AutoGen adapter for the Python SDK

[prashar32]

Adds an AutoGen adapter so an AutoGen agent (or team) can be bound to a governed run with no code change beyond wrapping the model client.

What it does

GovernedChatCompletionClient wraps any AutoGen v0.4+ ChatCompletionClient (the autogen_core.models protocol that autogen-agentchat agents use) and ticks one governed step per model request before delegating to the real client — one model call == one governed step. So a runaway agent is hard-stopped at its loop/time budget. With gate_tools=True, each tool call the model requests (a FunctionCall in the returned CreateResult) routes through the human-approval gate before the agent can run it. Every other protocol method delegates to the wrapped client, so it's a drop-in:

from riskkernel.adapters.autogen import GovernedChatCompletionClient
from autogen_agentchat.agents import AssistantAgent

client = GovernedChatCompletionClient(model_client, run)   # the one integration line
agent = AssistantAgent("assistant", model_client=client)   # otherwise unchanged
await agent.run(task="...")                                 # raises rk.BudgetExceeded at the cap

I picked the model-client wrapper as the hook because it's the cleanest no-app-change surface that meets all three needs at once: it ticks exactly one step per model request, it sees tool-call requests in the result for gating, and AutoGen does not catch model-client exceptions inside a single agent — so a BudgetExceeded raised there propagates out and halts the run.

Propagation (the asymmetry, verified against the real library)

• A single agent run directly (agent.run() / agent.on_messages()) does not wrap model-client exceptions — the typed BudgetExceeded propagates out unwrapped.
• Inside a team (RoundRobinGroupChat, SelectorGroupChat, …) the agent's container catches the exception, serializes it (SerializableException), and the team re-raises it to the caller as a plain RuntimeError(str(error)) whose first line reads "BudgetExceeded: run halted: loop_budget_exceeded". The run still halts — it is not swallowed — but the type is lost. governed_run_errors() (a context manager in the same module) wraps a team call and restores the typed BudgetExceeded / ApprovalDenied so callers can except on it as everywhere else.

Both paths were exercised against a real autogen-agentchat / autogen-core 0.7.5 install: the single-agent integration test halts a real runaway AssistantAgent with BudgetExceeded unwrapped, and the team path was confirmed to re-raise the wrapped RuntimeError that governed_run_errors() then converts back.

Supported versions

Targets the actively maintained v0.4+ line: autogen-agentchat / autogen-core >= 0.4, < 1. Not the legacy pyautogen 0.2 register_reply API. autogen is not a dependency of the SDK — the wrapper is duck-typed and imports nothing from autogen, so the SDK installs and imports fine without it. Added an [autogen] optional extra mirroring the existing [langchain] / [crewai] extras.

Included

• riskkernel/adapters/autogen.py — the wrapper + governed_run_errors() helper.
• tests/test_autogen.py — 20 stdlib-only unit tests (step ticking on create/create_stream, loop-budget halt surfaced-not-swallowed, tool gating off/on/denied, protocol delegation, the team-error unwrap) + a skip-unless-installed integration test that halts a real runaway agent.
• examples/autogen/ — a key-free runnable demo (a local looping model client) that a budget halts, with a README.
• README, CHANGELOG, pyproject.toml updates.

Full suite: python3 -m unittest discover -s tests -t . → all green (integration test skipped unless autogen is installed; passes when it is).

Closes #84