Skip to content

Build and run agents you can see, understand and trust.

fix(gemini): fix GeminiChatModel compatibility with Gemini-3-pro (#1206)

DavdGao6579dff68b47
diff --git a/src/agentscope/formatter/_gemini_formatter.py b/src/agentscope/formatter/_gemini_formatter.py
index 814838b04..2c297fc65 100644
--- a/src/agentscope/formatter/_gemini_formatter.py
+++ b/src/agentscope/formatter/_gemini_formatter.py
@@ -200,10 +200,11 @@ class GeminiChatFormatter(TruncatedFormatterBase):
parts.append(
{
"function_call": {
- "id": block["id"],
+ "id": None,
"name": block["name"],
"args": block["input"],
},
+ "thought_signature": block.get("id", None),
},
)
diff --git a/src/agentscope/model/_gemini_model.py b/src/agentscope/model/_gemini_model.py
index 635232aa0..6dfeddc8d 100644
--- a/src/agentscope/model/_gemini_model.py
+++ b/src/agentscope/model/_gemini_model.py
@@ -1,10 +1,11 @@
# -*- coding: utf-8 -*-
# mypy: disable-error-code="dict-item"
"""The Google Gemini model in agentscope."""
+import base64
import copy
-import json
import warnings
from datetime import datetime
+import json
from typing import (
AsyncGenerator,
Any,
@@ -333,39 +334,54 @@ class GeminiChatModel(ChatModelBase):
text = ""
thinking = ""
+ tool_calls: list[ToolUseBlock] = []
metadata: dict | None = None
async for chunk in response:
- content_block: list = []
-
- # Thinking parts
if (
chunk.candidates
and chunk.candidates[0].content
and chunk.candidates[0].content.parts
):
for part in chunk.candidates[0].content.parts:
- if part.thought and part.text:
- thinking += part.text
+ if part.text:
+ if part.thought:
+ thinking += part.text
+ else:
+ text += part.text
+
+ if part.function_call:
+ keyword_args = part.function_call.args or {}
+ # .. note:: Gemini API always returns None for
+ # function_call.id, so we use thought_signature
+ # as the unique identifier for tool
+ # calls when available. That maybe
+ # infeasible someday, but Gemini
+ # requires the thought_signature for some
+ # llms like gemini-3-pro
+
+ if part.thought_signature:
+ call_id = base64.b64encode(
+ part.thought_signature,
+ ).decode("utf-8")
+ else:
+ call_id = part.function_call.id
+
+ tool_calls.append(
+ ToolUseBlock(
+ type="tool_use",
+ id=call_id,
+ name=part.function_call.name,
+ input=keyword_args,
+ raw_input=json.dumps(
+ keyword_args,
+ ensure_ascii=False,
+ ),
+ ),
+ )
# Text parts
- if chunk.text:
- text += chunk.text
- if structured_model:
- metadata = _json_loads_with_repair(text)
-
- # Function calls
- tool_calls = []
- if chunk.function_calls:
- for function_call in chunk.function_calls:
- tool_calls.append(
- ToolUseBlock(
- type="tool_use",
- id=function_call.id,
- name=function_call.name,
- input=function_call.args or {},
- raw_input=json.dumps(function_call.args or {}),
- ),
- )
+ if text and structured_model:
+ metadata = _json_loads_with_repair(text)
usage = None
if chunk.usage_metadata:
@@ -376,8 +392,11 @@ class GeminiChatModel(ChatModelBase):
time=(datetime.now() - start_datetime).total_seconds(),
)
+ # The content blocks for the current chunk
+ content_blocks: list = []
+
if thinking:
- content_block.append(
+ content_blocks.append(
ThinkingBlock(
type="thinking",
thinking=thinking,
@@ -385,21 +404,18 @@ class GeminiChatModel(ChatModelBase):
)
if text:
- content_block.append(
+ content_blocks.append(
TextBlock(
type="text",
text=text,
),
)
- content_block.extend(tool_calls)
-
- parsed_chunk = ChatResponse(
- content=content_block,
+ yield ChatResponse(
+ content=content_blocks + tool_calls,
usage=usage,
metadata=metadata,
)
- yield parsed_chunk
def _parse_gemini_generation_response(
self,
@@ -429,6 +445,7 @@ class GeminiChatModel(ChatModelBase):
"""
content_blocks: List[TextBlock | ToolUseBlock | ThinkingBlock] = []
metadata: dict | None = None
+ tool_calls: list = []
if (
response.candidates
@@ -436,35 +453,54 @@ class GeminiChatModel(ChatModelBase):
and response.candidates[0].content.parts
):
for part in response.candidates[0].content.parts:
- if part.thought and part.text:
- content_blocks.append(
- ThinkingBlock(
- type="thinking",
- thinking=part.text,
+ if part.text:
+ if part.thought:
+ content_blocks.append(
+ ThinkingBlock(
+ type="thinking",
+ thinking=part.text,
+ ),
+ )
+ else:
+ content_blocks.append(
+ TextBlock(
+ type="text",
+ text=part.text,
+ ),
+ )
+
+ if part.function_call:
+ keyword_args = part.function_call.args or {}
+ # .. note:: Gemini API always returns None for
+ # function_call.id, so we use thought_signature
+ # as the unique identifier for tool
+ # calls when available. That maybe infeasible
+ # someday, but Gemini requires the thought_signature
+ # for some llms like gemini-3-pro
+
+ if part.thought_signature:
+ call_id = base64.b64encode(
+ part.thought_signature,
+ ).decode("utf-8")
+ else:
+ call_id = part.function_call.id
+
+ tool_calls.append(
+ ToolUseBlock(
+ type="tool_use",
+ id=call_id,
+ name=part.function_call.name,
+ input=keyword_args,
+ raw_input=json.dumps(
+ keyword_args,
+ ensure_ascii=False,
+ ),
),
)
- if response.text:
- content_blocks.append(
- TextBlock(
- type="text",
- text=response.text,
- ),
- )
- if structured_model:
- metadata = _json_loads_with_repair(response.text)
-
- if response.function_calls:
- for tool_call in response.function_calls:
- content_blocks.append(
- ToolUseBlock(
- type="tool_use",
- id=tool_call.id,
- name=tool_call.name,
- input=tool_call.args or {},
- raw_input=json.dumps(tool_call.args or {}),
- ),
- )
+ # For the structured output case
+ if response.text and structured_model:
+ metadata = _json_loads_with_repair(response.text)
if response.usage_metadata:
usage = ChatUsage(
@@ -478,7 +514,7 @@ class GeminiChatModel(ChatModelBase):
usage = None
return ChatResponse(
- content=content_blocks,
+ content=content_blocks + tool_calls,
usage=usage,
metadata=metadata,
)
diff --git a/tests/formatter_gemini_test.py b/tests/formatter_gemini_test.py
index cf0a27e72..bb4d49aae 100644
--- a/tests/formatter_gemini_test.py
+++ b/tests/formatter_gemini_test.py
@@ -271,12 +271,13 @@ class TestGeminiFormatter(IsolatedAsyncioTestCase):
"parts": [
{
"function_call": {
- "id": "1",
+ "id": None,
"name": "get_capital",
"args": {
"country": "Japan",
},
},
+ "thought_signature": "1",
},
],
},
@@ -353,12 +354,13 @@ class TestGeminiFormatter(IsolatedAsyncioTestCase):
"parts": [
{
"function_call": {
- "id": "1",
+ "id": None,
"name": "get_capital",
"args": {
"country": "Japan",
},
},
+ "thought_signature": "1",
},
],
},
@@ -405,12 +407,13 @@ class TestGeminiFormatter(IsolatedAsyncioTestCase):
"parts": [
{
"function_call": {
- "id": "1",
+ "id": None,
"name": "get_capital",
"args": {
"country": "Japan",
},
},
+ "thought_signature": "1",
},
],
},
@@ -490,12 +493,13 @@ class TestGeminiFormatter(IsolatedAsyncioTestCase):
"parts": [
{
"function_call": {
- "id": "1",
+ "id": None,
"name": "get_capital",
"args": {
"country": "Japan",
},
},
+ "thought_signature": "1",
},
],
},
@@ -532,12 +536,13 @@ class TestGeminiFormatter(IsolatedAsyncioTestCase):
"parts": [
{
"function_call": {
- "id": "2",
+ "id": None,
"name": "get_capital",
"args": {
"country": "South Korea",
},
},
+ "thought_signature": "2",
},
],
},
@@ -700,12 +705,13 @@ class TestGeminiFormatter(IsolatedAsyncioTestCase):
"parts": [
{
"function_call": {
- "id": "1",
+ "id": None,
"name": "get_capital",
"args": {
"country": "Japan",
},
},
+ "thought_signature": "1",
},
],
},
@@ -922,12 +928,13 @@ class TestGeminiFormatter(IsolatedAsyncioTestCase):
"parts": [
{
"function_call": {
- "id": "1",
+ "id": None,
"name": "get_capital",
"args": {
"country": "Japan",
},
},
+ "thought_signature": "1",
},
],
},
diff --git a/tests/model_gemini_test.py b/tests/model_gemini_test.py
index f62d63ffd..3ebcce594 100644
--- a/tests/model_gemini_test.py
+++ b/tests/model_gemini_test.py
@@ -25,7 +25,30 @@ class GeminiResponseMock:
self.usage_metadata = (
self._create_usage_mock(usage_metadata) if usage_metadata else None
)
- self.candidates = candidates or []
+
+ if candidates:
+ # Use provided candidates structure
+ self.candidates = candidates
+ else:
+ # Build default candidate structure
+ part = Mock()
+ part.text = text
+ part.thought = False
+ part.function_call = None
+
+ first_candidate = Mock()
+ first_candidate.content = Mock()
+ first_candidate.content.parts = [part]
+
+ for function_call in function_calls or []:
+ part = Mock()
+ part.text = None
+ part.thought = False
+ part.function_call = function_call
+ part.thought_signature = None
+ first_candidate.content.parts.append(part)
+
+ self.candidates = [first_candidate]
def _create_usage_mock(self, usage_data: dict) -> Mock:
usage_mock = Mock()
@@ -49,6 +72,8 @@ class GeminiPartMock:
def __init__(self, text: str = "", thought: bool = False):
self.text = text
self.thought = thought
+ self.function_call = None
+ self.thought_signature = None
class GeminiCandidateMock:
@@ -127,6 +152,7 @@ class TestGeminiChatModel(IsolatedAsyncioTestCase):
call_args = mock_client.aio.models.generate_content.call_args[1]
self.assertEqual(call_args["model"], "gemini-2.5-flash")
self.assertEqual(call_args["contents"], messages)
+ self.assertIn("config", call_args)
self.assertIsInstance(result, ChatResponse)
expected_content = [
TextBlock(type="text", text="Hello! How can I help you?"),
@@ -232,7 +258,11 @@ class TestGeminiChatModel(IsolatedAsyncioTestCase):
text="Let me analyze this step by step...",
thought=True,
)
- candidate = GeminiCandidateMock(parts=[thinking_part])
+ text_part = GeminiPartMock(
+ text="Here's my analysis",
+ thought=False,
+ )
+ candidate = GeminiCandidateMock(parts=[thinking_part, text_part])
mock_response = self._create_mock_response_with_thinking(
"Here's my analysis",
candidates=[candidate],