Beaver/beaver_project
6
0
Fork 0
Code Issues Pull Requests 1 Actions Packages Projects Releases Wiki Activity

feat(engine): 添加运行时上下文支持并重构工具迭代限制

Browse Source 
添加 RuntimeContext 类用于捕获模型运行时的日期时间信息，
包括UTC时间、本地时间和时区信息，并在系统提示中显示这些信息。

同时增加最大上下文消息数和工具迭代次数的配置选项，
将验证服务从引擎加载器中移除，并更新相关的数据结构和接口。

BREAKING CHANGE: 移除了验证服务，相关字段被替换为证据状态和接受状态。

- 添加 RuntimeContext 类和相关渲染方法
- 增加 max_context_messages 和 max_tool_iterations 配置
- 移除 ValidationService 相关代码
- 更新消息记录中的验证状态字段
- 添加原始工具调用检测和回退处理
This commit is contained in:
steven_li
2026-05-26 11:18:35 +08:00
parent 16347caf5e
commit 6e9e74d1ee
57 changed files with 5710 additions and 1582 deletions
Download Patch File Download Diff File Expand all files Collapse all files

128
app-instance/backend/beaver/engine/loop.py
View File

@ -4,12 +4,15 @@ from __future__ import annotations
import asyncio
import json
import os
import re
from dataclasses import dataclass, field
from datetime import datetime, timezone
from typing import Any
from uuid import uuid4
from zoneinfo import ZoneInfo, ZoneInfoNotFoundError
from beaver.engine.context import ContextBuildInput, SessionContext, SkillContext
from beaver.engine.context import ContextBuildInput, RuntimeContext, SessionContext, SkillContext
from beaver.memory.runs import RunRecord, SkillEffectRecord
from beaver.skills.learning import RunReceiptContext
from beaver.skills.catalog.utils import strip_frontmatter
@ -26,6 +29,17 @@ TOOL_FAILURE_GUIDANCE_PROMPT = (
"Use available materials, state uncertainty clearly, and provide partial confirmed results."
)
RAW_TOOL_CALL_FALLBACK = (
"The run reached the configured tool-call limit before producing a reliable final answer. "
"The model attempted another tool call instead of answering, so the raw tool call was suppressed. "
"Please request a revision to continue the task."
)
_RAW_TOOL_CALL_RE = re.compile(
r"^\s*<tool_call\b[\s\S]*?</tool_call>\s*$|^\s*<function=[^>]+>[\s\S]*?</function>\s*$",
re.IGNORECASE,
)
@dataclass(slots=True)
class AgentProfile:
@ -35,8 +49,9 @@ class AgentProfile:
system_prompt: str = ""
default_model: str = "gpt-4.1-mini"
max_tokens: int = 4096
max_context_messages: int = 1000
temperature: float = 0.2
max_tool_iterations: int = 8
max_tool_iterations: int = 30
@dataclass(slots=True)
@ -446,7 +461,7 @@ class AgentLoop:
*(pinned_skill_contexts or []),
*self._load_pinned_skill_contexts(skills_loader, pinned_skill_names or []),
]
if not include_skill_assembly or thinking_enabled is False:
if not include_skill_assembly:
activated_skills = self._merge_skill_contexts(pinned_skills, [])
else:
skill_query = skill_selection_context or task
@ -512,8 +527,6 @@ class AgentLoop:
if not include_tools:
selected_tool_specs = []
elif thinking_enabled is False:
selected_tool_specs = tool_registry.list_specs()
else:
selected_tool_specs = await tool_assembler.assemble(
task_description=task,
@ -543,7 +556,10 @@ class AgentLoop:
build_input = ContextBuildInput(
base_system_prompt=self.profile.system_prompt,
history=session_manager.get_history(resolved_session_id),
history=session_manager.get_history(
resolved_session_id,
max_messages=max(1, self.profile.max_context_messages),
),
current_user_input=task,
memory_snapshot=memory_snapshot,
activated_skills=activated_skills,
@ -554,6 +570,7 @@ class AgentLoop:
user_id=user_id,
parent_session_id=parent_session_id,
),
runtime_context=self._current_runtime_context(),
execution_context=execution_context,
extra_sections=[TOOL_FAILURE_GUIDANCE_PROMPT],
)
@ -693,6 +710,7 @@ class AgentLoop:
tool_calls=assistant_tool_calls or None,
finish_reason=response.finish_reason,
reasoning=response.reasoning_content,
context_visible=not bool(assistant_tool_calls),
source=source,
title=title,
model=final_model,
@ -707,7 +725,11 @@ class AgentLoop:
if not response.has_tool_calls:
final_text = response.content or ""
final_finish_reason = response.finish_reason or "stop"
if self._looks_like_raw_tool_call(final_text):
final_text = RAW_TOOL_CALL_FALLBACK
final_finish_reason = "invalid_tool_call_text"
else:
final_finish_reason = response.finish_reason or "stop"
break
if iterations >= resolved_max_tool_iterations:
@ -719,10 +741,7 @@ class AgentLoop:
temperature=resolved_temperature,
thinking_enabled=thinking_enabled,
)
final_text = finalized or (
"Tool loop stopped after reaching the configured iteration limit, "
"and no final answer was produced."
)
final_text = finalized or RAW_TOOL_CALL_FALLBACK
final_finish_reason = "max_tool_iterations_finalized" if finalized else "max_tool_iterations"
session_manager.append_message(
resolved_session_id,
@ -877,17 +896,14 @@ class AgentLoop:
temperature: float,
thinking_enabled: bool | None,
) -> str:
final_messages = [
*messages,
{
"role": "system",
"content": (
"The configured tool iteration budget is exhausted. Do not call tools. "
"Produce the best final answer from the existing conversation and tool results. "
"State uncertainty explicitly."
),
},
]
final_messages = AgentLoop._with_system_guidance(
messages,
(
"The configured tool iteration budget is exhausted. Do not call tools. "
"Produce the best final answer from the existing conversation and tool results. "
"State uncertainty explicitly."
),
)
kwargs: dict[str, Any] = {
"messages": final_messages,
"tools": None,
@ -898,7 +914,27 @@ class AgentLoop:
if thinking_enabled is not None:
kwargs["thinking_enabled"] = thinking_enabled
response = await provider.chat(**kwargs)
return (response.content or "").strip()
if response.has_tool_calls:
return ""
content = (response.content or "").strip()
if AgentLoop._looks_like_raw_tool_call(content):
return ""
return content
@staticmethod
def _looks_like_raw_tool_call(content: str | None) -> bool:
if not content:
return False
return bool(_RAW_TOOL_CALL_RE.match(content))
@staticmethod
def _with_system_guidance(messages: list[dict[str, Any]], guidance: str) -> list[dict[str, Any]]:
copied = [dict(message) for message in messages]
if copied and copied[0].get("role") == "system":
existing = str(copied[0].get("content") or "").strip()
copied[0]["content"] = "\n\n".join(part for part in (existing, guidance.strip()) if part)
return copied
return [{"role": "system", "content": guidance.strip()}, *copied]
@staticmethod
def _load_pinned_skill_contexts(skills_loader: Any, skill_names: list[str]) -> list[SkillContext]:
@ -1133,3 +1169,49 @@ class AgentLoop:
@staticmethod
def _utc_now() -> str:
return datetime.now(timezone.utc).isoformat()
@staticmethod
def _current_runtime_context() -> RuntimeContext:
utc_now = datetime.now(timezone.utc)
timezone_name = AgentLoop._configured_timezone_name()
local_now = datetime.now().astimezone()
rendered_timezone = local_now.tzname()
if timezone_name:
try:
local_now = utc_now.astimezone(ZoneInfo(timezone_name))
rendered_timezone = timezone_name
except ZoneInfoNotFoundError:
rendered_timezone = local_now.tzname() or timezone_name
return RuntimeContext(
utc_datetime=utc_now.isoformat(),
local_datetime=local_now.isoformat(),
timezone=rendered_timezone,
utc_offset=AgentLoop._format_utc_offset(local_now),
)
@staticmethod
def _configured_timezone_name() -> str | None:
for value in (os.getenv("BEAVER_RUNTIME_TIMEZONE"), os.getenv("TZ")):
cleaned = (value or "").strip()
if cleaned:
return cleaned
try:
timezone_file = "/etc/timezone"
if os.path.exists(timezone_file):
with open(timezone_file, encoding="utf-8") as file:
cleaned = file.read().strip()
if cleaned:
return cleaned
except OSError:
return None
return None
@staticmethod
def _format_utc_offset(value: datetime) -> str | None:
raw = value.strftime("%z")
if not raw:
return None
return f"{raw[:3]}:{raw[3:]}"