diff --git a/litellm/litellm_core_utils/prompt_templates/common_utils.py b/litellm/litellm_core_utils/prompt_templates/common_utils.py index d70530534da..2485896184e 100644 --- a/litellm/litellm_core_utils/prompt_templates/common_utils.py +++ b/litellm/litellm_core_utils/prompt_templates/common_utils.py @@ -6,8 +6,8 @@ import io import json import mimetypes import re -from collections.abc import Iterable, Iterator, Mapping, Sequence -from itertools import groupby +from collections.abc import Callable, Iterable, Iterator, Mapping, Sequence +from itertools import groupby, islice from os import PathLike from pathlib import Path from types import MappingProxyType @@ -1320,17 +1320,128 @@ def flatten_top_level_schema_combinators(schema: Mapping[str, object]) -> Mappin return _flatten_schema_against_root(schema, schema, frozenset(), 0, {}) # mutable-ok: fresh per-call $ref memo -def tool_with_flattened_parameters(tool: Mapping[str, object]) -> Mapping[str, object]: +_SUBSCHEMA_KEYWORDS: Final = frozenset( + { + "additionalItems", + "additionalProperties", + "contains", + "else", + "if", + "items", + "not", + "propertyNames", + "then", + "unevaluatedItems", + "unevaluatedProperties", + } +) +_SUBSCHEMA_LIST_KEYWORDS: Final = frozenset({"allOf", "anyOf", "items", "oneOf", "prefixItems"}) +_SUBSCHEMA_MAP_KEYWORDS: Final = frozenset( + {"$defs", "definitions", "dependentSchemas", "patternProperties", "properties"} +) + +_MAX_SCHEMA_NESTING: Final = 1024 + + +def drop_non_python_regex_patterns(schema: Mapping[str, object]) -> Mapping[str, object]: + """Drop every regex in a schema position that Python's ``re`` cannot compile. + + OpenAI validates tool ``parameters`` against the 2020-12 metaschema with + ``jsonschema``'s format checker, which hands each ``pattern`` value and each + ``patternProperties`` key to ``re.compile``, so a regex written for an + ECMA-262 engine (Unicode property escapes such as ``\\p{Cc}``, as in Claude + Code's ``Artifact`` tool) is refused with "'...' is not a 'regex'" by every + model family on both the chat and Responses wires. Only schema positions are + walked (properties, items, combinators, ``$defs`` and the other applicators), + so a ``pattern`` key inside ``default``, ``examples``, ``const`` or vendor + extensions is data and stays. Outside strict mode the keyword is only a + hint, so dropping it costs the model a constraint and the caller nothing. + Compilable regexes and everything else pass through, the input is never + mutated, and the same object comes back when nothing was dropped. The walk + is level-order rather than recursive, rebuilt deepest level first, and stops + at more schema levels than a JSON parser admits, so a cyclic schema built in + code cannot spin it. + """ + rebuilt: dict[int, Mapping[str, object]] = {} # mutable-ok: per-call memo of rewritten nodes, deepest level first + for level in reversed(tuple(islice(_schema_levels(schema), _MAX_SCHEMA_NESTING))): + rebuilt.update( + (id(node), rewritten) + for node in level + if (rewritten := _node_without_non_python_regex(node, rebuilt)) is not node + ) + return rebuilt.get(id(schema), schema) + + +def _schema_levels(schema: Mapping[str, object]) -> Iterator[tuple[Mapping[str, object], ...]]: + frontier: tuple[Mapping[str, object], ...] = (schema,) # rebind-ok: level-order cursor, one level a round + while frontier: + yield frontier + frontier = tuple(child for node in frontier for child in _subschemas(node)) + + +def _subschemas(node: Mapping[str, object]) -> Iterator[Mapping[str, object]]: + for key, value in node.items(): + if key in _SUBSCHEMA_MAP_KEYWORDS and isinstance(value, dict): + yield from (sub for sub in value.values() if isinstance(sub, dict)) + elif key in _SUBSCHEMA_LIST_KEYWORDS and isinstance(value, list): + yield from (sub for sub in value if isinstance(sub, dict)) + elif key in _SUBSCHEMA_KEYWORDS and isinstance(value, dict): + yield value + + +def _node_without_non_python_regex( + node: Mapping[str, object], rebuilt: Mapping[int, Mapping[str, object]] +) -> Mapping[str, object]: + kept: Final = { # mutable-ok: tool parameters are JSON dicts + key: _keyword_value_rebuilt(key, value, rebuilt) + for key, value in node.items() + if key != "pattern" or not isinstance(value, str) or _is_python_regex(value) + } + return node if len(kept) == len(node) and all(kept[key] is node[key] for key in kept) else kept + + +def _keyword_value_rebuilt(key: str, value: object, rebuilt: Mapping[int, Mapping[str, object]]) -> object: + if key in _SUBSCHEMA_MAP_KEYWORDS and isinstance(value, dict): + kept: Final = { # mutable-ok: tool parameters are JSON dicts + name: rebuilt.get(id(sub), sub) + for name, sub in value.items() + if key != "patternProperties" or not isinstance(name, str) or _is_python_regex(name) + } + return value if len(kept) == len(value) and all(kept[name] is value[name] for name in kept) else kept + if key in _SUBSCHEMA_LIST_KEYWORDS and isinstance(value, list): + items: Final = [rebuilt.get(id(sub), sub) for sub in value] # mutable-ok: tool parameters are JSON lists + return value if all(new is old for new, old in zip(items, value, strict=True)) else items + if key in _SUBSCHEMA_KEYWORDS and isinstance(value, dict): + return rebuilt.get(id(value), value) + return value + + +def _is_python_regex(pattern: str) -> bool: + try: + re.compile(pattern) + except (re.error, RecursionError): + return False + return True + + +def flatten_combinators_and_drop_non_python_regex_patterns(schema: Mapping[str, object]) -> Mapping[str, object]: + return flatten_top_level_schema_combinators(drop_non_python_regex_patterns(schema)) + + +def tool_with_sanitized_parameters( + tool: Mapping[str, object], + sanitize: Callable[[Mapping[str, object]], Mapping[str, object]], +) -> Mapping[str, object]: function: Final = tool.get("function") if not isinstance(function, dict): return tool parameters: Final = function.get("parameters") if not isinstance(parameters, dict): return tool - flattened: Final = flatten_top_level_schema_combinators(parameters) - if flattened is parameters: + sanitized: Final = sanitize(parameters) + if sanitized is parameters: return tool - return {**tool, "function": {**function, "parameters": flattened}} # mutable-ok: request tools are JSON dicts + return {**tool, "function": {**function, "parameters": sanitized}} # mutable-ok: request tools are JSON dicts def _get_image_mime_type_from_url(url: str) -> str | None: diff --git a/litellm/llms/azure/chat/gpt_transformation.py b/litellm/llms/azure/chat/gpt_transformation.py index 880a51eb584..ed16d7f3de0 100644 --- a/litellm/llms/azure/chat/gpt_transformation.py +++ b/litellm/llms/azure/chat/gpt_transformation.py @@ -7,8 +7,9 @@ from httpx._models import Headers, Response import litellm from litellm.litellm_core_utils.prompt_templates.common_utils import ( drop_tool_reference_parts_from_tool_messages, + flatten_combinators_and_drop_non_python_regex_patterns, hoist_images_from_tool_messages, - tool_with_flattened_parameters, + tool_with_sanitized_parameters, ) from litellm.litellm_core_utils.prompt_templates.factory import ( convert_to_azure_openai_messages, @@ -39,14 +40,17 @@ else: _NO_TOOLS_UPDATE: Final[Mapping[str, object]] = MappingProxyType({}) -def flattened_tools_update(optional_params: Mapping[str, object]) -> Mapping[str, object]: +def sanitized_tools_update(optional_params: Mapping[str, object]) -> Mapping[str, object]: tools: Final = optional_params.get("tools") if not isinstance(tools, list): return _NO_TOOLS_UPDATE - flattened: Final = [ # mutable-ok: request tools are a JSON list - tool_with_flattened_parameters(tool) if isinstance(tool, dict) else tool for tool in tools + sanitized: Final = [ # mutable-ok: request tools are a JSON list + tool_with_sanitized_parameters(tool, flatten_combinators_and_drop_non_python_regex_patterns) + if isinstance(tool, dict) + else tool + for tool in tools ] - return MappingProxyType({"tools": flattened}) + return MappingProxyType({"tools": sanitized}) class AzureOpenAIConfig(BaseConfig): @@ -278,7 +282,7 @@ class AzureOpenAIConfig(BaseConfig): "model": model, "messages": azure_messages, **optional_params, - **flattened_tools_update(optional_params), + **sanitized_tools_update(optional_params), } def transform_response( diff --git a/litellm/llms/azure/chat/o_series_transformation.py b/litellm/llms/azure/chat/o_series_transformation.py index 246bf69cb5f..09d8075e857 100644 --- a/litellm/llms/azure/chat/o_series_transformation.py +++ b/litellm/llms/azure/chat/o_series_transformation.py @@ -20,7 +20,7 @@ from litellm.types.llms.openai import AllMessageValues from litellm.utils import get_model_info, supports_reasoning from ...openai.chat.o_series_transformation import OpenAIOSeriesConfig -from .gpt_transformation import flattened_tools_update +from .gpt_transformation import sanitized_tools_update class AzureOpenAIO1Config(OpenAIOSeriesConfig): @@ -111,6 +111,6 @@ class AzureOpenAIO1Config(OpenAIOSeriesConfig): model = model.replace("o_series/", "") # handle o_series/my-random-deployment-name flattened_params: Final = { # mutable-ok: transform_request's contract takes a plain JSON params dict **optional_params, - **flattened_tools_update(optional_params), + **sanitized_tools_update(optional_params), } return super().transform_request(model, messages, flattened_params, litellm_params, headers) diff --git a/litellm/llms/openai/chat/gpt_transformation.py b/litellm/llms/openai/chat/gpt_transformation.py index 9afc6331d96..9b410cf073e 100644 --- a/litellm/llms/openai/chat/gpt_transformation.py +++ b/litellm/llms/openai/chat/gpt_transformation.py @@ -19,10 +19,12 @@ from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response impo _should_convert_tool_call_to_json_mode, ) from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_non_python_regex_patterns, drop_tool_reference_parts_from_tool_messages, + flatten_combinators_and_drop_non_python_regex_patterns, get_tool_call_names, hoist_images_from_tool_messages, - tool_with_flattened_parameters, + tool_with_sanitized_parameters, ) from litellm.litellm_core_utils.prompt_templates.image_handling import ( async_convert_url_to_base64, @@ -432,7 +434,7 @@ class OpenAIGPTConfig(BaseLLMModelInfo, BaseConfig): custom_llm_provider, api_base ) - def _flattened_tools_update_for_openai( + def _sanitized_tools_update_for_openai( self, optional_params: Mapping[str, object], litellm_params: Mapping[str, object], @@ -440,22 +442,26 @@ class OpenAIGPTConfig(BaseLLMModelInfo, BaseConfig): """ OpenAI's chat completions validator rejects tool `parameters` carrying 'oneOf'/'anyOf'/'allOf'/'enum'/'const'/'not' at the top level for every - model family, unlike the Responses API, where GPT-5+ accepts them. + model family, unlike the Responses API, where GPT-5+ accepts them, and + a `pattern` Python's `re` cannot compile for every model family on both. + A custom api_base on the `openai` provider is usually a proxy in front of + the same validator, so regexes are dropped there too, while the lossier + combinator flattening stays limited to api.openai.com hosts. """ tools: Final = optional_params.get("tools") - if not isinstance(tools, list): - return _NO_TOOLS_UPDATE provider: Final = litellm_params.get("custom_llm_provider") - raw_api_base: Final = litellm_params.get("api_base") - if not self._targets_openai_hosted_endpoint( - provider if isinstance(provider, str) else None, - raw_api_base if isinstance(raw_api_base, str) else None, - ): + if not isinstance(tools, list) or provider != "openai": return _NO_TOOLS_UPDATE - flattened: Final = [ # mutable-ok: request tools are a JSON list - tool_with_flattened_parameters(tool) if isinstance(tool, dict) else tool for tool in tools + raw_api_base: Final = litellm_params.get("api_base") + sanitize: Final = ( + flatten_combinators_and_drop_non_python_regex_patterns + if self._targets_openai_hosted_endpoint(provider, raw_api_base if isinstance(raw_api_base, str) else None) + else drop_non_python_regex_patterns + ) + sanitized: Final = [ # mutable-ok: request tools are a JSON list + tool_with_sanitized_parameters(tool, sanitize) if isinstance(tool, dict) else tool for tool in tools ] - return MappingProxyType({"tools": flattened}) + return MappingProxyType({"tools": sanitized}) def transform_request( self, @@ -489,7 +495,7 @@ class OpenAIGPTConfig(BaseLLMModelInfo, BaseConfig): "model": model, "messages": messages, **optional_params, - **self._flattened_tools_update_for_openai(optional_params, litellm_params), + **self._sanitized_tools_update_for_openai(optional_params, litellm_params), } async def async_transform_request( @@ -521,7 +527,7 @@ class OpenAIGPTConfig(BaseLLMModelInfo, BaseConfig): "model": model, "messages": transformed_messages, **optional_params, - **self._flattened_tools_update_for_openai(optional_params, litellm_params), + **self._sanitized_tools_update_for_openai(optional_params, litellm_params), } else: ## allow for any object specific behaviour to be handled diff --git a/litellm/llms/openai/responses/transformation.py b/litellm/llms/openai/responses/transformation.py index 97e7bcc60b7..666e9b32011 100644 --- a/litellm/llms/openai/responses/transformation.py +++ b/litellm/llms/openai/responses/transformation.py @@ -1,4 +1,4 @@ -from collections.abc import Mapping, Sequence +from collections.abc import Callable, Mapping, Sequence from functools import lru_cache from types import MappingProxyType from typing import TYPE_CHECKING, Any, Final, Protocol, cast, get_type_hints @@ -15,6 +15,10 @@ from litellm.litellm_core_utils.get_model_cost_map import GetModelCostMap from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response import ( _safe_convert_created_field, ) +from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_non_python_regex_patterns, + flatten_combinators_and_drop_non_python_regex_patterns, +) from litellm.litellm_core_utils.safe_json_loads import safe_json_loads from litellm.litellm_core_utils.url_utils import encode_url_path_segment from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig @@ -40,7 +44,7 @@ else: _NO_TOOL_UPDATE: Final[Mapping[str, object]] = MappingProxyType({}) _MODEL_FAMILIES_REJECTING_TOP_LEVEL_SCHEMA_COMBINATORS: Final = ("gpt-4", "gpt-3.5", "chatgpt-4o", "o1", "o3", "o4") -_PROVIDERS_WITH_COMBINATOR_REJECTING_VALIDATOR: Final = frozenset({LlmProviders.AZURE, LlmProviders.OPENAI}) +_PROVIDERS_WITH_OPENAI_SCHEMA_VALIDATOR: Final = frozenset({LlmProviders.AZURE, LlmProviders.OPENAI}) _PROVIDERS_VALIDATING_TOOL_CALL_ITEM_IDS: Final = frozenset({LlmProviders.AZURE, LlmProviders.OPENAI}) @@ -293,7 +297,7 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig): model=model, input=validated_input, tools=tools ) object_schema_tools: Final = self._tools_with_object_parameters(model=model, tools=stripped_tools) - sanitized_tools: Final = self._flatten_tool_schema_combinators_for_openai( + sanitized_tools: Final = self._sanitized_tool_schemas_for_openai( model=model, tools=object_schema_tools, litellm_params=litellm_params ) return self._drop_foreign_tool_call_item_ids(stripped_input), sanitized_tools @@ -378,35 +382,35 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig): return item return {key: value for key, value in item.items() if key != "id"} # mutable-ok: outgoing JSON request item - def _flatten_tool_schema_combinators_for_openai( + def _sanitized_tool_schemas_for_openai( self, model: str, tools: Sequence[ALL_RESPONSES_API_TOOL_PARAMS] | None, litellm_params: GenericLiteLLMParams, ) -> Sequence[ALL_RESPONSES_API_TOOL_PARAMS] | None: - """Flatten top-level schema combinators only where OpenAI's validator rejects them. + """Rewrite tool schemas only where OpenAI's validator rejects them. - OpenAI-compatible backends reusing this config (and the ChatGPT backend - Codex talks to natively) accept them, and so do GPT-5 and later models, - which also call tools better with the union intact. Codex wraps MCP tools - inside namespace entries, so nested ``tools`` arrays are walked too. - Azure OpenAI shares the validator but names deployments arbitrarily, so - the router's declared ``model_info.base_model`` wins over the deployment - name and an unrecognized name without one is left untouched. + Every model family refuses a ``pattern`` Python's ``re`` cannot compile, + while top-level schema combinators are flattened only for the families + whose validator rejects them: OpenAI-compatible backends reusing this + config (and the ChatGPT backend Codex talks to natively) accept them, + and so do GPT-5 and later models, which also call tools better with the + union intact. Codex wraps MCP tools inside namespace entries, so nested + ``tools`` arrays are walked too. Azure OpenAI shares the validator but + names deployments arbitrarily, so the router's declared + ``model_info.base_model`` wins over the deployment name and an + unrecognized name without one keeps its combinators. """ - if tools is None or self.custom_llm_provider not in _PROVIDERS_WITH_COMBINATOR_REJECTING_VALIDATOR: + if tools is None or self.custom_llm_provider not in _PROVIDERS_WITH_OPENAI_SCHEMA_VALIDATOR: return tools gate_model: Final = self._combinator_gate_model(model=model, litellm_params=litellm_params) - if not self._rejects_top_level_schema_combinators(gate_model): - return tools - flattened: Final = [ # mutable-ok: request tools are a JSON list - self._flattened_tool_or_passthrough(tool) for tool in tools - ] - return cast("Sequence[ALL_RESPONSES_API_TOOL_PARAMS]", flattened) # cast-ok: spread keeps each tool's shape - - @staticmethod - def _flattened_tool_or_passthrough(tool: object) -> object: - return OpenAIResponsesAPIConfig._flattened_tool_entry(tool) if isinstance(tool, dict) else tool + sanitize: Final = ( + flatten_combinators_and_drop_non_python_regex_patterns + if self._rejects_top_level_schema_combinators(gate_model) + else drop_non_python_regex_patterns + ) + sanitized: Final = self._sanitized_tools(tools, sanitize) + return cast("Sequence[ALL_RESPONSES_API_TOOL_PARAMS]", sanitized) # cast-ok: spread keeps each tool's shape @staticmethod def _rejects_top_level_schema_combinators(model: str) -> bool: @@ -421,35 +425,42 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig): return base_model if isinstance(base_model, str) and base_model else model @staticmethod - def _flattened_tool_entry( + def _sanitized_tool_entry( entry: Mapping[str, object], - ) -> dict[str, object]: # mutable-ok: request tools are JSON dicts - from litellm.litellm_core_utils.prompt_templates.common_utils import ( - flatten_top_level_schema_combinators, - ) - + sanitize: Callable[[Mapping[str, object]], Mapping[str, object]], + ) -> Mapping[str, object]: parameters: Final = entry.get("parameters") nested_tools: Final = entry.get("tools") + sanitized_parameters: Final = sanitize(parameters) if isinstance(parameters, dict) else parameters + sanitized_nested_tools: Final = ( + OpenAIResponsesAPIConfig._sanitized_tools(nested_tools, sanitize) + if isinstance(nested_tools, list) + else nested_tools + ) parameters_update: Final = ( - MappingProxyType({"parameters": flatten_top_level_schema_combinators(parameters)}) - if isinstance(parameters, dict) + MappingProxyType({"parameters": sanitized_parameters}) + if sanitized_parameters is not parameters else _NO_TOOL_UPDATE ) tools_update: Final = ( - MappingProxyType({"tools": OpenAIResponsesAPIConfig._flattened_nested_tools(nested_tools)}) - if isinstance(nested_tools, list) + MappingProxyType({"tools": sanitized_nested_tools}) + if sanitized_nested_tools is not nested_tools else _NO_TOOL_UPDATE ) + if not parameters_update and not tools_update: + return entry return {**entry, **parameters_update, **tools_update} # mutable-ok: request tools are JSON dicts @staticmethod - def _flattened_nested_tools( - nested_tools: Sequence[object], - ) -> list[object]: # mutable-ok: namespace tools are a JSON list - return [ # mutable-ok: namespace tools are a JSON list - OpenAIResponsesAPIConfig._flattened_tool_entry(item) if isinstance(item, dict) else item - for item in nested_tools + def _sanitized_tools( + tools: Sequence[object], + sanitize: Callable[[Mapping[str, object]], Mapping[str, object]], + ) -> Sequence[object]: + sanitized: Final = [ # mutable-ok: request tools are a JSON list + OpenAIResponsesAPIConfig._sanitized_tool_entry(item, sanitize) if isinstance(item, dict) else item + for item in tools ] + return tools if all(new is old for new, old in zip(sanitized, tools, strict=True)) else sanitized def _validate_input_param(self, input: str | ResponseInputParam) -> str | ResponseInputParam: """ diff --git a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py index 7c1445d79c9..b5890d1a5b0 100644 --- a/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py +++ b/tests/test_litellm/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py @@ -2,11 +2,12 @@ import copy import functools import json import os +import sys +from typing import Final from unittest.mock import MagicMock, patch import pytest - from litellm.litellm_core_utils.prompt_templates.common_utils import ( ENCRYPTED_REASONING_SIGNATURE_PREFIX, TOOL_RESULT_IMAGE_BOUNDARY, @@ -25,6 +26,8 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import ( update_messages_with_model_file_ids, ) +_ARTIFACT_FIELD_PATTERN: Final = r'^(?!__.*__$)[^\p{Cc}\p{Cf}\p{Zl}\p{Zp}"\\./[\]]{1,200}$' + def test_get_format_from_file_id(): unified_file_id = "litellm_proxy:application/pdf;unified_id,cbbe3534-8bf8-4386-af00-f5f6b7e370bf" @@ -1442,7 +1445,7 @@ class TestFlattenTopLevelSchemaCombinators: assert schema == snapshot -class TestToolWithFlattenedParameters: +class TestToolWithSanitizedParameters: def _anyof_tool(self): return { "type": "function", @@ -1469,11 +1472,12 @@ class TestToolWithFlattenedParameters: def test_flattens_anyof_parameters_into_new_tool(self): from litellm.litellm_core_utils.prompt_templates.common_utils import ( - tool_with_flattened_parameters, + flatten_combinators_and_drop_non_python_regex_patterns, + tool_with_sanitized_parameters, ) tool = self._anyof_tool() - result = tool_with_flattened_parameters(tool) + result = tool_with_sanitized_parameters(tool, flatten_combinators_and_drop_non_python_regex_patterns) assert result is not tool parameters = result["function"]["parameters"] @@ -1486,7 +1490,8 @@ class TestToolWithFlattenedParameters: def test_clean_parameters_return_the_same_tool_object(self): from litellm.litellm_core_utils.prompt_templates.common_utils import ( - tool_with_flattened_parameters, + flatten_combinators_and_drop_non_python_regex_patterns, + tool_with_sanitized_parameters, ) tool = { @@ -1497,7 +1502,23 @@ class TestToolWithFlattenedParameters: }, } - assert tool_with_flattened_parameters(tool) is tool + assert tool_with_sanitized_parameters(tool, flatten_combinators_and_drop_non_python_regex_patterns) is tool + + def test_pattern_only_sanitizer_drops_the_regex_and_keeps_the_union(self): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_non_python_regex_patterns, + tool_with_sanitized_parameters, + ) + + tool = self._anyof_tool() + tool["function"]["parameters"]["properties"]["id"]["pattern"] = _ARTIFACT_FIELD_PATTERN + + result = tool_with_sanitized_parameters(tool, drop_non_python_regex_patterns) + + parameters = result["function"]["parameters"] + assert parameters["properties"]["id"] == {"type": "string"} + assert parameters["anyOf"] == self._anyof_tool()["function"]["parameters"]["anyOf"] + assert tool["function"]["parameters"]["properties"]["id"]["pattern"] == _ARTIFACT_FIELD_PATTERN @pytest.mark.parametrize( "tool", @@ -1510,10 +1531,127 @@ class TestToolWithFlattenedParameters: ) def test_non_dict_function_or_parameters_return_the_same_tool_object(self, tool): from litellm.litellm_core_utils.prompt_templates.common_utils import ( - tool_with_flattened_parameters, + flatten_combinators_and_drop_non_python_regex_patterns, + tool_with_sanitized_parameters, ) - assert tool_with_flattened_parameters(tool) is tool + assert tool_with_sanitized_parameters(tool, flatten_combinators_and_drop_non_python_regex_patterns) is tool + + +class TestDropNonPythonRegexPatterns: + """Claude Code's Artifact tool declares ECMA-262 ``\\p{..}`` escapes that OpenAI's + validator, which compiles ``pattern`` values and ``patternProperties`` keys with + Python ``re``, refuses as "not a 'regex'".""" + + def _schema(self, pattern): + return { + "type": "object", + "properties": { + "field": {"type": "string", "pattern": pattern}, + "writes": { + "type": "array", + "items": {"properties": {"doc_id": {"type": "string", "pattern": pattern}}}, + }, + "query": {"anyOf": [{"type": "string", "pattern": pattern}, {"type": "null"}]}, + "pair": {"type": "array", "prefixItems": [{"type": "string", "pattern": pattern}]}, + "extra": {"type": "object", "additionalProperties": {"type": "string", "pattern": pattern}}, + "tagged": { + "type": "object", + "patternProperties": {pattern: {"type": "string"}, "^x_": {"type": "integer"}}, + }, + }, + "$defs": {"segment": {"type": "string", "pattern": pattern}}, + "required": ["field"], + } + + def test_drops_every_regex_python_re_rejects_from_every_schema_position(self): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_non_python_regex_patterns, + ) + + schema = self._schema(_ARTIFACT_FIELD_PATTERN) + + result = drop_non_python_regex_patterns(schema) + + assert '"pattern"' not in json.dumps(result) + properties = result["properties"] + assert properties["field"] == {"type": "string"} + assert properties["writes"]["items"]["properties"]["doc_id"] == {"type": "string"} + assert properties["query"]["anyOf"] == [{"type": "string"}, {"type": "null"}] + assert properties["pair"]["prefixItems"] == [{"type": "string"}] + assert properties["extra"]["additionalProperties"] == {"type": "string"} + assert properties["tagged"]["patternProperties"] == {"^x_": {"type": "integer"}} + assert result["$defs"]["segment"] == {"type": "string"} + assert result["required"] == ["field"] + assert schema == self._schema(_ARTIFACT_FIELD_PATTERN) + + def test_keeps_regexes_python_re_compiles_and_returns_the_same_object(self): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_non_python_regex_patterns, + ) + + schema = self._schema(r'^(?!__.*__$)[^"\\./[\]]{1,200}$') + + assert drop_non_python_regex_patterns(schema) is schema + + def test_pattern_keys_inside_data_positions_are_not_regexes(self): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_non_python_regex_patterns, + ) + + schema = { + "type": "object", + "properties": { + "pattern": {"type": "string"}, + "template": {"type": "object", "default": {"pattern": _ARTIFACT_FIELD_PATTERN}}, + "samples": {"type": "array", "examples": [{"pattern": _ARTIFACT_FIELD_PATTERN}]}, + "fixed": {"const": {"pattern": _ARTIFACT_FIELD_PATTERN}}, + "vendor": {"type": "string", "x-litellm": {"pattern": _ARTIFACT_FIELD_PATTERN}}, + }, + "required": ["pattern"], + } + + assert drop_non_python_regex_patterns(schema) is schema + + def test_regex_nested_past_what_python_re_can_parse_is_dropped_not_raised(self): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_non_python_regex_patterns, + ) + + schema = { + "type": "object", + "properties": {"deep": {"type": "string", "pattern": "(" * 2000 + "a" + ")" * 2000}}, + } + + assert drop_non_python_regex_patterns(schema)["properties"]["deep"] == {"type": "string"} + + def test_walks_schemas_deeper_than_the_interpreter_recursion_limit(self): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_non_python_regex_patterns, + ) + + depth = sys.getrecursionlimit() + leaf = {"type": "string", "pattern": _ARTIFACT_FIELD_PATTERN} + schema = functools.reduce( + lambda inner, _: {"type": "object", "properties": {"child": inner}}, range(depth), leaf + ) + + result = drop_non_python_regex_patterns(schema) + + assert functools.reduce(lambda node, _: node["properties"]["child"], range(depth), result) == {"type": "string"} + assert functools.reduce(lambda node, _: node["properties"]["child"], range(depth), schema) is leaf + + def test_leaves_levels_past_the_json_nesting_limit_alone(self): + from litellm.litellm_core_utils.prompt_templates.common_utils import ( + drop_non_python_regex_patterns, + ) + + leaf = {"type": "string", "pattern": _ARTIFACT_FIELD_PATTERN} + schema = functools.reduce( + lambda inner, _: {"type": "object", "properties": {"child": inner}}, range(1100), leaf + ) + + assert drop_non_python_regex_patterns(schema) is schema class TestRequestContainsImageContent: diff --git a/tests/test_litellm/llms/azure/chat/test_azure_chat_gpt_transformation.py b/tests/test_litellm/llms/azure/chat/test_azure_chat_gpt_transformation.py index 4e6b9ed0188..bc6cb0c0fed 100644 --- a/tests/test_litellm/llms/azure/chat/test_azure_chat_gpt_transformation.py +++ b/tests/test_litellm/llms/azure/chat/test_azure_chat_gpt_transformation.py @@ -198,6 +198,9 @@ def test_azure_gpt_5_takes_the_reasoning_path() -> None: assert "reasoning_effort" in supported +_ARTIFACT_FIELD_PATTERN: Final = r'^(?!__.*__$)[^\p{Cc}\p{Cf}\p{Zl}\p{Zp}"\\./[\]]{1,200}$' + + class TestAzureToolSchemaCombinatorFlattening: """ Regression tests for LIT-6510: Azure's chat completions validator rejects @@ -259,6 +262,26 @@ class TestAzureToolSchemaCombinatorFlattening: self._transform(AzureOpenAIConfig(), "gpt-4o", [tool]) assert tool == self._anyof_tool() + def test_transform_request_drops_non_python_regex_pattern(self): + tool = { + "type": "function", + "function": { + "name": "Artifact", + "parameters": { + "type": "object", + "properties": {"field": {"type": "string", "pattern": _ARTIFACT_FIELD_PATTERN}}, + }, + }, + } + + request = self._transform(AzureOpenAIConfig(), "gpt-4o", [tool]) + + assert request["tools"][0]["function"]["parameters"] == { + "type": "object", + "properties": {"field": {"type": "string"}}, + } + assert tool["function"]["parameters"]["properties"]["field"]["pattern"] == _ARTIFACT_FIELD_PATTERN + def test_clean_object_schema_passes_through_as_same_object(self): tool = { "type": "function", diff --git a/tests/test_litellm/llms/azure/response/test_azure_transformation.py b/tests/test_litellm/llms/azure/response/test_azure_transformation.py index 0cac2705ab0..726c9f65681 100644 --- a/tests/test_litellm/llms/azure/response/test_azure_transformation.py +++ b/tests/test_litellm/llms/azure/response/test_azure_transformation.py @@ -1,4 +1,5 @@ from copy import deepcopy +from typing import Final from unittest.mock import MagicMock, patch import pytest @@ -243,6 +244,9 @@ def test_provider_config_manager_o_series_selection(): assert not isinstance(default_config, AzureOpenAIOSeriesResponsesAPIConfig) +_ARTIFACT_FIELD_PATTERN: Final = r'^(?!__.*__$)[^\p{Cc}\p{Cf}\p{Zl}\p{Zp}"\\./[\]]{1,200}$' + + class TestAzureResponsesAPIConfig: def setup_method(self): self.config = AzureOpenAIResponsesAPIConfig() @@ -599,6 +603,31 @@ class TestAzureResponsesAPIConfig: assert result["tools"][0] is tool assert "anyOf" in result["tools"][0]["parameters"] + def test_azure_drops_non_python_regex_pattern_while_keeping_gpt5_combinators(self): + tool = { + "type": "function", + "name": "Artifact", + "parameters": { + "type": "object", + "anyOf": [{"properties": {"field": {"type": "string", "pattern": _ARTIFACT_FIELD_PATTERN}}}], + "properties": {"field": {"type": "string", "pattern": _ARTIFACT_FIELD_PATTERN}}, + }, + } + + result = self.config.transform_responses_api_request( + model="my-eastus-deployment", + input="hi", + response_api_optional_request_params={"tools": [tool]}, + litellm_params=GenericLiteLLMParams(model_info={"base_model": "azure/gpt-5.4-mini"}), + headers={}, + ) + + assert result["tools"][0]["parameters"] == { + "type": "object", + "anyOf": [{"properties": {"field": {"type": "string"}}}], + "properties": {"field": {"type": "string"}}, + } + def test_azure_keeps_combinators_for_unrecognized_deployment_without_base_model(self): tool = self._anyof_tool() diff --git a/tests/test_litellm/llms/openai/chat/test_openai_gpt_transformation.py b/tests/test_litellm/llms/openai/chat/test_openai_gpt_transformation.py index 9737d63cc26..b110586ae5b 100644 --- a/tests/test_litellm/llms/openai/chat/test_openai_gpt_transformation.py +++ b/tests/test_litellm/llms/openai/chat/test_openai_gpt_transformation.py @@ -4,6 +4,7 @@ Tests for OpenAI GPT transformation (litellm/llms/openai/chat/gpt_transformation import pytest +from typing import Final import litellm @@ -1168,6 +1169,9 @@ class TestOpenAIPromptCacheBreakpointChatPath: assert "prompt_cache_options" not in request +_ARTIFACT_FIELD_PATTERN: Final = r'^(?!__.*__$)[^\p{Cc}\p{Cf}\p{Zl}\p{Zp}"\\./[\]]{1,200}$' + + class TestToolSchemaCombinatorFlatteningForOpenAI: """ Regression tests for LIT-6488: OpenAI's chat completions validator rejects @@ -1281,3 +1285,50 @@ class TestToolSchemaCombinatorFlatteningForOpenAI: parameters = request["tools"][0]["function"]["parameters"] assert "anyOf" not in parameters assert set(parameters["properties"]) == {"id", "enabled", "schedule"} + + @staticmethod + def _artifact_tool(): + return { + "type": "function", + "function": { + "name": "Artifact", + "parameters": { + "type": "object", + "properties": {"field": {"type": "string", "pattern": _ARTIFACT_FIELD_PATTERN}}, + "required": ["field"], + }, + }, + } + + def test_drops_non_python_regex_pattern_for_hosted_openai(self): + tool = self._artifact_tool() + + request = self._transform(self.config, "gpt-4o", {"custom_llm_provider": "openai", "api_base": None}, [tool]) + + assert request["tools"][0]["function"]["parameters"] == { + "type": "object", + "properties": {"field": {"type": "string"}}, + "required": ["field"], + } + assert tool == self._artifact_tool() + + def test_custom_api_base_drops_non_python_regex_pattern_but_keeps_union(self): + tool = self._anyof_tool() + tool["function"]["parameters"]["properties"]["id"]["pattern"] = _ARTIFACT_FIELD_PATTERN + + request = self._transform( + self.config, "gpt-4o", {"custom_llm_provider": "openai", "api_base": "http://localhost:8000/v1"}, [tool] + ) + + parameters = request["tools"][0]["function"]["parameters"] + assert parameters["properties"]["id"] == {"type": "string"} + assert parameters["anyOf"] == self._anyof_tool()["function"]["parameters"]["anyOf"] + + def test_non_openai_provider_keeps_non_python_regex_pattern(self): + tool = self._artifact_tool() + + request = self._transform( + self.config, "some-oss-model", {"custom_llm_provider": "groq", "api_base": None}, [tool] + ) + + assert request["tools"][0] is tool diff --git a/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py b/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py index c5902b32a06..4cf8767764b 100644 --- a/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py +++ b/tests/test_litellm/llms/openai/responses/test_openai_responses_transformation.py @@ -1,5 +1,6 @@ import json from types import SimpleNamespace +from typing import Final from unittest.mock import AsyncMock, MagicMock, Mock, patch import httpx @@ -20,6 +21,8 @@ from litellm.types.llms.openai import ( ) from litellm.types.router import GenericLiteLLMParams +_ARTIFACT_FIELD_PATTERN: Final = r'^(?!__.*__$)[^\p{Cc}\p{Cf}\p{Zl}\p{Zp}"\\./[\]]{1,200}$' + class TestOpenAIResponsesAPIConfig: def setup_method(self): @@ -2022,6 +2025,86 @@ class TestFlattenToolSchemaCombinatorsWiring: assert "anyOf" not in result["tools"][1]["parameters"] +class TestToolSchemaRegexPatternWiring: + """Claude Code's Artifact tool reaches /v1/responses (the /v1/messages bridge) with an + ECMA-262 ``pattern``; OpenAI compiles patterns with Python ``re`` and 400s + "'...' is not a 'regex'" for every model family, so the keyword is dropped. + """ + + def _artifact_tool(self): + return { + "type": "function", + "name": "Artifact", + "parameters": { + "type": "object", + "properties": { + "field": {"type": "string", "pattern": _ARTIFACT_FIELD_PATTERN}, + "doc_id": {"type": "string", "pattern": r"^(?!\.\.?(?:/|$))[A-Za-z0-9_\-.~:@+]{1,200}$"}, + }, + "required": ["field"], + }, + } + + @pytest.mark.parametrize("model", ["gpt-5.6", "gpt-4o", "o3"]) + def test_openai_drops_only_the_pattern_python_re_rejects_for_every_family(self, model): + tool = self._artifact_tool() + + result = OpenAIResponsesAPIConfig().transform_responses_api_request( + model=model, + input="hi", + response_api_optional_request_params={"tools": [tool]}, + litellm_params=GenericLiteLLMParams(), + headers={}, + ) + + properties = result["tools"][0]["parameters"]["properties"] + assert properties["field"] == {"type": "string"} + assert properties["doc_id"] == tool["parameters"]["properties"]["doc_id"] + assert result["tools"][0]["parameters"]["required"] == ["field"] + assert tool["parameters"]["properties"]["field"]["pattern"] == _ARTIFACT_FIELD_PATTERN + assert json.loads(json.dumps(result["tools"])) == result["tools"] + + def test_openai_drops_patterns_inside_codex_namespace_tools(self): + namespace = {"type": "namespace", "name": "mcp__claude", "tools": [self._artifact_tool()]} + + result = OpenAIResponsesAPIConfig().transform_responses_api_request( + model="gpt-5.6", + input="hi", + response_api_optional_request_params={"tools": [namespace]}, + litellm_params=GenericLiteLLMParams(), + headers={}, + ) + + assert result["tools"][0]["tools"][0]["parameters"]["properties"]["field"] == {"type": "string"} + + def test_openai_compact_request_drops_patterns(self): + _, data = OpenAIResponsesAPIConfig().transform_compact_response_api_request( + model="gpt-5.6", + input="hi", + response_api_optional_request_params={"tools": [self._artifact_tool()]}, + api_base="https://api.openai.com/v1/responses", + litellm_params=GenericLiteLLMParams(), + headers={}, + ) + + assert data["tools"][0]["parameters"]["properties"]["field"] == {"type": "string"} + + def test_non_openai_subclass_keeps_patterns(self): + from litellm.llms.hosted_vllm.responses.transformation import HostedVLLMResponsesAPIConfig + + tool = self._artifact_tool() + + result = HostedVLLMResponsesAPIConfig().transform_responses_api_request( + model="hosted_vllm/qwen", + input="hi", + response_api_optional_request_params={"tools": [tool]}, + litellm_params=GenericLiteLLMParams(), + headers={}, + ) + + assert result["tools"][0] is tool + + class TestReasoningFollowsModelSupport: """Responses API clients like Codex send `reasoning` on every request, and OpenAI 400s it on non-reasoning models like gpt-4o. drop_params must strip it there, the same way the