|
| 1 | +"""Tests for OpenAIProvider streaming chat completions (issue #653). |
| 2 | +
|
| 3 | +Anthropic-backed OpenAI-compatible endpoints reject non-streaming requests |
| 4 | +that run longer than 10 minutes. With ``params.stream: true`` on the model |
| 5 | +node, the provider must issue a streaming request and reassemble the chunks |
| 6 | +into the classic chat.completion shape the rest of the pipeline expects. |
| 7 | +""" |
| 8 | + |
| 9 | +from types import SimpleNamespace |
| 10 | +from unittest.mock import MagicMock |
| 11 | + |
| 12 | +import openai |
| 13 | +import pytest |
| 14 | + |
| 15 | +from entity.configs import AgentConfig |
| 16 | +from entity.messages import Message, MessageRole |
| 17 | +from runtime.node.agent.providers.openai_provider import OpenAIProvider |
| 18 | + |
| 19 | + |
| 20 | +# --------------------------------------------------------------------------- |
| 21 | +# Helpers |
| 22 | +# --------------------------------------------------------------------------- |
| 23 | + |
| 24 | +def _make_provider(params=None) -> OpenAIProvider: |
| 25 | + config = MagicMock(spec=AgentConfig) |
| 26 | + config.base_url = None |
| 27 | + config.api_key = "test-key" |
| 28 | + config.name = "test-model" |
| 29 | + config.provider = "openai" |
| 30 | + config.params = params or {} |
| 31 | + config.token_tracker = None |
| 32 | + return OpenAIProvider(config) |
| 33 | + |
| 34 | + |
| 35 | +def _chunk(content=None, role=None, tool_calls=None, finish_reason=None, usage=None): |
| 36 | + delta = SimpleNamespace(role=role, content=content, tool_calls=tool_calls) |
| 37 | + choice = SimpleNamespace(index=0, delta=delta, finish_reason=finish_reason) |
| 38 | + return SimpleNamespace( |
| 39 | + id="chatcmpl-123", |
| 40 | + model="test-model", |
| 41 | + created=1, |
| 42 | + choices=[choice] if (content or role or tool_calls or finish_reason) else [], |
| 43 | + usage=usage, |
| 44 | + ) |
| 45 | + |
| 46 | + |
| 47 | +def _tool_call_delta(index, call_id=None, name=None, arguments=None): |
| 48 | + return SimpleNamespace( |
| 49 | + index=index, |
| 50 | + id=call_id, |
| 51 | + function=SimpleNamespace(name=name, arguments=arguments), |
| 52 | + ) |
| 53 | + |
| 54 | + |
| 55 | +def _make_client(chunks): |
| 56 | + client = MagicMock() |
| 57 | + client.chat.completions.create.return_value = iter(chunks) |
| 58 | + return client |
| 59 | + |
| 60 | + |
| 61 | +# --------------------------------------------------------------------------- |
| 62 | +# Non-streaming default is preserved |
| 63 | +# --------------------------------------------------------------------------- |
| 64 | + |
| 65 | +def test_default_call_is_non_streaming(): |
| 66 | + provider = _make_provider() |
| 67 | + client = MagicMock() |
| 68 | + provider._create_chat_completion(client, {"model": "test-model", "messages": []}) |
| 69 | + |
| 70 | + kwargs = client.chat.completions.create.call_args.kwargs |
| 71 | + assert "stream" not in kwargs |
| 72 | + assert "stream_options" not in kwargs |
| 73 | + |
| 74 | + |
| 75 | +# --------------------------------------------------------------------------- |
| 76 | +# Streaming opt-in |
| 77 | +# --------------------------------------------------------------------------- |
| 78 | + |
| 79 | +def test_stream_param_enables_streaming_request(): |
| 80 | + provider = _make_provider(params={"stream": True}) |
| 81 | + client = _make_client([_chunk(content="hi", finish_reason="stop")]) |
| 82 | + provider._create_chat_completion(client, {"model": "test-model", "messages": []}) |
| 83 | + |
| 84 | + kwargs = client.chat.completions.create.call_args.kwargs |
| 85 | + assert kwargs["stream"] is True |
| 86 | + assert kwargs["stream_options"] == {"include_usage": True} |
| 87 | + |
| 88 | + |
| 89 | +def test_stream_chunks_reassembled_into_classic_shape(): |
| 90 | + provider = _make_provider(params={"stream": True}) |
| 91 | + usage = SimpleNamespace( |
| 92 | + prompt_tokens=10, completion_tokens=5, total_tokens=15 |
| 93 | + ) |
| 94 | + client = _make_client([ |
| 95 | + _chunk(role="assistant", content="Hello"), |
| 96 | + _chunk(content=", "), |
| 97 | + _chunk(content="world"), |
| 98 | + _chunk(finish_reason="stop"), |
| 99 | + _chunk(usage=usage), |
| 100 | + ]) |
| 101 | + |
| 102 | + response = provider._create_chat_completion( |
| 103 | + client, {"model": "test-model", "messages": []} |
| 104 | + ) |
| 105 | + |
| 106 | + assert response.choices[0].message.content == "Hello, world" |
| 107 | + assert response.choices[0].finish_reason == "stop" |
| 108 | + assert response.choices[0].message.tool_calls is None |
| 109 | + assert response.usage.total_tokens == 15 |
| 110 | + |
| 111 | + # The reassembled response must satisfy the existing deserializer. |
| 112 | + message = provider._deserialize_chat_response(response) |
| 113 | + assert message.role is MessageRole.ASSISTANT |
| 114 | + assert message.text_content() == "Hello, world" |
| 115 | + |
| 116 | + # And token tracking must read the streamed usage payload. |
| 117 | + token_usage = provider.extract_token_usage(response) |
| 118 | + assert token_usage.input_tokens == 10 |
| 119 | + assert token_usage.output_tokens == 5 |
| 120 | + |
| 121 | + |
| 122 | +def test_streamed_tool_calls_are_accumulated(): |
| 123 | + provider = _make_provider(params={"stream": True}) |
| 124 | + client = _make_client([ |
| 125 | + _chunk(tool_calls=[_tool_call_delta(0, call_id="call_1", name="search")]), |
| 126 | + _chunk(tool_calls=[_tool_call_delta(0, arguments='{"query": ')]), |
| 127 | + _chunk(tool_calls=[_tool_call_delta(0, arguments='"chatdev"}')]), |
| 128 | + _chunk(finish_reason="tool_calls"), |
| 129 | + ]) |
| 130 | + |
| 131 | + response = provider._create_chat_completion( |
| 132 | + client, {"model": "test-model", "messages": []} |
| 133 | + ) |
| 134 | + |
| 135 | + calls = response.choices[0].message.tool_calls |
| 136 | + assert len(calls) == 1 |
| 137 | + assert calls[0].id == "call_1" |
| 138 | + assert calls[0].function.name == "search" |
| 139 | + assert calls[0].function.arguments == '{"query": "chatdev"}' |
| 140 | + |
| 141 | + message = provider._deserialize_chat_response(response) |
| 142 | + assert message.tool_calls[0].function_name == "search" |
| 143 | + assert message.tool_calls[0].arguments == '{"query": "chatdev"}' |
| 144 | + |
| 145 | + |
| 146 | +def test_stream_true_in_payload_params_also_streams(): |
| 147 | + """`stream` passed through node params ends up in the payload; it must not |
| 148 | + reach the SDK as a bare kwarg on the reassembly path.""" |
| 149 | + provider = _make_provider() |
| 150 | + client = _make_client([_chunk(content="ok", finish_reason="stop")]) |
| 151 | + response = provider._create_chat_completion( |
| 152 | + client, {"model": "test-model", "messages": [], "stream": True} |
| 153 | + ) |
| 154 | + assert response.choices[0].message.content == "ok" |
| 155 | + |
| 156 | + |
| 157 | +def test_stream_options_rejection_falls_back_without_usage(): |
| 158 | + provider = _make_provider(params={"stream": True}) |
| 159 | + client = MagicMock() |
| 160 | + bad_request = openai.BadRequestError( |
| 161 | + message="stream_options not supported", |
| 162 | + response=MagicMock(status_code=400, headers={}), |
| 163 | + body=None, |
| 164 | + ) |
| 165 | + client.chat.completions.create.side_effect = [ |
| 166 | + bad_request, |
| 167 | + iter([_chunk(content="ok", finish_reason="stop")]), |
| 168 | + ] |
| 169 | + |
| 170 | + response = provider._create_chat_completion( |
| 171 | + client, {"model": "test-model", "messages": []} |
| 172 | + ) |
| 173 | + |
| 174 | + assert response.choices[0].message.content == "ok" |
| 175 | + retry_kwargs = client.chat.completions.create.call_args.kwargs |
| 176 | + assert retry_kwargs["stream"] is True |
| 177 | + assert "stream_options" not in retry_kwargs |
| 178 | + |
| 179 | + |
| 180 | +def test_call_model_streams_end_to_end(): |
| 181 | + provider = _make_provider(params={"stream": True, "protocol": "chat"}) |
| 182 | + client = _make_client([ |
| 183 | + _chunk(role="assistant", content="done"), |
| 184 | + _chunk(finish_reason="stop"), |
| 185 | + ]) |
| 186 | + |
| 187 | + timeline = [] |
| 188 | + result = provider.call_model( |
| 189 | + client, |
| 190 | + conversation=[Message(role=MessageRole.USER, content="go")], |
| 191 | + timeline=timeline, |
| 192 | + ) |
| 193 | + |
| 194 | + assert result.message.text_content() == "done" |
| 195 | + assert timeline[-1]["content"] == "done" |
| 196 | + kwargs = client.chat.completions.create.call_args.kwargs |
| 197 | + assert kwargs["stream"] is True |
0 commit comments