fix(flows): 步数预算与图硬上限分离,预算耗尽时优雅收尾
max_steps 以前同时充当软预算和 recursion_limit,于是预算耗尽的那一刻 必然触发 GraphRecursionError:模型连一句收尾都来不及说,用户只收到 「处理消息时出错,请稍后再试」,已产出的内容也不会落库。 - 新增 StepBudgetMiddleware:按模型调用轮数计账(agent_rounds = max_steps // 2),剩余 <=4 轮时注入一次「立刻停止调用工具、给出结论」 的收尾提示,只提醒一次 - 新增 STEP_GRACE(20),recursion_limit = max_steps + STEP_GRACE,把余量 留给收尾动作 - GraphRecursionError 统一转成 StepBudgetExceeded,ainvoke 与流式两条 路径都覆盖 - AgentReply 把「步数用尽」渲染成可操作说明,接在已产出内容之后一并 下发并落库;用 _reported 标记让 on_error 不再补发通用错误文案,任务 状态仍如实记 failed - 同步 README / docs/architecture.md / db/schema.py 中 max_steps 的语义 测试:新增 tests/test_budget.py(7 个用例),全量 241 passed。
This commit is contained in:
1 parent
fbb01c4c83
commit
5da55f8d17
12 files changed
+523
-32
No files matched your search
@@ -154,8 +154,10 @@ bot serve --channel wecom
|
||||
规则:
|
||||
|
||||
- 未设置 `BOT042_MODEL` 时,使用 `models` 表中 `is_default=1` 的模型;
|
||||
- 单次任务步数上限同理:`settings.agent.max_steps`(默认 200),`BOT042_MAX_STEPS` 优先;
|
||||
联网检索类任务步数消耗快,触顶会抛 `GraphRecursionError`,表现为「处理消息时出错」;
|
||||
- 单次任务步数预算同理:`settings.agent.max_steps`(默认 200),`BOT042_MAX_STEPS` 优先;
|
||||
预算将尽时智能体先收到一条「立刻收尾」提示,把已完成的部分总结出来;交给 LangGraph
|
||||
的硬上限是 `max_steps + 20` 个图步,真撞上时用户看到的是「本轮执行步数用尽(上限 N 步)」
|
||||
加已产出的内容,而不是笼统的「处理消息时出错」;
|
||||
- `OPENAI_API_KEY` / `OPENAI_BASE_URL` 为空时,从该模型所属 `providers` 行补齐;
|
||||
- 环境变量指定的模型与库中默认模型**不一致**时,不会套用库里的密钥
|
||||
(错用别的 key 只会得到 401);
|
||||
@@ -169,7 +171,7 @@ bot serve --channel wecom
|
||||
| `BOT042_MODEL` | langchain 模型标识(`provider:model`) | `openai:gpt-4o-mini` |
|
||||
| `OPENAI_API_KEY` | OpenAI 兼容 API Key | — |
|
||||
| `OPENAI_BASE_URL` | 兼容网关地址(DeepSeek / Kimi 等均可) | 官方地址 |
|
||||
| `BOT042_MAX_STEPS` | 单次任务最大步数(recursion_limit) | `200` |
|
||||
| `BOT042_MAX_STEPS` | 单次任务步数预算(软上限,硬上限为其 +20) | `200` |
|
||||
| `BOT042_LOG_LEVEL` | 日志级别 | `INFO` |
|
||||
|
||||
使用 DeepSeek 等 OpenAI 兼容服务示例:
|
||||
|
||||
+13
-4
@@ -457,13 +457,22 @@ langchain 从环境变量取凭据。
|
||||
一次 `ainvoke`,但缓存由持有者(`ChannelServices`)决定生命周期,不做模块级单例,
|
||||
避免配置不同或测试替换 `create_agent` 时串味。
|
||||
|
||||
`run_task()` 把历史转成 `{"messages": [...]}`,以 `recursion_limit=settings.max_steps` 调用 `ainvoke`,
|
||||
`run_task()` 把历史转成 `{"messages": [...]}`,以 `recursion_limit=settings.recursion_limit` 调用 `ainvoke`,
|
||||
返回最终 state;`last_message_text()` 负责从多模态 content 中只取文本块(按行拼接)。
|
||||
|
||||
`max_steps` 默认 200,优先级 **环境变量 `BOT042_MAX_STEPS` > `settings.agent.max_steps` > 默认**:
|
||||
联网检索类任务一轮要连续消耗十几个图步数(搜索 → 抓取 → 再搜索 → 汇总),上限过低时
|
||||
`GraphRecursionError` 会被渠道兜底成用户看到的「处理消息时出错,请稍后再试。」,
|
||||
所以它和模型配置一起由 `resolve_model_settings()` 解析进 `Settings`。
|
||||
联网检索类任务一轮要连续消耗十几个图步数(搜索 → 抓取 → 再搜索 → 汇总)。它是**软预算**,
|
||||
不是 `recursion_limit` 本身——
|
||||
|
||||
- `StepBudgetMiddleware`(`agent/middlewares/budget.py`)按模型调用轮数计账(`agent_rounds = max_steps // 2`,
|
||||
一次「模型 → 工具」往返约耗 2 个图步),剩余 ≤4 轮时注入一条「立刻停止调用工具、给出结论」
|
||||
的提示,且只提醒一次;
|
||||
- 交给 LangGraph 的硬上限是 `recursion_limit = max_steps + STEP_GRACE(20)`,留给收尾动作;
|
||||
- 真撞上硬上限(模型不听话、或子智能体的步数不在计数内)时,`GraphRecursionError` 会被转成
|
||||
`StepBudgetExceeded`;`AgentReply` 把它渲染成「本轮执行步数用尽(上限 N 步)」+ 已产出的部分内容,
|
||||
一并落库为 assistant 消息,并标记 `_reported` 让 `on_error` 不再补发通用错误文案。
|
||||
|
||||
两者以前是同值,于是预算耗尽的那一刻必然抛异常,模型连一句话都来不及说——这是 v0.1 的 bug。
|
||||
|
||||
## 7. 存储层(`db/`)
|
||||
|
||||
|
||||
@@ -11,6 +11,7 @@ from deepagents.backends import LocalShellBackend
|
||||
from langchain.agents.middleware.types import AgentMiddleware
|
||||
from langgraph.graph.state import CompiledStateGraph
|
||||
|
||||
from bot042.agent.middlewares.budget import StepBudgetMiddleware
|
||||
from bot042.agent.middlewares.steering import SteeringMiddleware
|
||||
from bot042.agent.prompts import SYSTEM_PROMPT
|
||||
from bot042.agent.tools import get_tools
|
||||
@@ -90,12 +91,13 @@ def create_agent(
|
||||
"openai:gpt-4o-mini"、"openai:deepseek-chat"(配合 OPENAI_BASE_URL)等。
|
||||
"""
|
||||
settings = settings or Settings.from_env()
|
||||
# SteeringMiddleware 放最前:每轮模型调用前都先消化插话
|
||||
# SteeringMiddleware 放最前:每轮模型调用前都先消化插话;
|
||||
# StepBudgetMiddleware 紧随其后:预算将尽时注入收尾提示,避免撞硬上限
|
||||
return create_deep_agent(
|
||||
model=_build_model(settings),
|
||||
tools=get_tools(),
|
||||
system_prompt=SYSTEM_PROMPT,
|
||||
backend=_build_backend(settings),
|
||||
middleware=[SteeringMiddleware(), *middleware],
|
||||
middleware=[SteeringMiddleware(), StepBudgetMiddleware(), *middleware],
|
||||
interrupt_on=interrupt_on,
|
||||
)
|
||||
@@ -0,0 +1,126 @@
|
||||
"""步数预算:在硬上限之前提醒模型收尾。
|
||||
|
||||
LangGraph 的 ``recursion_limit`` 是**硬中断**——撞上就抛
|
||||
``GraphRecursionError``,模型没有任何机会整理已做的工作、给出结论。所以这里
|
||||
在触到硬上限之前先注入一条收尾提示,让模型自己把半成品收口:说清做到哪一步、
|
||||
结论是什么、还差什么。
|
||||
|
||||
预算按**模型调用轮数**计(一个「模型 → 工具」往返约耗 2 个图步),轮数预算由
|
||||
:attr:`bot042.config.Settings.agent_rounds` 从 ``max_steps`` 换算而来;硬上限
|
||||
则是 :attr:`bot042.config.Settings.recursion_limit`,比预算多留
|
||||
:data:`bot042.config.STEP_GRACE` 步给收尾动作。
|
||||
|
||||
与插话一样用 :mod:`contextvars` 绑定:``AgentCache`` 让同一张图被多个会话复用,
|
||||
轮数记在实例属性上会串台。
|
||||
|
||||
限制:中间件只数**主智能体**的模型调用;子智能体(``task`` 工具派发的)消耗的
|
||||
步数不在计数内。所以硬上限不能省——它是唯一能兜住子智能体跑飞的东西。
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import contextvars
|
||||
from dataclasses import dataclass
|
||||
from typing import Any
|
||||
|
||||
from langchain.agents.middleware.types import AgentMiddleware
|
||||
from langchain_core.messages import HumanMessage
|
||||
|
||||
__all__ = [
|
||||
"BUDGET_PREFIX",
|
||||
"BUDGET_PROMPT",
|
||||
"BUDGET_WARN_MARGIN",
|
||||
"StepBudget",
|
||||
"StepBudgetExceeded",
|
||||
"StepBudgetMiddleware",
|
||||
"bind",
|
||||
"current_budget",
|
||||
"unbind",
|
||||
]
|
||||
|
||||
BUDGET_WARN_MARGIN = 4
|
||||
"""剩余轮数 ≤ 该值时开始提醒收尾(留够 4 轮写下结论)。"""
|
||||
|
||||
BUDGET_PREFIX = "[系统 · 步数预算] "
|
||||
"""收尾提示的前缀,让模型能识别这不是用户的新指令。"""
|
||||
|
||||
BUDGET_PROMPT = (
|
||||
"本次任务的执行预算即将用尽(剩余约 {remaining} 轮模型调用)。"
|
||||
"请立刻停止调用工具,基于已经完成的工作给出最终答复:"
|
||||
"说明做到了哪一步、结论是什么、还差什么没做。不要再发起新的探索或修改。"
|
||||
)
|
||||
|
||||
|
||||
class StepBudgetExceeded(RuntimeError):
|
||||
"""图步数硬上限(``recursion_limit``)被耗尽。
|
||||
|
||||
这是**兜底**路径:正常情况下预算中间件会先提醒模型收尾;走到这里说明模型
|
||||
没有遵守提醒,或子智能体吃掉了大量步数。
|
||||
"""
|
||||
|
||||
def __init__(self, max_steps: int) -> None:
|
||||
self.max_steps = max_steps
|
||||
super().__init__(f"任务步数用尽(max_steps={max_steps})")
|
||||
|
||||
|
||||
@dataclass
|
||||
class StepBudget:
|
||||
"""一次运行的轮数预算(绑定单个 asyncio 任务,不加锁)。"""
|
||||
|
||||
rounds: int
|
||||
"""允许的模型调用轮数。"""
|
||||
|
||||
used: int = 0
|
||||
"""已发生的模型调用次数。"""
|
||||
|
||||
warned: bool = False
|
||||
"""是否已经注入过收尾提示(只提醒一次,避免反复打断)。"""
|
||||
|
||||
@property
|
||||
def remaining(self) -> int:
|
||||
"""剩余轮数(不为负)。"""
|
||||
return max(0, self.rounds - self.used)
|
||||
|
||||
def take(self) -> int:
|
||||
"""登记一次模型调用,返回登记后的剩余轮数。"""
|
||||
self.used += 1
|
||||
return self.remaining
|
||||
|
||||
|
||||
_current: contextvars.ContextVar[StepBudget | None] = contextvars.ContextVar(
|
||||
"bot042_step_budget", default=None
|
||||
)
|
||||
|
||||
|
||||
def current_budget() -> StepBudget | None:
|
||||
"""当前 asyncio 上下文的步数预算(无运行中任务时为 ``None``)。"""
|
||||
return _current.get()
|
||||
|
||||
|
||||
def bind(budget: StepBudget) -> contextvars.Token[StepBudget | None]:
|
||||
"""把预算绑定到当前上下文,返回用于还原的 token。"""
|
||||
return _current.set(budget)
|
||||
|
||||
|
||||
def unbind(token: contextvars.Token[StepBudget | None]) -> None:
|
||||
"""还原上下文绑定。"""
|
||||
_current.reset(token)
|
||||
|
||||
|
||||
class StepBudgetMiddleware(AgentMiddleware):
|
||||
"""在预算将尽时注入一次收尾提示。"""
|
||||
|
||||
async def abefore_model(self, state: Any, runtime: Any) -> dict[str, Any] | None:
|
||||
_ = state, runtime
|
||||
budget = _current.get()
|
||||
if budget is None or budget.warned:
|
||||
return None
|
||||
remaining = budget.take()
|
||||
if remaining > BUDGET_WARN_MARGIN:
|
||||
return None
|
||||
budget.warned = True
|
||||
return {
|
||||
"messages": [
|
||||
HumanMessage(content=BUDGET_PREFIX + BUDGET_PROMPT.format(remaining=remaining))
|
||||
]
|
||||
}
|
||||
+27
-3
@@ -8,6 +8,13 @@ from pathlib import Path
|
||||
|
||||
from bot042.home import db_path, resolve_home
|
||||
|
||||
STEP_GRACE = 20
|
||||
"""图步数硬上限相对 ``max_steps`` 的余量。
|
||||
|
||||
留给「预算将尽」提示被遵守的空间:预算用尽与硬上限如果相等,撞上限的那一刻
|
||||
直接抛 ``GraphRecursionError``,模型连一句话都来不及说。
|
||||
"""
|
||||
|
||||
|
||||
@dataclass(frozen=True)
|
||||
class Settings:
|
||||
@@ -21,10 +28,13 @@ class Settings:
|
||||
|
||||
model: str = "openai:gpt-4o-mini"
|
||||
max_steps: int = 200
|
||||
"""单次任务的图步数上限(recursion_limit)。
|
||||
"""单次任务的**软**步数预算。
|
||||
|
||||
联网检索类任务一轮要跑很多次工具调用(搜索 → 抓取 → 再搜索),50 步很容易
|
||||
在给出答案前耗尽并抛 ``GraphRecursionError``,因此默认给到 200。
|
||||
这是给模型看的预算:预算将尽时中间件会注入一条「立刻收尾」提示
|
||||
(见 :mod:`bot042.agent.middlewares.budget`),交给 LangGraph 的硬上限是
|
||||
:attr:`recursion_limit`(比它多 :data:`STEP_GRACE` 步)。联网检索类任务一轮
|
||||
要跑很多次工具调用(搜索 → 抓取 → 再搜索),50 步很容易在给出答案前耗尽,
|
||||
因此默认给到 200。
|
||||
"""
|
||||
log_level: str = "INFO"
|
||||
home: Path = field(default_factory=resolve_home)
|
||||
@@ -40,6 +50,20 @@ class Settings:
|
||||
"""home 目录下的 bot.db 路径。"""
|
||||
return db_path(self.home)
|
||||
|
||||
@property
|
||||
def recursion_limit(self) -> int:
|
||||
"""交给 LangGraph 的硬步数上限。
|
||||
|
||||
硬上限必须**大于**软预算:预算用尽时模型需要几步把结论写完,直接等值会
|
||||
让收尾提示和中断同时到达,退化回「抛异常、无结论」。
|
||||
"""
|
||||
return self.max_steps + STEP_GRACE
|
||||
|
||||
@property
|
||||
def agent_rounds(self) -> int:
|
||||
"""模型调用轮数预算:一次「模型 → 工具」往返约耗 2 个图步。"""
|
||||
return max(1, self.max_steps // 2)
|
||||
|
||||
@classmethod
|
||||
def from_env(cls) -> Settings:
|
||||
return cls(
|
||||
|
||||
@@ -100,7 +100,7 @@ _CHANNEL_SEEDS = [
|
||||
|
||||
_SETTING_SEEDS = [
|
||||
("agent.default_model", "openai:gpt-4o-mini", "默认模型"),
|
||||
("agent.max_steps", "200", "单次任务最大步数"),
|
||||
("agent.max_steps", "200", "单次任务步数预算"),
|
||||
("log.level", "INFO", "日志级别"),
|
||||
]
|
||||
|
||||
|
||||
@@ -8,9 +8,15 @@ from dataclasses import dataclass, field
|
||||
from typing import TYPE_CHECKING, Any
|
||||
|
||||
from langchain_core.messages import AIMessageChunk
|
||||
from langgraph.errors import GraphRecursionError
|
||||
|
||||
from bot042.agent import create_agent
|
||||
from bot042.agent.middlewares.steering import SteeringBuffer, bind, unbind
|
||||
from bot042.agent.middlewares.budget import StepBudget, StepBudgetExceeded
|
||||
from bot042.agent.middlewares.budget import bind as bind_budget
|
||||
from bot042.agent.middlewares.budget import unbind as unbind_budget
|
||||
from bot042.agent.middlewares.steering import SteeringBuffer
|
||||
from bot042.agent.middlewares.steering import bind as bind_steering
|
||||
from bot042.agent.middlewares.steering import unbind as unbind_steering
|
||||
from bot042.config import Settings
|
||||
|
||||
if TYPE_CHECKING:
|
||||
@@ -37,6 +43,11 @@ def _content_text(content: object) -> str:
|
||||
return ""
|
||||
|
||||
|
||||
def _run_config(settings: Settings) -> dict[str, Any]:
|
||||
"""图调用配置:硬步数上限 = 软预算 + 余量(见 ``Settings.recursion_limit``)。"""
|
||||
return {"recursion_limit": settings.recursion_limit}
|
||||
|
||||
|
||||
def last_message_text(result: dict[str, Any]) -> str:
|
||||
"""从智能体最终状态中提取最后一条消息的文本(多模态只取文本块)。"""
|
||||
messages = result.get("messages") or []
|
||||
@@ -84,6 +95,9 @@ async def run_task(
|
||||
history: 历史消息(不含本轮)。
|
||||
cache: 智能体缓存;给定则复用其中的实例,缺省每次新建(一次性调用用)。
|
||||
steering: 插话缓冲;给定则该次运行期间的插话会在下一轮模型调用前注入。
|
||||
|
||||
Raises:
|
||||
StepBudgetExceeded: 图步数硬上限被耗尽(软预算的收尾提示没起作用)。
|
||||
"""
|
||||
settings = settings or Settings.from_env()
|
||||
agent = cache.get(settings) if cache is not None else create_agent(settings)
|
||||
@@ -91,15 +105,20 @@ async def run_task(
|
||||
messages.append({"role": "user", "content": task})
|
||||
logger.debug("执行任务:%s", task)
|
||||
|
||||
token = bind(steering) if steering is not None else None
|
||||
token = bind_steering(steering) if steering is not None else None
|
||||
budget_token = bind_budget(StepBudget(settings.agent_rounds))
|
||||
try:
|
||||
return await agent.ainvoke(
|
||||
{"messages": messages},
|
||||
config={"recursion_limit": settings.max_steps},
|
||||
config=_run_config(settings),
|
||||
)
|
||||
except GraphRecursionError as exc:
|
||||
# 硬上限只是兜底:走到这里说明收尾提示没被遵守,或子智能体吃掉了步数
|
||||
raise StepBudgetExceeded(settings.max_steps) from exc
|
||||
finally:
|
||||
unbind_budget(budget_token)
|
||||
if token is not None:
|
||||
unbind(token)
|
||||
unbind_steering(token)
|
||||
|
||||
|
||||
@dataclass
|
||||
@@ -133,6 +152,9 @@ async def stream_task(
|
||||
history: 历史消息(不含本轮)。
|
||||
cache: 智能体缓存;给定则复用其中的实例。
|
||||
steering: 插话缓冲;给定则该次运行期间的插话会在下一轮模型调用前注入。
|
||||
|
||||
Raises:
|
||||
StepBudgetExceeded: 图步数硬上限被耗尽;调用方可用已产出的部分文本拼提示。
|
||||
"""
|
||||
settings = settings or Settings.from_env()
|
||||
agent = cache.get(settings) if cache is not None else create_agent(settings)
|
||||
@@ -140,13 +162,14 @@ async def stream_task(
|
||||
messages.append({"role": "user", "content": task})
|
||||
logger.debug("流式执行任务:%s", task)
|
||||
|
||||
token = bind(steering) if steering is not None else None
|
||||
token = bind_steering(steering) if steering is not None else None
|
||||
budget_token = bind_budget(StepBudget(settings.agent_rounds))
|
||||
rounds: dict[str, _Round] = {}
|
||||
order: list[str] = []
|
||||
try:
|
||||
async for chunk, _meta in agent.astream(
|
||||
{"messages": messages},
|
||||
config={"recursion_limit": settings.max_steps},
|
||||
config=_run_config(settings),
|
||||
stream_mode="messages",
|
||||
):
|
||||
# messages 模式同时会吐出 ToolMessage(工具返回值)与 HumanMessage,
|
||||
@@ -166,6 +189,10 @@ async def stream_task(
|
||||
visible = "".join(rounds[k].text for k in order if not rounds[k].tool)
|
||||
if visible:
|
||||
yield visible
|
||||
except GraphRecursionError as exc:
|
||||
# 已产出的文本由调用方持有(累积式产出),这里只把异常换成可读的错误类型
|
||||
raise StepBudgetExceeded(settings.max_steps) from exc
|
||||
finally:
|
||||
unbind_budget(budget_token)
|
||||
if token is not None:
|
||||
unbind(token)
|
||||
unbind_steering(token)
|
||||
@@ -20,6 +20,7 @@ from dataclasses import dataclass, field
|
||||
from pathlib import Path
|
||||
from typing import Any
|
||||
|
||||
from bot042.agent.middlewares.budget import StepBudgetExceeded
|
||||
from bot042.agent.middlewares.steering import SteeringBuffer
|
||||
from bot042.channels.manager import ChannelManager
|
||||
from bot042.channels.models import (
|
||||
@@ -62,10 +63,24 @@ ACK_TEXT = "收到,正在处理…"
|
||||
|
||||
ERROR_REPLY = "处理消息时出错,请稍后再试。"
|
||||
|
||||
BUDGET_REPLY = (
|
||||
"本轮执行步数用尽(上限 {steps} 步),我先停在这里。\n"
|
||||
"要继续的话,把需求拆小一点重新发一条,或者直接说「接着做」,我从当前进度往下推。"
|
||||
)
|
||||
"""步数用尽时的说明(接在已完成的部分回答之后)。"""
|
||||
|
||||
REPLY_META_KEYS = ("_frame", "response_url", "chat_id", "_ws_client")
|
||||
"""入站 metadata 中需要回传给渠道的回复上下文(IM 被动回复用)。"""
|
||||
|
||||
|
||||
def _with_notice(answer: str, aborted: StepBudgetExceeded | None) -> str:
|
||||
"""把「步数用尽」的说明接到部分回答后面;没有部分回答时只给说明。"""
|
||||
if aborted is None:
|
||||
return answer
|
||||
notice = BUDGET_REPLY.format(steps=aborted.max_steps)
|
||||
return f"{answer}\n\n{notice}" if answer.strip() else notice
|
||||
|
||||
|
||||
def reply_metadata(message: InboundMessage) -> dict[str, Any]:
|
||||
"""挑出入站消息里的回复上下文,随出站 RPC 回传给渠道。
|
||||
|
||||
@@ -205,6 +220,8 @@ class AgentReply(Task):
|
||||
super().__init__(context)
|
||||
self._steering = SteeringBuffer()
|
||||
self._steered: list[str] = []
|
||||
self._reported = False
|
||||
"""是否已自行回话并落库(失败时 ``on_error`` 不再发通用文案)。"""
|
||||
|
||||
@property
|
||||
def steering(self) -> SteeringBuffer:
|
||||
@@ -294,6 +311,9 @@ class AgentReply(Task):
|
||||
)
|
||||
|
||||
if stream is None:
|
||||
answer = ""
|
||||
aborted: StepBudgetExceeded | None = None
|
||||
try:
|
||||
result = await run_task(
|
||||
text,
|
||||
services.settings,
|
||||
@@ -302,18 +322,25 @@ class AgentReply(Task):
|
||||
steering=self._steering,
|
||||
)
|
||||
answer = last_message_text(result)
|
||||
except StepBudgetExceeded as exc:
|
||||
aborted = exc
|
||||
await self._persist_steered(services, session_id, channel)
|
||||
if answer:
|
||||
reply = _with_notice(answer, aborted)
|
||||
if reply:
|
||||
await repo.add_message(
|
||||
services.db,
|
||||
session_id=session_id,
|
||||
role="assistant",
|
||||
content=answer,
|
||||
content=reply,
|
||||
channel=channel,
|
||||
)
|
||||
await services.send_text(
|
||||
payload.channel_id, payload.subject_id, answer or EMPTY_REPLY, metadata=meta
|
||||
payload.channel_id, payload.subject_id, reply or EMPTY_REPLY, metadata=meta
|
||||
)
|
||||
if aborted is not None:
|
||||
# 已回话并落库:任务仍记失败(步数用尽是真的没做完),但不再发通用错误
|
||||
self._reported = True
|
||||
raise aborted
|
||||
return
|
||||
|
||||
# 支持流式:先给回执,再边生成边覆盖同一条消息。
|
||||
@@ -321,6 +348,7 @@ class AgentReply(Task):
|
||||
await stream.update(ACK_TEXT, force=True)
|
||||
answer = ""
|
||||
first = True
|
||||
aborted = None
|
||||
try:
|
||||
async for partial in stream_task(
|
||||
text,
|
||||
@@ -333,20 +361,33 @@ class AgentReply(Task):
|
||||
# 首段跳过节流:让用户在 ack 之后立刻看到内容开始出现
|
||||
await stream.update(partial, force=first)
|
||||
first = False
|
||||
except StepBudgetExceeded as exc:
|
||||
aborted = exc
|
||||
finally:
|
||||
await stream.finish(answer or EMPTY_REPLY)
|
||||
# 步数用尽时把说明补在已经流出的内容后面,不必再单独发一条
|
||||
final = _with_notice(answer, aborted)
|
||||
await stream.finish(final or EMPTY_REPLY)
|
||||
await self._persist_steered(services, session_id, channel)
|
||||
if answer:
|
||||
if final:
|
||||
await repo.add_message(
|
||||
services.db,
|
||||
session_id=session_id,
|
||||
role="assistant",
|
||||
content=answer,
|
||||
content=final,
|
||||
channel=channel,
|
||||
)
|
||||
if aborted is not None:
|
||||
self._reported = True
|
||||
raise aborted
|
||||
|
||||
async def on_error(self, exc: BaseException) -> None:
|
||||
"""失败时尽量回一句可读提示(发送本身失败只记日志)。"""
|
||||
"""失败时尽量回一句可读提示(发送本身失败只记日志)。
|
||||
|
||||
已经自行回话并落库的失败(如步数用尽)跳过——否则用户会先收到准确说明,
|
||||
再收到一句没信息量的「处理消息时出错」。
|
||||
"""
|
||||
if self._reported:
|
||||
return
|
||||
services = self.deps
|
||||
payload = self.context.payload
|
||||
if not (isinstance(services, ChannelServices) and isinstance(payload, ChannelMessage)):
|
||||
|
||||
@@ -0,0 +1,139 @@
|
||||
"""步数预算测试:预算换算、收尾提示注入、硬上限转成可读失败。"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import asyncio
|
||||
from typing import Any
|
||||
|
||||
import pytest
|
||||
from langgraph.errors import GraphRecursionError
|
||||
|
||||
from bot042.agent.middlewares.budget import (
|
||||
BUDGET_PREFIX,
|
||||
BUDGET_WARN_MARGIN,
|
||||
StepBudget,
|
||||
StepBudgetExceeded,
|
||||
StepBudgetMiddleware,
|
||||
bind,
|
||||
current_budget,
|
||||
unbind,
|
||||
)
|
||||
from bot042.config import STEP_GRACE, Settings
|
||||
from bot042.flows import agent_task as agent_task_module
|
||||
from bot042.flows.agent_task import run_task, stream_task
|
||||
|
||||
|
||||
class RecordingAgent:
|
||||
"""记录调用配置,并顺带把当时绑定的预算读出来。"""
|
||||
|
||||
def __init__(self) -> None:
|
||||
self.config: dict[str, Any] | None = None
|
||||
self.remaining: int | None = None
|
||||
|
||||
async def ainvoke(
|
||||
self, state: dict[str, Any], config: dict[str, Any] | None = None
|
||||
) -> dict[str, Any]:
|
||||
_ = state
|
||||
self.config = config
|
||||
budget = current_budget()
|
||||
self.remaining = budget.remaining if budget is not None else None
|
||||
return {"messages": []}
|
||||
|
||||
|
||||
class RecursionBoomAgent:
|
||||
"""模拟 LangGraph 撞上硬上限。"""
|
||||
|
||||
async def ainvoke(
|
||||
self, state: dict[str, Any], config: dict[str, Any] | None = None
|
||||
) -> dict[str, Any]:
|
||||
_ = state, config
|
||||
raise GraphRecursionError("Recursion limit of 27 reached")
|
||||
|
||||
async def astream(self, *args: Any, **kwargs: Any) -> Any:
|
||||
_ = args, kwargs
|
||||
raise GraphRecursionError("Recursion limit of 27 reached")
|
||||
yield # pragma: no cover - 让本方法成为异步生成器
|
||||
|
||||
|
||||
def test_settings_derive_recursion_limit_and_rounds() -> None:
|
||||
"""硬上限必须大于软预算,轮数预算按「一次往返约 2 步」换算。"""
|
||||
settings = Settings(max_steps=200)
|
||||
assert settings.recursion_limit == 200 + STEP_GRACE
|
||||
assert settings.agent_rounds == 100
|
||||
# 极小预算也要留出至少一轮,否则中间件永远来不及提醒
|
||||
assert Settings(max_steps=1).agent_rounds == 1
|
||||
|
||||
|
||||
def test_budget_take_counts_down() -> None:
|
||||
budget = StepBudget(rounds=2)
|
||||
assert budget.take() == 1
|
||||
assert budget.take() == 0
|
||||
assert budget.take() == 0 # 不为负
|
||||
assert budget.used == 3
|
||||
assert budget.remaining == 0
|
||||
|
||||
|
||||
def test_middleware_noop_without_budget() -> None:
|
||||
middleware = StepBudgetMiddleware()
|
||||
assert asyncio.run(middleware.abefore_model({}, None)) is None # type: ignore[arg-type]
|
||||
|
||||
|
||||
def test_middleware_warns_once_when_budget_runs_low() -> None:
|
||||
budget = StepBudget(rounds=BUDGET_WARN_MARGIN + 2)
|
||||
token = bind(budget)
|
||||
try:
|
||||
middleware = StepBudgetMiddleware()
|
||||
# 剩余 5 → 安静;剩余 4 → 注入收尾提示
|
||||
assert asyncio.run(middleware.abefore_model({}, None)) is None # type: ignore[arg-type]
|
||||
update = asyncio.run(middleware.abefore_model({}, None)) # type: ignore[arg-type]
|
||||
assert update is not None
|
||||
message = update["messages"][0]
|
||||
assert isinstance(message.content, str)
|
||||
assert message.content.startswith(BUDGET_PREFIX)
|
||||
assert "剩余约 4 轮" in message.content
|
||||
assert budget.warned is True
|
||||
# 只提醒一次:反复打断只会让模型更答不完
|
||||
assert asyncio.run(middleware.abefore_model({}, None)) is None # type: ignore[arg-type]
|
||||
finally:
|
||||
unbind(token)
|
||||
|
||||
|
||||
def test_run_task_passes_derived_limit_and_binds_budget(monkeypatch) -> None:
|
||||
agent = RecordingAgent()
|
||||
monkeypatch.setattr(
|
||||
agent_task_module, "create_agent", lambda settings=None: agent
|
||||
)
|
||||
|
||||
settings = Settings(max_steps=7)
|
||||
asyncio.run(run_task("hi", settings))
|
||||
|
||||
assert agent.config == {"recursion_limit": 7 + STEP_GRACE}
|
||||
assert agent.remaining == settings.agent_rounds # 预算已绑定且未被计数
|
||||
# 运行结束必须解绑,否则下一个任务会读到上一轮的预算
|
||||
assert current_budget() is None
|
||||
|
||||
|
||||
def test_run_task_translates_recursion_error(monkeypatch) -> None:
|
||||
monkeypatch.setattr(
|
||||
agent_task_module, "create_agent", lambda settings=None: RecursionBoomAgent()
|
||||
)
|
||||
|
||||
with pytest.raises(StepBudgetExceeded) as info:
|
||||
asyncio.run(run_task("hi", Settings(max_steps=7)))
|
||||
assert info.value.max_steps == 7
|
||||
assert "max_steps=7" in str(info.value)
|
||||
assert current_budget() is None
|
||||
|
||||
|
||||
def test_stream_task_translates_recursion_error(monkeypatch) -> None:
|
||||
monkeypatch.setattr(
|
||||
agent_task_module, "create_agent", lambda settings=None: RecursionBoomAgent()
|
||||
)
|
||||
|
||||
async def _consume() -> None:
|
||||
async for _ in stream_task("hi", Settings(max_steps=7)):
|
||||
pass
|
||||
|
||||
with pytest.raises(StepBudgetExceeded):
|
||||
asyncio.run(_consume())
|
||||
assert current_budget() is None
|
||||
+57
-1
@@ -9,6 +9,7 @@ from typing import Any
|
||||
|
||||
import pytest
|
||||
|
||||
from bot042.agent.middlewares.budget import StepBudgetExceeded
|
||||
from bot042.channels.base import BaseChannel
|
||||
from bot042.channels.manager import ChannelManager
|
||||
from bot042.channels.models import (
|
||||
@@ -20,7 +21,13 @@ from bot042.channels.models import (
|
||||
from bot042.config import Settings
|
||||
from bot042.db import repo
|
||||
from bot042.flows import builtin
|
||||
from bot042.flows.builtin import ERROR_REPLY, RPC_SEND, AgentReply, ChannelServices
|
||||
from bot042.flows.builtin import (
|
||||
BUDGET_REPLY,
|
||||
ERROR_REPLY,
|
||||
RPC_SEND,
|
||||
AgentReply,
|
||||
ChannelServices,
|
||||
)
|
||||
from bot042.flows.bus import EventBus
|
||||
from bot042.flows.rpc import RpcHub
|
||||
from bot042.flows.scheduler import CN_TZ
|
||||
@@ -209,3 +216,52 @@ def test_agent_reply_sends_error_hint_on_failure(
|
||||
sent, statuses = asyncio.run(_run())
|
||||
assert sent == [ERROR_REPLY]
|
||||
assert statuses == [RunStatus.FAILED]
|
||||
|
||||
|
||||
def test_agent_reply_explains_step_budget_without_generic_error(
|
||||
monkeypatch: pytest.MonkeyPatch, tmp_path: Path
|
||||
) -> None:
|
||||
"""步数用尽:给出可操作说明并落库,且不再补发通用错误文案。"""
|
||||
|
||||
async def exhausted(task, settings=None, history=None, *, cache=None, steering=None):
|
||||
raise StepBudgetExceeded(200)
|
||||
|
||||
monkeypatch.setattr(builtin, "run_task", exhausted)
|
||||
|
||||
async def _run() -> tuple[list[str], list[RunStatus]]:
|
||||
bus = EventBus()
|
||||
bus.bind_loop()
|
||||
hub = RpcHub(bus)
|
||||
manager = ChannelManager()
|
||||
channel = EchoChannel()
|
||||
await manager.add_channel(channel)
|
||||
await manager.start()
|
||||
_ = hub.serve(RPC_SEND, manager.rpc_send)
|
||||
|
||||
services = ChannelServices(
|
||||
channels=manager, settings=Settings(), db=tmp_path / "bot.db", hub=hub
|
||||
)
|
||||
runner = TaskRunner(bus, deps=services)
|
||||
runner.register(AgentReply)
|
||||
_ = runner.add_task(AgentReply.name, EventTrigger(event="channel.inbound"))
|
||||
await runner.start()
|
||||
|
||||
bus.emit("channel.inbound", _envelope("把整站重构一遍"))
|
||||
await runner.wait_idle(timeout=2)
|
||||
await bus.wait_for_complete()
|
||||
statuses = [run.status for run in runner.runs(AgentReply.name)]
|
||||
await runner.stop()
|
||||
hub.close()
|
||||
await manager.stop()
|
||||
return channel.sent, statuses
|
||||
|
||||
sent, statuses = asyncio.run(_run())
|
||||
notice = BUDGET_REPLY.format(steps=200)
|
||||
assert sent == [notice] # 只有说明,没有多余的「处理消息时出错」
|
||||
assert statuses == [RunStatus.FAILED]
|
||||
# 失败轮也要落库:用户已经看到这段内容,历史缺了它会前后不接
|
||||
rows = asyncio.run(repo.list_messages(tmp_path / "bot.db", session_id="s1"))
|
||||
assert [(r["role"], r["content"]) for r in rows] == [
|
||||
("user", "把整站重构一遍"),
|
||||
("assistant", notice),
|
||||
]
|
||||
@@ -8,7 +8,7 @@ from typing import Any
|
||||
|
||||
from bot042.channels.base import BaseChannel
|
||||
from bot042.channels.models import ChannelSubject, InboundMessage, TextContent
|
||||
from bot042.config import Settings
|
||||
from bot042.config import STEP_GRACE, Settings
|
||||
from bot042.flows.agent_task import last_message_text, run_task
|
||||
from bot042.flows.session import run_session
|
||||
|
||||
@@ -86,7 +86,8 @@ def test_run_task_uses_async_ainvoke(monkeypatch) -> None:
|
||||
assert last_message_text(result) == "done"
|
||||
state, config = fake.calls[0]
|
||||
assert state["messages"][-1] == {"role": "user", "content": "hi"}
|
||||
assert config == {"recursion_limit": 7}
|
||||
# 硬步数上限 = 软预算 + 收尾余量(见 Settings.recursion_limit)
|
||||
assert config == {"recursion_limit": 7 + STEP_GRACE}
|
||||
|
||||
|
||||
def test_last_message_text_multimodal_content() -> None:
|
||||
|
||||
+66
-2
@@ -9,6 +9,7 @@ from typing import Any, ClassVar
|
||||
|
||||
import pytest
|
||||
|
||||
from bot042.agent.middlewares.budget import StepBudgetExceeded
|
||||
from bot042.channels.base import BaseChannel
|
||||
from bot042.channels.manager import ChannelManager
|
||||
from bot042.channels.models import (
|
||||
@@ -21,12 +22,13 @@ from bot042.channels.models import (
|
||||
)
|
||||
from bot042.channels.registry import register_channel
|
||||
from bot042.config import Settings
|
||||
from bot042.db import repo
|
||||
from bot042.flows import builtin
|
||||
from bot042.flows.builtin import RPC_STREAM, AgentReply, ChannelServices
|
||||
from bot042.flows.builtin import BUDGET_REPLY, RPC_STREAM, AgentReply, ChannelServices
|
||||
from bot042.flows.bus import EventBus
|
||||
from bot042.flows.rpc import RpcHub
|
||||
from bot042.flows.stream import ReplyStream
|
||||
from bot042.flows.tasks import EventTrigger, TaskRunner
|
||||
from bot042.flows.tasks import EventTrigger, RunStatus, TaskRunner
|
||||
|
||||
|
||||
def test_stream_sends_ack_and_throttles() -> None:
|
||||
@@ -220,6 +222,68 @@ def test_agent_reply_acks_then_streams(
|
||||
assert frames[-1] == ("你好", True)
|
||||
|
||||
|
||||
def test_agent_reply_streams_budget_notice_after_partial(
|
||||
monkeypatch: pytest.MonkeyPatch, tmp_path: Path
|
||||
) -> None:
|
||||
"""步数用尽:说明补在已流出的内容后面,落库,且不再补发通用错误文案。"""
|
||||
StreamChannel.instances.clear()
|
||||
|
||||
async def exhausted(task, settings=None, history=None, *, cache=None, steering=None):
|
||||
yield "做到一半"
|
||||
raise StepBudgetExceeded(200)
|
||||
|
||||
monkeypatch.setattr(builtin, "stream_task", exhausted)
|
||||
|
||||
async def _run() -> tuple[list[tuple[str, bool]], list[RunStatus]]:
|
||||
bus = EventBus()
|
||||
bus.bind_loop()
|
||||
hub = RpcHub(bus)
|
||||
manager = ChannelManager()
|
||||
channel = StreamChannel(channel_id="s1")
|
||||
await manager.add_channel(channel)
|
||||
await manager.start()
|
||||
_ = hub.serve(RPC_STREAM, manager.rpc_stream)
|
||||
|
||||
services = ChannelServices(
|
||||
channels=manager,
|
||||
settings=Settings(),
|
||||
db=tmp_path / "bot.db",
|
||||
hub=hub,
|
||||
)
|
||||
runner = TaskRunner(bus, deps=services)
|
||||
runner.register(AgentReply)
|
||||
_ = runner.add_task(AgentReply.name, EventTrigger(event="channel.inbound"))
|
||||
await runner.start()
|
||||
|
||||
bus.emit(
|
||||
"channel.inbound",
|
||||
ChannelMessage.of(
|
||||
InboundMessage(
|
||||
channel_id="s1",
|
||||
channel_type="stream-test",
|
||||
channel_subject=ChannelSubject(subject_id="u1"),
|
||||
content=[TextContent(text="把整站重构一遍")],
|
||||
)
|
||||
),
|
||||
)
|
||||
await asyncio.sleep(0.05)
|
||||
await runner.wait_idle()
|
||||
statuses = [run.status for run in runner.runs(AgentReply.name)]
|
||||
await runner.stop()
|
||||
hub.close()
|
||||
await manager.stop()
|
||||
return channel.frames, statuses
|
||||
|
||||
frames, statuses = asyncio.run(_run())
|
||||
final = "做到一半\n\n" + BUDGET_REPLY.format(steps=200)
|
||||
assert frames[0] == (builtin.ACK_TEXT, False)
|
||||
assert frames[-1] == (final, True) # 收尾帧保留了部分内容 + 说明
|
||||
assert statuses == [RunStatus.FAILED]
|
||||
rows = asyncio.run(repo.list_messages(tmp_path / "bot.db", session_id="s1:u1"))
|
||||
assert rows[-1]["role"] == "assistant"
|
||||
assert rows[-1]["content"] == final
|
||||
|
||||
|
||||
def test_outbound_stream_payload_defaults() -> None:
|
||||
payload = OutboundStream(
|
||||
channel_id="c", subject_id="u", stream_id="s", text="x"
|
||||
|
||||
Reference in new issue
Block a user