kernelcad-repair

kernelCAD — trace-guided repair

evaluate_script tells you a feature broke. why_did_this_fail now also tells you which lines you may edit and what edits are on the table, and repair_script applies them and re-checks. Use this loop instead of re-reading the whole script and hand-authoring a guess — the guess is what regresses unrelated geometry.

The loop

1. Evaluate. evaluate_script({ file }) → diagnostics + featureHealth. A non-empty featureHealth matters even when ok: true: it names features that silently degraded to a passthrough.
2. Explain. why_did_this_fail({ file, feature_id }) → the upstream chain, the full trace, a repairRegion, and ordered candidates.
3. Preview. repair_script({ file, diagnostic: <targetDiagnosticId>, strategy: 'dry-run' }) → every candidate patch as a diff, nothing evaluated, nothing applied. Read the diffs. If none is the fix you want, edit the region yourself.
4. Apply. repair_script({ file, diagnostic: <id>, strategy: 'try-all' }) → applies candidates in order, re-evaluates after each, and keeps the first that clears the diagnostic without new errors.
5. Persist and report. Write new_code back to the file, then state the change: the diagnostic, the candidate that cleared it, and the diff. Do not report "repaired" without naming what changed.

Repeat from step 1 while errors remain. Each pass targets ONE diagnostic.

Worked example: examples/repair/ has deliberately broken models (an oversized fillet, a hole anchored off its plate, a cutter placed away from its body, a revolve that crosses its axis, a tangent circle too small to sit outside two circles, a 90° draft, a shell thicker than the solid, an emboss parked over a hole) and a walkthrough that runs this exact loop on them: npx tsx examples/repair/run-repair-example.ts. Its README includes the captured output. lookup_cookbook({ query: 'repair failing fillet' }) returns the repaired fillet pattern.