Open topics¶

Points to settle in the first implementation concept – none of them is a decision yet.

Workflow¶

Targets are decisions. Raising a target should be logged with date and reason, later as a deciwa decision.
Gap list as work queue. The measures between reached and target are the natural to-do list for agents (“bring the overlay to level 3”).
Gaming by agents. Agents working towards green write tests that test nothing or docs that only exist; mutation testing and AI-rated checks are in the model from the start so the trick becomes visible.
Snapshot per release or customer. A maturity profile can be part of an agreement (“you get 2, security 4”); the snapshot per release is the evidence.
Model upgrades. A new model version makes levels drop without the code changing; projects pin a version and upgrade on purpose, quawa shows “with danwos@2 you would be at 2”.

Scope mapping. Modules and features need code, tests and docs assigned to them: paths first, feature ↔ test links from deciwa.
Adapters. One result schema, many inputs: JUnit XML, lcov, cobertura, SARIF, CycloneDX, deciwa.
Cost and cadence. Linting runs on save, E2E and mutation tests take minutes to hours; every check has a cadence and a maximum age.
Flaky checks. A check that changes state without a code change must be recognisable, or every regression alarm loses its worth.
Runs locally. The runner works as a CLI without CI; CI and apps only call it.
Where results live. Configuration in the repository; results as history outside Git or as a snapshot per release.

deciwa first – it is public – then quawa itself and monsterbrain.