Xi Zhang
7ccfe68f3f
* feat: add scheduler functionality with cron-style task management - Implemented a new scheduler subagent to automate recurring tasks using cron expressions. - Enhanced the subagent factory to include the skill manager and auxiliary chat model for the scheduler. - Created a YAML configuration for the scheduler with a detailed system prompt and toolset. - Updated README files to include documentation on scheduled tasks and usage examples. - Added tests for the scheduler, including command execution, scheduling tools, and middleware integration. - Introduced new dependencies for timezone handling and ensured compatibility in the project configuration. * fix(async-notifier): ensure fallback hint is used for unknown notification kinds * feat: enhance scheduling functionality and improve system message handling
63 lines
2.4 KiB
Python
63 lines
2.4 KiB
Python
"""Shared utilities for EvoScientist middleware.
|
|
|
|
Functions here are used by multiple middleware modules (memory, tool_selector)
|
|
and should not depend on any specific middleware class.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
from typing import Any
|
|
|
|
from langchain_core.language_models import BaseChatModel
|
|
from langchain_core.messages import SystemMessage
|
|
|
|
|
|
def disable_thinking(model: BaseChatModel) -> BaseChatModel:
|
|
"""Return a copy of the model with thinking/reasoning disabled.
|
|
|
|
Anthropic's API does not allow extended thinking when ``tool_choice``
|
|
forces tool use (as ``with_structured_output`` does). Similarly,
|
|
OpenAI reasoning can conflict. Strip these settings so structured
|
|
output calls work reliably.
|
|
|
|
Uses ``model_copy()`` to produce a real new instance — ``bind()`` only
|
|
wraps the model in a ``RunnableBinding`` whose kwargs do NOT override
|
|
first-class Pydantic fields like ``thinking`` on ``ChatAnthropic``.
|
|
"""
|
|
updates: dict[str, Any] = {}
|
|
model_kwargs = getattr(model, "model_kwargs", {}) or {}
|
|
|
|
if getattr(model, "thinking", None) or "thinking" in model_kwargs:
|
|
updates["thinking"] = None
|
|
if getattr(model, "reasoning", None) or "reasoning" in model_kwargs:
|
|
updates["reasoning"] = None
|
|
|
|
if not updates:
|
|
return model
|
|
|
|
# Prefer Pydantic model_copy (creates a true new instance with the
|
|
# field cleared) over bind() which only adds invocation kwargs.
|
|
try:
|
|
return model.model_copy(update=updates)
|
|
except Exception:
|
|
# Fallback for non-Pydantic or unusual model classes
|
|
# Note: bind() may not effectively override first-class Pydantic fields
|
|
return model.bind(**updates)
|
|
|
|
|
|
def append_to_system_message(
|
|
system_message: SystemMessage | None, text: str
|
|
) -> SystemMessage:
|
|
"""Append a text block to a system message, preserving its metadata.
|
|
|
|
Used by the memory and scheduler middleware. Unlike building a fresh
|
|
``SystemMessage``, ``model_copy`` keeps ``additional_kwargs`` (e.g.
|
|
``cache_control`` prompt-cache breakpoints), ``id``, ``name`` and
|
|
``response_metadata`` from the original message.
|
|
"""
|
|
existing_blocks = list(system_message.content_blocks) if system_message else []
|
|
new_blocks = [*existing_blocks, {"type": "text", "text": text}]
|
|
if system_message is None:
|
|
return SystemMessage(content=new_blocks)
|
|
return system_message.model_copy(update={"content": new_blocks})
|