Codex writes better fixes when it sees production. Praxen Inc. hands it the logs, traces and recent changes.

// Which deploy slowed checkout-edge?
const deploys = await tools.deployHistory({ service: "checkout-edge", hours: 24 });
const { points } = await tools.queryMetrics({
  service: "checkout-edge",
  metric: "latency.p99",
  hours: 24,
});

const median = (xs) => xs.sort((a, b) => a - b)[xs.length >> 1] ?? 0;
const p99 = (t, side) =>
  median(points.filter((p) => (side < 0 ? p.t < t : p.t > t)).map((p) => p.v));

// Score every deploy by the p99 shift across its timestamp.
const [worst] = deploys
  .map((d) => ({ ...d, shift: p99(d.at, 1) / p99(d.at, -1) }))
  .sort((a, b) => b.shift - a.shift);

// Pull the error logs right after the suspect deploy.
const errors = await tools.queryLogs({
  service: "checkout-edge",
  level: "error",
  after: worst.at,
  limit: 5,
});

return { deploy: worst.sha, shift: worst.shift, errors };

One call. The raw data stays in the sandbox and Codex reads the answer.