I actually took a lot of inspiration from sqlx, which is really nice. The main differences are:
- in JS/TS you don't have compile-time scripts that you can run like with Rust's macros, so you need to run a codegen command before running the type checks (disadvantage)
- I had to create a TS parser that goes and finds the tagged template functions with the sql statements, while sqlx has them "for free" because sql statements are the input to the macro itself (disadvantage)
- I use an in-memory Postgres (PGLite) to describe the queries, instead of requiring a running pg instance (advantage)
- I don't cache the statements and codegen for now like sqlx does, something that can be added later
I think they are similar in that they both substitute the dynamic params with no-ops like $1, $2, etc. before handing the sql statement to the pg's DESCRIBE function
I was targeting Bun because I really like its built-in SQL module. I can tweak the TS parser to look for e.g. postgres.js tagged template functions and make it work for that as well. I don't really see any blockers
I write Bun.sql with raw SQL and no ORM, and the one thing I kept missing was types. You write a query, get back `any[]`, and hand-write a row type that silently drifts from the actual columns. Drizzle/Kysely fix this by moving the query into TypeScript, but then you're not really writing SQL anymore.
bun-sqlgen goes the other way. You keep writing raw SQL queries, just give each one a name.
A codegen step reads your migration `.sql` files, stands up a throwaway Postgres via PGlite (so no Docker) or SQLite, prepares every tagged query against it, and writes a `.d.ts` that maps each query name to its real result type. After that, plain `tsc` does the rest: `user.notExistingField` won't compile, and `display_name.length` gets flagged because the column is nullable.
Nullability was the annoying part. Postgres's describe doesn't hand you per-column nullability, so I infer it from the query plan plus the catalog, with manual overrides for the cases that genuinely can't be inferred. SQLite works too.
The runtime stays 100% Bun.sql, the generated file is the only artifact (commit it), and codegen is fast enough to rerun on save.
It's early (v0.1, built it for my own projects) so I'd mostly like to hear where it falls over.
In the last couple of companies I've worked in, I've felt both overwhelmed by PRs to review and disappointed by my teammates that were just rubber-stamping my PRs
That's a good framework and I like that the creator took the same design choices as the FastAPI framework.
Although I think a type-safe language is a better fit for CLIs, to let coding agents iterate faster and have less runtime bugs (help messages not updated, unknown flags, etc.)
Great idea! Are you also considering fine-tuning models as a "side-effect"? Basically, you collect traces of the agent's mistakes from your proxy, and you fine-tune a smaller model that you can offer specifically for that agent as a cheaper alternative?
It produces an arithmetic program but with wrong operands. The frozen LLM's hidden states for "two" and "2" are nearly orthogonal (cosine sim 0.09) in this context, so the head can't extract the right numbers. "2 plus 3" works fine and draws 5. The model understands the task structure but can't bridge word-form to digit-form without token generation