Custom integration

The reliable integration boundary is an explicit argv. It works with any framework that has a CLI and keeps AgentDiff independent from agent self-reporting.

CLI boundary

agentdiff run --task "Fix authentication" -- your-agent --project .

For automation, add --format json and choose --fail-on never|review|deny explicitly.

Python boundary

from agentdiff import AgentRunTransaction, load_policy_file

transaction = AgentRunTransaction(
    root="/workspace/project",
    policy=load_policy_file("/workspace/project/agentdiff.yaml"),
    task="Fix authentication",
)
result = transaction.run(
    ["your-agent", "--project", "."],
    timeout_seconds=900,
)

print(result.to_dict())

This path produces the same capsule, policy decisions, score, and recovery evidence as the CLI.

Experimental compatibility session

AgentDiffSession remains available for frameworks that want to record trajectory steps and calculate the original cleanliness metric:

from agentdiff import AgentDiffConfig, AgentDiffSession

config = AgentDiffConfig(root="/workspace/project", target_paths=["src/auth.py"])
with AgentDiffSession("Fix authentication", config=config) as session:
    run_your_agent()
    session.record(
        thought="Update authentication",
        tool_name="write_file",
        tool_args={"path": "src/auth.py"},
        observation="updated",
    )

evaluation = session.evaluate()

The compatibility session is separate from the transaction/recovery core. There is no plugin entry-point discovery or built-in CrewAI/AutoGen adapter in 0.1.0.