This is the SDK-side change for langgenius/dify#35772.

Problem

When the plugin daemon dispatches a model invocation (LLM, Embedding, Rerank, etc.), it creates a Session that already contains app_id — the daemon uses a shared InvokePluginRequest struct for both tool and model dispatch, so app_id arrives in the Session correctly.

However, plugin_executor.py does not forward this session to model plugin instances. It passes session to tool plugins but not to model plugins:

# Tool invocation — session IS passed
def invoke_tool(self, session: Session, request):
    tool = tool_cls(
        runtime=ToolRuntime(...),
        session=session,  # ← passed
    )
    yield from tool.invoke(request.tool_parameters)

# Model invocation — session is NOT passed
def invoke_llm(self, session: Session, data):
    model_instance = self.registration.get_model_instance(...)
    return model_instance.invoke(
        data.model,
        data.credentials,
        data.prompt_messages,
        data.model_parameters,
        data.tools,
        data.stop,
        data.stream,
        data.user_id,
        # ← session is not forwarded
    )

This means model plugins cannot access session.app_id, which is needed for provider-side cost attribution (e.g., Bedrock requestMetadata, OpenAI metadata, Vertex AI labels).

Proposed change

1. In plugin_executor.py, pass session to model plugin instances for all model invocation methods (invoke_llm, invoke_text_embedding, invoke_rerank, invoke_tts, invoke_speech_to_text, invoke_moderation, etc.)
2. Add an optional session parameter to model base class _invoke() methods (LargeLanguageModel, TextEmbeddingModel, RerankModel, etc.) so plugin authors can access session.app_id when available

This is backward compatible — session would be optional, and existing model plugins that don't use it require no changes.

Parent issue

langgenius/dify#35772

We have a PR ready and can submit it upon agreement on the design direction.