fix(responses): strip thought-signature and enforce <=64 length on tool call item IDs

Prevent Gemini thought-signatures embedded in tool_call.id from leaking
into Responses API input[N].id during multi-turn router fallbacks to OpenAI/Azure.
OpenAI strictly limits input[N].id to 64 characters and rejects longer IDs with
HTTP 400.

- In custom_tools.py: strip THOUGHT_SIGNATURE_SEPARATOR in openai_shaped_tool_call_item_id
  before applying fc_/ctc_ prefix, preserving raw call_id for Gemini continuity.
- In transformation.py: strip THOUGHT_SIGNATURE_SEPARATOR, validate genuine prefix
  and len <= 64, and drop non-conforming or oversized id fields.
- Add unit tests in test_openai_responses_transformation.py and test_custom_tool_call.py.

Fixes #41534

Signed-off-by: Ansh Rohilla <anshrohilla.b81@gmail.com>
This commit is contained in:
Ansh Rohilla 2026-09-17 16:25:46 +05:30
parent 4b368bf066
commit 116cee15cb
4 changed files with 161 additions and 4 deletions

View file

@ -381,8 +381,19 @@ class OpenAIResponsesAPIConfig(BaseResponsesAPIConfig):
item_type: Final = item.get("type")
item_id: Final = item.get("id")
genuine_prefix: Final = TOOL_CALL_ITEM_ID_PREFIX_BY_TYPE.get(item_type) if isinstance(item_type, str) else None
if genuine_prefix is None or not isinstance(item_id, str) or item_id.startswith(genuine_prefix):
if genuine_prefix is None or not isinstance(item_id, str):
return item
from litellm.litellm_core_utils.prompt_templates.factory import (
THOUGHT_SIGNATURE_SEPARATOR,
)
clean_id: Final = (
item_id.split(THOUGHT_SIGNATURE_SEPARATOR, 1)[0] if THOUGHT_SIGNATURE_SEPARATOR in item_id else item_id
)
if clean_id.startswith(genuine_prefix) and len(clean_id) <= 64:
if clean_id == item_id:
return item
return {**item, "id": clean_id} # mutable-ok: outgoing JSON request item
return {key: value for key, value in item.items() if key != "id"} # mutable-ok: outgoing JSON request item
def _sanitized_tool_schemas_for_openai(

View file

@ -15,6 +15,7 @@ extending this module without touching the streaming iterator or transformation
logic.
"""
import hashlib
import json
from collections.abc import Mapping, Sequence
from types import MappingProxyType
@ -34,9 +35,22 @@ TOOL_CALL_ITEM_ID_PREFIX_BY_TYPE: Final = MappingProxyType({"function_call": "fc
def openai_shaped_tool_call_item_id(item_type: str, tool_id: str) -> str:
prefix: Final = TOOL_CALL_ITEM_ID_PREFIX_BY_TYPE.get(item_type)
if prefix is None or not tool_id or tool_id.startswith(prefix):
if prefix is None or not tool_id:
return tool_id
return f"{prefix}_{tool_id}"
from litellm.litellm_core_utils.prompt_templates.factory import (
THOUGHT_SIGNATURE_SEPARATOR,
)
clean_id: Final = (
tool_id.split(THOUGHT_SIGNATURE_SEPARATOR, 1)[0] if THOUGHT_SIGNATURE_SEPARATOR in tool_id else tool_id
)
if not clean_id:
return tool_id
candidate: Final = clean_id if clean_id.startswith(prefix) else f"{prefix}_{clean_id}"
if len(candidate) <= 64:
return candidate
digest: Final = hashlib.sha256(candidate.encode("utf-8")).hexdigest()[:16]
return f"{candidate[:47]}_{digest}"
class _ToolNameFields(BaseModel):

View file

@ -273,6 +273,79 @@ class TestOpenAIResponsesAPIConfig:
assert replayed_input[1]["id"] == "toolu_01Foreign"
assert replayed_input[3]["id"] == "srvtoolu_01Foreign"
def test_transform_strips_thought_signature_from_tool_call_item_ids(self):
"""Replayed tool call items containing Gemini thought signatures in their ids
(e.g., fc_call_xxx__thought__<base64>) must have the signature stripped so that
the resulting id conforms to OpenAI's 64-char maximum, while call_id is preserved."""
from litellm.litellm_core_utils.prompt_templates.factory import (
THOUGHT_SIGNATURE_SEPARATOR,
)
long_sig: Final = "A" * 900
raw_call_id: Final = f"call_gemini_123{THOUGHT_SIGNATURE_SEPARATOR}{long_sig}"
replayed_input: Final = [
{
"type": "function_call",
"id": f"fc_{raw_call_id}",
"call_id": raw_call_id,
"name": "get_weather",
"arguments": '{"city": "Paris"}',
},
{
"type": "custom_tool_call",
"id": f"ctc_{raw_call_id}",
"call_id": raw_call_id,
"name": "run_terminal",
"input": "ls",
},
{
"type": "function_call_output",
"call_id": raw_call_id,
"output": "sunny",
},
]
result: Final = self.config.transform_responses_api_request(
model=self.model,
input=replayed_input,
response_api_optional_request_params={},
litellm_params={},
headers={},
)
assert result["input"][0]["id"] == "fc_call_gemini_123"
assert len(result["input"][0]["id"]) <= 64
assert result["input"][0]["call_id"] == raw_call_id
assert result["input"][1]["id"] == "ctc_call_gemini_123"
assert len(result["input"][1]["id"]) <= 64
assert result["input"][1]["call_id"] == raw_call_id
assert result["input"][2]["call_id"] == raw_call_id
def test_transform_drops_oversized_tool_call_item_ids(self):
"""Replayed tool call items with genuine prefixes whose ids exceed OpenAI's 64-char
maximum (without thought signatures) must have the id dropped to prevent 400."""
oversized_id: Final = "fc_" + "x" * 70
replayed_input: Final = [
{
"type": "function_call",
"id": oversized_id,
"call_id": "call_123",
"name": "get_weather",
"arguments": "{}",
}
]
result: Final = self.config.transform_responses_api_request(
model=self.model,
input=replayed_input,
response_api_optional_request_params={},
litellm_params={},
headers={},
)
assert "id" not in result["input"][0]
assert result["input"][0]["call_id"] == "call_123"
def test_transform_keeps_foreign_tool_call_item_ids_for_other_providers(self):
"""Providers reusing this config that do not enforce OpenAI's id
shapes must keep replayed ids untouched."""

View file

@ -10,7 +10,7 @@ LiteLLM bridge correctly:
import json
import pytest
from typing import Dict, Any, List
from typing import Any, Dict, Final, List
from openai.types.responses import ResponseFunctionToolCall
@ -160,6 +160,65 @@ class TestCustomToolUtilities:
assert openai_shaped_tool_call_item_id("function_call", "") == ""
assert openai_shaped_tool_call_item_id("message", "toolu_01Abc") == "toolu_01Abc"
def test_openai_shaped_tool_call_item_id_strips_thought_signature(self):
"""Gemini thought signatures encoded into tool ids (call_xxx__thought__<sig>)
must be stripped from the emitted item id (producing fc_call_xxx <= 64 chars)
so that subsequent turns / fallbacks to OpenAI do not 400 with 'string too long'."""
from litellm.litellm_core_utils.prompt_templates.factory import (
THOUGHT_SIGNATURE_SEPARATOR,
)
long_sig: Final = "a" * 900
raw_id: Final = f"call_12345{THOUGHT_SIGNATURE_SEPARATOR}{long_sig}"
shaped_fc_id: Final = openai_shaped_tool_call_item_id("function_call", raw_id)
assert shaped_fc_id == "fc_call_12345"
assert len(shaped_fc_id) <= 64
shaped_ctc_id: Final = openai_shaped_tool_call_item_id("custom_tool_call", raw_id)
assert shaped_ctc_id == "ctc_call_12345"
assert len(shaped_ctc_id) <= 64
already_prefixed: Final = f"fc_call_12345{THOUGHT_SIGNATURE_SEPARATOR}{long_sig}"
assert openai_shaped_tool_call_item_id("function_call", already_prefixed) == "fc_call_12345"
long_provider_id: Final = "a" * 62
shaped_long_id: Final = openai_shaped_tool_call_item_id("function_call", long_provider_id)
assert len(shaped_long_id) == 64
assert shaped_long_id.startswith("fc_")
only_sig: Final = f"{THOUGHT_SIGNATURE_SEPARATOR}{long_sig}"
assert openai_shaped_tool_call_item_id("function_call", only_sig) == only_sig
shared_prefix_id1: Final = ("x" * 61) + "1"
shared_prefix_id2: Final = ("x" * 61) + "2"
shortened1: Final = openai_shaped_tool_call_item_id("function_call", shared_prefix_id1)
shortened2: Final = openai_shaped_tool_call_item_id("function_call", shared_prefix_id2)
assert len(shortened1) == 64
assert len(shortened2) == 64
assert shortened1.startswith("fc_")
assert shortened2.startswith("fc_")
assert shortened1 != shortened2
def test_without_foreign_tool_call_item_id_strips_thought_signature(self):
from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
from litellm.litellm_core_utils.prompt_templates.factory import (
THOUGHT_SIGNATURE_SEPARATOR,
)
long_sig: Final = "a" * 900
raw_id: Final = f"call_12345{THOUGHT_SIGNATURE_SEPARATOR}{long_sig}"
item: Final = {"type": "function_call", "id": f"fc_{raw_id}", "call_id": raw_id}
sanitized: Final = OpenAIResponsesAPIConfig._without_foreign_tool_call_item_id(item)
assert sanitized["id"] == "fc_call_12345"
oversized_item: Final = {"type": "function_call", "id": "fc_" + ("x" * 70), "call_id": "call_123"}
dropped: Final = OpenAIResponsesAPIConfig._without_foreign_tool_call_item_id(oversized_item)
assert "id" not in dropped
normal_item: Final = {"type": "function_call", "id": "fc_normal", "call_id": "call_normal"}
kept: Final = OpenAIResponsesAPIConfig._without_foreign_tool_call_item_id(normal_item)
assert kept["id"] == "fc_normal"
def test_build_tool_call_item_kwargs_normalizes_item_id_keeps_call_id(self):
"""The streaming item id gets the OpenAI shape while call_id stays raw
so tool_result pairing (which keys off call_id) keeps working."""