Practical notes for an eventual migration of `packages/opencode` server routes from the current Hono handlers to Effect `HttpApi`, either as a full replacement or as a parallel surface.
## Goal
Use Effect `HttpApi` where it gives us a better typed contract for:
- route definition
- request decoding and validation
- typed success and error responses
- OpenAPI generation
- handler composition inside Effect
This should be treated as a later-stage HTTP boundary migration, not a prerequisite for ongoing service, route-handler, or schema work.
## Core model
`HttpApi` is definition-first.
-`HttpApi` is the root API
-`HttpApiGroup` groups related endpoints
-`HttpApiEndpoint` defines a single route and its request / response schemas
- handlers are implemented separately from the contract
This is a better fit once route inputs and outputs are already moving toward Effect Schema-first models.
## Why it is relevant here
The current route-effectification work is already pushing handlers toward:
- one `AppRuntime.runPromise(Effect.gen(...))` body
- yielding services from context
- using typed Effect errors instead of Promise wrappers
That work is a good prerequisite for `HttpApi`. Once the handler body is already a composed Effect, the remaining migration is mostly about replacing the Hono route declaration and validator layer.
## What HttpApi gives us
### Contracts
Request params, query, payload, success payloads, and typed error payloads are declared in one place using Effect Schema.
### Validation and decoding
Incoming data is decoded through Effect Schema instead of hand-maintained Zod validators per route.
### OpenAPI
`HttpApi` can derive OpenAPI from the API definition, which overlaps with the current `describeRoute(...)` and `resolver(...)` pattern.
### Typed errors
`Schema.TaggedErrorClass` maps naturally to endpoint error contracts.
## Likely fit for opencode
Best fit first:
- JSON request / response endpoints
- route groups that already mostly delegate into services
- endpoints whose request and response models can be defined with Effect Schema
Harder / later fit:
- SSE endpoints
- websocket endpoints
- streaming handlers
- routes with heavy Hono-specific middleware assumptions
## Current blockers and gaps
### Schema split
Many route boundaries still use Zod-first validators. That does not block all experimentation, but full `HttpApi` adoption is easier after the domain and boundary types are more consistently Schema-first with `.zod` compatibility only where needed.
Many current `server/routes/instance/*.ts` handlers still mix composed Effect code with smaller Promise- or ALS-backed seams. Migrating those to consistent `Effect.gen(...)` handlers is the low-risk step to do first.
The server currently includes SSE, websocket, and streaming-style endpoints. Those should not be the first `HttpApi` targets.
### Existing Hono integration
The current server composition, middleware, and docs flow are Hono-centered today. That suggests a parallel or incremental adoption plan is safer than a flag day rewrite.
Do not re-architect business logic during the HTTP migration. `HttpApi` handlers should call the same Effect services already used by the Hono handlers.
The bridge is gated behind `OPENCODE_EXPERIMENTAL_HTTPAPI` (or `OPENCODE_EXPERIMENTAL`). When the flag is off (default), all requests go through the original Hono handlers unchanged.
The Hono route handlers are always registered (after the bridge) so `hono-openapi` generates the OpenAPI spec entries that feed SDK codegen. When the flag is on, these handlers are dead code — the `.all()` bridge matches first.
### 5. Observability
The `webHandler` provides `Observability.layer` via `Layer.provideMerge`. Since the `memoMap` is shared with `AppRuntime`, the tracing provider is deduplicated — no extra initialization cost.
This gives:
- **spans**: `Effect.fn("QuestionHttpApi.list")` etc. appear in traces alongside service-layer spans
- **HTTP logs**: `HttpMiddleware.logger` emits structured `Effect.log` entries with `http.method`, `http.url`, `http.status` annotations, flowing to motel via `OtlpLogger`
Every `HttpApi` slice should follow `specs/effect/schema.md` and the Schema -> Zod interop rule in `specs/effect/migration.md`.
Default rule:
- Effect Schema owns the type
-`.zod` exists only as a compatibility surface
- do not introduce a new hand-written Zod schema for a type that is already migrating to Effect Schema
Practical implication for `HttpApi` migration:
- if a route boundary already depends on a shared DTO, ID, input, output, or tagged error, migrate that model to Effect Schema first or in the same change
- if an existing Hono route or tool still needs Zod, derive it with `@/util/effect-zod`
- avoid maintaining parallel Zod and Effect definitions for the same request or response type
Ordering for a route-group migration:
1. move implicated shared `schema.ts` leaf types to Effect Schema first
- every schema migration must preserve the generated SDK output byte-for-byte **unless the new ref is intentional** (see Schema.Class vs Schema.Struct below)
- if an unintended diff appears in `packages/sdk/js/src/v2/gen/types.gen.ts`, the migration introduced an unintended API surface change — fix it before merging
### Schema.Class vs Schema.Struct
The pattern choice determines whether a schema becomes a **named** export in the SDK or stays **anonymous inline**.
**Schema.Class** emits a named `$ref` in OpenAPI via its identifier → produces a named `export type Foo = ...` in `types.gen.ts`:
```ts
export class Info extends Schema.Class<Info>("FooConfig")({ ... }) {
static readonly zod = zod(this)
}
```
**Schema.Struct** stays anonymous and is inlined everywhere it is referenced:
```ts
export const Info = Schema.Struct({ ... }).pipe(
withStatics((s) => ({ zod: zod(s) })),
)
export type Info = Schema.Schema.Type<typeofInfo>
```
When to use each:
- Use **Schema.Class** when:
- the original Zod had `.meta({ ref: ... })` (preserve the existing named SDK type byte-for-byte)
- the schema is a top-level endpoint request or response (SDK consumers benefit from a stable importable name)
- Use **Schema.Struct** when:
- the type is only used as a nested field inside another named schema
- the original Zod was anonymous and promoting it would bloat SDK types with no import value
Promoting a previously-anonymous schema to Schema.Class is acceptable when it is top-level or endpoint-facing, but call it out in the PR — it is an additive SDK change (`export type Foo = ...` newly appears) even if it preserves the JSON shape.
Schemas that are **not** pure objects (enums, unions, records, tuples) cannot use Schema.Class. For those, add `.annotate({ identifier: "FooName" })` to get the same named-ref behavior:
- it is acceptable to keep a route-local Zod schema for the first spike only when the type is boundary-local and migrating it would create unrelated churn
- if that happens, leave a short note so the type does not become a permanent second source of truth
## First vertical slice
The first `HttpApi` spike should be intentionally small and repeatable.
Chosen slice:
- group: `question`
- endpoints: `GET /question` and `POST /question/:requestID/reply`
Non-goals:
- no `session` routes
- no SSE or websocket routes
- no auth redesign
- no broad service refactor
Behavior rule:
- preserve current runtime behavior first
- treat semantic changes such as introducing new `404` behavior as a separate follow-up unless they are required to make the contract honest
Add `POST /question/:requestID/reject` only after the first two endpoints work cleanly.
## Repeatable slice template
Use the same sequence for each route group.
1. Pick one JSON-only route group that already mostly delegates into services.
2. Identify the shared DTOs, IDs, and errors implicated by that slice.
3. Apply the schema migration ordering above so those types are Effect Schema-first.
4. Define the `HttpApi` contract separately from the handlers.
5. Implement handlers by yielding the existing service from context.
- the bridged `HttpApi` layer resolves instance context via an `HttpRouter.middleware` that reads `x-opencode-directory` headers and `directory` query params
-`Schema.Class` works well for route DTOs such as `Question.Request`, `Question.Info`, and `Question.Reply`.
- scalar or collection schemas such as `Question.Answer` should stay as schemas and use helpers like `withStatics(...)` instead of being forced into classes.
- if an `HttpApi` success schema uses `Schema.Class`, the handler or underlying service needs to return real schema instances rather than plain objects.
- internal event payloads can stay anonymous when we want to avoid adding extra named OpenAPI component churn for non-route shapes.
-`Schema.Class` emits named `$ref` in OpenAPI — only use it for types that already had `.meta({ ref })` in the old Zod schema. Inner/nested types should stay as `Schema.Struct` to avoid SDK shape changes.
### Integration
-`HttpRouter.toWebHandler` with the shared `memoMap` from `run-service.ts` cleanly bridges Effect routes into Hono — one process, one port, shared layer instances.
-`Observability.layer` must be explicitly provided via `Layer.provideMerge` in the routes layer for OTEL spans and HTTP logs to flow. The `memoMap` deduplicates it with `AppRuntime` — no extra cost.
-`HttpMiddleware.logger` (enabled by default when `disableLogger` is not set) emits structured `Effect.log` entries with `http.method`, `http.url`, `http.status` — these flow through `OtlpLogger` to motel.
- Hono OpenAPI stubs must remain registered for SDK codegen until the SDK pipeline reads from the Effect OpenAPI spec instead.
- the `OPENCODE_EXPERIMENTAL_HTTPAPI` flag gates the bridge at the Hono router level — default off, no behavior change unless opted in.
If `HttpApi` is adopted here, it should arrive after the handler body is already Effect-native and after the relevant request / response models have moved to Effect Schema.
