mirror of
https://github.com/ThearchyHelios/AgentLab.git
synced 2026-11-18 23:12:01 +00:00
协作团队真的能并行了,界面上也看得出来
先说一个和名字不符的事实:这个节点叫「多 Agent 协作」,但在此之前它是一条
严格的串行流水线。调度协议写死成 `next: string`——"一次只能派一个"被编进了
协议本身,模型再想并行也表达不出来。三个专家各花 10 秒,总是 30 秒。
(真并行一直在图级 fan-out:实测 3 个各 sleep 2 秒的节点连在同一个上游,
全部 +0.00s 开始、+2.37s 结束,墙钟 2.6 秒。只是协作团队走不到那条路。)
**协议改成 assignments 数组**,一轮可以派多个互不依赖的专家,asyncio.gather
并发执行,总耗时按最慢的那个算。max_parallel 可配,设成 1 即退回严格串行。
**并发只在任务真的互不依赖时才成立**,这件事的成本必须说清楚:同一轮发出去的
专家拿到的是**同一份进展快照**,谁也看不见谁这一轮干了什么。派错了不报错,
只会让两个人基于一样的旧信息重复劳动,而这在结果里很难看出来。所以判断责任
压在调度者身上,提示词里把反例写出来:
可以同时派:「查 A 的资料」和「查 B 的资料」
不能同时派:「查资料」和「根据资料写报告」
不能同时派:「写初稿」和「校对初稿」
并明确告诉它拿不准就分两轮——代价只是慢一点。
**每个人单独计时**,而不是事后拿整轮墙钟去摊。界面要显示"并行省下多少",
那个数必须是量出来的:按 N×墙钟 推算等于假设每个人都跑满了最慢那条,
而实际上快的那个可能只花了三分之一。测试里专门有个快慢不对称的用例守这条。
**调度日志改成结构化字段**(agents / done / reason / parallel)。以前只有一句
中文消息串,解码层得拿正则去拆「调度 → X(理由)」——后端文案一改就散架,
而这次正是改了文案。老运行的事件没有这些字段,正则那条路留着兜底。
**界面加泳道。** 每个成员一行、每轮一格,条形宽度按耗时归一化,所以一眼看得出
谁和谁是同时干的、谁是这一轮的瓶颈。顶上一行汇总"3 名成员 · 3 轮 · 2 轮并行,
省下 11.8s";点任意一格展开那个人这一轮的任务和回复。全程串行时不装作有并行,
而是说"本次任务互相依赖,逐轮进行"。
泳道排在子步骤**前面**:先给一眼看清的形状,再往下是每一步的内容。反过来的话,
得把十几行步骤读完才知道这个团队是怎么分工的。窄栏(画布右栏 360px)下隐去
"N 人同时"的文字——并行与否看条形本来就一目了然,而缩写成"并2"没人看得懂。
验证:后端 276 passed(新增 test_supervisor.py 7 条,核心那条用**墙钟**判定
并发而不是看事件顺序——顺序只能说明"发出去了",说明不了"同时跑");
check-stream 新增 10 条泳道断言,含"同一轮里慢的那个条更长"和两种栏宽不溢出;
tsc 干净,三个 check 不回归。
This commit is contained in:
@@ -1,5 +1,6 @@
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import json
|
||||
import time
|
||||
from typing import Any
|
||||
@@ -64,17 +65,39 @@ async def run_supervisor(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
f"- {n}:{agent_map[n].get('description') or agent_map[n].get('system', '')[:100]}"
|
||||
for n in names
|
||||
)
|
||||
# 一轮可以派给多个专家。协议用 assignments 数组而不是单个 next:
|
||||
# 后者把"一次只能派一个"编进了协议本身,模型再想并行也表达不出来。
|
||||
#
|
||||
# 但并行只在任务**互不依赖**时才成立——并发的几个专家看到的是同一份进展
|
||||
# 快照,谁也看不见谁。派错了不会报错,只会让两个人基于同样的旧信息
|
||||
# 重复劳动,而这件事在结果里很难看出来。所以判断责任明确压在调度者身上,
|
||||
# 提示词里把反例写出来。
|
||||
route_schema = {
|
||||
"title": "route", # langchain 靠 title 识别 JSON Schema dict
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"next": {"type": "string", "enum": names + ["FINISH"]},
|
||||
"instruction": {"type": "string"},
|
||||
"assignments": {
|
||||
"type": "array",
|
||||
"description": "这一轮要派出去的任务。互不依赖时可以给多个,它们会同时执行",
|
||||
"items": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"agent": {"type": "string", "enum": names},
|
||||
"instruction": {"type": "string"},
|
||||
},
|
||||
"required": ["agent", "instruction"],
|
||||
},
|
||||
},
|
||||
"done": {"type": "boolean", "description": "目标已达成,结束协作"},
|
||||
"reason": {"type": "string"},
|
||||
},
|
||||
"required": ["next"],
|
||||
"required": ["assignments"],
|
||||
}
|
||||
|
||||
#: 一轮最多同时派几个。再多的话调度者多半只是在把任务拆碎,
|
||||
#: 而每多一路就多一份"基于陈旧进展"的风险
|
||||
max_parallel = max(1, min(int(ctx.cfg("max_parallel", 3) or 3), len(names)))
|
||||
|
||||
transcript: list[dict[str, Any]] = []
|
||||
usage_total = {"input_tokens": 0, "output_tokens": 0, "cost_usd": 0.0, "calls": 0}
|
||||
|
||||
@@ -94,17 +117,27 @@ async def run_supervisor(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
f"你是团队调度者。目标:\n{goal}\n\n"
|
||||
f"可用成员:\n{roster}\n\n"
|
||||
f"当前进展:\n{progress or '(还没有任何进展)'}\n\n"
|
||||
"决定下一步交给谁处理,并给出一句明确的指令。"
|
||||
"如果目标已经达成,next 返回 FINISH。"
|
||||
"决定这一轮把什么任务交给谁,每个任务给一句明确的指令。\n\n"
|
||||
f"**互不依赖的任务可以放在同一轮**(最多 {max_parallel} 个),它们会同时执行,"
|
||||
"总耗时按最慢的那个算。判断依据只有一条:**后一个任务需不需要看到前一个的结果**。\n"
|
||||
" 可以同时派:「查 A 产品的资料」和「查 B 产品的资料」——两边互不相干\n"
|
||||
" 不能同时派:「查资料」和「根据资料写报告」——后者要等前者\n"
|
||||
" 不能同时派:「写初稿」和「校对初稿」——同上\n\n"
|
||||
"同时派出去的成员看到的是**同一份进展快照**,谁也看不见谁这一轮干了什么。"
|
||||
"拿不准是否独立时就分两轮派,代价只是慢一点;派错了则是两个人基于同样的"
|
||||
"旧信息重复劳动,而这在结果里很难看出来。\n\n"
|
||||
"目标已经达成时,done 填 true、assignments 给空数组。"
|
||||
)
|
||||
try:
|
||||
value = await supervisor_model.with_structured_output(route_schema).ainvoke(prompt)
|
||||
if not isinstance(value, dict):
|
||||
value = json.loads(message_text(value))
|
||||
except Exception: # noqa: BLE001
|
||||
# 不支持结构化输出的模型退回文本:从回复里认出一个成员名,按单人派
|
||||
raw = message_text(await supervisor_model.ainvoke(prompt))
|
||||
picked = next((n for n in names if n in raw), "FINISH")
|
||||
value = {"next": picked, "instruction": raw[:500], "reason": ""}
|
||||
picked = next((n for n in names if n in raw), None)
|
||||
value = ({"assignments": [{"agent": picked, "instruction": raw[:500]}]}
|
||||
if picked else {"assignments": [], "done": True, "reason": raw[:200]})
|
||||
return value
|
||||
|
||||
@task
|
||||
@@ -179,42 +212,99 @@ async def run_supervisor(state: GraphState, ctx: NodeContext) -> dict[str, Any]:
|
||||
|
||||
progress_lines: list[str] = []
|
||||
final_text = ""
|
||||
rounds_meta: list[dict[str, Any]] = []
|
||||
for round_no in range(max_rounds):
|
||||
progress = "\n\n".join(progress_lines)
|
||||
decision = await route(round_no, progress)
|
||||
_accumulate(AIMessage(content=""), sup_model_id) # route 用了一次调用
|
||||
nxt = str(decision.get("next") or "FINISH")
|
||||
reason = str(decision.get("reason", ""))
|
||||
ctx.emit(
|
||||
EventType.LOG, level="info", round=round_no,
|
||||
message=f"调度 → {nxt}" + (f"({reason})" if reason else ""),
|
||||
)
|
||||
if nxt == "FINISH" or nxt not in agent_map:
|
||||
final_text = progress_lines[-1].split(":", 1)[-1] if progress_lines else ""
|
||||
|
||||
# 只留认得出的成员,并按上限截断。模型偶尔会重复派同一个人或者编一个
|
||||
# 不存在的名字——这两种在并发里都是纯浪费
|
||||
batch: list[tuple[str, str]] = []
|
||||
seen: set[str] = set()
|
||||
for item in decision.get("assignments") or []:
|
||||
name = str((item or {}).get("agent") or "")
|
||||
if name in agent_map and name not in seen:
|
||||
seen.add(name)
|
||||
batch.append((name, str(item.get("instruction") or goal)))
|
||||
if len(batch) >= max_parallel:
|
||||
break
|
||||
|
||||
if decision.get("done") or not batch:
|
||||
# agents / done 是结构化字段,界面照着它渲染。以前只有一句中文
|
||||
# 消息串,解码层得拿正则去拆「调度 → X(理由)」——文案一改就散架
|
||||
ctx.emit(EventType.LOG, level="info", round=round_no,
|
||||
agents=[], done=True,
|
||||
message="调度 → 结束协作" + (f"({reason})" if reason else ""))
|
||||
final_text = progress_lines[-1].split("】", 1)[-1] if progress_lines else ""
|
||||
break
|
||||
|
||||
instruction = str(decision.get("instruction") or goal)
|
||||
ctx.emit(EventType.AGENT_STEP_START, agent=nxt, instruction=instruction[:300],
|
||||
round=round_no)
|
||||
started = time.perf_counter()
|
||||
outcome = await work(round_no, nxt, instruction, progress)
|
||||
elapsed = int((time.perf_counter() - started) * 1000)
|
||||
ctx.emit(EventType.AGENT_STEP_END, agent=nxt, duration_ms=elapsed,
|
||||
round=round_no, preview=outcome["text"][:500])
|
||||
|
||||
progress_lines.append(f"【{nxt}】{outcome['text']}")
|
||||
transcript.append(
|
||||
{"round": round_no, "agent": nxt, "instruction": instruction, **outcome,
|
||||
"duration_ms": elapsed}
|
||||
who = "、".join(n for n, _ in batch)
|
||||
ctx.emit(
|
||||
EventType.LOG, level="info", round=round_no, parallel=len(batch),
|
||||
agents=[n for n, _ in batch], done=False, reason=reason,
|
||||
message=(f"调度 → {who}" + (f"({reason})" if reason else "")
|
||||
+ (f" · {len(batch)} 人同时进行" if len(batch) > 1 else "")),
|
||||
)
|
||||
final_text = outcome["text"]
|
||||
|
||||
for name, instruction in batch:
|
||||
ctx.emit(EventType.AGENT_STEP_START, agent=name, instruction=instruction[:300],
|
||||
round=round_no, parallel=len(batch))
|
||||
|
||||
# 并发执行。它们拿到的是**同一份** progress 快照——这正是"任务必须
|
||||
# 互不依赖"的技术含义,也是上面提示词里那几条反例的由来。
|
||||
#
|
||||
# 每个人单独计时,而不是事后拿整轮墙钟去摊:界面要显示"并行省了多少
|
||||
# 时间",那个数必须是量出来的。按 N×墙钟 推算等于假设每个人都跑满了
|
||||
# 最慢那条,而实际上快的那个可能只花了三分之一
|
||||
async def _timed(name: str, instruction: str) -> tuple[str, Any, int]:
|
||||
t0 = time.perf_counter()
|
||||
try:
|
||||
out: Any = await work(round_no, name, instruction, progress)
|
||||
except BaseException as exc: # noqa: BLE001 - 交给下面统一处置
|
||||
out = exc
|
||||
return name, out, int((time.perf_counter() - t0) * 1000)
|
||||
|
||||
started = time.perf_counter()
|
||||
results = await asyncio.gather(*(_timed(n, i) for n, i in batch))
|
||||
wall_ms = int((time.perf_counter() - started) * 1000)
|
||||
|
||||
sum_ms = 0
|
||||
for (name, instruction), (_, outcome, each_ms) in zip(batch, results):
|
||||
if isinstance(outcome, BaseException):
|
||||
# 一个专家挂了不该拖垮整轮:把失败当成它的产出交回调度者,
|
||||
# 让它决定改派还是收工
|
||||
text = f"({name} 执行失败:{type(outcome).__name__}: {outcome})"
|
||||
outcome = {"text": text, "tool_calls": []}
|
||||
ctx.emit(EventType.LOG, level="warn", round=round_no,
|
||||
message=f"{name} 这一轮失败了:{text}")
|
||||
sum_ms += each_ms
|
||||
ctx.emit(EventType.AGENT_STEP_END, agent=name, duration_ms=each_ms,
|
||||
round=round_no, parallel=len(batch), preview=outcome["text"][:500])
|
||||
progress_lines.append(f"【{name}】{outcome['text']}")
|
||||
transcript.append(
|
||||
{"round": round_no, "agent": name, "instruction": instruction, **outcome,
|
||||
"duration_ms": each_ms}
|
||||
)
|
||||
final_text = outcome["text"]
|
||||
|
||||
rounds_meta.append({
|
||||
"round": round_no, "agents": [n for n, _ in batch],
|
||||
"parallel": len(batch), "wall_ms": wall_ms, "sum_ms": sum_ms,
|
||||
})
|
||||
|
||||
usage_total["total_tokens"] = usage_total["input_tokens"] + usage_total["output_tokens"]
|
||||
# 并行省下的时间:各轮"串行本该花的"减去"实际花的"。只有并发过才有差值
|
||||
saved_ms = sum(r["sum_ms"] - r["wall_ms"] for r in rounds_meta)
|
||||
result = {
|
||||
"text": final_text,
|
||||
"rounds": len(transcript),
|
||||
"rounds": len(rounds_meta),
|
||||
"steps": len(transcript),
|
||||
"transcript": transcript,
|
||||
"agents": names,
|
||||
"schedule": rounds_meta,
|
||||
"parallel_saved_ms": saved_ms,
|
||||
}
|
||||
updates: dict[str, Any] = {
|
||||
"nodes": {ctx.node.id: result},
|
||||
|
||||
@@ -0,0 +1,310 @@
|
||||
"""协作团队:一轮可以同时派多个互不依赖的专家。
|
||||
|
||||
在此之前调度协议是 `next: string`——"一次只能派一个"被编进了协议本身,模型
|
||||
再想并行也表达不出来。于是这个节点叫「多 Agent 协作」,实际是一条严格的串行
|
||||
流水线:三个专家各花 10 秒,总是 30 秒。
|
||||
|
||||
改成 assignments 数组之后,互不依赖的任务同一轮发出去,总耗时按最慢的那个算。
|
||||
|
||||
**但并发只在任务真的互不依赖时才成立。** 同一轮发出去的专家拿到的是同一份
|
||||
进展快照,谁也看不见谁这一轮干了什么。派错了不会报错,只会让两个人基于一样的
|
||||
旧信息重复劳动——所以判断责任压在调度者身上,提示词里写了反例,而这里守的是
|
||||
机制本身:说好并发就真的并发、串行的路径一步没坏、单个专家挂掉不拖垮整轮。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
import json
|
||||
|
||||
import pytest
|
||||
from langchain_core.messages import AIMessage
|
||||
from sqlalchemy import select
|
||||
|
||||
from app.db.base import SessionLocal
|
||||
from app.db.models import Run, RunEvent
|
||||
from app.engine.runner import run_manager
|
||||
|
||||
#: 每个专家假装干这么久。并发时整轮 ≈ 这个数,串行时 ≈ 这个数 × 人数
|
||||
WORK_SECONDS = 0.4
|
||||
|
||||
|
||||
@pytest.fixture
|
||||
async def runner():
|
||||
await run_manager.setup()
|
||||
yield run_manager
|
||||
await run_manager.shutdown()
|
||||
|
||||
|
||||
def _graph(*, max_parallel: int = 3, members=("researcher", "analyst", "writer")) -> dict:
|
||||
agents = [
|
||||
{"name": n, "description": f"{n} 的职责", "system": f"你是 {n}", "max_steps": 1}
|
||||
for n in members
|
||||
]
|
||||
return {
|
||||
"nodes": [
|
||||
{"id": "in", "type": "input", "data": {"config": {}}},
|
||||
{"id": "team", "type": "supervisor", "data": {"config": {
|
||||
"goal": "比较三种向量库",
|
||||
"agents": agents,
|
||||
"max_rounds": 4,
|
||||
"max_parallel": max_parallel,
|
||||
}}},
|
||||
{"id": "out", "type": "output", "data": {"config": {}}},
|
||||
],
|
||||
"edges": [
|
||||
{"source": "in", "target": "team"},
|
||||
{"source": "team", "target": "out"},
|
||||
],
|
||||
}
|
||||
|
||||
|
||||
def script(monkeypatch, decisions: list[dict]) -> None:
|
||||
"""让调度者按剧本派活,让专家干活花真实时间。
|
||||
|
||||
专家那一边必须是**异步**等待:同步 sleep 会把事件循环堵死,于是并发测试
|
||||
永远"通过"(因为根本没并发过,墙钟也就成了串行时间)。
|
||||
"""
|
||||
from app.providers import mock_model
|
||||
|
||||
turns = {"n": 0}
|
||||
|
||||
def _decide(self, messages):
|
||||
# response_format 有值 = 调度者在做结构化路由;没有 = 专家在干活
|
||||
if self.response_format:
|
||||
i = turns["n"]
|
||||
turns["n"] += 1
|
||||
payload = decisions[i] if i < len(decisions) else {"assignments": [], "done": True}
|
||||
return AIMessage(content=json.dumps(payload, ensure_ascii=False))
|
||||
return AIMessage(content="(干完了)")
|
||||
|
||||
async def _agenerate(self, messages, stop=None, run_manager=None, **kw):
|
||||
if not self.response_format:
|
||||
await asyncio.sleep(WORK_SECONDS)
|
||||
return self._generate(messages, stop, **kw)
|
||||
|
||||
monkeypatch.setattr(mock_model.MockChatModel, "_decide", _decide)
|
||||
monkeypatch.setattr(mock_model.MockChatModel, "_agenerate", _agenerate)
|
||||
|
||||
|
||||
async def _run(graph: dict):
|
||||
run = await run_manager.start(graph=graph, input_payload={"question": "比较三种向量库"})
|
||||
for _ in range(300):
|
||||
await asyncio.sleep(0.05)
|
||||
async with SessionLocal() as session:
|
||||
row = (await session.execute(select(Run).where(Run.id == run.id))).scalar_one()
|
||||
if row.status in ("succeeded", "failed", "cancelled"):
|
||||
break
|
||||
else:
|
||||
pytest.fail("运行没有在 15 秒内结束")
|
||||
async with SessionLocal() as session:
|
||||
events = list((await session.execute(
|
||||
select(RunEvent).where(RunEvent.run_id == run.id).order_by(RunEvent.seq)
|
||||
)).scalars())
|
||||
return events, row
|
||||
|
||||
|
||||
def steps(events, kind):
|
||||
return [e.data for e in events if e.type == kind]
|
||||
|
||||
|
||||
# --------------------------------------------------------------------------
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_independent_tasks_really_run_at_the_same_time(runner, monkeypatch):
|
||||
"""这块改动的全部意义:说好并发就真的并发。
|
||||
|
||||
用墙钟判定而不是看事件顺序——顺序只能说明"发出去了",说明不了"同时跑"。
|
||||
"""
|
||||
script(monkeypatch, [{
|
||||
"assignments": [
|
||||
{"agent": "researcher", "instruction": "查 A 的资料"},
|
||||
{"agent": "analyst", "instruction": "查 B 的资料"},
|
||||
],
|
||||
"reason": "两边互不相干",
|
||||
}, {"assignments": [], "done": True}])
|
||||
|
||||
events, run = await _run(_graph())
|
||||
assert run.status == "succeeded", run.error
|
||||
|
||||
starts = steps(events, "agent.step.start")
|
||||
assert {d["agent"] for d in starts} == {"researcher", "analyst"}
|
||||
assert all(d["parallel"] == 2 for d in starts), starts
|
||||
|
||||
# 两个人都开跑了,第一个才结束——这是"同时"的事件层证据
|
||||
first_end = next(e.ts for e in events if e.type == "agent.step.end")
|
||||
assert all(e.ts <= first_end for e in events if e.type == "agent.step.start")
|
||||
|
||||
# 而这是墙钟证据:串行要 2×WORK_SECONDS,并发约 1×
|
||||
sched = (run.output or {}).get("_schedule") or _schedule_of(events)
|
||||
assert sched, "调度信息要能拿得到"
|
||||
r0 = sched[0]
|
||||
assert r0["parallel"] == 2
|
||||
assert r0["wall_ms"] < WORK_SECONDS * 2000 * 0.8, f"看着像串行:{r0}"
|
||||
assert r0["sum_ms"] > r0["wall_ms"], "并发就该省出时间"
|
||||
|
||||
|
||||
def _schedule_of(events):
|
||||
"""schedule 在节点输出里,测试从事件流兜一份等价的出来。"""
|
||||
from collections import defaultdict
|
||||
rounds = defaultdict(lambda: {"agents": [], "parallel": 0, "wall_ms": 0, "sum_ms": 0})
|
||||
for d in (e.data for e in events if e.type == "agent.step.end"):
|
||||
r = rounds[d["round"]]
|
||||
r["round"] = d["round"]
|
||||
r["agents"].append(d["agent"])
|
||||
r["parallel"] = d.get("parallel", 1)
|
||||
r["wall_ms"] = max(r["wall_ms"], d["duration_ms"])
|
||||
r["sum_ms"] += d["duration_ms"]
|
||||
return [rounds[k] for k in sorted(rounds)]
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_saved_time_is_measured_not_assumed(runner, monkeypatch):
|
||||
"""界面要显示"并行省了 X 秒",那个数必须是量出来的。
|
||||
|
||||
两个人快慢不同时才区分得出来:按"整轮墙钟 × 人数"推算的话,快的那个会被
|
||||
记成和慢的一样长,省下的时间就被夸大了。所以让一个人明显更快,
|
||||
再看各自的耗时是不是真的不一样。
|
||||
"""
|
||||
from app.providers import mock_model
|
||||
|
||||
script(monkeypatch, [{
|
||||
"assignments": [
|
||||
{"agent": "researcher", "instruction": "慢活"},
|
||||
{"agent": "analyst", "instruction": "快活"},
|
||||
],
|
||||
}, {"assignments": [], "done": True}])
|
||||
|
||||
async def _uneven(self, messages, stop=None, run_manager=None, **kw):
|
||||
if not self.response_format:
|
||||
text = " ".join(str(getattr(m, "content", "")) for m in messages)
|
||||
await asyncio.sleep(0.05 if "快活" in text else WORK_SECONDS)
|
||||
return self._generate(messages, stop, **kw)
|
||||
|
||||
monkeypatch.setattr(mock_model.MockChatModel, "_agenerate", _uneven)
|
||||
|
||||
events, run = await _run(_graph())
|
||||
assert run.status == "succeeded", run.error
|
||||
|
||||
took = {d["agent"]: d["duration_ms"] for d in steps(events, "agent.step.end")}
|
||||
assert took["analyst"] < took["researcher"] / 2, f"快慢该有差别:{took}"
|
||||
# 整轮墙钟按最慢的那个算,而不是两个人加起来
|
||||
assert took["researcher"] < WORK_SECONDS * 2000 * 0.8, took
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_dependent_tasks_stay_sequential(runner, monkeypatch):
|
||||
"""分两轮派就该老老实实跑两轮——并行是可选项,不是强加的。"""
|
||||
script(monkeypatch, [
|
||||
{"assignments": [{"agent": "researcher", "instruction": "查资料"}]},
|
||||
{"assignments": [{"agent": "writer", "instruction": "根据资料写报告"}]},
|
||||
{"assignments": [], "done": True},
|
||||
])
|
||||
|
||||
events, run = await _run(_graph())
|
||||
assert run.status == "succeeded", run.error
|
||||
|
||||
starts = steps(events, "agent.step.start")
|
||||
assert [d["agent"] for d in starts] == ["researcher", "writer"]
|
||||
assert all(d["parallel"] == 1 for d in starts)
|
||||
|
||||
# 第二个人开跑必须在第一个人结束之后
|
||||
first_end = next(e.ts for e in events if e.type == "agent.step.end")
|
||||
second_start = [e for e in events if e.type == "agent.step.start"][1].ts
|
||||
assert second_start >= first_end
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_max_parallel_caps_the_batch(runner, monkeypatch):
|
||||
"""上限设成 1 就退回严格串行——给不放心的人一个明确的开关。"""
|
||||
script(monkeypatch, [{
|
||||
"assignments": [
|
||||
{"agent": "researcher", "instruction": "a"},
|
||||
{"agent": "analyst", "instruction": "b"},
|
||||
{"agent": "writer", "instruction": "c"},
|
||||
],
|
||||
}, {"assignments": [], "done": True}])
|
||||
|
||||
events, run = await _run(_graph(max_parallel=1))
|
||||
assert run.status == "succeeded", run.error
|
||||
starts = steps(events, "agent.step.start")
|
||||
assert len(starts) == 1, "上限 1 就只该派一个"
|
||||
assert starts[0]["parallel"] == 1
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_unknown_and_duplicate_agents_are_dropped(runner, monkeypatch):
|
||||
"""模型偶尔会重复派同一个人或者编一个不存在的名字,两种在并发里都是纯浪费。"""
|
||||
script(monkeypatch, [{
|
||||
"assignments": [
|
||||
{"agent": "researcher", "instruction": "a"},
|
||||
{"agent": "researcher", "instruction": "又派一次"},
|
||||
{"agent": "根本没这个人", "instruction": "c"},
|
||||
],
|
||||
}, {"assignments": [], "done": True}])
|
||||
|
||||
events, run = await _run(_graph())
|
||||
assert run.status == "succeeded", run.error
|
||||
starts = steps(events, "agent.step.start")
|
||||
assert [d["agent"] for d in starts] == ["researcher"]
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_one_failing_member_does_not_sink_the_round(runner, monkeypatch):
|
||||
"""并发里一个人挂了,另一个的成果不该跟着丢。"""
|
||||
from app.providers import mock_model
|
||||
|
||||
script(monkeypatch, [{
|
||||
"assignments": [
|
||||
{"agent": "researcher", "instruction": "会挂的那个"},
|
||||
{"agent": "analyst", "instruction": "正常的那个"},
|
||||
],
|
||||
}, {"assignments": [], "done": True}])
|
||||
|
||||
original = mock_model.MockChatModel._agenerate
|
||||
|
||||
async def _boom(self, messages, stop=None, run_manager=None, **kw):
|
||||
text = " ".join(str(getattr(m, "content", "")) for m in messages)
|
||||
if "会挂的那个" in text:
|
||||
raise RuntimeError("这个专家炸了")
|
||||
return await original(self, messages, stop, run_manager, **kw)
|
||||
|
||||
monkeypatch.setattr(mock_model.MockChatModel, "_agenerate", _boom)
|
||||
|
||||
events, run = await _run(_graph())
|
||||
assert run.status == "succeeded", run.error
|
||||
|
||||
# 失败要说出来,而不是安静地少一个人的产出
|
||||
warns = [d["message"] for d in steps(events, "log") if d.get("level") == "warn"]
|
||||
assert any("researcher" in w for w in warns), warns
|
||||
# 另一个人的结束事件照常有
|
||||
ends = {d["agent"] for d in steps(events, "agent.step.end")}
|
||||
assert "analyst" in ends
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_schedule_is_reported_for_the_ui(runner, monkeypatch):
|
||||
"""界面要画泳道,就得拿得到每轮谁干了活、并发几路、省了多少时间。"""
|
||||
script(monkeypatch, [{
|
||||
"assignments": [
|
||||
{"agent": "researcher", "instruction": "a"},
|
||||
{"agent": "analyst", "instruction": "b"},
|
||||
],
|
||||
}, {
|
||||
"assignments": [{"agent": "writer", "instruction": "汇总"}],
|
||||
}, {"assignments": [], "done": True}])
|
||||
|
||||
events, run = await _run(_graph())
|
||||
assert run.status == "succeeded", run.error
|
||||
|
||||
rounds = sorted({d["round"] for d in steps(events, "agent.step.end")})
|
||||
assert rounds == [0, 1]
|
||||
by_round = {r: [d["agent"] for d in steps(events, "agent.step.end") if d["round"] == r]
|
||||
for r in rounds}
|
||||
assert set(by_round[0]) == {"researcher", "analyst"}
|
||||
assert by_round[1] == ["writer"]
|
||||
|
||||
# 调度决策本身也要在事件流里,否则界面上这个节点就是个跑了 N 秒的黑盒
|
||||
dispatches = [d["message"] for d in steps(events, "log") if d.get("round") is not None]
|
||||
assert any("同时进行" in m for m in dispatches), dispatches
|
||||
Reference in New Issue
Block a user