code editor: Agent Studio improvement iteration 2 for invoice-guard-studio-v1. Goal

agent.py

@@ -44,6 +44,7 @@ MAX_INVOICES = 50
 MAX_UPLOADS = 5
 MAX_UPLOAD_BYTES = 256 * 1024
 MAX_TEXT_FIELD_LENGTH = 180
+AGENT_VERSION = "0.1.3"
 ACCEPTED_MEDIA_TYPES = {"application/json", "text/json", "text/csv", "text/plain"}
 DB_CONNECT_OPTIONS = "-c statement_timeout=5000 -c lock_timeout=3000 -c idle_in_transaction_session_timeout=10000"
 
@@ -81,7 +82,7 @@ class InvoiceGuardStudioV1(A2AAgent[InvoiceGuardStudioV1Config, PlatformUserAuth
         "One-page InvoiceGuard app for deterministic duplicate invoice number "
         "and stated-total mismatch review with user-scoped Postgres persistence."
     )
-    version = "0.1.2"
+    version = AGENT_VERSION
 
     config_model = InvoiceGuardStudioV1Config
     auth_model = PlatformUserAuth
@@ -124,7 +125,7 @@ class InvoiceGuardStudioV1(A2AAgent[InvoiceGuardStudioV1Config, PlatformUserAuth
             str,
             Field(min_length=1, max_length=120, pattern=r"^[A-Za-z0-9_.:-]+$"),
         ],
-        invoices: Annotated[list[dict[str, Any]], Field(min_length=1, max_length=MAX_INVOICES)],
+        invoices: Annotated[list[InvoiceInput], Field(min_length=1, max_length=MAX_INVOICES)],
     ) -> dict[str, Any]:
         tenant = _tenant_key(ctx)
         result = _review_payload(case_id, invoices)
@@ -300,7 +301,7 @@ def _money_json(value: Decimal) -> float:
     return float(value.quantize(Decimal("0.01"), rounding=ROUND_HALF_UP))
 
 
-def _review_payload(case_id: str, raw_invoices: list[dict[str, Any]]) -> dict[str, Any]:
+def _review_payload(case_id: str, raw_invoices: list[InvoiceInput | dict[str, Any]]) -> dict[str, Any]:
     try:
         clean_case_id = _clean_case_id(case_id)
     except ValueError as exc:
@@ -311,16 +312,29 @@ def _review_payload(case_id: str, raw_invoices: list[dict[str, Any]]) -> dict[st
         return _validation_error("too_many_invoices", f"At most {MAX_INVOICES} invoices are allowed per review.", case_id=clean_case_id)
 
     invoices: list[dict[str, Any]] = []
-    for index, item in enumerate(raw_invoices):
-        if not isinstance(item, dict):
+    for index, raw_item in enumerate(raw_invoices):
+        if isinstance(raw_item, InvoiceInput):
+            item = raw_item.model_dump()
+        elif isinstance(raw_item, dict):
+            try:
+                item = InvoiceInput.model_validate(raw_item).model_dump()
+            except ValidationError:
+                item = raw_item
+        else:
             return _validation_error("invalid_invoice", f"Invoice at index {index} must be an object.", case_id=clean_case_id)
         missing = [name for name in ("invoice_number", "vendor", "subtotal", "tax", "total") if item.get(name) in (None, "")]
+        invoice_number = str(item.get("invoice_number", "")).strip()
+        vendor = str(item.get("vendor", "")).strip()
+        if not invoice_number or not vendor:
+            missing.extend(name for name, value in (("invoice_number", invoice_number), ("vendor", vendor)) if not value and name not in missing)
         if missing:
             return _validation_error(
                 "incomplete_invoice",
                 f"Invoice at index {index} is missing required field(s): {', '.join(missing)}.",
                 case_id=clean_case_id,
             )
+        if len(invoice_number) > 120 or len(vendor) > MAX_TEXT_FIELD_LENGTH:
+            return _validation_error("invalid_invoice", f"Invoice at index {index} exceeds text field length limits.", case_id=clean_case_id)
         subtotal = _money(item.get("subtotal"))
         tax = _money(item.get("tax"))
         total = _money(item.get("total"))
@@ -329,8 +343,8 @@ def _review_payload(case_id: str, raw_invoices: list[dict[str, Any]]) -> dict[st
         invoices.append(
             {
                 "index": index,
-                "invoice_number": str(item["invoice_number"]).strip(),
-                "vendor": str(item["vendor"]).strip(),
+                "invoice_number": invoice_number,
+                "vendor": vendor,
                 "subtotal": _money_json(subtotal),
                 "tax": _money_json(tax),
                 "total": _money_json(total),
@@ -473,7 +487,7 @@ def _receipt_for(tenant: str, case: dict[str, Any], source: str, saved_at: str)
         "receipt_id": f"invoice_guard:{case['case_id']}:{digest[:16]}",
         "kind": "production_execution_receipt",
         "agent": "invoice-guard-studio-v1",
-        "agent_version": "0.1.1",
+        "agent_version": AGENT_VERSION,
         "skill": "review_invoices",
         "tenant_hash": hashlib.sha256(tenant.encode("utf-8")).hexdigest()[:16],
         "input_hash": digest,

tests/test_full_stack_contract.py

@@ -76,6 +76,15 @@ def test_incomplete_invoice_failure_fixture():
     assert result["code"] == "incomplete_invoice"
     assert "vendor" in result["message"] and "total" in result["message"]
 
+    blank = _review_payload("studio-invoice-invalid", [{"invoice_number": "   ", "vendor": "   ", "subtotal": 10, "tax": 1, "total": 11}])
+    assert blank["ok"] is False
+    assert blank["code"] == "incomplete_invoice"
+    assert "invoice_number" in blank["message"] and "vendor" in blank["message"]
+
+    oversized = _review_payload("studio-invoice-invalid", [{"invoice_number": "I" * 121, "vendor": "Acme", "subtotal": 10, "tax": 1, "total": 11}])
+    assert oversized["ok"] is False
+    assert oversized["code"] == "invalid_invoice"
+
 
 def test_browser_upload_bridge_bounds_and_parses_json():
     payload = json.dumps({"invoices": SUCCESS_INVOICES}).encode()