attach-receipts

attach-receipts

Turns a batch of candidate edges — each proposed alongside the metadata of the run that produced it — into written edges that carry a full receipt, or into a refusal with a stated reason. No edge gets written with a guessed or blank receipt field.

Instructions

You are a Claude Code skill implementing the receipt-per-edge pattern. A receipt is not optional metadata to fill in when convenient — it's a precondition for writing the edge at all. Follow these steps in order:

1. Read the schema first. Open schema.example.json and note the fixed edge shape: subject, predicate, object, plus the three receipt fields source_doc, extraction_run_id, schema_version. Note what v1 and v2 each additionally require of a caused-by edge (v2 adds impact_scope). Treat this shape as fixed for the run.
2. Read the source material. Default to sample-input.md in this kit unless the user names a different file. Pull out every write request it lists, each with its candidate edge and the run metadata attached to it.
3. Check the schema-version-specific fields first. For each candidate edge, confirm it carries every field its stated schema_version requires (e.g. a v2 caused-by edge must carry impact_scope). If a schema-version-specific field is missing, refuse the edge and report which field and which schema version required it.
4. Check all three receipt fields next — this is the step that must never be skipped. For every candidate edge that passed step 3, confirm source_doc, extraction_run_id, and schema_version are all present and non-empty. If any one of the three is missing, blank, or null, refuse to write that edge. Do not substitute a placeholder, do not infer a run id from context, and do not write the edge with the field left out "for now" — a missing receipt field is a hard stop, not a warning.
5. Write every candidate edge that passed both checks, with its subject/predicate/object, any schema-version-specific fields, and all three receipt fields populated exactly as given. Write the result to output.json (or the path the user requested) in this kit's root, following schema.example.json's edge shape.
6. Report every refused candidate edge by its subject/predicate/object and the specific missing field, alongside the written edges. A run that silently drops a refused edge without saying why has failed this step just as much as one that wrote the edge anyway.
7. Confirm distinguishability before finishing: if more than one written edge shares the same subject/predicate/object (as write requests A and B do in this kit's scenario), check that their receipt fields alone are enough to tell them apart — different extraction_run_id, different schema_version. State this check explicitly in your output rather than assuming it.