feat: add MiniMax provider support

[octo-patch]

Description

This PR adds a new model/minimax package that implements the model.LLM interface for MiniMax models using the OpenAI-compatible API.

Motivation

MiniMax is a leading AI provider offering high-performance models. Adding native support enables ADK users to build agents powered by MiniMax models with the same clean interface used for Gemini.

Changes

• New package google.golang.org/adk/model/minimax implementing model.LLM
• Supports MiniMax-M2.7 (default) and MiniMax-M2.7-highspeed models
• Converts genai.Content/genai.Part types to/from OpenAI message format, handling:
  • Text parts
  • Function calls → OpenAI tool_calls
  • Function responses → OpenAI tool messages
  • System instructions from GenerateContentConfig
• Converts genai.Schema to OpenAI-compatible JSON Schema for function declarations
• Clamps temperature to MiniMax's valid range (0.0, 1.0]
• Reads API key from MINIMAX_API_KEY environment variable
• No new external dependencies — uses only net/http and encoding/json

Usage

import "google.golang.org/adk/model/minimax"

m, err := minimax.NewModel("MiniMax-M2.7")
// or with options:
m, err := minimax.NewModel("MiniMax-M2.7",
    minimax.WithAPIKey("your-api-key"),
    minimax.WithBaseURL("https://api.minimax.io"),
)

API Reference

• Chat (OpenAI Compatible): https://platform.minimax.io/docs/api-reference/text-openai-api

Testing Plan

• Unit tests cover: model creation, text generation, tool use, function call/response conversion, schema conversion, temperature clamping, error handling, and custom base URL
• Integration test verified against the live MiniMax API: model returns correct text response with FinishReason: STOP
• All 13 unit tests pass: go test ./model/minimax/... -v

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[gemini-code-assist]

Code Review

This pull request introduces a new MiniMax model implementation for the LLM interface, including support for OpenAI-compatible API requests, tool calling, and schema conversion. While the implementation provides a solid foundation, there are several critical issues to address. Specifically, the GenerateContent method currently lacks actual streaming support despite accepting a stream parameter, which will cause JSON parsing failures if streaming is enabled. Additionally, the temperature clamping logic incorrectly maps deterministic requests (0.0) to maximum randomness (1.0), and there is an unhandled error when unmarshaling tool call arguments in the response conversion logic.

[gemini-code-assist]

The current implementation of GenerateContent does not support streaming, although it accepts a stream parameter and returns an iterator. The internal generate method (line 219) performs a blocking io.ReadAll on the response body and returns a single *model.LLMResponse. For streaming to work, the response should be processed as a stream of Server-Sent Events (SSE), and each chunk should be yielded through the iterator. As it stands, if stream is set to true, the API will return an SSE stream which json.Unmarshal (line 254) will fail to parse.

[gemini-code-assist]

Clamping a temperature of 0.0 to 1.0 is likely incorrect. In LLM APIs, 0.0 usually signifies a request for deterministic output. If MiniMax requires the temperature to be in the range (0.0, 1.0], mapping 0.0 to the maximum randomness (1.0) contradicts the typical user intent. It would be better to clamp 0.0 to a small positive value like 0.01.

Suggested change

	if temp <= 0 {
	temp = 1.0
	if temp <= 0 {
	temp = 0.01

[gemini-code-assist]

The error from json.Unmarshal when parsing tool call arguments is ignored. It is better to handle this error to ensure that malformed arguments from the API are caught and reported, rather than returning a FunctionCall with potentially nil or empty arguments.