One page, three promises. The map below shows every shipped subsystem, every feature in flight, and the releases that come next — sliced two ways so you can read it by where in the engine or by when.
version
0.1.0
models in graph
0
ESM bundle
0.00 KB
indexed lookup
0.00ms
days · since first commit
—
anchored to source
schema.zanith
0102030405060708091011
AST
Hover to pause · click a stage to jump · this is the actual pipeline, not a sketch
Test breakdown
765 / 765 · 2.14s
Build artifact
88.97 KB · ESM
Type defs
35.44 KB · .d.ts
Coverage
94.2% lines
compiler: cache parameter index across joins
● head · ci passed
graph: precompute relation reverse-lookup map
test(scale): add 5k-model benchmark suite
adapter/pg: wire onQuery debug hook
expression: collapse nested AND blocks
03 — The map
Six lanes — the four engine layers plus tooling and site/docs. Filter by status, search anything, share the URL.
lex · parse · validate
Token stream → CST → AST in one pass. ~22ms / 1000 models.
Duplicate models, dangling relations, unknown types caught at parse.
Map AST nodes back to source positions for better error messages.
runtime structure
Models, fields, enums, relations, indexes — in memory.
0.73ms for 1000 lookups against a 1000-model graph.
Old graph vs new graph → set of structural changes. Powers the migration generator that shipped in v0.2.
generateTypes(graph) emits TypeScript .d.ts from the runtime graph. measureTypegenSize() reports byte cost. Optional — the typed client doesn't need it.
serializeSchemaGraph / deserializeSchemaGraph + the snapshots table. listSnapshots / findSnapshot read history programmatically.
AST → SQL
Every CRUD path compiles to parameterized SQL.
14 operators across 5 groups, all parameterized.
Eager-fetch into a single statement, no N+1. 72% complete.
Compiler primitives done; ModelAPI wiring next. 55%.
ROW_NUMBER / RANK / OVER. 40% — type-safe column ref pending.
EXISTS, IN, correlated subqueries on the relational builder. 30%.
Surface query plans through the typed client (Studio already has it).
union/intersect/except, lateralJoin, listen/notify/unlisten — all on the relational builder. Postgres-only paths throw NotImplementedError on SQLite.
search() composes ranked text, trigram, and vector queries in a single statement. tsvectorSetupSql() emits the column + trigger; facets() emits filter UIs.
Typed wrappers for pgvector (cosine/L2/inner-product/L1), pg_trgm (similarity/word_similarity), pgcrypto (digest/hmac/pgp), pg_cron (schedule/list/history).
caseWhen, coalesce, nullIf, fn, plus the full JSON / array / range operator family as standalone helpers.
wire driver
The default. Pool-aware. Battle-tested.
Lighter, faster startup. Same five-method interface.
File-backed and :memory: modes. Sync I/O for tests and edge contexts.
Five TLS modes plus bastion-tunneled Postgres. Same connection API.
Auto-rollback on throw, isolation levels, savepoints.
Dialect already shipped (MysqlDialect / mysqlDialect) — only the adapter is missing. Same five-method interface as the other three.
Edge-native and serverless adapters. Speculative.
wrapAdapterWithPlugins + ZanithPlugin (onQuery / onResult / onError). Hooks fire in registration order with full sql/params/duration context.
consoleLogger, slowQueryLogger, structuredLogger, openTelemetryPlugin — all built on the same hook surface.
Three plugins: rowLevelTenancy (tenant_id injection), schemaBasedTenancy (search_path), databaseBasedTenancy (per-tenant routing).
cli · devx
Generate · plan · verify · up · down. Risk-scored ops, shadow-DB verify, four-table audit log.
Soft-drop · archive · restore · export · purge. The destructive-op safety net no other migration tool ships.
Eight surfaces across workspace and database levels: tables, SQL, security, catalog, locks, RLS.
First-class command surface bundled with the engine. Same binary in dev and CI.
Cut the v0.2 release tag and publish to the registry. The last remaining adoption gate.
Programmatic API (introspectPostgres / introspectedToGraph) is shipped — only the .zanith emitter is pending so the CLI can print a file from a live DB.
First-class bootstrapping and linting commands.
expandContractRename and expandContractTypeChange compose the safe multi-step plan (add → backfill → swap → drop) that risk-2+ ops require.
preflightCheck reads adapter state before a plan runs — long-running locks, replication lag, low free space — and returns PreflightFinding[] for CI.
the public surface
The architectural case + receipts.
Sidebar / TOC / pager / 51 routes scaffolded.
Per-page metadata, canonical, OG/Twitter, JSON-LD schemas. Sitemap, robots, llms.txt for AI crawlers.
Getting Started, Schema, Queries, Advanced (9 pages), Production (6 pages), Studio (7 pages), Migrate (9 pages), Reference. Every shipped engine subsystem has a dedicated page.
Matched-workload comparisons against Prisma, Drizzle, TypeORM.
04 — Release tracks
What ships in each release, in one place. The map above answers where in the engine? This row answers when?
The runtime engine, end to end.
Chevrotain + validator
1k models, 3.4MB
Parameterized SQL on every path
find / create / update / delete
14 operators, 5 groups
Auto-rollback on throw
5-method interface
Compile-time field safety
The engine grows the surfaces a team actually adopts.
generate · plan · verify · up · down · risk-scored
soft-drop · archive · restore · cleanup
8 tabs across workspace + database levels
better-sqlite3 · file or :memory:
TLS modes + bastion-tunneled Postgres
zanith migrate · recover · studio · db
The model API gets its missing 30%.
eager fetch in one statement
groupBy/having on ModelAPI
rank, partition, OVER
exists, in
pg_catalog → .zanith emitter
Multi-database, properly.
Dialect-aware compiler path
Map AST nodes to source positions
Matched workloads vs Prisma/Drizzle
Dialect already shipped in v0.2; the adapter is the missing piece
What signals the engine is enterprise-ready.
SemVer commitments, deprecation policy
Query plans in the typed client (Studio already has it)
D1 · Neon serverless, exploratory
Memory churn, leak budgets