Draagon AI: 17 runtime schema patches to zero

The challenge: three copies of every shape — and the drift was executable

Every agent, telemetry record, and config shape lived in three hand-kept declarations: the TypeScript Drizzle schema, the Python persistence layer, and the shared Zod validators and TS types. The cross-language contract between the engine and the platform was maintained by hand and documented in a comment.

The drift wasn't theoretical — it was executable. The Python side carried 17 ALTER TABLE … ADD COLUMN IF NOT EXISTS statements run on every database connect: Python defensively re-patching the schema on startup because it ran ahead of the TypeScript migrations. The two languages reconciled their schema at runtime, at the first connection.

Then it shipped a bug. Hand-written validators duplicated the data model and silently diverged — marking optional config fields as required — and legitimate partial agent-config writes started returning 400 in production. Meanwhile the 100+ LLM prompts were string constants formatted inside a try/except that silently degraded when a payload key went missing, and a ~190-line hand-written mapper re-mapped every field one by one — primed to drop one whenever the model grew.

Hand-syncing the third copy of every shape, with a runtime patcher papering over the gap, was the decision point: the shapes themselves had to become data.

The solution: declare it once, generate every consumer

The fix wasn't a better sync process — it was making the model the artifact. 44 entities, 16 value objects, 13 projection views, and 106 prompt templates are now declared once in YAML, and every Python and TypeScript consumer is generated from that spine and drift-gated in CI. One field declaration becomes the table column and the validator that used to be hand-kept:

# meta.platform.yaml
- field.int:
    name: maxTokens
    column: max_tokens

// Drizzle table (TypeScript)
maxTokens: integer("max_tokens"),

// Zod validator — was hand-kept, now generated
maxTokens: z.number().int().optional(),

From that one model, meta gen produces 471 committed files (~19.6k lines): Python record models, value objects, and prompt payloads with output parsers on the engine side; Drizzle tables, typed query helpers, Zod/TS contracts, TanStack hooks, admin UI, and API routes on the platform side. Business logic and the orchestration engine stay hand-written — the generated layer is the plumbing that used to drift.

Four meta verify gates run in CI, plus a pre-commit hook and a dedicated drift workflow on every push — so a model change that isn't reflected in both languages fails the build, not the first INSERT. The migration took five weeks and 97 commits, alongside normal feature work.

Results

• The runtime schema patcher is deleted. All 17 ALTER TABLEs are gone; both languages generate from one canonical schema, so "Python thinks the table has a column TypeScript doesn't" now fails the build instead of reaching production.
• The validator-drift bug class is dead. Write validators are recomposed from the generated schemas with the real invariants encoded once in metadata — the partial write that used to 400 now succeeds, and a genuinely invalid write still fails, verified end-to-end.
• 106 prompts became typed contracts. Each is a typed payload plus an externalized template with deterministic render, drift-checked at build time. Every migration was byte-parity-gated against the original before the hand-written constant was deleted — 54 cases, 0 mismatches.
• Plumbing deleted, not refactored: the ~190-line field mapper, the hand web data layer, and the hand admin tables were replaced by generated equivalents and removed.

The schema patcher that ran 17 ALTER TABLEs on every connect wasn't refactored — it was deleted. The canonical model owns the shape now, in both languages.

— verified in-repo: schema patcher removed (−75 lines), June 2026

Draagon's source is private — for a clone-and-read version of the same pattern, run the 10-second drift demo in Wizards of Odd or read smon.

Where it is now

Since 2026-09-09 it has run MetaObjects 1.0.0, every package it consumes pinned to the stable release, with its drift gate enforced in CI.