Why it exists

A dashboard that stops receiving data and doesn't say so is worse than having no dashboard at all: it still looks fine, and people keep deciding based on it. This is the one that says so.

The whole suite runs on resident processes that watch folders, clean up files, and push them into the integration layer. They're invisible while they work and catastrophic when they fail silently. This dashboard puts all of them on one screen with a readable status:

• Up to date, attention, stuck, late. Based on how long something has gone unprocessed and how far it strays from its usual cadence.
• Off. A hard fact from the agent — the process isn't alive —, not an inference from lack of activity.
• Learning, log-only, dormant. The cases that aren't failures: not enough runs yet to measure cadence, a process that never archives anything and whose log is its only sign of life, or something that's been quiet for over a month because it's legacy.

Where the data comes from

A Python agent walks the machine where the watchers live and reports facts, not judgments: how many files are waiting and since when, when it last archived something, when it last wrote to its log, what cadence was measured for it and with how many samples, whether the process is alive, and what its last error was.

The traffic-light status isn't decided by the agent: it's decided by the dashboard. That's a design rule of the system, not an aesthetic choice — it means adjusting a threshold is a one-line change in the dashboard, not logging into a production machine to touch a process that's running.

Build decisions

• Cadence is measured, not declared. The agent infers it from actual runs; until it has enough samples the status is «learning», not «late», so it doesn't cry wolf over a process that just came online.
• Watchers with a fixed schedule are measured against the clock, not their average cadence: with uneven runs, the average lies.
• Every watcher states which dashboard it feeds. The status alone doesn't trigger anything; knowing who's left without data does.
• Watchers are standardized on a common template and get configured, not rewritten. Adding a new one is a config file, and it shows up here on its own.

What's different in this demo

Same production file. The fixture delivers the raw facts just like the real agent does, and the dashboard still calculates the traffic-light status: changing one last-processed date is enough for it to recompute the status on its own, which is exactly the property this is meant to show.

• The 14 watchers deliberately cover every status the dashboard knows how to render, including the rare ones.
• Host, process, and company names are fictional.