Multi-modal schema source

functional

Layer local files over the live database so the model can include the changes you are actively shaping.

Keep the live database as the baseline, then bring your checked-in DDL along for the ride. SQLly's model is built from more than one source at once, one or more connected databases, plus a folder of local SQL files - and merged into a single picture of your schema.

How the two sources are merged Collection runs in parallel: database sources through their connection pools, files through a bounded pool of reader threads. Each collected object becomes a fragment tagged with its provenance, either the profile it came from, or the file path, content hash, and byte span it was parsed from. Fragments are staged against a generation counter, joined at a barrier, and swapped in as one immutable snapshot, so you never see a half-built model. When both sources define the same object, the rule is deliberate and simple:

The file wins; the database is kept as a shadow. Your local definition is what completion and validation use, and the server's version is retained alongside it rather than discarded - so the difference is recorded rather than lost. It is not a “newest timestamp wins” guess, because a file's mtime is a poor witness and a silent coin-flip is worse than a stated rule.

Drift is a first-class answer

Because the losing definition is kept, the model can tell you exactly which objects differ between your working tree and the live server. Every object with a shadow is drift, listed by name, which turns “is staging up to date?” into a question with an answer instead of a hunch. It stays live File changes, file deletions, and database changes each notify the engine, and only the affected sources re-collect. An object removed and re-added in the same generation, which is what saving an edited file looks like from the outside, is reported as a change , not as a delete followed by an add. Work against checked-in DDL before it is deployed . Get completion for a table that only exists in your branch so far. The query can still fail against the live database until the schema catches up. That distinction stays explicit; the overlay is helpful, not make-believe. Rough, not-yet-valid DDL is handled too, see
Incomplete-model IntelliSense . Snapshots are cached locally per connection and database, so a reconnect does not start from nothing.