Generating DRTML

This page is for external agents / IDE tools. The Framework does not embed an LLM. DRTML and Framework primitives are designed to be machine-generatable and machine-composable; production NL→DRTML agents live outside this package.

Recommended agent loop

1. Discover — inspect --format json or build_catalog().
2. Select primitives — filter by kind; read contracts for params/I/O.
3. Check composition — can_compose / compatible_consumers.
4. Draft DRTML — emit YAML using only known ids and namespaces.
5. Validate — validate_manifest_path(..., compile=True).
6. Repair — apply structured errors[].code/path/available/fixes.
7. Execute — run_from_manifest (or Airflow/local runner) when valid.

Worked example (no LLM)

See examples/agent-pipeline-loop in the Framework repo: a script simulates “NL intent → discovery → bad draft → structured repair → compile” using only Framework APIs.

Constraints

• Param namespaces must be in the allowlist (runtime, domain, compute, …) — Namespaces.
• Dataset backends must be registered (parquet, postgres, …).
• Patterns / run types must exist in the catalog.
• PostgreSQL datasets use inline datasets.*.postgres only (no top-level postgres.tables).
• Do not invent proprietary Ontobuilder step ids inside Framework core docs.