AutoGen Integration

TheseusAutoGenMiddleware hooks into AutoGen's ConversableAgent.register_reply() to validate incoming messages and log outgoing ones.

Requirements: pyautogen >= 0.4

pip install "ct-toolkit[autogen]"

Quick start

from ct_toolkit import TheseusWrapper
from ct_toolkit.middleware.autogen import TheseusAutoGenMiddleware
from autogen import ConversableAgent

wrapper = TheseusWrapper(provider="openai", kernel_name="defense")
assistant = ConversableAgent("assistant", llm_config={...})

TheseusAutoGenMiddleware.apply_to_agent(assistant, wrapper)

What it does

• Incoming: Validates every message against the kernel before the agent processes it. Axiomatic violations return a [CT Toolkit BLOCKED] response.
• Outgoing: Runs divergence analysis on every generated reply and records it in the Provenance Log. Also supports Passive Compression Guard to monitor if the conversation history is silently summarized by the LLM provider.

Inject headers for HTTP sub-agents

config_list = TheseusAutoGenMiddleware.wrap_config_list(config_list, wrapper)

Injects X-CT-Kernel and X-CT-Parent-Provider headers into every model config entry.