ai workbench transformations toolkit: skill body steps drop the .schema/<cdm-name>/ folder their own contracts declare

[rstover-fo]

Filing here because issues are disabled on dlt-hub/dlthub-ai-workbench — happy to move this wherever it belongs.

Summary

In the transformations toolkit (0.1.1, installed via dlt ai toolkit install transformations), the create-ontology, generate-cdm, and create-transformation skills declare all artifacts under .schema/<cdm-name>/ in their Requires/Output sections, but several body steps reference bare .schema/ paths. An agent following the body text literally reads/writes artifacts outside the CDM folder, so each skill in the chain can fail to find (or misplace) what the previous one produced.

Affected lines (dlthub-ai-workbench: workbench/transformations/skills/)

create-ontology/SKILL.md

• Step 1: Read .schema/taxonomy.json → should be .schema/<cdm-name>/taxonomy.json (the header says to read _name from it to resolve <cdm-name>; all files live under that folder)
• Step 6: Write .schema/ontology.ison → should be .schema/<cdm-name>/ontology.ison (Output section says <cdm-name>)
• Review prompt: Please review .schema/ontology.md → .schema/<cdm-name>/ontology.md

generate-cdm/SKILL.md

• Step 1: Read .schema/ontology.ison → .schema/<cdm-name>/ontology.ison (Requires says <cdm-name>)
• Step 6: Write .schema/CDM.dbml → .schema/<cdm-name>/CDM.dbml (Output says <cdm-name>)

create-transformation/SKILL.md

• Step 2 input list: .schema/annotated-sources.dbml, .schema/taxonomy.json, .schema/CDM.dbml → the Requires section (correctly) says .schema/<cdm-name>/<pipeline_name>.dbml (per pipeline), .schema/<cdm-name>/taxonomy.json, .schema/<cdm-name>/CDM.dbml. Note annotated-sources.dbml is never produced by annotate-sources at all — it emits per-pipeline DBML files.
• Step 3 cross-check references annotated-sources.dbml (same issue)
• Naming-convention section: .schema/CDM.dbml, .schema/ontology.ison, .schema/taxonomy.json → <cdm-name> variants

Impact

Multiple PR-review bots independently flagged these as P1s in our repo after the toolkit vendored the skills (Formentera-Operations/dlt-pipelines#18). We patched the vendored copies locally, but a toolkit reinstall reverts the patch — hence this report.

Suggested fix

Align every body-step path with the <cdm-name> contract already declared in each skill's Requires/Output sections (a find/replace of the bare .schema/ references listed above).

[rstover-fo]

One more instance of the same class, in create-transformation/SKILL.md step 6: the script template declares @dlt.source def <dataset_name>_to_cdm(...) but the __main__ entrypoint calls <business_domain>_to_cdm(source_dataset) — the two placeholders should match (<dataset_name>_to_cdm).