Python: feat: Add Astraflow (UCloud) AI provider connector

[ucloudnb666]

Summary

This PR adds a new AI provider connector for Astraflow (by UCloud / 优刻得), an OpenAI-compatible AI model aggregation platform that supports 200+ models.

What's added

• python/semantic_kernel/connectors/ai/astraflow/ — full connector package following the exact same structure as the existing nvidia connector:
  • Settings (AstraflowSettings) — reads ASTRAFLOW_API_KEY (global) or ASTRAFLOW_CN_API_KEY (China) from environment, with sensible defaults for both endpoints.
  • Prompt execution settings (AstraflowPromptExecutionSettings, AstraflowChatPromptExecutionSettings, AstraflowEmbeddingPromptExecutionSettings)
  • Handler (AstraflowHandler) — thin async wrapper around openai.AsyncOpenAI
  • Chat completion service (AstraflowChatCompletion)
  • Text embedding service (AstraflowTextEmbedding)
  • README with quick-start examples
• .env.example — adds ASTRAFLOW_API_KEY and ASTRAFLOW_CN_API_KEY entries

Endpoints

Region	Base URL	Env var
Global (US/CA)	https://api-us-ca.umodelverse.ai/v1	ASTRAFLOW_API_KEY
China	https://api.modelverse.cn/v1	ASTRAFLOW_CN_API_KEY

Usage

from semantic_kernel.connectors.ai.astraflow import AstraflowChatCompletion, AstraflowTextEmbedding

# Chat completion (global endpoint, set ASTRAFLOW_API_KEY env var)
chat = AstraflowChatCompletion(ai_model_id="deepseek-ai/DeepSeek-V3")

# Text embedding
embed = AstraflowTextEmbedding(ai_model_id="BAAI/bge-m3")

# China endpoint
chat_cn = AstraflowChatCompletion(
    ai_model_id="deepseek-ai/DeepSeek-V3",
    base_url="https://api.modelverse.cn/v1",
    api_key="...",
)

Sign up at https://astraflow.ucloud.cn/

[github-actions]

Automated Code Review

Reviewers: 4 | Confidence: 93%

✗ Correctness

The Astraflow connector has one blocking bug: streaming chat completions will always crash with an AssertionError because the code asserts that the OpenAI AsyncStream response is an instance of collections.abc.AsyncGenerator, which it is not (verified: issubclass(AsyncStream, AsyncGenerator) is False). Additionally, the _prepare_chat_history_for_request method unnecessarily overrides the base class's robust implementation with a naive one that loses tool-call and structured content information.

✗ Security Reliability

The new Astraflow connector introduces three reliability concerns: (1) the custom _prepare_chat_history_for_request override bypasses ChatMessageContent.to_dict(), which strips tool_calls, tool_call_id, and multi-modal content from the request payload — breaking function/tool calling despite the connector explicitly implementing tool-call parsing; (2) AstraflowTextEmbedding.from_dict omits the base_url parameter, silently losing custom endpoint configuration (e.g. China endpoint) during deserialization; (3) the embedding service only warns when no API key is found instead of failing fast, deferring errors to runtime with confusing messages.

✗ Test Coverage

This PR adds a new Astraflow AI connector with chat completion, text embedding, prompt execution settings, a handler, and configuration settings — but includes zero unit tests. Every other AI connector in the project (Nvidia, Ollama, Mistral AI, Anthropic, Azure AI Inference, etc.) has a corresponding test directory under python/tests/unit/connectors/ai/ with tests covering initialization, settings validation, chat completion (mocked), text embedding (mocked), error handling, and serialization round-trips. The complete absence of tests for all new classes is a blocking gap in test coverage.

✗ Design Approach

The connector is a reasonable 'OpenAI-compatible with custom base_url' implementation that follows the existing NVIDIA connector pattern. The most critical design flaw is that _prepare_chat_history_for_request is overridden to use message.content (plain text only) instead of message.to_dict(), which the base class already provides correctly. This means any conversation involving function/tool calls will silently lose the tool_call_id and structured content from TOOL-role messages, breaking function calling entirely. A secondary issue is that the response_format field is declared in the settings class but explicitly excluded from prepare_settings_dict, making JSON mode and structured output silently non-functional.

Flagged Issues

• _prepare_chat_history_for_request (lines 282-293 of astraflow_chat_completion.py) overrides the base class to build a simple {role, content} dict using message.content (plain text) instead of message.to_dict(). This drops tool_calls on assistant messages, tool_call_id on tool-result messages, multi-modal content items, message name, and ignores the instruction_role setting. Since the connector supports function calling (parses tool_calls in responses, exposes tools in settings), any multi-turn function-calling conversation will fail. The base class ChatCompletionClientBase._prepare_chat_history_for_request already handles OpenAI-compatible serialization correctly via message.to_dict(). The fix is to remove this override entirely.

---

Automated review by ucloudnb666's agents

[sphenry]

@ucloudnb666 These connectors are better supported from from the Astraflow repo