Validate a WrenAI semantic model's references before an agent queries through it

The Stack

Steps

1. 1

   Parse the MDL and assert every reference resolves

   Build the model-to-columns map from the MDL, then check each relationship's endpoints and join condition, and each metric's measure and dimension, against columns that actually exist. Rename orders.amount and the metric's measure stops resolving here, loudly, instead of the agent returning a plausible wrong total. CI runs this against a fixture MDL. Pointing an agent at the layer is fenced.

   cat > mdl.json <<'JSON'
   {
     "models": [
       { "name": "orders", "columns": [{"name":"id"},{"name":"customer_id"},{"name":"amount"}] },
       { "name": "customers", "columns": [{"name":"id"},{"name":"region"}] }
     ],
     "relationships": [
       { "name": "orders_customer", "models": ["orders","customers"], "condition": "orders.customer_id = customers.id" }
     ],
     "metrics": [
       { "name": "revenue_by_region", "baseObject": "orders", "measure": {"column":"amount","expression":"sum"}, "dimension": {"model":"customers","column":"region"} }
     ]
   }
   JSON
   python3 - <<'PY'
   import json, sys
   
   with open("mdl.json") as f:
       mdl = json.load(f)
   
   cols = {}
   for m in mdl.get("models", []):
       cols[m["name"]] = set(c["name"] for c in m.get("columns", []))
   
   def resolve(model, column):
       if model not in cols:
           return "model '" + str(model) + "' does not exist"
       if column is not None and column not in cols[model]:
           return "column '" + str(model) + "." + str(column) + "' does not exist"
       return None
   
   errors = []
   
   for rel in mdl.get("relationships", []):
       for mm in rel.get("models", []):
           e = resolve(mm, None)
           if e:
               errors.append("relationship '" + rel.get("name", "?") + "': " + e)
       cond = rel.get("condition", "")
       if " = " in cond:
           for side in cond.split(" = "):
               side = side.strip()
               if "." in side:
                   a, b = side.split(".", 1)
                   e = resolve(a, b)
                   if e:
                       errors.append("relationship '" + rel.get("name", "?") + "' condition: " + e)
   
   for met in mdl.get("metrics", []):
       meas = met.get("measure", {})
       e = resolve(met.get("baseObject"), meas.get("column"))
       if e:
           errors.append("metric '" + met.get("name", "?") + "' measure: " + e)
       dim = met.get("dimension")
       if dim:
           e = resolve(dim.get("model"), dim.get("column"))
           if e:
               errors.append("metric '" + met.get("name", "?") + "' dimension: " + e)
   
   if errors:
       print("BAD: semantic model has dangling references (the wrong-but-plausible-number risk):")
       for x in errors:
           print("  - " + x)
       sys.exit(1)
   
   print("MDL OK: " + str(len(mdl.get("models", []))) + " models, " + str(len(mdl.get("relationships", []))) + " relationship, " + str(len(mdl.get("metrics", []))) + " metric; all refs resolve (orders.customer_id -> customers.id, revenue_by_region measures orders.amount by customers.region)")
   PY

2. 2

   Point the agent at the governed layer (the model step, not checked by CI)

   Install WrenAI's skills into your coding agent (npx skills add Canner/WrenAI --skill '*'), start a session and run the wren-onboarding skill to wire up the MDL and connectors, then ask questions in plain English. The agent's translation of a question into SQL through the layer is the fenced step, no green check can promise a model's query is correct.