Microloop Python SDK

Python bindings for the Microloop local reliability runtime.

from microloop import InterventionAction, Monitor, Policy

policy = Policy(
    stalled=InterventionAction.Replan,
    regressing=InterventionAction.Stop,
    cooldown_steps=5,
)
monitor = Monitor(policy=policy)

for step in agent.steps():
    decision = monitor.observe(
        action=step.action,
        observation=step.result,
        state=step.state,
        metrics={"exit_code": step.exit_code},
    )
    if decision.should_intervene:
        agent.inject(decision.recovery_context)

Monitor.observe(...) returns a Decision with status, reasons, intervention, severity, verified_progress and feedback. The default policy only observes; automatic recovery requires explicit opt-in.

You own the agent loop. Monitor classifies and advises; it never runs tools, calls a model, or stops a process.

CLI

microloop inspect trajectory.jsonl   # detection only, default observe-only policy
microloop replay trajectory.jsonl --json  # re-run recorded events through the engine
microloop monitor trajectory.jsonl --follow
microloop doctor

replay re-runs recorded events through the current engine. It does not reproduce the original agent execution. A trajectory with an incompatible schema_version major version is rejected rather than analyzed.

Building

pip install maturin
maturin develop --manifest-path python/microloop/Cargo.toml

No external credentials or network connections are required. All trajectory evaluation runs locally in Rust via PyO3.