Context Witness Pattern

Choose between witness (existence) and capability (behavior) patterns for Context Tags.

Coupling: Hard vs Soft

Some coupling is necessary and good - but move it from hard to soft coupling.

Hard Coupling (Schema)

Field exists in the schema - tightly coupled to domain model:

// ❌ HARD COUPLING - Serial is part of the schema
export const PaymentIntent = Schema.Struct({
  id: Schema.String,
  serial: Schema.String,  // In schema = hard coupled
  amount: Schema.BigInt
})

// Every PaymentIntent MUST have a serial
// Serialization/validation requires serial
// Cannot create without providing serial
// Schema change needed to remove/change serial

Soft Coupling (Witness)

Field removed from schema, only injected in code:

// ✅ SOFT COUPLING - Serial not in schema
export const PaymentIntent = Schema.Struct({
  id: Schema.String,
  amount: Schema.BigInt
  // No serial field!
})

// Serial is a witness - required but injected via Context
class Serial extends Context.Tag("Serial")<Serial, string>() {}

const createPaymentIntent = (amount: bigint) =>
  Effect.gen(function* () {
    const serial = yield* Serial  // Injected from context

    // Use serial in business logic, logging, etc.
    // but it's not part of the persisted data
    yield* Logger.info(`Creating payment intent ${serial}`)

    return PaymentIntent.make({ id: generateId(), amount })
  })

// Type: Effect<PaymentIntent, never, Serial>

Key insight: schema (hard coupling) => witness (soft coupling)

By removing the field from the schema and injecting it only where needed, you:

• Keep domain models minimal
• Avoid unnecessary persistence
• Easy to test (provide test serial)
• Easy to remove/change (just change injection)
• Explicit dependencies in type signature

When to use witnesses:

• Correlation IDs (for tracing, not persistence)
• Request IDs (for logging, not data)
• Transaction contexts (for coordination, not storage)
• Tenant/Region markers (for routing, not schema)

Witness: Existence Only

Use when you only need to know something exists in the environment:

// Witness - a serial number exists
export class Serial extends Context.Tag("Serial")<Serial, string>() {}

const createPaymentIntent = Effect.gen(function* () {
  const serial = yield* Serial  // Pull from environment
  return PaymentIntent.make({ serial, ...other })
})

// Type: Effect<PaymentIntent, never, Serial>

Capability: Behavior

Use when you need operations:

// Capability - can generate/validate
export class SerialService extends Context.Tag("SerialService")<
  SerialService,
  {
    readonly next: () => string
    readonly validate: (s: string) => boolean
  }
>() {}

const createPaymentIntent = Effect.gen(function* () {
  const svc = yield* SerialService
  const serial = svc.next()  // Behavior
  return PaymentIntent.make({ serial, ...other })
})

// Type: Effect<PaymentIntent, never, SerialService>

Decision Framework

Need	Pattern
Just presence/value	Witness
Operations/generation	Capability
Precondition marker	Witness
Side effects	Capability
Multiple implementations	Capability
Mocking behavior	Capability
Correlation ID	Witness
Transaction context	Witness
Logger	Capability
Database	Capability

When to Use Witness

Good fits:

• Request ID - must exist for tracing
• Transaction context - must be established
• Tenant/Region - required for data boundary
• Pre-validated tokens - already verified

When to Use Capability

Good fits:

• Serial generation - create/validate operations
• Clock - now() operation
• Logger - structured logging methods
• Database - query/transact operations
• HTTP clients - fetch/post operations

Testing Implications

Witnesses are trivial to provide:

const test = myProgram.pipe(
  Effect.provideService(Serial, "test-serial-123")
)

Capabilities need implementation:

const test = myProgram.pipe(
  Effect.provideService(SerialService, {
    next: () => "test-serial-123",
    validate: () => true
  })
)

Coupling Strategy

Rule of thumb: Remove non-essential fields from schema, inject via witness instead.

Ask yourself: Does this need to be persisted/serialized?

• No → Remove from schema, inject via witness
• Yes → Keep in schema

// ✅ Domain model - only persisted data
export const Order = Schema.Struct({
  id: Schema.String,
  items: Schema.Array(LineItem),
  total: Schema.BigInt
  // No correlationId - not persisted!
  // No timestamp - derived from system!
})

// Witnesses for runtime context
class CorrelationId extends Context.Tag("CorrelationId")<CorrelationId, string>() {}
class RequestId extends Context.Tag("RequestId")<RequestId, string>() {}

// Use in code, not in data
const createOrder = (items: Array<LineItem>) =>
  Effect.gen(function* () {
    const correlationId = yield* CorrelationId  // For tracing
    const requestId = yield* RequestId          // For logging
    const clock = yield* Clock                  // For timestamp

    yield* Logger.info({
      message: "Creating order",
      correlationId,    // Used for tracing
      requestId,        // Used for logging
      timestamp: Clock.currentTimeMillis(clock)
    })

    // Data only contains what's persisted
    return Order.make({
      id: generateId(),
      items,
      total: calculateTotal(items)
    })
  })

// Type: Effect<Order, never, CorrelationId | RequestId | Clock>

Benefits:

• Minimal schemas (only persisted data)
• Context values available when needed
• Easy to test with different context
• Can add/remove context without schema changes
• Explicit dependencies in type signatures

Choose witness for simplicity, capability for flexibility.