Know before you swap

The doctor scans your codebase and reports exactly which Prisma constructs the compat layer doesn't translate — file and line, before you change anything. Ten seconds of scanning replaces a week of surprises.

#Run it

SHELLterminal

5 lines


1zanith compat scan ./src
2 
3# machine-readable, and a CI gate that fails on unsupported findings:
4zanith compat scan ./src --json
5zanith compat scan ./src --strict

The scan walks your tree (skipping node_modules, dist, build output), counts every Prisma model-method call site, and matches the remainder against the compat layer's refusal vocabulary. Findings carry two severities:

Severity	Meaning
unsupported	Will throw ZanithCompatError at runtime — a named, deliberate refusal.
review	Heuristic match that usually means an untranslated construct, but wants a human glance (could be a same-named field).

TXTsample output

17 lines


1Scanned 412 file(s) under ./src
2Prisma call sites seen: 1,387 (the call surface itself is translated)
3 
4Findings: 2 unsupported (throw at runtime), 3 to review (heuristic match)
5 
6  ✗ $transaction isolationLevel — 1 site(s)
7      src/payments/ledger.ts:84  await prisma.$transaction(fn, { isolationLevel: 'Serializable' });
8 
9  ✗ JSON path filter — 1 site(s)
10      src/search/contracts.ts:31  where: { core_data: { path: ['health'], equals: 'red' } }
11 
12  ? nested write "set" (replace memberships) — 3 site(s)
13      src/teams/sync.ts:112  data: { members: { set: [] } }
14      … and 2 more
15 
16This is a heuristic text scan — "review" findings may be same-named fields.
17Runtime behavior is the contract: untranslated constructs throw ZanithCompatError, never approximate.

#The rule that keeps it honest

The doctor's marker list mirrors the runtime's refusal list, with a standing rule: when the compat layer learns a construct, its marker is deleted in the same commit. The doctor must never claim less than the runtime delivers. Three markers have already died this way — case-insensitive mode, array filters, and many-to-many writes all graduated from "flagged" to "translated".

#The adoption loop

The intended workflow, end to end:

TSfour steps, each reversible

10 lines


1# 1. Measure — zero changes, zero risk
2zanith compat scan ./src
3 
4# 2. If clean (most codebases are): swap db.ts on a branch
5#    new PrismaClient(...)  →  new ZanithPrismaClient({ adapter })
6 
7# 3. Run YOUR test suite. Every divergence is either a named
8#    ZanithCompatError or a bug we want to hear about.
9 
10# 4. Ship — or revert one file and walk away. Both are fine.

Findings that block you go on the roadmap as named line items — several constructs were built precisely because scans of real codebases said they'd be hit. A request that lands on an existing line item is a vote, not a discovery.