Skip to content

Framework for orchestrating role-playing, autonomous AI agents. By fostering collaborative intelligence, CrewAI empowers agents to work together seamlessly, tackling complex tasks.

fix: sanitize tool schemas for strict mode

Greyson LaLonde8de44217055a
diff --git a/lib/crewai/src/crewai/llms/providers/anthropic/completion.py b/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
index 0d666a63c..c0e854ce0 100644
--- a/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/anthropic/completion.py
@@ -11,10 +11,14 @@ from crewai.events.types.llm_events import LLMCallType
from crewai.llms.base_llm import BaseLLM, JsonResponseFormat, llm_call_context
from crewai.llms.hooks.base import BaseInterceptor
from crewai.llms.hooks.transport import AsyncHTTPTransport, HTTPTransport
+from crewai.llms.providers.utils.common import safe_tool_conversion
from crewai.utilities.agent_utils import is_context_length_exceeded
from crewai.utilities.exceptions.context_window_exceeding_exception import (
LLMContextLengthExceededError,
)
+from crewai.utilities.pydantic_schema_utils import (
+ sanitize_tool_params_for_anthropic_strict,
+)
from crewai.utilities.types import LLMMessage
@@ -473,10 +477,8 @@ class AnthropicCompletion(BaseLLM):
continue
try:
- from crewai.llms.providers.utils.common import safe_tool_conversion
-
name, description, parameters = safe_tool_conversion(tool, "Anthropic")
- except (ImportError, KeyError, ValueError) as e:
+ except (KeyError, ValueError) as e:
logging.error(f"Error converting tool to Anthropic format: {e}")
raise e
@@ -485,8 +487,15 @@ class AnthropicCompletion(BaseLLM):
"description": description,
}
+ func_info = tool.get("function", {})
+ strict_enabled = bool(func_info.get("strict"))
+
if parameters and isinstance(parameters, dict):
- anthropic_tool["input_schema"] = parameters
+ anthropic_tool["input_schema"] = (
+ sanitize_tool_params_for_anthropic_strict(parameters)
+ if strict_enabled
+ else parameters
+ )
else:
anthropic_tool["input_schema"] = {
"type": "object",
@@ -494,8 +503,7 @@ class AnthropicCompletion(BaseLLM):
"required": [],
}
- func_info = tool.get("function", {})
- if func_info.get("strict"):
+ if strict_enabled:
anthropic_tool["strict"] = True
anthropic_tools.append(anthropic_tool)
diff --git a/lib/crewai/src/crewai/llms/providers/bedrock/completion.py b/lib/crewai/src/crewai/llms/providers/bedrock/completion.py
index 0747b7006..75fe712b5 100644
--- a/lib/crewai/src/crewai/llms/providers/bedrock/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/bedrock/completion.py
@@ -12,11 +12,15 @@ from typing_extensions import Required
from crewai.events.types.llm_events import LLMCallType
from crewai.llms.base_llm import BaseLLM, llm_call_context
+from crewai.llms.providers.utils.common import safe_tool_conversion
from crewai.utilities.agent_utils import is_context_length_exceeded
from crewai.utilities.exceptions.context_window_exceeding_exception import (
LLMContextLengthExceededError,
)
-from crewai.utilities.pydantic_schema_utils import generate_model_description
+from crewai.utilities.pydantic_schema_utils import (
+ generate_model_description,
+ sanitize_tool_params_for_bedrock_strict,
+)
from crewai.utilities.types import LLMMessage
@@ -1949,8 +1953,6 @@ class BedrockCompletion(BaseLLM):
tools: list[dict[str, Any]],
) -> list[ConverseToolTypeDef]:
"""Convert CrewAI tools to Converse API format following AWS specification."""
- from crewai.llms.providers.utils.common import safe_tool_conversion
-
converse_tools: list[ConverseToolTypeDef] = []
for tool in tools:
@@ -1962,12 +1964,19 @@ class BedrockCompletion(BaseLLM):
"description": description,
}
+ func_info = tool.get("function", {})
+ strict_enabled = bool(func_info.get("strict"))
+
if parameters and isinstance(parameters, dict):
- input_schema: ToolInputSchema = {"json": parameters}
+ schema_params = (
+ sanitize_tool_params_for_bedrock_strict(parameters)
+ if strict_enabled
+ else parameters
+ )
+ input_schema: ToolInputSchema = {"json": schema_params}
tool_spec["inputSchema"] = input_schema
- func_info = tool.get("function", {})
- if func_info.get("strict"):
+ if strict_enabled:
tool_spec["strict"] = True
converse_tool: ConverseToolTypeDef = {"toolSpec": tool_spec}
diff --git a/lib/crewai/src/crewai/llms/providers/openai/completion.py b/lib/crewai/src/crewai/llms/providers/openai/completion.py
index 668794e83..e9509e7d2 100644
--- a/lib/crewai/src/crewai/llms/providers/openai/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/openai/completion.py
@@ -32,11 +32,15 @@ from crewai.events.types.llm_events import LLMCallType
from crewai.llms.base_llm import BaseLLM, JsonResponseFormat, llm_call_context
from crewai.llms.hooks.base import BaseInterceptor
from crewai.llms.hooks.transport import AsyncHTTPTransport, HTTPTransport
+from crewai.llms.providers.utils.common import safe_tool_conversion
from crewai.utilities.agent_utils import is_context_length_exceeded
from crewai.utilities.exceptions.context_window_exceeding_exception import (
LLMContextLengthExceededError,
)
-from crewai.utilities.pydantic_schema_utils import generate_model_description
+from crewai.utilities.pydantic_schema_utils import (
+ generate_model_description,
+ sanitize_tool_params_for_openai_strict,
+)
from crewai.utilities.types import LLMMessage
@@ -764,8 +768,6 @@ class OpenAICompletion(BaseLLM):
"function": {"name": "...", "description": "...", "parameters": {...}}
}
"""
- from crewai.llms.providers.utils.common import safe_tool_conversion
-
responses_tools = []
for tool in tools:
@@ -1548,11 +1550,6 @@ class OpenAICompletion(BaseLLM):
self, tools: list[dict[str, BaseTool]]
) -> list[dict[str, Any]]:
"""Convert CrewAI tool format to OpenAI function calling format."""
- from crewai.llms.providers.utils.common import safe_tool_conversion
- from crewai.utilities.pydantic_schema_utils import (
- force_additional_properties_false,
- )
-
openai_tools = []
for tool in tools:
@@ -1571,8 +1568,9 @@ class OpenAICompletion(BaseLLM):
params_dict = (
parameters if isinstance(parameters, dict) else dict(parameters)
)
- params_dict = force_additional_properties_false(params_dict)
- openai_tool["function"]["parameters"] = params_dict
+ openai_tool["function"]["parameters"] = (
+ sanitize_tool_params_for_openai_strict(params_dict)
+ )
openai_tools.append(openai_tool)
return openai_tools
diff --git a/lib/crewai/src/crewai/utilities/pydantic_schema_utils.py b/lib/crewai/src/crewai/utilities/pydantic_schema_utils.py
index 62536cbe7..4c69c9bf6 100644
--- a/lib/crewai/src/crewai/utilities/pydantic_schema_utils.py
+++ b/lib/crewai/src/crewai/utilities/pydantic_schema_utils.py
@@ -19,7 +19,7 @@ from collections.abc import Callable
from copy import deepcopy
import datetime
import logging
-from typing import TYPE_CHECKING, Annotated, Any, Final, Literal, TypedDict, Union
+from typing import TYPE_CHECKING, Annotated, Any, Final, Literal, TypedDict, Union, cast
import uuid
import jsonref # type: ignore[import-untyped]
@@ -417,6 +417,119 @@ def strip_null_from_types(schema: dict[str, Any]) -> dict[str, Any]:
return schema
+_STRICT_METADATA_KEYS: Final[tuple[str, ...]] = (
+ "title",
+ "default",
+ "examples",
+ "example",
+ "$comment",
+ "readOnly",
+ "writeOnly",
+ "deprecated",
+)
+
+_CLAUDE_STRICT_UNSUPPORTED: Final[tuple[str, ...]] = (
+ "minimum",
+ "maximum",
+ "exclusiveMinimum",
+ "exclusiveMaximum",
+ "multipleOf",
+ "minLength",
+ "maxLength",
+ "pattern",
+ "minItems",
+ "maxItems",
+ "uniqueItems",
+ "minContains",
+ "maxContains",
+ "minProperties",
+ "maxProperties",
+ "patternProperties",
+ "propertyNames",
+ "dependentRequired",
+ "dependentSchemas",
+)
+
+
+def _strip_keys_recursive(d: Any, keys: tuple[str, ...]) -> Any:
+ """Recursively delete a fixed set of keys from a schema."""
+ if isinstance(d, dict):
+ for key in keys:
+ d.pop(key, None)
+ for v in d.values():
+ _strip_keys_recursive(v, keys)
+ elif isinstance(d, list):
+ for i in d:
+ _strip_keys_recursive(i, keys)
+ return d
+
+
+def lift_top_level_anyof(schema: dict[str, Any]) -> dict[str, Any]:
+ """Unwrap a top-level anyOf/oneOf/allOf wrapping a single object variant.
+
+ Anthropic's strict ``input_schema`` rejects top-level union keywords. When
+ exactly one variant is an object schema, lift it so the root is a plain
+ object; otherwise leave the schema alone.
+ """
+ for key in ("anyOf", "oneOf", "allOf"):
+ variants = schema.get(key)
+ if not isinstance(variants, list):
+ continue
+ object_variants = [
+ v for v in variants if isinstance(v, dict) and v.get("type") == "object"
+ ]
+ if len(object_variants) == 1:
+ lifted = deepcopy(object_variants[0])
+ schema.pop(key)
+ schema.update(lifted)
+ break
+ return schema
+
+
+def _common_strict_pipeline(params: dict[str, Any]) -> dict[str, Any]:
+ """Shared strict sanitization: inline refs, close objects, require all properties."""
+ sanitized = resolve_refs(deepcopy(params))
+ sanitized.pop("$defs", None)
+ sanitized = convert_oneof_to_anyof(sanitized)
+ sanitized = ensure_type_in_schemas(sanitized)
+ sanitized = force_additional_properties_false(sanitized)
+ sanitized = ensure_all_properties_required(sanitized)
+ return cast(dict[str, Any], _strip_keys_recursive(sanitized, _STRICT_METADATA_KEYS))
+
+
+def sanitize_tool_params_for_openai_strict(
+ params: dict[str, Any],
+) -> dict[str, Any]:
+ """Sanitize a JSON schema for OpenAI strict function calling."""
+ if not isinstance(params, dict):
+ return params
+ return cast(
+ dict[str, Any], strip_unsupported_formats(_common_strict_pipeline(params))
+ )
+
+
+def sanitize_tool_params_for_anthropic_strict(
+ params: dict[str, Any],
+) -> dict[str, Any]:
+ """Sanitize a JSON schema for Anthropic strict tool use."""
+ if not isinstance(params, dict):
+ return params
+ sanitized = lift_top_level_anyof(_common_strict_pipeline(params))
+ sanitized = _strip_keys_recursive(sanitized, _CLAUDE_STRICT_UNSUPPORTED)
+ return cast(dict[str, Any], strip_unsupported_formats(sanitized))
+
+
+def sanitize_tool_params_for_bedrock_strict(
+ params: dict[str, Any],
+) -> dict[str, Any]:
+ """Sanitize a JSON schema for Bedrock Converse strict tool use.
+
+ Bedrock Converse uses the same grammar compiler as the underlying Claude
+ model, so the constraints match Anthropic's.
+ """
+ return sanitize_tool_params_for_anthropic_strict(params)
+
+
def generate_model_description(
model: type[BaseModel],
*,