fix: render SDK-surface list/get as structured JSON

json.dumps(pydantic_model, default=str) collapses the model to a single
repr string, so `list`/`get` on the default SDK surface printed unreadable
output while REST printed proper JSON. Add a _dump() helper that unwraps
pydantic results via model_dump(mode="json") and passes REST dicts through.
Regression tests cover both surfaces.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: fschlz

scripts/online_eval_rules/create_online_eval_rules.py

@@ -134,6 +134,15 @@ def _dispatch_create(args, dry_run: bool) -> int:
     return 0
 
 
+def _dump(result) -> str:
+    """JSON-print a result from either surface. REST returns a dict; the SDK returns a
+    pydantic model, which json.dumps can't serialise (default=str would collapse it to a
+    single repr string), so unwrap it with model_dump first."""
+    if hasattr(result, "model_dump"):
+        result = result.model_dump(mode="json")
+    return json.dumps(result, indent=2, default=str)
+
+
 def _dispatch_manage(args, dry_run: bool) -> int:
     if dry_run:
         print(f"[DRY RUN] would '{args.command}' via {getattr(args, 'surface', 'sdk')} "
@@ -149,12 +158,12 @@ def _dispatch_manage(args, dry_run: bool) -> int:
     if args.command == "list":
         result = (via_sdk(client, "list", project_id=project_id) if args.surface == "sdk"
                   else via_rest("list", project_id=project_id))
-        print(json.dumps(result, indent=2, default=str))
+        print(_dump(result))
         return 0
     if args.command == "get":
         result = (via_sdk(client, "get", rule_id=args.id, project_id=project_id) if args.surface == "sdk"
                   else via_rest("get", rule_id=args.id, project_id=project_id))
-        print(json.dumps(result, indent=2, default=str))
+        print(_dump(result))
         return 0
     if args.command == "update":
         current = via_rest("get", rule_id=args.id, project_id=project_id)

scripts/online_eval_rules/test_create_online_eval_rules.py

@@ -182,6 +182,21 @@ def test_build_sdk_request_returns_correct_variant():
     assert req.code.schema_[0].name == "relevance_score"
 
 
+def test_dump_unwraps_pydantic_model_to_structured_json():
+    # SDK-surface list/get return pydantic models; _dump must emit structured JSON,
+    # not a single repr string (json.dumps(model, default=str) would do the latter).
+    req = cli.build_sdk_request(
+        cli.build_payload(cli.RULE_LLM_JUDGE, name="r", project_id="pid",
+                          sampling_rate=1.0, model="gpt-4o")
+    )
+    parsed = json.loads(cli._dump(req))
+    assert isinstance(parsed, dict) and parsed["name"] == "r"
+
+
+def test_dump_passes_through_rest_dict():
+    assert json.loads(cli._dump({"content": [], "ok": True})) == {"content": [], "ok": True}
+
+
 class _FakeEvaluators:
     def __init__(self):
         self.created = None