Fix #4509: Strip markdown code-block fences from LLM JSON responses in Converter

When LLMs return JSON wrapped in markdown code blocks (e.g. ```json...```),
the Converter.to_pydantic() method failed to parse the response, causing
Pydantic validation errors when saving to long-term memory.

Changes:
- Add _strip_markdown_fences() helper to remove markdown code-block delimiters
- Apply fence stripping before model_validate_json() in both function-calling
  and non-function-calling paths of Converter.to_pydantic()
- Add handle_partial_json() fallback in the function-calling path
- Apply fence stripping in convert_to_model() before json.loads()
- Add 11 tests covering the fix

Co-Authored-By: João <joao@crewai.com>

lib/crewai/src/crewai/utilities/converter.py

@@ -21,9 +21,32 @@ if TYPE_CHECKING:
     from crewai.llms.base_llm import BaseLLM
 
 _JSON_PATTERN: Final[re.Pattern[str]] = re.compile(r"({.*})", re.DOTALL)
+_MARKDOWN_FENCE_PATTERN: Final[re.Pattern[str]] = re.compile(
+    r"^\s*```(?:json)?\s*\n(.*)\n\s*```\s*$", re.DOTALL
+)
 _I18N = get_i18n()
 
 
+def _strip_markdown_fences(text: str) -> str:
+    """Strip markdown code-block fences that some LLMs wrap around JSON.
+
+    Handles patterns like:
+        ```json\n{...}\n```
+        ```\n{...}\n```
+
+    Args:
+        text: Raw LLM response that may contain markdown fences.
+
+    Returns:
+        The inner content with fences removed, or the original text if
+        no fences are detected.
+    """
+    match = _MARKDOWN_FENCE_PATTERN.match(text)
+    if match:
+        return match.group(1)
+    return text
+
+
 class ConverterError(Exception):
     """Error raised when Converter fails to parse the input."""
 
@@ -65,7 +88,25 @@ class Converter(OutputConverter):
                 if isinstance(response, BaseModel):
                     result = response
                 else:
-                    result = self.model.model_validate_json(response)
+                    cleaned = _strip_markdown_fences(response)
+                    try:
+                        result = self.model.model_validate_json(cleaned)
+                    except (ValidationError, ValueError):
+                        result = handle_partial_json(
+                            result=response,
+                            model=self.model,
+                            is_json_output=False,
+                            agent=None,
+                        )
+                        if not isinstance(result, BaseModel):
+                            if isinstance(result, dict):
+                                result = self.model.model_validate(result)
+                            elif isinstance(result, str):
+                                result = self.model.model_validate_json(result)
+                            else:
+                                raise ConverterError(
+                                    "handle_partial_json returned an unexpected type."
+                                ) from None
             else:
                 response = self.llm.call(
                     [
@@ -74,9 +115,9 @@ class Converter(OutputConverter):
                     ]
                 )
                 try:
-                    # Try to directly validate the response JSON
+                    cleaned = _strip_markdown_fences(response)
-                    result = self.model.model_validate_json(response)
+                    result = self.model.model_validate_json(cleaned)
-                except ValidationError:
+                except (ValidationError, ValueError):
                     # If direct validation fails, attempt to extract valid JSON
                     result = handle_partial_json(  # type: ignore[assignment]
                         result=response,
@@ -185,7 +226,8 @@ def convert_to_model(
         )
 
     try:
-        escaped_result = json.dumps(json.loads(result, strict=False))
+        cleaned_result = _strip_markdown_fences(result)
+        escaped_result = json.dumps(json.loads(cleaned_result, strict=False))
         return validate_model(
             result=escaped_result, model=model, is_json_output=bool(output_json)
         )

lib/crewai/tests/utilities/test_converter.py

@@ -8,6 +8,7 @@ from crewai.llm import LLM
 from crewai.utilities.converter import (
     Converter,
     ConverterError,
+    _strip_markdown_fences,
     convert_to_model,
     convert_with_instructions,
     create_converter,
@@ -952,3 +953,162 @@ def test_internal_instructor_real_unsupported_provider() -> None:
 
     # Verify it's a configuration error about unsupported provider
     assert "Unsupported provider" in str(exc_info.value) or "unsupported" in str(exc_info.value).lower()
+
+
+# ---------------------------------------------------------------------------
+# Tests for _strip_markdown_fences  (issue #4509)
+# ---------------------------------------------------------------------------
+
+
+def test_strip_markdown_fences_with_json_tag() -> None:
+    raw = '```json\n{"name": "Alice", "age": 30}\n```'
+    assert _strip_markdown_fences(raw) == '{"name": "Alice", "age": 30}'
+
+
+def test_strip_markdown_fences_without_language_tag() -> None:
+    raw = '```\n{"name": "Bob", "age": 25}\n```'
+    assert _strip_markdown_fences(raw) == '{"name": "Bob", "age": 25}'
+
+
+def test_strip_markdown_fences_multiline_json() -> None:
+    raw = '```json\n{\n  "name": "Charlie",\n  "age": 35\n}\n```'
+    result = _strip_markdown_fences(raw)
+    parsed = json.loads(result)
+    assert parsed == {"name": "Charlie", "age": 35}
+
+
+def test_strip_markdown_fences_plain_json_unchanged() -> None:
+    raw = '{"name": "Dave", "age": 40}'
+    assert _strip_markdown_fences(raw) == raw
+
+
+def test_strip_markdown_fences_with_surrounding_whitespace() -> None:
+    raw = '  ```json\n{"key": "value"}\n```  '
+    assert json.loads(_strip_markdown_fences(raw)) == {"key": "value"}
+
+
+# ---------------------------------------------------------------------------
+# Tests for Converter.to_pydantic with markdown-wrapped JSON (issue #4509)
+# ---------------------------------------------------------------------------
+
+
+def test_converter_to_pydantic_strips_markdown_fences_function_calling() -> None:
+    """When LLM supports function calling but returns markdown-wrapped JSON string,
+    the converter should strip fences and parse successfully."""
+    llm = Mock(spec=LLM)
+    llm.supports_function_calling.return_value = True
+    llm.call.return_value = '```json\n{"name": "Alice", "age": 30}\n```'
+
+    converter = Converter(
+        llm=llm,
+        text="Name: Alice, Age: 30",
+        model=SimpleModel,
+        instructions="Convert to JSON.",
+    )
+
+    output = converter.to_pydantic()
+    assert isinstance(output, SimpleModel)
+    assert output.name == "Alice"
+    assert output.age == 30
+
+
+def test_converter_to_pydantic_strips_markdown_fences_no_function_calling() -> None:
+    """When LLM does NOT support function calling and returns markdown-wrapped JSON,
+    the converter should strip fences and parse successfully."""
+    llm = Mock(spec=LLM)
+    llm.supports_function_calling.return_value = False
+    llm.call.return_value = '```json\n{"name": "Bob", "age": 25}\n```'
+
+    converter = Converter(
+        llm=llm,
+        text="Name: Bob, Age: 25",
+        model=SimpleModel,
+        instructions="Convert to JSON.",
+    )
+
+    output = converter.to_pydantic()
+    assert isinstance(output, SimpleModel)
+    assert output.name == "Bob"
+    assert output.age == 25
+
+
+def test_converter_to_pydantic_strips_fences_multiline_function_calling() -> None:
+    """Multiline markdown-fenced JSON should be parsed correctly (function calling path)."""
+    from crewai.utilities.evaluators.task_evaluator import Entity, TaskEvaluation
+
+    evaluation_json = json.dumps({
+        "suggestions": ["Improve research depth"],
+        "quality": 8.5,
+        "entities": [
+            {
+                "name": "AI",
+                "type": "Technology",
+                "description": "Artificial Intelligence",
+                "relationships": ["Healthcare"],
+            }
+        ],
+    })
+    llm = Mock(spec=LLM)
+    llm.supports_function_calling.return_value = True
+    llm.call.return_value = f"```json\n{evaluation_json}\n```"
+
+    converter = Converter(
+        llm=llm,
+        text="Evaluate this task.",
+        model=TaskEvaluation,
+        instructions="Convert to JSON.",
+    )
+
+    output = converter.to_pydantic()
+    assert isinstance(output, TaskEvaluation)
+    assert output.quality == 8.5
+    assert len(output.suggestions) == 1
+    assert len(output.entities) == 1
+    assert output.entities[0].name == "AI"
+
+
+def test_converter_to_pydantic_strips_fences_multiline_no_function_calling() -> None:
+    """Multiline markdown-fenced JSON should be parsed correctly (non-function-calling path)."""
+    from crewai.utilities.evaluators.task_evaluator import Entity, TaskEvaluation
+
+    evaluation_json = json.dumps({
+        "suggestions": ["Better analysis needed"],
+        "quality": 7.0,
+        "entities": [],
+    })
+    llm = Mock(spec=LLM)
+    llm.supports_function_calling.return_value = False
+    llm.call.return_value = f"```json\n{evaluation_json}\n```"
+
+    converter = Converter(
+        llm=llm,
+        text="Evaluate this task.",
+        model=TaskEvaluation,
+        instructions="Convert to JSON.",
+    )
+
+    output = converter.to_pydantic()
+    assert isinstance(output, TaskEvaluation)
+    assert output.quality == 7.0
+
+
+# ---------------------------------------------------------------------------
+# Tests for convert_to_model with markdown-wrapped JSON (issue #4509)
+# ---------------------------------------------------------------------------
+
+
+def test_convert_to_model_with_markdown_fenced_json() -> None:
+    """convert_to_model should handle markdown-fenced JSON strings."""
+    result = '```json\n{"name": "Eve", "age": 28}\n```'
+    output = convert_to_model(result, SimpleModel, None, None)
+    assert isinstance(output, SimpleModel)
+    assert output.name == "Eve"
+    assert output.age == 28
+
+
+def test_convert_to_model_with_markdown_fenced_json_output() -> None:
+    """convert_to_model should handle markdown-fenced JSON when output_json is set."""
+    result = '```json\n{"name": "Frank", "age": 45}\n```'
+    output = convert_to_model(result, None, SimpleModel, None)
+    assert isinstance(output, dict)
+    assert output == {"name": "Frank", "age": 45}