add or update tests for any behavior change.
- Prefer honest, measured claims in docs. If something is experimental, say so.
OpenAdapt is open-core. Do not add private crown-jewel artifacts (grown hardening corpus, tuned adversary params, deployment-derived thresholds, per-system-of-record oracle/connector recipes, real-EMR datasets) to this public repository. Interfaces and mechanisms are public; data, recipes, and empirical tuning are private. See the OpenAdapt Source-Availability Boundary policy.
Do not file security problems as public issues; see SECURITY.md if present, or
email security@openadapt.ai.