Data

Postgres via Drizzle, how migrations work, and the tables you will actually touch.

Railway Postgres, accessed through Drizzle ORM, reached from Workers through Hyperdrive. The schema lives in lib/db/drizzle/schema.ts (49 tables), and data access is grouped by concern in lib/db/.

Migrations

Schema changes happen one way: a SQL file in db/migrations/, then bun run db:migrate. There are 133 of them. Drizzle is used as a query builder and type source, not as a migration generator.

bun run db:status          # what has run
bun run db:migrate         # apply pending
bun run db:sql "SELECT 1"  # ad-hoc query

Production migrations run through a GitHub Action on push to main. That workflow is dormant because main is dormant, but it is the launch-day migration path. Do not delete it.

Naming

camelCase in code, DTOs, and API payloads. snake_case only for actual database column names. The Drizzle schema is the boundary where the two meet.

The tables that matter

Table	Holds
`offer`	A product or offer, scraped and analyzed. Also carries `offer_type` (affiliate vs ecom_brand), `product_disclosure`, and `brand_assets` (real fonts, colours, type sheet, swipe refs)
`persona`	Buyer personas, linked to an offer
`outputs`	Every saved agent output, by module and offer
`documents`	Editable long-form documents with revision support
`campaigns`, `campaign_steps`, `campaign_messages`	The multi-stage campaign builder

Table	Holds
`chat_sessions`	One conversation
`chat_messages`	Messages, with the content blocks in `metadata.v2.blocks` and a per-session `sequence` that orders them
`background_jobs`	Anything asynchronous, surfaced in the notification bell

Table	Holds
`images`, `videos`	Generated assets, their status, and their R2 location
`image_queue`, `video_queue`	In-flight generation tasks
`static_ad_batches`	One ad batch: mode and disclosure frozen at concept time, `concepts_json`, `approved_concepts`, `plan_json`
`image_folders`	Media library folders, shared across media types
`user_uploads`	Files the user uploaded, including chat drops
`characters`	A character identity plus its generated character map
`elements`	Locations, props, and products used as visual references
`storyboards`, `video_pipelines`	Scene-by-scene production state
`image_presets`	Style presets, admin-managed

Table	Holds
`user`	People. `clerk_user_id` joins to Clerk, which owns identity only
`admin_session`, `admin_verification`	Elevated admin actions
`prompts`	The runtime system and user prompt per module
`ai_models`	Chat, image, and video models with a JSONB `config`
`app_settings`	Key/value settings, including the default model per type
`subscription_plans`, `pending_customers`	Polar billing
`api_request_logs`, `api_usage_logs`	Every external call, with agent attribution
`email_templates`, `page_templates`, `themes`	Admin-editable content

The module table is not in the Drizzle schema file by that name but is central: it holds one row per AI module, keyed to lib/registry.ts. See Prompts.

Media storage

One R2 bucket, adstack-media, served at media.tryadstack.com, shared by staging and production.

Request and usage logging

Every external call is logged to api_request_logs with provider, agent attribution, status, timing, and sanitized payloads. Providers are cloudflare_ai_gateway, firecrawl, and kie_ai. The agent column identifies the caller: host, a specific agent slug, image-prompt, output-verifier, and so on. The admin UI is at /admin/logs.

sanitizeApiLogValue() redacts API keys, tokens, passwords, signed URLs, and data URLs before anything is written, truncates strings over 12K characters, and stops at depth 8.