|
| 1 | +"""The brain → publisher hand-off must preserve tool calls. |
| 2 | +
|
| 3 | +`stream_for_publish` is what both publisher entry points (`zhub up` and |
| 4 | +examples/multi_brain_publisher.py) use to turn a brain adapter's stream |
| 5 | +into the chunks a zhub chat_handler yields. Every bespoke adapter was |
| 6 | +taught to surface OpenAI-shape tool_call deltas + a |
| 7 | +``finish_reason="tool_calls"`` terminator, but a handler that yields only |
| 8 | +``chunk.delta`` throws all of that away — so a tool-call turn reaches the |
| 9 | +hub as empty text with finish_reason "stop" and auto-resolution never |
| 10 | +fires. These tests pin the hand-off on both the streaming and the |
| 11 | +non-streaming publisher paths. |
| 12 | +""" |
| 13 | + |
| 14 | +from __future__ import annotations |
| 15 | + |
| 16 | +import json |
| 17 | + |
| 18 | +import pytest |
| 19 | + |
| 20 | +from zhub.brains import stream_for_publish |
| 21 | +from zhub.brains.base import BrainAdapter, ChatChunk |
| 22 | +from zhub.client import ( |
| 23 | + _accumulate_tool_call, |
| 24 | + _chunk_fields, |
| 25 | + _finalize_accumulated, |
| 26 | + _serialize_stream_chunk, |
| 27 | +) |
| 28 | + |
| 29 | + |
| 30 | +class _ToolCallBrain(BrainAdapter): |
| 31 | + """A brain that answers a turn with a single tool call: one text |
| 32 | + preamble delta, a tool_call opener (empty delta), an argument |
| 33 | + fragment (empty delta), then a final empty chunk carrying |
| 34 | + finish_reason="tool_calls". Records the tools it was handed.""" |
| 35 | + |
| 36 | + name = "toolfake" |
| 37 | + label = "tool-call fake brain" |
| 38 | + |
| 39 | + def __init__(self) -> None: |
| 40 | + self.seen_tools = None |
| 41 | + |
| 42 | + @classmethod |
| 43 | + def try_init(cls): |
| 44 | + return cls() |
| 45 | + |
| 46 | + async def stream(self, messages, *, system=None, temperature=0.7, |
| 47 | + max_tokens=2048, tools=None): |
| 48 | + self.seen_tools = tools |
| 49 | + yield ChatChunk(delta="let me check ") |
| 50 | + yield ChatChunk(delta="", tool_call_delta={ |
| 51 | + "index": 0, "id": "call_1", "type": "function", |
| 52 | + "function": {"name": "get_weather"}, |
| 53 | + }) |
| 54 | + yield ChatChunk(delta="", tool_call_delta={ |
| 55 | + "index": 0, "function": {"arguments": '{"city":"Paris"}'}, |
| 56 | + }) |
| 57 | + yield ChatChunk(delta="", done=True, finish_reason="tool_calls") |
| 58 | + |
| 59 | + |
| 60 | +async def _collect(brain, **kw): |
| 61 | + return [c async for c in stream_for_publish(brain, [{"role": "user", "content": "hi"}], **kw)] |
| 62 | + |
| 63 | + |
| 64 | +@pytest.mark.asyncio |
| 65 | +async def test_tool_call_chunks_are_not_dropped(): |
| 66 | + brain = _ToolCallBrain() |
| 67 | + chunks = await _collect(brain) |
| 68 | + |
| 69 | + # The text delta survives. |
| 70 | + assert any(c.delta == "let me check " for c in chunks) |
| 71 | + # Both tool_call deltas survive (the bug dropped these — empty delta). |
| 72 | + tcds = [c.tool_call_delta for c in chunks if c.tool_call_delta] |
| 73 | + assert len(tcds) == 2 |
| 74 | + assert tcds[0]["function"]["name"] == "get_weather" |
| 75 | + assert tcds[1]["function"]["arguments"] == '{"city":"Paris"}' |
| 76 | + # The tool_calls finish_reason survives. |
| 77 | + assert any(c.finish_reason == "tool_calls" for c in chunks) |
| 78 | + |
| 79 | + |
| 80 | +@pytest.mark.asyncio |
| 81 | +async def test_tools_are_forwarded_to_the_brain(): |
| 82 | + brain = _ToolCallBrain() |
| 83 | + tools = [{"type": "function", "function": {"name": "get_weather"}}] |
| 84 | + await _collect(brain, tools=tools) |
| 85 | + assert brain.seen_tools == tools |
| 86 | + |
| 87 | + |
| 88 | +@pytest.mark.asyncio |
| 89 | +async def test_streaming_serialization_carries_tool_calls_to_hub(): |
| 90 | + """Serialize each yielded chunk the way the publisher's streaming |
| 91 | + path does, then run the envelopes through the hub's streaming |
| 92 | + accumulator logic — a tool call + tool_calls finish must come out.""" |
| 93 | + brain = _ToolCallBrain() |
| 94 | + chunks = await _collect(brain) |
| 95 | + |
| 96 | + accumulated: dict[int, dict] = {} |
| 97 | + final_finish = "stop" |
| 98 | + for chunk in chunks: |
| 99 | + payload = json.loads(_serialize_stream_chunk(chunk, "req"))["payload"] |
| 100 | + tcd = payload.get("tool_call_delta") |
| 101 | + if tcd: |
| 102 | + _accumulate_tool_call(accumulated, tcd) |
| 103 | + if payload.get("done"): |
| 104 | + final_finish = payload.get("finish_reason") or "stop" |
| 105 | + |
| 106 | + assert final_finish == "tool_calls" |
| 107 | + assert accumulated[0]["function"]["name"] == "get_weather" |
| 108 | + assert accumulated[0]["function"]["arguments"] == '{"city":"Paris"}' |
| 109 | + |
| 110 | + |
| 111 | +@pytest.mark.asyncio |
| 112 | +async def test_nonstreaming_accumulation_carries_tool_calls_to_hub(): |
| 113 | + """The non-streaming publisher path folds the same chunks into one |
| 114 | + chat-response; tool_calls and the real finish_reason must be on it.""" |
| 115 | + brain = _ToolCallBrain() |
| 116 | + chunks = await _collect(brain) |
| 117 | + |
| 118 | + text_parts: list[str] = [] |
| 119 | + slots: dict[int, dict] = {} |
| 120 | + finish = None |
| 121 | + for chunk in chunks: |
| 122 | + text, tcd, fin = _chunk_fields(chunk) |
| 123 | + if text: |
| 124 | + text_parts.append(text) |
| 125 | + if tcd: |
| 126 | + _accumulate_tool_call(slots, tcd) |
| 127 | + if fin: |
| 128 | + finish = fin |
| 129 | + payload = _finalize_accumulated(text_parts, slots, finish) |
| 130 | + |
| 131 | + assert payload["finish_reason"] == "tool_calls" |
| 132 | + assert payload["tool_calls"][0]["function"]["name"] == "get_weather" |
| 133 | + assert payload["tool_calls"][0]["function"]["arguments"] == '{"city":"Paris"}' |
| 134 | + assert payload["text"] == "let me check " |
0 commit comments