Effect Program Design

Fold Effect Program Design

Build deep Fold modules: a small, domain-shaped public interface hides descriptor lowering, service wiring, provider details, persistence, resource ownership, and workflow coordination. Keep expected failures in Effect's error channel, dependencies in R, resources in scopes, and external values at explicit schema boundaries.

Fold is an Effect v4 (4.0.0-rc.109) Bun monorepo. Read the installed declarations first; when they do not settle an API, read ~/projects/effect, not Effect v3 documentation or examples.

Fold's architecture

• Public Fold APIs are descriptor-facing. Hosts use defineAgent, model descriptors, tool descriptors, and event-log descriptors without learning Layer, Toolkit, or runtime wiring. See README.md and packages/fold-core/src/Api/Provisioning.ts.
• Provisioning owns lowering. The provisioner turns descriptors into the required services and layers once per runtime/session. Do not make callers build or pass Fold's internal clients, layers, toolsets, or runtime services.
• Services are internal capabilities. Use Context.Service and Layer where a capability varies by runtime, implementation, or test seam. One implementation is enough when the seam is valuable for real tests.
• Event schemas are durable contracts. Model persisted and wire-visible log entries with Schema.TaggedStruct and Schema.Union; decode before projecting or dispatching. See packages/fold-core/src/EventLog/Schemas.ts.

The defaults

1. Deep modules. A module's interface is the cost; hidden behavior is the benefit. A caller supplies the domain values it holds, not credentials, provider clients, decoded rows, layers, or other internals.
2. Effects retain their channels. Expected failure stays in E; dependencies stay in R; resources have an owning scope. Do not pass errors, services, layers, or effects around as ordinary data just to compose them later.
3. Typed failures recover in-channel. Classify an untrusted/throwing boundary once, then use Effect.catchTag or Effect.catchTags to recover or narrow. Do not recover typed failures with instanceof or manual _tag checks.
4. Normal tagged data dispatches explicitly. For decoded or otherwise trusted union values, use Predicate for reusable narrowing and Match for complete transformations. See PREDICATE-MATCH.md.
5. Schemas parse boundaries. Decode JSON, files, provider payloads, host input, and durable records at their edge. Do not cast or ad hoc-narrow unknown data through the core.
6. Tests cross real seams. Test services through their public interface with @effect/vitest and substitute layers or real test adapters. Do not use vi.mock, module patching, or method spies.

Module depth and seams

• Give public operations named domain inputs and outputs. A descriptor API should hide the provider/layer graph it needs to realize the descriptor.
• Use the deletion test: deleting a useful module should spread its orchestration and invariants across callers, not simply delete a pass-through.
• Keep pure domain decisions, projections, and formatting separate from the I/O shell. Pure helpers accept and return domain values; layers/adapters sequence Effects and own I/O.
• Do not add a wrapper solely to mirror an SDK or data structure. A module earns its seam by concentrating meaningful lifecycle, policy, parsing, or orchestration.
• Do not make all dependencies ordinary parameters. Runtime-varying capabilities belong in Effect's environment; pure values legitimately held by the caller stay explicit inputs.

Services, layers, and resources

• Prefer Context.Service for an application capability and Layer.effect or Layer.scoped for its implementation. Small capabilities may collocate shape, tag, and live layer; split an external adapter, persistence, or workflow only when that separation improves locality.
• Dependencies normally remain ambient in R. Yield them where the operation needs them rather than forwarding them through public signatures.
• Build runtime-specific layer graphs at the provisioner/facade seam. Fold's agent provisioner owns memo-map and scope semantics; callers do not recreate that graph.
• Acquire resources in a scope and make background work supervised/owned. Bound concurrency for unbounded fan-out and keep external calls outside authoritative transactions.
• Use Effect.fn or Effect.withSpan for meaningful public or I/O operation boundaries when observability is configured. Add safe context only; never log secrets or unrestricted provider payloads.

Errors and boundaries

• Model expected failures as tagged errors with fields a caller can use. Data.TaggedError is appropriate for an internal failure; use a schema-backed tagged error when the error itself crosses an encoded boundary.
• Wrap a throwing SDK or native API once at its adapter edge with Effect.try or Effect.tryPromise. Its catch maps the unknown cause into the module's typed error vocabulary. Do not repeat provider-specific inspection in callers.
• Preserve rich internal failures until the module boundary, then narrow to the small set of outcomes callers can actually act on. A fallback is only correct when it is part of the operation's contract.
• Capture actionable or unexpected failures with the repository's configured logging/observability before swallowing or narrowing them. Do not add Sentry or other product-specific dependencies unless Fold provides and configures one.
• Use catchCause only for a deliberate top-level safety net that must include defects, such as a best-effort operation. It must make the failure observable before it is swallowed.

Schema and domain data

• Use Schema.Struct for ordinary records and Schema.TaggedStruct plus Schema.Union for encoded tagged variants. Derive the TypeScript type from the schema with typeof X.Type.
• Use Schema.Literals for closed scalar vocabularies. Branded IDs are valuable where Fold must prevent same-typed identity mix-ups, but do not introduce brands by default without a concrete boundary or misuse risk.
• Use Schema.optionalKey, Schema.optional, and Schema.NullOr to preserve the distinction between absent, undefined, and null values.
• Decode untrusted encoded values with Schema.decodeUnknownEffect; use make/makeEffect only for trusted type-side construction. Do not cast decoded JSON.
• Keep schemas that define durable log data backward compatible. Add an entry schema/version and upcast at the decode boundary instead of mutating an incompatible persisted format.

Predicate and Match

The error channel and ordinary tagged values require different tools:

• Use Effect.catchTag / Effect.catchTags for typed errors in Effect<A, E, R>.
• Use Predicate.isTagged for a reusable one-tag guard or a named multi-tag refinement used by find, filter, or a guard clause.
• Use Match.type<T>() and Match.tagsExhaustive for a reusable transformation over a Fold-owned closed union. Use Match.value(value) for a one-off dispatch.
• Keep a simple direct _tag guard in a local stateful loop when a matcher or extracted predicate would add ceremony. Do not mechanically replace every conditional. Repeated comparisons and complete domain transformations should not drift across ad hoc conditionals.
• Neither Predicate nor Match validates an unknown provider/file/JSON value. Decode it first.

Read PREDICATE-MATCH.md before writing or reviewing tagged normal-value control flow.

Tests

• Use @effect/vitest and it.effect; provide the service/layer graph to the program under test.
• Substitute an external provider, filesystem, clock, or other true boundary with a narrow fake Effect layer. Use real ephemeral infrastructure when behavior depends on its constraints or persistence semantics.
• Prefer deterministic TestClock, Deferred, Queue, Latch, and Ref coordination over sleeps and timing races.
• Assert both the returned value/error and the relevant observable end state: durable entries, emitted events, files, requests recorded by a fake, or released resources.
• A fake should fail loudly for unexpected methods. If a dependency cannot be replaced at a layer seam, improve the module boundary rather than reaching for a module mock.

References

• PREDICATE-MATCH.md for tagged normal-value dispatch.
• SCHEMA-DOMAIN-PATTERNS.md for Fold schema, durable-data, and error-model choices.
• REFERENCE.md for canonical Fold modules and the expected service/layer shapes.
• codebase-design for shared vocabulary on depth, interface, seam, adapter, leverage, and locality.

Review checklist

• The public interface is small, domain-shaped, and hides Fold's descriptor-lowering and runtime wiring.
• Expected failures remain typed in E; dependencies remain declared in R; resources have an owner.
• Unknown data is decoded at the edge and raw provider/file data does not leak through a public contract.
• Typed errors use catchTag / catchTags; ordinary trusted tags use Predicate / Match when appropriate.
• Fold-owned complete unions use an exhaustive transformation where a new variant must force a review.
• Provider calls, concurrency, retries, and resource lifetime are bounded and owned by the module that needs them.
• Tests use @effect/vitest, real seams, deterministic coordination, and no module mocks or spies.