Build Agentic workflows, RAG pipelines, with rich AI model and tool support on one collaborative workspace. Deploy on cloud, VPC, or self-hosted, so teams move from prototype to production without rebuilding the stack.
Python agentagentic-aiagentic-frameworkagentic-workflowaiautomationclaudedeepseekgenaigptllmlow-codemcpnextjsno-codeopenaipythonskillsworkflow
"""Conversation persistence in JSON state; run instructions remain transient."""
from collections.abc import Sequence
from dataclasses import replace
from pydantic import BaseModel, ConfigDict, Field
from pydantic_ai.capabilities import AbstractCapability
from pydantic_ai.messages import ModelMessage, ModelRequest, ModelResponse
from dify_agent.runtime.context import Deps
class Config(BaseModel):
model_config = ConfigDict(extra="forbid")
class State(BaseModel):
messages: list[ModelMessage] = Field(default_factory=list)
model_config = ConfigDict(extra="forbid")
class Capability(AbstractCapability[Deps]):
def __init__(self, name: str):
self.id = name
self.name = name
def load_messages(self, deps: Deps) -> list[ModelMessage]:
return State.model_validate(deps.layers[self.name]["state"]).messages
def save_messages(self, deps: Deps, messages: Sequence[ModelMessage], *, interrupted: bool = False) -> None:
"""Save captured history, retaining the native interruption state.
Pydantic AI records interruption on the in-progress request when a complete
response's tool calls were not followed by a model request. Older captured
histories may instead leave the response itself complete, so mark that
trailing tool-call response interrupted for native history repair.
"""
persistent = [replace(m, instructions=None) if isinstance(m, ModelRequest) else m for m in messages]
if interrupted and persistent:
last = persistent[-1]
if isinstance(last, ModelResponse) and last.state == "complete" and last.tool_calls:
persistent[-1] = replace(last, state="interrupted")
deps.layers[self.name]["state"] = State(messages=persistent).model_dump(mode="json")