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

2
app-instance/backend/beaver/engine/context/__init__.py
View File

@ -4,6 +4,7 @@ from .builder import (
ContextBuildInput,
ContextBuildResult,
ContextBuilder,
RuntimeContext,
SessionContext,
SkillContext,
)
@ -12,6 +13,7 @@ __all__ = [
"ContextBuildInput",
"ContextBuildResult",
"ContextBuilder",
"RuntimeContext",
"SessionContext",
"SkillContext",
]

47
app-instance/backend/beaver/engine/context/builder.py
View File

@ -80,6 +80,16 @@ class SessionContext:
parent_session_id: str | None = None
@dataclass(slots=True)
class RuntimeContext:
"""Per-run runtime facts that should be visible to the model."""
utc_datetime: str
local_datetime: str
timezone: str | None = None
utc_offset: str | None = None
@dataclass(slots=True)
class ContextBuildInput:
"""一次上下文构建所需的全部输入。
@ -103,6 +113,7 @@ class ContextBuildInput:
memory_snapshot: MemorySnapshot | None = None
activated_skills: list[SkillContext] = field(default_factory=list)
session_context: SessionContext | None = None
runtime_context: RuntimeContext | None = None
execution_context: str | None = None
extra_sections: list[str] = field(default_factory=list)
@ -143,9 +154,10 @@ class ContextBuilder:
1. Beaver user-facing assistant identity
2. base system prompt
3. session metadata
4. execution context
5. frozen memory snapshot
6. extra sections
4. runtime date/time
5. execution context
6. frozen memory snapshot
7. extra sections
这样设计的原因：
- 身份与总规则要最靠前
@ -164,6 +176,10 @@ class ContextBuilder:
if session_section:
sections.append(session_section)
runtime_section = self._render_runtime_section(build_input.runtime_context)
if runtime_section:
sections.append(runtime_section)
execution_context = (build_input.execution_context or "").strip()
if execution_context:
sections.append(f"# Execution Context\n\n{execution_context}")
@ -347,6 +363,31 @@ class ContextBuilder:
return None
return "# Current Session\n\n" + "\n".join(rows)
def _render_runtime_section(self, runtime_context: RuntimeContext | None) -> str | None:
"""Render date/time facts captured for the current model run."""
if runtime_context is None:
return None
rows: list[str] = []
if runtime_context.utc_datetime:
rows.append(f"Current UTC time: {runtime_context.utc_datetime}")
if runtime_context.local_datetime:
rows.append(f"Current local time: {runtime_context.local_datetime}")
if runtime_context.timezone:
rows.append(f"Local timezone: {runtime_context.timezone}")
if runtime_context.utc_offset:
rows.append(f"Local UTC offset: {runtime_context.utc_offset}")
if not rows:
return None
return (
"# Current Date and Time\n\n"
+ "\n".join(rows)
+ "\n\nUse this section as authoritative for relative date/time references such as "
'"today", "tomorrow", "now", "this week", and "next month".'
)
def build_skill_activation_messages(self, activated_skills: list[SkillContext]) -> list[dict[str, str]]:
"""把已激活 skill 转成显式消息。

7
app-instance/backend/beaver/engine/loader.py
View File

@ -24,7 +24,7 @@ from beaver.skills.learning.eval import SkillDraftEvaluator
from beaver.skills.publisher import SkillPublisher
from beaver.skills.reviews import ReviewService
from beaver.skills.specs import SkillSpecStore
from beaver.tasks import TaskExecutionPlanner, TaskService, ValidationService
from beaver.tasks import TaskExecutionPlanner, TaskService
from beaver.tasks.skill_resolver import TaskSkillResolver
from beaver.skills import SkillAssembler, SkillsLoader
from beaver.tools import ObjectBackedTool, ToolAssembler, ToolExecutor, ToolRegistry
@ -91,7 +91,6 @@ class EngineLoadResult:
task_skill_resolver: TaskSkillResolver | None = None
task_service: TaskService | None = None
task_execution_planner: TaskExecutionPlanner | None = None
validation_service: ValidationService | None = None
mcp_manager: MCPConnectionManager | None = None
mcp_report: dict[str, dict] = field(default_factory=dict)
closeables: list[tuple[str, Callable[[], None]]] = field(default_factory=list, repr=False)
@ -166,7 +165,6 @@ class EngineLoader:
task_skill_resolver: TaskSkillResolver | None = None,
task_service: TaskService | None = None,
task_execution_planner: TaskExecutionPlanner | None = None,
validation_service: ValidationService | None = None,
) -> None:
self.config = config or load_config(workspace=workspace, config_path=config_path)
configured_workspace = self.config.agents_defaults.workspace
@ -192,7 +190,6 @@ class EngineLoader:
self._task_skill_resolver = task_skill_resolver
self._task_service = task_service
self._task_execution_planner = task_execution_planner
self._validation_service = validation_service
def load(self) -> EngineLoadResult:
"""装配当前主链需要的最小 runtime 对象。"""
@ -276,7 +273,6 @@ class EngineLoader:
)
task_service = self._task_service or TaskService(workspace / "tasks")
task_execution_planner = self._task_execution_planner or TaskExecutionPlanner(task_skill_resolver=task_skill_resolver)
validation_service = self._validation_service or ValidationService()
mcp_manager = MCPConnectionManager(
self.config.tools.mcp_servers,
authz_config=self.config.authz,
@ -311,7 +307,6 @@ class EngineLoader:
task_skill_resolver=task_skill_resolver,
task_service=task_service,
task_execution_planner=task_execution_planner,
validation_service=validation_service,
mcp_manager=mcp_manager,
)
if self._session_manager is None:

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:]}"

10
app-instance/backend/beaver/engine/providers/litellm.py
View File

@ -119,13 +119,23 @@ class LiteLLMProvider(LLMProvider):
@staticmethod
def _sanitize_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
sanitized = []
system_contents: list[str] = []
for message in messages:
clean = {key: value for key, value in message.items() if key in _ALLOWED_MSG_KEYS}
if clean.get("role") == "system":
content = clean.get("content")
if isinstance(content, str) and content.strip():
system_contents.append(content.strip())
elif content is not None:
system_contents.append(str(content))
continue
if clean.get("role") == "assistant" and "content" not in clean:
clean["content"] = None
if isinstance(clean.get("tool_calls"), list):
clean["tool_calls"] = LiteLLMProvider._sanitize_tool_calls(clean["tool_calls"])
sanitized.append(clean)
if system_contents:
sanitized.insert(0, {"role": "system", "content": "\n\n".join(system_contents)})
return sanitized
@staticmethod

6
app-instance/backend/beaver/engine/session/models.py
View File

@ -84,8 +84,10 @@ class MessageRecord:
payload["task_id"] = self.event_payload.get("task_id")
if self.event_payload.get("task_status"):
payload["task_status"] = self.event_payload.get("task_status")
if self.event_payload.get("validation_status"):
payload["validation_status"] = self.event_payload.get("validation_status")
if self.event_payload.get("evidence_status"):
payload["evidence_status"] = self.event_payload.get("evidence_status")
if self.event_payload.get("acceptance_state"):
payload["acceptance_state"] = self.event_payload.get("acceptance_state")
if self.event_payload.get("feedback_state"):
payload["feedback_state"] = self.event_payload.get("feedback_state")
if self.event_payload.get("feedback_error"):