Overview & result
DeciLabel helps repository maintainers triage issues and pull requests. In its direct TypeSafe mode, Jev judges whether existing labels fit the available text and diff evidence. Code validates the returned probabilities, applies thresholds and label policies, and adds accepted labels. It does not generate patches or invent new labels.
From the original post
DeciLabel classifies GitHub Issues and pull requests with a typed Decisions API, then adds only labels supported by the available evidence.
How Jev fits in the loop
- Builds Noul questions for independent labels and Choice questions with a none option for configured exclusive groups.
- Direct TypeSafe mode calls /v1/systemone with jev-latest; the default workflow instead uses OpenRouter.
- Deterministic policy checks reject missing or invalid answers, apply thresholds and a Choice margin, and limit additional labels.
- Provides a dry-run and a run summary so maintainers can inspect decisions before allowing writes.
How to reproduce
- Start with the repository’s examples/minimal.yml workflow and an existing set of labels.
- For direct Jev, set provider: typesafe and typesafe-api-key to your TYPESAFE_API_KEY Actions secret; review the repository’s workflow permissions and event policies.
- Set dry-run: true to inspect proposed labels before enabling label writes.
- Review the Actions run summary and adjust the label allow-list and threshold for your own repository.
Why this build matters
Turns repository triage into a configurable, reviewable workflow with an explicit route from semantic label decisions to controlled GitHub writes.
Limitations
- Source and configuration were reviewed; the action was not installed or independently run by us.
- Issue/PR text and available diff excerpts are sent to the selected provider. Repeated runs may incur another paid model request.
- Partial diffs limit classification coverage. Thresholds and author-reported calibration are specific to the tested taxonomy and are not a general accuracy guarantee.
- Existing labels are preserved; a concurrent human removal can still race the final additive write.