Effect

Ground every recommendation in the target project. Inspect repository instructions, the pinned effect version, installed package source, and local conventions before choosing APIs. Prefer project evidence over remembered signatures.

Choose the Relevant Guidance

Read only the references needed for the task, and combine them when concerns overlap:

• Schema, records, variants, brands, and typed errors: references/SCHEMA.md
• Services, layers, modules, scope ownership, and Effect.fn: references/SERVICES_LAYERS.md
• Configuration and providers: references/CONFIG.md
• Schedules, retry, repeat, polling, and pacing: references/SCHEDULING.md
• Cache, memoization, in-flight dedupe, and request batching: references/CACHING.md
• Streams, queues, PubSub, SubscriptionRef, and long-lived consumers: references/STREAMS.md
• HTTP clients, decoding, retries, and rate limits: references/HTTP_CLIENTS.md
• Effect-aware tests, TestClock, layers, and synchronization: references/TESTING.md

Model Data and Failures

• Use Schema.Struct(...) for application records.
• Use brands for scalar IDs and constrained values.
• Use tagged Schema variants and unions for values that cross boundaries.
• Decode unknown boundary data with Schema.decodeUnknownEffect(...).
• Model expected failures with Schema.TaggedErrorClass.
• Never use casts or throwing construction to skip validation at untrusted boundaries.

Build Explicit Services and Layers

• Keep business rules in named workflows and services.
• Represent dependencies explicitly in the Effect environment.
• Build implementations with layers that own resource lifetime.
• Keep HTTP and transport handlers focused on decoding, context, invocation, and response mapping.
• Wrap public or non-trivial service methods with named Effect.fn calls.
• Do not hide required authority, credentials, persistence, or transports behind defaults.

Choose Runtime Primitives by Meaning

• Use Config for runtime configuration and override it with providers in tests.
• Use Schedule for retry, repeat, polling, pacing, and backoff.
• Use Cache or Effect memoization when their ownership and eviction model fit.
• Use request batching only when the backend exposes a real batch operation.
• Use Queue, PubSub, SubscriptionRef, and Stream according to producer, consumer, broadcast, and backpressure needs.
• Keep long-lived fibers scoped to the layer or resource that owns them.

Keep Boundaries Truthful

• Wrap HTTP clients, SDKs, CLIs, and external systems in named Effect boundaries.
• Decode persisted and external values before treating them as domain data.
• Preserve distinct transport, status, decode, and domain failures.
• Retry only proven transient failures on idempotent operations with a bounded policy.
• Keep exhausted failures visible unless the boundary has a real fallback.
• Keep provider and network calls outside authoritative database transactions.

Test Deterministically

• Use Effect-aware tests and explicit test layers.
• Fork time-dependent work before advancing TestClock.
• Coordinate concurrent work with Deferred, Queue, Latch, Ref, fibers, or explicit hooks.
• Never add arbitrary real sleeps when deterministic synchronization can express the condition.

Preserve Types and Dependencies

Solve missing requirements through explicit service and layer composition. Do not use as any, non-null assertions, broad cause recovery, Layer.mergeAll, provideMerge, or default authority merely to make code compile.

Before Editing

1. Identify every Effect concern in the task.
2. Read the matching references and project-local source.
3. Choose APIs supported by the pinned version.
4. Make the smallest architecture-preserving change.
5. Run the project's typecheck and focused tests.