refactor: replace genAI model sdk by okHttp

[Aias00]

AgentScope-Java Version

[The version of AgentScope-Java you are working on, e.g. 1.0.2, check your pom.xml dependency version or run mvn dependency:tree | grep agentscope-parent:pom(only mac/linux)]

Description

[Please describe the background, purpose, changes made, and how to test this PR]
Introduce dedicated DTOs for Gemini API request and response structures and update related components.
fixes #96

Checklist

Please check the following items before code is ready to be reviewed.

• Code has been formatted with mvn spotless:apply
• All tests are passing (mvn test)
• Javadoc comments are complete and follow project conventions
• Related documentation has been updated (e.g. links, examples, etc.)
• Code is ready for review

[codecov]

Codecov Report

❌ Patch coverage is 78.31776% with 348 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...cope/core/model/GeminiStructuredOutputHandler.java	72.03%	49 Missing and 97 partials ⚠️
...ope/core/formatter/gemini/GeminiChatFormatter.java	64.36%	0 Missing and 31 partials ⚠️
...pe/core/formatter/gemini/GeminiResponseParser.java	74.19%	5 Missing and 19 partials ⚠️
.../core/formatter/gemini/GeminiMessageConverter.java	68.65%	14 Missing and 7 partials ⚠️
.../agentscope/core/model/GeminiRequestAssembler.java	74.32%	6 Missing and 13 partials ⚠️
...java/io/agentscope/core/model/GeminiChatModel.java	76.31%	11 Missing and 7 partials ⚠️
...java/io/agentscope/core/model/GeminiTransport.java	70.49%	7 Missing and 11 partials ⚠️
.../io/agentscope/core/model/GeminiResponseGuard.java	64.44%	8 Missing and 8 partials ⚠️
...atter/gemini/GeminiMultiAgentMessageConverter.java	78.78%	6 Missing and 8 partials ⚠️
...pe/core/agent/accumulator/ThinkingAccumulator.java	41.66%	3 Missing and 4 partials ⚠️
... and 12 more

📢 Thoughts on this report? Let us know!

[Copilot]

Pull request overview

This PR replaces the Google GenAI Java SDK with dedicated DTOs and direct HTTP calls to the Gemini API, simplifying the implementation and removing dependencies on the Google SDK and Vertex AI support.

• Custom DTOs introduced for all Gemini API request/response structures
• Direct HTTP implementation using OkHttp with SSE streaming support
• All formatters, converters, and parsers updated to use new DTOs

Reviewed changes

Copilot reviewed 32 out of 32 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
GeminiChatModel.java	Replaced SDK client with OkHttpclient for direct API calls with streaming support
GeminiRequest.java, GeminiResponse.java	New DTOs for API request/response structures
GeminiContent.java, GeminiPart.java	DTOs for message content and parts
GeminiTool.java, GeminiToolConfig.java	DTOs for tool definitions and configuration
GeminiGenerationConfig.java	DTO for generation parameters
GeminiChatFormatter.java	Updated to work with DTOs instead of SDK types
GeminiMessageConverter.java	Updated conversion logic for new DTOs
GeminiMediaConverter.java	Updated media handling to use DTO Blob type
GeminiToolsHelper.java	Simplified tool conversion without SDK schema conversion
GeminiResponseParser.java	Updated response parsing for new DTO structures
ModelProviderType.java	Simplified Gemini config removing Vertex AI parameters
AgentScopeProducer.java	Removed Vertex AI configuration logic
pom.xml (core)	Removed google-genai dependency
pom.xml (mem0)	Added jackson-datatype-jsr310 dependency
Test files	Updated all tests to use new DTOs instead of SDK types

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Aias00]

gemini-2.0-flush

gemini-3-pro-preview

[Copilot]

Pull request overview

Copilot reviewed 38 out of 38 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (1)

agentscope-extensions/agentscope-extensions-mem0/pom.xml:1

• The jackson-datatype-jsr310 version 2.15.2 may be outdated. Consider using a more recent version that matches the Jackson version used elsewhere in the project, or verify if 2.15.2 is the intended version for compatibility reasons.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[AlbumenJ]

Please make sure all the E2E Tests pass

[AlbumenJ]

Your current logic for handling the Base URL differs from the GenAI SDK's approach, specifically regarding the handling of the GOOGLE_API_BASE_URL key.

[github-actions]

Closing due to inactivity. Feel free to reopen when ready.

[Aias00]

@AlbumenJ @LearningGp hi, pls help me reopen this pr

[LearningGp]

For Gemini 3, are we supposed to use thinkingLevel instead of thinkingBudget now? Curious about the specific reason why "thinking" is restricted on Gemini 3 Flash.

[Aias00]

For Gemini 3, are we supposed to use thinkingLevel instead of thinkingBudget now? Curious about the specific reason why "thinking" is restricted on Gemini 3 Flash.

For Gemini 3, yes, Google now recommends using thinkingLevel rather than thinkingBudget. thinkingBudget is still accepted for backward compatibility, but the official guidance is: use thinkingLevel for Gemini 3 and later, and use thinkingBudget for Gemini 2.5. Google also notes that using thinkingBudget with Gemini 3 Pro can lead to unexpected performance, so it should not be treated as the preferred control going forward.

For the Gemini 3 Flash + structured output case, the current attempt to suppress thinking is not because Gemini Flash is officially defined as incompatible with thinking. The reason is that, in our earlier implementation, we observed that once thinking was enabled on gemini-3-flash, the tool-based structured output path became less stable and less deterministic. Since AgentScope currently implements structured output through the framework’s unified generate_response tool flow, we introduced this workaround to make Gemini 3 Flash align more reliably with the framework-defined structured output capability.

In other words, the current “disable thinking” behavior is a framework-compatibility mitigation: it was added to make gemini-3-flash + structured output tool behave more predictably within the existing framework contract, rather than because Gemini’s native design requires thinking to be turned off.

Going forward, we should evaluate whether Gemini structured output can be handled in a more native way, for example through responseSchema, while still adapting the result back into the framework’s common structured output semantics. If that works well, this workaround may no longer be necessary.

Official references:

• Gemini Thinking
• Gemini Structured Output