Welcome

Convalesce watches the parts of a data platform that usually go unwatched, and tells you when one of them breaks. When it can work out why, it proposes a fix and sends it to you as a pull request.

What it reads

We read metadata all the time: table shapes, run outcomes, row counts, lineage.

Once something has failed, an investigation may read a little from a connection that holds data, to confirm a cause. Each read is small and capped at 200 rows, personal columns are masked, nothing read is stored, and what comes back is sent to the AI model.

This is on by default. Each connection has its own switch: open the connection in Integrations, go to its Settings tab, and find Reading its data. The switch reads The agent may read from this connection. Turn it off for data you would not share with the AI model.

Column profiling is separate. It is off until you turn it on, per table.

Three things to do first

1. Connect a tool. Start with whatever orchestrates your work: Airflow, or dbt. A tool we read from asks for its credentials; one that sends to us, like Airflow, gets an ingest key for the deployment it runs in.
2. Check we can reach it. Nothing is wrong until something has been read. Press Test the connection. A tool we read from is tried there and then; for one that sends to us, start a run first and it looks for what arrived.
3. Connect GitHub. This is how a proposed fix reaches you. Until it is connected we can find the cause of a failure but not offer you a change to the code.

What happens after that

We build a picture of how your platform fits together, and watch it. When something breaks, or quietly stops, an agent reads the failure, checks what changed, works back through what depends on it, and either explains what it found or tells you it could not.

If it can propose a change, you get a pull request with the evidence attached. Nothing on your systems changes until somebody merges it.