How it uses Jev
Stanley routes a natural-language code task to an eligible workflow, gathers bounded evidence, and asks Jev fixed-choice questions about selected pieces. In the included TODO-audit workflow, each candidate comment is judged against the task; code applies a fixed confidence threshold and reports the result.
What Jev decides
Is this TODO comment still actionable?
A code task and one TODO comment found in `src/app.ts`; the embedded workflow asks one fixed-choice question for each of up to 20 comments.
- actionable
- stale
- cannot_tell
App workflow
- Route the request
Select an eligible built-in or trusted repository workflow based on the request and deterministic availability gates.
- Gather bounded evidence
Read selected diffs, logs, or repository facts and prioritize what can be checked.
- Ask Jev fixed-choice questions
Classify small pieces of evidence with bounded options.
- Apply code rules
Use deterministic policy and confidence thresholds to form findings.
- Report checked scope
Return findings and identify evidence the workflow did not check.
Why it is interesting
The system separates evidence gathering and thresholds from semantic classification. Its built-in workflow example returns checked findings and explicitly reports the scope it did not inspect.