Env Vars

Environment Variables

Everything here documents the repository as it is on main. If main disagrees with this file, main wins: follow it and flag the drift.

Configuration follows a two-tier pattern: a framework env holding only the vars framework code reads, and an app env holding every var (framework vars plus product vars).

Architecture

Framework env — app/framework/env.server.ts

Holds only what framework code needs — today, the database path. Framework files (db.server.ts, scheduler.server.ts) import env from ./env.server with a relative path. This is what keeps app/framework/ extractable: it validates only its own surface and never learns a product var exists.

App env — app/env.server.ts

Holds all vars: the framework vars again, plus the product vars (DISCORD_BOT_TOKEN, DISCORD_OWNER_USER_ID, DISCORD_GUILD_ID). It imports makeTypedEnv from make-typed-env and defines its own schema.

Product code — app/business/, app/mcp/, app/ingest/, app/db/, app/test/ — imports env from ~/env.server.

Why two independent singletons

Each env() parses process.env through its own makeTypedEnv instance. Caching is built into makeTypedEnv, so the () => getEnvironment(process.env) wrapper is enough. The two share no state, which is exactly the point: the framework validates what it needs, and the app validates everything.

Read configuration through env()

Application code never reads process.env directly. Every value arrives through env(), already validated and camelCased, so a missing or malformed variable fails loudly at startup instead of surfacing as undefined deep inside a query or a Discord call.

Adding a new env var

Product var (a new Discord setting, a new tuning knob)

1. Add the Zod field to app/env.server.ts only
2. Add the key, with an explanatory comment, to .env.example
3. Add the value to the job-level env: block in .github/workflows/ci.yml
4. Update the setup documentation if the owner has to obtain the value from somewhere
5. Import env from ~/env.server in the consuming file

Framework var (a new database or scheduler option)

1. Add the Zod field to both app/framework/env.server.ts and app/env.server.ts
2. Add the key to .env.example
3. Add the value to the job-level env: block in .github/workflows/ci.yml
4. Framework files import from ./env.server; product files import from ~/env.server

CI environment variables

The CI workflow lives at .github/workflows/ci.yml. Values go in the workflow-level env: block, and the seed smoke step carries its own overrides on top of it. CI has no .env, so a required variable added to a schema but not to the workflow passes every local gate and red-builds the PR that adds it. Use placeholder values for services CI never actually calls ('placeholder' for the bot token).

The E2E harness holds a third copy: serverEnvironment in tests/mcp-client.ts hardcodes the environment the spawned server receives, so a variable the server requires lands there in the same change.

Import conventions

// framework code, inside app/framework/
import { env } from './env.server'

// product code, anywhere else under app/
import { env } from '~/env.server'

Never import env from ~/framework/env.server in product code — that singleton carries only framework vars and silently lacks the product fields.

The makeTypedEnv factory

Provided by the make-typed-env package:

import { makeTypedEnv } from 'make-typed-env'
import { camelKeys } from 'string-ts'
import { z } from 'zod'

const getEnvironment = makeTypedEnv(
  z.object({ /* ... */ }),
  { transform: camelKeys },
)
const env = () => getEnvironment(process.env)

• Schema-agnostic — it accepts any Standard Schema; this project uses Zod
• transform: camelKeys from string-ts converts SNAKE_CASE keys to camelCase, so DISCORD_GUILD_ID reads as env().discordGuildId
• Caches by default, so the thin wrapper is all that is needed
• Every call to the factory creates an independent instance, so the two tiers never interfere

Schema duplication is intentional

Framework vars appear in both schemas. That duplication is what decouples the folder from the app. When a framework var's validation changes — a new default, a tighter format — update both files in the same change.

Secrets

DISCORD_BOT_TOKEN is a credential. It is read through env() at the point of use and never copied into a stored row, an error message, a telemetry payload, or an MCP tool result. .env stays out of version control; .env.example carries the key and the instructions for obtaining a value, never a value.