fix: normalize optional fields in tool invocation identity (#4289)

examples/financial_research_agent/agents/verifier_agent.py

@@ -12,9 +12,12 @@ VERIFIER_PROMPT = (
     "URLs. Judge the report only against that supplied evidence; do not reject or approve claims "
     "based on your own memory. Check that material numeric and time-sensitive claims are supported "
     "by the evidence, that citations use supplied URLs, that the report is internally consistent, "
-    "and that uncertainty is appropriately caveated. Treat information published on or before the "
-    "research cutoff as potentially available. Mark unsupported claims separately from claims that "
-    "the evidence directly contradicts."
+    "and that uncertainty is appropriately caveated. The payload includes allowed_source_urls; "
+    "compare citation URL strings exactly against that list. If a citation exactly matches an "
+    "allowed URL, accept the URL even when it contains a tracking parameter or another allowed "
+    "variant exists. Treat information published on or before the research cutoff as potentially "
+    "available. Mark unsupported claims separately from claims that the evidence directly "
+    "contradicts."
 )
 
 

examples/financial_research_agent/agents/writer_agent.py

@@ -17,7 +17,9 @@ WRITER_PROMPT = (
 REVISION_PROMPT = (
     f"{WRITER_PROMPT} You are revising an existing report after evidence verification. Address "
     "every verification issue, remove claims that cannot be supported, preserve valid analysis, "
-    "and return a complete replacement report rather than a patch or commentary."
+    "and return a complete replacement report rather than a patch or commentary. Copy replacement "
+    "citation URLs verbatim from each verification issue's source_urls; do not change path casing, "
+    "language segments, or query parameters."
 )
 
 

examples/financial_research_agent/manager.py

@@ -237,18 +237,31 @@ class FinancialResearchManager:
         search_results: Sequence[FinancialSearchEvidence],
     ) -> VerificationResult:
         self.printer.update_item("verifying", "Verifying report...")
-        input_data = json.dumps(
+        result = await Runner.run(
+            verifier_agent,
+            self._verification_input(query, report, search_results),
+        )
+        self.printer.mark_item_done("verifying")
+        return result.final_output_as(VerificationResult)
+
+    def _verification_input(
+        self,
+        query: str,
+        report: FinancialReportData,
+        search_results: Sequence[FinancialSearchEvidence],
+    ) -> str:
+        return json.dumps(
             {
                 "original_query": query,
                 "research_cutoff": self.research_cutoff,
                 "report": report.model_dump(mode="json"),
                 "evidence": [item.model_dump(mode="json") for item in search_results],
+                "allowed_source_urls": sorted(
+                    {source.url for item in search_results for source in item.sources}
+                ),
             },
             ensure_ascii=False,
         )
-        result = await Runner.run(verifier_agent, input_data)
-        self.printer.mark_item_done("verifying")
-        return result.final_output_as(VerificationResult)
 
     def _report_input(
         self,

examples/run_examples.py

@@ -79,6 +79,8 @@ DEFAULT_AUTO_SKIP = {
     "examples/sandbox/docker/mounts/gcs_mount_read_write.py",
     "examples/sandbox/docker/mounts/s3_files_mount_read_write.py",
     "examples/sandbox/docker/mounts/s3_mount_read_write.py",
+    # Blaxel 0.3.2 still imports an MCP v1 module that was removed in MCP v2.
+    "examples/sandbox/extensions/blaxel_runner.py",
     "examples/sandbox/extensions/daytona/usaspending_text2sql/setup_db.py",
     "examples/sandbox/extensions/temporal/temporal_sandbox_agent.py",
     # Temporarily disabled due to credential issues.

src/agents/_tool_invocation.py

@@ -72,15 +72,16 @@ def _as_mapping(value: Any) -> Mapping[str, Any] | None:
     return None
 
 
-def _normalize_value(value: Any) -> Any:
+def _normalize_value(value: Any, *, exclude_none: bool = False) -> Any:
     mapping = _as_mapping(value)
     if mapping is not None:
         return {
-            str(key): _normalize_value(item)
+            str(key): _normalize_value(item, exclude_none=exclude_none)
             for key, item in sorted(mapping.items(), key=lambda pair: str(pair[0]))
+            if not (exclude_none and item is None)
         }
     if isinstance(value, Sequence) and not isinstance(value, str | bytes | bytearray):
-        return [_normalize_value(item) for item in value]
+        return [_normalize_value(item, exclude_none=exclude_none) for item in value]
     if value is None or isinstance(value, str | int | float | bool):
         return value
     return str(value)
@@ -197,8 +198,12 @@ def tool_invocation_identity_and_scope(
         if field_name not in mapping:
             continue
         value = mapping[field_name]
+        if value is None:
+            continue
         semantic_payload[field_name] = (
-            _normalize_arguments(value) if field_name == "arguments" else _normalize_value(value)
+            _normalize_arguments(value)
+            if field_name == "arguments"
+            else _normalize_value(value, exclude_none=True)
         )
 
     return (

tests/test_example_workflows.py

@@ -279,6 +279,30 @@ def test_financial_report_input_includes_cutoff_and_evidence() -> None:
     }
 
 
+def test_financial_verification_input_lists_exact_allowed_source_urls() -> None:
+    manager = object.__new__(FinancialResearchManager)
+    manager.research_cutoff = "2026-07-11"
+    source_url = "https://example.com/report?utm_source=openai"
+    evidence = FinancialSearchEvidence(
+        query="company annual report",
+        reason="Ground annual metrics",
+        summary="Revenue increased.",
+        sources=[FinancialSource(title="Annual report", url=source_url)],
+        retrieved_at="2026-07-11",
+    )
+    report = FinancialReportData(
+        short_summary="Summary",
+        markdown_report=f"Revenue increased ([source]({source_url})).",
+        follow_up_questions=[],
+    )
+
+    payload = json.loads(manager._verification_input("Analyze the company", report, [evidence]))
+
+    assert payload["allowed_source_urls"] == [source_url]
+    assert payload["report"] == report.model_dump(mode="json")
+    assert payload["evidence"] == [evidence.model_dump(mode="json")]
+
+
 def test_sandbox_basic_direct_run_imports_external_docker_sdk(
     monkeypatch: pytest.MonkeyPatch,
     tmp_path: Path,

tests/test_run_examples_script.py

@@ -11,6 +11,7 @@ def test_default_auto_skip_excludes_prerequisite_bound_examples() -> None:
         "examples/sandbox/docker/mounts/gcs_mount_read_write.py",
         "examples/sandbox/docker/mounts/s3_files_mount_read_write.py",
         "examples/sandbox/docker/mounts/s3_mount_read_write.py",
+        "examples/sandbox/extensions/blaxel_runner.py",
         "examples/sandbox/extensions/daytona/usaspending_text2sql/setup_db.py",
         "examples/sandbox/extensions/temporal/temporal_sandbox_agent.py",
         "examples/sandbox/extensions/vercel_runner.py",

tests/test_tool_approval_call_id_reuse.py

@@ -79,6 +79,41 @@ def test_canonical_shell_identity_ignores_stripped_provider_metadata() -> None:
     assert tool_invocation_identity(provider_call) == tool_invocation_identity(persisted_call)
 
 
+def test_canonical_shell_identity_treats_optional_nulls_as_omitted() -> None:
+    provider_call = {
+        "type": "shell_call",
+        "call_id": "shell_0",
+        "action": {
+            "commands": ["echo safe"],
+            "max_output_length": None,
+            "timeout_ms": None,
+        },
+        "environment": None,
+    }
+    normalized_call = dict(provider_call)
+    normalized_call["action"] = {"commands": ["echo safe"]}
+    normalized_call.pop("environment")
+
+    assert tool_invocation_identity(provider_call) == tool_invocation_identity(normalized_call)
+
+
+def test_canonical_function_identity_preserves_null_argument_values() -> None:
+    call_with_null = {
+        "type": "function_call",
+        "call_id": "call_0",
+        "name": "lookup",
+        "arguments": '{"value": null}',
+    }
+    call_without_value = {
+        "type": "function_call",
+        "call_id": "call_0",
+        "name": "lookup",
+        "arguments": "{}",
+    }
+
+    assert tool_invocation_identity(call_with_null) != tool_invocation_identity(call_without_value)
+
+
 @pytest.mark.asyncio
 @pytest.mark.parametrize("fallback_type", ["custom", "function"])
 async def test_completed_apply_patch_fallback_run_state_round_trip(fallback_type: str) -> None: