Put safety & cost checks around your agent.

Create an environment and install the wheel

curl -fLO https://github.com/runlog-cc/loop-guard/releases/download/v0.3.1/runlog_loopguard-0.3.1-py3-none-any.whl
python3 -m venv .venv
.venv/bin/python -m pip install ./runlog_loopguard-0.3.1-py3-none-any.whl

Save as guard.py

from loopguard import CircuitBreaker, CircuitTrippedError, TripPolicy

guard = CircuitBreaker(policy=TripPolicy(max_consecutive_retries=3))
for attempt in range(3):
    try:
        guard.before_tool_call("lookup", {"id": 7})
    except CircuitTrippedError:
        print("Stopped repeated lookup.")
        break
    print("Lookup ran.")
    guard.after_tool_call("lookup", {"id": 7}, result="ok")

Run it

.venv/bin/python guard.py

Install the compiled package in your project

curl -fLO https://github.com/runlog-cc/loop-guard/releases/download/v0.3.1/runlog-loopguard-0.3.1.tgz
npm install ./runlog-loopguard-0.3.1.tgz

The download includes CommonJS JavaScript and TypeScript declarations; no build step is needed to use it.

Save as guard.cjs

const { CircuitBreaker, CircuitTrippedError } = require("@runlog/loopguard");
const guard = new CircuitBreaker({ policy: { maxConsecutiveRetries: 3 } });

for (const attempt of [1, 2, 3]) {
  try {
    guard.beforeToolCall("lookup", { id: 7 });
  } catch (error) {
    if (!(error instanceof CircuitTrippedError)) throw error;
    console.log("Stopped repeated lookup.");
    break;
  }
  console.log("Lookup ran.");
  guard.afterToolCall({ toolName: "lookup", args: { id: 7 }, result: "ok" });
}

Run it

node guard.cjs

Put your version 1 scenario suite in scenarios.json and its recorded tool trace in candidate.json. The release README defines their exact shapes. This command calls the local evaluator directly:

python3 -c 'from runner import evaluate, load_json; print(evaluate(load_json("scenarios.json"), load_json("candidate.json")))'

Passing prints a dictionary of scenario, step, and cost totals; invalid records raise an error. The files are read locally and no gate request is sent.

Supply RUNLOG_TOKEN securely in your environment, then run this from the extracted directory with your own evaluated files:

python3 runner.py --scenarios scenarios.json --trace candidate.json

This command sends the trace, scenario assertions, and token to the configured HTTPS gate. This low-level runner requests a trace-policy decision; use the evidence bridge in the README for signature verification. The default is https://runlog.cc/api/v1/gate. It needs network access and a valid RunLog API key; local checks alone do not grant release approval. The signed decision scope is trace-loop-policy, not independent evaluation of answer quality. The optional evidence bridge verifies the gate signature and exact artifact and commit bindings.