Add Tool Failure Evaluator for IntermediateSteps

packages/nvidia_nat_core/src/nat/utils/atif_converter.py

@@ -101,6 +101,29 @@ def _safe_str(value: Any) -> str:
     return str(value)
 
 
+def _extract_tool_error(output: Any) -> dict[str, str] | None:
+    """Extract error metadata from a tool output for ``step.extra["tool_errors"]``."""
+    # TODO: return a model instead of a plain dict once ATIF spec adds error support
+    status: str | None = getattr(output, "status", None) or (output.get("status") if isinstance(output, dict) else None)
+    if status != "error":
+        return None
+    content: str = (getattr(output, "content", None) or (output.get("content") if isinstance(output, dict) else None)
+                    or _safe_str(output))
+    error_type: str = "Unknown"
+    error_message: str = content
+    if ":" in content:
+        candidate: str = content.split(":", 1)[0].strip()
+        if candidate.isidentifier():
+            error_type = candidate
+            error_message = content.split(":", 1)[1].strip()
+    return {
+        "error": content,
+        "error_type": error_type,
+        "error_message": error_message,
+        "status": "error",
+    }
+
+
 def _extract_user_input(value: Any) -> str:
     """Extract the user-facing input text from a workflow start payload.
 
@@ -334,18 +357,31 @@ def _flush_pending() -> None:
                 tool_name = ist.name or "unknown_tool"
                 tool_input: dict[str, Any] = {}
                 tool_output = ""
+                raw_output: Any = None
+
                 if ist.data:
                     tool_input = _parse_tool_arguments(ist.data.input)
-                    tool_output = _safe_str(ist.data.output)
+                    raw_output = ist.data.output
+                    tool_output = _safe_str(raw_output)
                 call_id = f"call_{ist.UUID}"
                 tc = ATIFToolCall(tool_call_id=call_id, function_name=tool_name, arguments=tool_input)
                 obs = ATIFObservationResult(source_call_id=call_id, content=tool_output)
+                tool_error: dict[str, str] | None = _extract_tool_error(raw_output)
+
+                if tool_error is not None:
+                    tool_error["tool"] = tool_name
+                extra: dict[str, Any] | None = ({"tool_errors": [tool_error]} if tool_error else None)
+
                 if pending is not None:
                     pending.tool_calls.append(tc)
                     pending.observations.append(obs)
+                    if tool_error:
+                        pending.extra.setdefault("tool_errors", []).append(tool_error)
                     pending.tool_ancestry.append(_atif_ancestry_from_ist(ist))
                 else:
                     extra = _atif_step_extra_model_from_ist(ist).model_dump(exclude_none=True)
+                    if tool_error:
+                        extra.setdefault("tool_errors", []).append(tool_error)
                     atif_steps.append(
                         ATIFStep(
                             step_id=step_id,
@@ -508,19 +544,28 @@ def push(self, ist: IntermediateStep) -> ATIFStep | None:
             tool_name = ist.name or "unknown_tool"
             tool_input: dict[str, Any] = {}
             tool_output = ""
+            raw_output: Any = None
             if ist.data:
                 tool_input = _parse_tool_arguments(ist.data.input)
-                tool_output = _safe_str(ist.data.output)
+                raw_output = ist.data.output
+                tool_output = _safe_str(raw_output)
             call_id = f"call_{ist.UUID}"
             tc = ATIFToolCall(tool_call_id=call_id, function_name=tool_name, arguments=tool_input)
             obs = ATIFObservationResult(source_call_id=call_id, content=tool_output)
+            tool_error: dict[str, str] | None = _extract_tool_error(raw_output)
+            if tool_error is not None:
+                tool_error["tool"] = tool_name
             if self._pending is not None:
                 self._pending.tool_calls.append(tc)
                 self._pending.observations.append(obs)
+                if tool_error:
+                    self._pending.extra.setdefault("tool_errors", []).append(tool_error)
                 self._pending.tool_ancestry.append(_atif_ancestry_from_ist(ist))
                 return None
 
             extra = _atif_step_extra_model_from_ist(ist).model_dump(exclude_none=True)
+            if tool_error:
+                extra.setdefault("tool_errors", []).append(tool_error)
             orphan_step = ATIFStep(
                 step_id=self._step_id,
                 source="agent",

packages/nvidia_nat_core/tests/nat/utils/test_atif_converter.py

@@ -17,6 +17,7 @@
 import datetime
 
 import pytest
+from langchain_core.messages import ToolMessage
 
 from nat.builder.framework_enum import LLMFrameworkEnum
 from nat.data_models.atif import ATIFTrajectory
@@ -705,3 +706,129 @@ def test_stream_matches_batch(
             assert s_step.message == b_step.message
             if b_step.tool_calls:
                 assert len(s_step.tool_calls) == len(b_step.tool_calls)
+
+
+# ---------------------------------------------------------------------------
+# Tool error → ATIF conversion tests
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture(name="error_trajectory")
+def fixture_error_trajectory() -> list[IntermediateStep]:
+    """Trajectory with one successful and one failed tool call."""
+    error_output: ToolMessage = ToolMessage(
+        content="ValueError: bad input",
+        name="failing_tool",
+        tool_call_id="failing_tool",
+        status="error",
+    )
+    return [
+        _make_step(IntermediateStepType.WORKFLOW_START, input_data="Do something", timestamp_offset=0.0),
+        _make_step(IntermediateStepType.LLM_END,
+                   name="gpt-4",
+                   output_data="calling tools",
+                   timestamp_offset=1.0,
+                   usage=_make_usage(100, 20)),
+        _make_step(IntermediateStepType.TOOL_END,
+                   name="good_tool",
+                   input_data={"q": "hello"},
+                   output_data="success",
+                   timestamp_offset=2.0,
+                   step_uuid="tool-good"),
+        _make_step(IntermediateStepType.TOOL_END,
+                   name="failing_tool",
+                   input_data={"q": "fail"},
+                   output_data=error_output,
+                   timestamp_offset=3.0,
+                   step_uuid="tool-fail"),
+        _make_step(IntermediateStepType.WORKFLOW_END, output_data="partial", timestamp_offset=4.0),
+    ]
+
+
+class TestToolErrorATIFConversion:
+    """Verify tool errors in IntermediateStepPayload are converted to ATIF step.extra['tool_errors']."""
+
+    def test_error_dict_has_all_required_keys(
+        self,
+        batch_converter: IntermediateStepToATIFConverter,
+        error_trajectory: list[IntermediateStep],
+    ):
+        """Each tool_errors entry contains exactly the expected keys."""
+        result: ATIFTrajectory = batch_converter.convert(error_trajectory)
+        agent_step = result.steps[1]
+        errors: list = agent_step.extra["tool_errors"]
+        assert len(errors) == 1
+        assert set(errors[0].keys()) == {"tool", "error", "error_type", "error_message", "status"}
+
+    def test_error_dict_values_are_parsed_from_content(
+        self,
+        batch_converter: IntermediateStepToATIFConverter,
+        error_trajectory: list[IntermediateStep],
+    ):
+        """The error dict splits the exception type from the message and preserves the full error string."""
+        result: ATIFTrajectory = batch_converter.convert(error_trajectory)
+        entry: dict = result.steps[1].extra["tool_errors"][0]
+        assert entry["tool"] == "failing_tool"
+        assert entry["status"] == "error"
+        assert entry["error"] == "ValueError: bad input"
+        assert entry["error_type"] == "ValueError"
+        assert entry["error_message"] == "bad input"
+
+    def test_error_dict_falls_back_to_unknown_type(self):
+        """Error content without a parseable exception type defaults to 'Unknown'."""
+        error_output: ToolMessage = ToolMessage(
+            content="something went wrong",
+            name="broken_tool",
+            tool_call_id="broken_tool",
+            status="error",
+        )
+        trajectory: list[IntermediateStep] = [
+            _make_step(IntermediateStepType.WORKFLOW_START, input_data="q", timestamp_offset=0.0),
+            _make_step(IntermediateStepType.LLM_END,
+                       name="gpt-4",
+                       output_data="calling",
+                       timestamp_offset=1.0,
+                       usage=_make_usage(10, 5)),
+            _make_step(IntermediateStepType.TOOL_END,
+                       name="broken_tool",
+                       input_data={},
+                       output_data=error_output,
+                       timestamp_offset=2.0,
+                       step_uuid="tool-broken"),
+            _make_step(IntermediateStepType.WORKFLOW_END, output_data="done", timestamp_offset=3.0),
+        ]
+        result: ATIFTrajectory = IntermediateStepToATIFConverter().convert(trajectory)
+        entry: dict = result.steps[1].extra["tool_errors"][0]
+        assert entry["error_type"] == "Unknown"
+        assert entry["error_message"] == "something went wrong"
+
+    def test_successful_tool_has_no_tool_errors(
+        self,
+        batch_converter: IntermediateStepToATIFConverter,
+        simple_trajectory: list[IntermediateStep],
+    ):
+        """Successful tool calls do not produce tool_errors entries in the ATIF output."""
+        result: ATIFTrajectory = batch_converter.convert(simple_trajectory)
+        for step in result.steps:
+            assert not (step.extra or {}).get("tool_errors")
+
+    def test_stream_and_batch_produce_same_errors(
+        self,
+        batch_converter: IntermediateStepToATIFConverter,
+        error_trajectory: list[IntermediateStep],
+    ):
+        """Both converter code paths produce identical tool_errors for the same input trajectory."""
+        batch_result: ATIFTrajectory = batch_converter.convert(error_trajectory)
+        stream_conv: ATIFStreamConverter = ATIFStreamConverter()
+        for ist in error_trajectory:
+            stream_conv.push(ist)
+        stream_conv.finalize()
+        stream_result: ATIFTrajectory = stream_conv.get_trajectory()
+
+        def _collect_errors(trajectory: ATIFTrajectory) -> list[dict]:
+            errors: list[dict] = []
+            for step in trajectory.steps:
+                errors.extend((step.extra or {}).get("tool_errors", []))
+            return errors
+
+        assert _collect_errors(batch_result) == _collect_errors(stream_result)

packages/nvidia_nat_eval/src/nat/plugins/eval/register.py

@@ -22,3 +22,6 @@
 from .dataset_loader.register import register_jsonl_dataset_loader
 from .dataset_loader.register import register_parquet_dataset_loader
 from .dataset_loader.register import register_xls_dataset_loader
+
+# Evaluators
+from .tool_failure_evaluator.register import register_tool_failure_evaluator

packages/nvidia_nat_eval/src/nat/plugins/eval/tool_failure_evaluator/__init__.py

@@ -0,0 +1,26 @@
+# SPDX-FileCopyrightText: Copyright (c) 2025-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .evaluator import ToolFailureEvaluator
+from .models import ToolFailureReasoning
+from .models import ToolSummary
+from .register import ToolFailureEvaluatorConfig
+
+__all__ = [
+    "ToolFailureEvaluator",
+    "ToolFailureEvaluatorConfig",
+    "ToolFailureReasoning",
+    "ToolSummary",
+]