Production monitoring

The earlier monitoring design uses saved tests. Scheduled execution is currently blocked by a safety stop.

A code-level safety stop blocks scheduled monitoring. Configuration flags cannot remove it. This reference describes the earlier Job workflow. It does not start website mapping or operate on mapping tests. Use the dashboard for that workflow.

One dispatcher, durable sweeps

One Vercel Cron request finds each due schedule and starts a Workflow. The Workflow groups the active tests and sends them to an isolated Playwright worker. The worker runs Chromium, returns results, and does not call AI.

Cron only starts the work

Proof stores schedules in its database. Do not add a Vercel Cron definition for each customer.

Create a monitor

bash

curl "$PROOF_PUBLIC_URL/api/monitors" \
  --request POST \
  --header "Authorization: Bearer $PROOF_API_KEY" \
  --header "Content-Type: application/json" \
  --data '{
    "projectId": "project_storefront",
    "name": "daily production",
    "cadenceMinutes": 1440,
    "maxTests": 50,
    "maxParallelism": 4
  }'

The API key determines which organization owns the monitor. The monitor uses the Project test URL and test settings.

The default schedule runs daily

The repository schedules the dispatcher once per day. The default minimum cadence is 1,440 minutes. A deployment operator must change both the dispatcher schedule and minimum cadence to support more frequent runs. The active deployment may use different settings.

Retry before escalation

Proof repeats every failure once, including a worker timeout. If the failure repeats, it records logs, screenshots, and paths. It then sends the result to the optional signed alert webhook. Proof does not use AI unless both AI settings are enabled.

Hard spend boundary

Each test reserves two cents before it runs. That covers one worker task and one retry. The shared spending guard combines reported AI Gateway spend with active reservations. Its default limit is $10, and a deployment can set a ceiling up to $50. A platform admin can save a shared admission limit. The lower of both limits applies. Proof blocks work when a new reservation would exceed that limit.

Provider billing is separate

These reservations are estimates. The guard does not measure every provider charge or establish the deployment’s billing plan. Deployment operators must also configure and check provider spending limits.