liaise 0.0.0 → 5.0.1

package/MIGRATION.md

@@ -0,0 +1,971 @@
+# Migration Guide
+
+Upgrade notes for `liaise` (published as `@iremlopsum/apify` up to 4.4.x). Only releases that need action appear
+here — if a version isn't listed, upgrading to it requires no changes.
+
+For the full record of what changed in each release, see [CHANGELOG.md](./CHANGELOG.md).
+
+---
+
+## Upgrading to 5.0.1
+
+No code changes for most callers. Six things you may observe.
+
+**Binary bodies arrive as binary.** A `Uint8Array`, other typed array, `DataView`
+or `Buffer` used to be sent as a JSON index map (`{"0":1,"1":2}`); it is now the
+raw bytes, with `Content-Type: application/octet-stream`. A `ReadableStream` is
+sent as a streaming upload and can be sent once: `retryMiddleware` and
+`result.retry()` return an error Result for it. If a call with a stream may be
+retried, read it into a `Blob` or `ArrayBuffer` first. If your server decoded the
+old index map, it needs to read bytes now.
+
+**Calls that used to send nothing now send, or fail with a Result.** A `Map` with
+string keys and a class with only `toJSON()` are now sent. A `Set`, a bare `Date`,
+a `Map` with non-string keys and a class with no fields return an error Result
+(`kind: 'network'`, `body` a `TypeError` naming the type). A typed array, stream
+or `toJSON`-only class on a GET is refused. If you were relying on the empty body,
+pass the object you meant: `{ ids: [...set] }`, `{ since: date.toISOString() }`.
+
+**`cacheMiddleware` entries are per URL and per header set.** The key now includes
+method, full URL (query string included, pairs sorted by name) and every request
+header except `Content-Type`. A different `Authorization` or other header, base
+URL, path or query value gets its own entry; the same query params in a different
+order share one. A warm cache is cold once after
+upgrading. A middleware placed before `cacheMiddleware` that adds a per-call
+unique header (a request ID) makes every call a miss; put it after
+`cacheMiddleware` in the `middleware` array.
+
+**Repeated header names within one source are joined.** A header-pairs array
+with two entries for one name now sends `a, b`, where the last used to win; so
+does a record with case-variant duplicates (`{ Accept: 'a', accept: 'b' }`). A
+later source (per-call over request over config) still replaces an earlier one.
+
+**After a call settles, a later abort no longer reaches it.** Aborting the
+caller's signal no longer reaches that call's `ctx.request.signal`, so
+fire-and-forget middleware work still holding it is not cancelled by the caller;
+keep your own controller if you need that.
+
+**A path token with non-word characters now fails.** Path tokens are
+`[a-zA-Z0-9_]`. A template like `/x/:a-b` with a key `a-b` used to resolve by
+accident; it is now read as the token `:a` followed by `-b`, and the call returns
+an error Result (a `TypeError`, "Unresolved path parameter :a…"). Use only
+`[a-zA-Z0-9_]` in path token names and their keys.
+
+---
+
+## Upgrading to 5.0.0
+
+The package has a new name: **`@iremlopsum/apify` is now `liaise`**. Nothing
+else about the API changes — every function, option, type and behaviour is the
+same as 4.4.3.
+
+**1. Swap the package.**
+
+```bash
+npm uninstall @iremlopsum/apify
+npm install liaise
+```
+
+**2. Update the import paths** — a find-and-replace across your code:
+
+| Before | After |
+|---|---|
+| `@iremlopsum/apify` | `liaise` |
+| `@iremlopsum/apify/middleware` | `liaise/middleware` |
+| `@iremlopsum/apify/testing` | `liaise/testing` |
+
+Replacing `@iremlopsum/apify` with `liaise` everywhere covers all three.
+
+**One behaviour change, only if you read the logs.** `logMiddleware` and
+`cacheMiddleware({ debug: true })` print `[liaise]` and `[liaise cache]` where
+they printed `[apify]` and `[apify cache]`. If a log filter, alert or test
+matches on the old prefix, update it.
+
+---
+
+## Upgrading to 4.4.3
+
+No code changes. Two things you may observe after upgrading.
+
+**More requests where there were wrongly fewer.** `share: true` endpoints and
+`cacheMiddleware` keyed params by their enumerable keys. A `Date`, `Map`, `Set`,
+`ArrayBuffer`, BigInt or private-state class instance anywhere below the top
+level therefore made every such call look identical, and one caller could
+receive the response meant for another caller's params. Those calls now key by
+content and go out separately. If request counts rise after upgrading, that is
+the fix working: the previous count was wrong responses, not savings. A value
+that cannot be keyed soundly (a BigInt, an `ArrayBuffer`, `Blob`, `FormData`
+or `URLSearchParams`, an object with no enumerable state) is now never shared
+or cached at any depth, where before only the top level was checked.
+
+**`{ a: undefined }` and `{}` are now the same key.** An `undefined` member is
+dropped, as `JSON.stringify` drops it on the wire, so equivalent calls hit the
+same cache entry and the same in-flight share. There is no way to make
+`undefined` mean "a different request", and the wire never carried one.
+
+One smaller change at the top level only: a `Map`, `Set` or `Date` param was
+previously never shared or cached; it now is, by content, so two identical such
+params coalesce. `FormData`, `Blob`, `ArrayBuffer` and `URLSearchParams` still
+never do.
+
+---
+
+## Upgrading to 4.4.2
+
+No action is needed for almost everyone. This release makes `timeout` and
+`CallOptions.signal` do what they were documented to do: settle a call even
+when a middleware is stuck awaiting work that ignores the signal.
+
+**The rule that changed:** a call is now bounded by its deadline no matter what
+it is waiting on. Before, the deadline only reached `fetch` and whatever read
+`ctx.request.signal`; anything else could run past it, and the call waited.
+So the outcome changes for any call that was still running **after** its
+`timeout` fired (or its signal aborted) on work the signal does not reach:
+
+- a middleware awaiting something that never settles — before: the call never
+  settled; now: `kind: 'timeout'` / `'abort'`.
+- response-side work that crosses the deadline — a middleware post-processing
+  a success, an async Standard Schema validator — or a `fetch` implementation
+  (a hand-rolled mock, a polyfill) that ignores its signal. Before: the late
+  result was delivered; now: `kind: 'timeout'` / `'abort'`, the same as if the
+  work had finished one moment later than it did.
+
+A call that finishes within its deadline is unaffected, and so is any work that
+already responds to the abort. If you have a test mock that ignores
+`init.signal` and resolves *after* a `timeout` you set, that test now sees a
+timeout — which is what the configuration asked for.
+
+### If you use `mockFetch` from `./testing`
+
+`mock.fetch` now honours `init.signal`. A mocked call whose signal is aborted —
+already, or while its route handler is still pending — rejects with
+`signal.reason` instead of resolving, exactly as real `fetch` does. A test that
+aborted a call and still expected the mocked response is the only thing this
+changes; it now sees the abort.
+
+### If you wrapped calls in your own `Promise.race` against a timer
+
+You can delete the wrapper and use `timeout` (or pass your `AbortSignal`)
+instead. A call whose middleware stalls now settles with `kind: 'timeout'` (or
+`'abort'`), `status: 0`.
+
+### If a middleware answers an abort slowly
+
+This is the one case above where a middleware's own handling of the deadline
+is overruled, so it gets its own section.
+
+Once the signal aborts, the chain gets one macrotask to answer by itself. A
+middleware that responds to a timeout by doing *more* I/O before returning its
+own `Result` — reading a fallback from IndexedDB, say — used to have that
+`Result` delivered, however long it took. Now, if it has not answered within
+that macrotask, the caller gets the timeout `Result` and the middleware's later
+answer is discarded.
+
+A fallback that answers from memory, or from anything already in hand, is
+unaffected — it settles within microtasks. If yours needs real I/O after the
+deadline, give it a deadline of its own that fires earlier than the call's:
+
+```ts
+const withFallback: Middleware = async (ctx, next) => {
+  ctx.request.signal = AbortSignal.timeout(4_000)   // inside the call's 5_000
+  const result = await next()
+  return result.error?.kind === 'timeout' ? await readFallback(ctx) : result
+}
+```
+
+Like the per-attempt example in the README, this *replaces* the signal, so the
+caller's own `AbortSignal` no longer reaches `fetch` — aborting it still settles
+the call (the backstop watches it), but the socket stays open until the 4-second
+signal fires. If that matters, merge the two instead of replacing, with
+`AbortSignal.any([ctx.request.signal, own])` where your runtimes support it.
+
+### If a middleware calls `next()` long after the call timed out
+
+It no longer sends a request. `next()` returns the `Result` the caller already
+received. Before, a middleware that installed a fresh signal of its own could
+still send one nobody was waiting for.
+
+Under `dedupe: true`, a request still in flight when the call is settled this
+way is aborted, not left running: nothing else can cancel it once the call has
+given up its dedupe slot.
+
+---
+
+## Upgrading to 4.4.0
+
+One change needs action, and only if you use `defineRequest` with a `path`
+containing a `#`.
+
+### A fragment in a `defineRequest` path literal is now a compile error
+
+If your build goes red on a line like this, that is this change:
+
+```ts
+defineRequest<Doc>()({ method: 'GET', path: '/docs#section' })
+//                                          ^ Property '__fragmentInPath' is missing:
+//                                            a URL fragment is never sent to the server
+```
+
+**Your code was already broken.** A URL fragment is never transmitted — `fetch`
+strips it — so this endpoint could never reach `/docs#section`. Since 4.2.1 it
+has been returning an error `Result` on **every call**. The only thing that
+changed in 4.4.0 is *when* you find out: at compile time, where you declared it,
+instead of at runtime on every request.
+
+**The fix is to delete the fragment:**
+
+```ts
+// before
+defineRequest<Doc>()({ method: 'GET', path: '/docs#section' })
+
+// after
+defineRequest<Doc>()({ method: 'GET', path: '/docs' })
+```
+
+If the fragment was carrying information you need on the server, it has to move
+into the path or the query string, because the server never received it:
+
+```ts
+defineRequest<Doc>()({ method: 'GET', path: '/docs/:section' })
+// or
+defineRequest<Doc>()({ method: 'GET', path: '/docs' })  // then pass { section } as a param
+```
+
+### What is not affected
+
+The check reads the **path literal**, so these all still compile and behave
+exactly as before — they are caught by the runtime error instead:
+
+```ts
+const path: string = loadFromConfig()
+defineRequest<Doc>()({ method: 'GET', path })          // not a literal
+
+const cfg: RequestConfig = { method: 'GET', path: '/docs#section' }
+defineRequest<Doc>()({ ...cfg, path: cfg.path })       // widened to string
+
+new Request<P, Doc>({ method: 'GET', path: '/docs#section' })  // no path literal to read
+```
+
+`new Request` has no equivalent check because its constructor takes no path type
+parameter to read a literal from. That asymmetry is the reason `defineRequest`
+exists.
+
+---
+
+## Upgrading to 4.2.1
+
+One change needs action, and only if a `path` or `baseUrl` of yours contains a
+`#`.
+
+### A URL fragment is now refused
+
+A fragment is never transmitted — `fetch` strips it — so one in a request URL
+could never do what it looked like it did. Worse, it silently swallowed the
+query string:
+
+```ts
+new Request({ method: 'GET', path: '/docs#section' })
+
+// 4.2.0 and earlier
+await api.docs({ page: 2 })
+// URL built:     '/docs#section?page=2'
+// what was sent: path '/docs', NO query at all -- page=2 was dropped, silently
+
+// 4.2.1
+await api.docs({ page: 2 })
+// error.kind === 'network'
+// "A URL fragment is never sent to the server, so it cannot appear in a path.
+//  Remove "#section" from "/docs#section"."
+```
+
+**What to do:** delete the fragment from the template. Nothing else changes —
+the params that were being dropped now reach the server.
+
+If neither your `baseUrl` nor any `path` contains a `#`, this release is a
+no-op for you, and the `baseUrl` fix below needs nothing from you either.
+
+### A `baseUrl` with a query string now works
+
+No action required — this only replaces broken output with correct output. If
+your `baseUrl` carried a query string (`https://api.example.com/v1?key=abc`),
+the path was previously appended *inside* the query value, so requests resolved
+to the base path and went to the wrong endpoint. They now go where they should,
+with the base's params merged ahead of the call's.
+
+## Upgrading to 4.0.0
+
+One rule changes: **a success must carry data.** If you declared
+`responseType: 'none'` on the endpoints 3.1.0's warning named, this release is
+a no-op for you.
+
+### 1. An empty body under `responseType: 'json'` is now an error
+
+```ts
+// A DELETE that answers 204 No Content, responseType left at the default:
+
+// 3.x
+const { data, error } = await api.deleteUser({ id: '42' })
+// error === null, data === null  -- and `data.deleted` compiles, then throws
+
+// 4.0.0
+const { data, error } = await api.deleteUser({ id: '42' })
+// error.kind === 'parse', error.status === 204, data === null
+```
+
+This is what makes `SuccessResult.data: TResponse` true. 3.0.0 made `Result<T>`
+a discriminated union so `if (error) return` narrows `data` — but an empty body
+produced `null` behind a non-null `TResponse`, so the narrowing was a lie for
+exactly the endpoints least likely to be checked.
+
+**The fix, on every endpoint that answers with no body:**
+
+```ts
+const deleteUser = new Request<{ id: string }, undefined>({
+  method: 'DELETE',
+  path: '/users/:id',
+  responseType: 'none',   // available since 3.1.0
+})
+```
+
+`'none'` reads no body on success, so it never reaches the rule. Non-2xx
+responses are unaffected — their body is still read and parsed for
+`error.body`, on `'none'` as on `'json'`.
+
+**There is no 204 special case.** One rule — declared JSON, got no JSON —
+applies at every status. A `200` with an empty body behaves identically to a
+`204`.
+
+**A literal `null` body still succeeds.** `JSON.parse("null")` is valid JSON, so
+a server that sends the body `null` is sending data. Only a zero-length body is
+an error.
+
+**Which endpoints are affected?** 3.1.0 told you, by name, once per request:
+any endpoint that logged `[apify] <name>: server returned an empty body`. If you
+are coming from 3.1.0 and never saw that warning in development or staging, no
+endpoint of yours hits this path.
+
+### 2. A GraphQL response with no `data` is now an error
+
+The same rule, at the GraphQL client's own seam. A 2xx response carrying
+neither `data` nor `errors` used to resolve as a success with `data: null`:
+
+```ts
+// Server returns 200 with body {}  (or "", or {"data": null})
+
+// 3.x
+const { data, error } = await client.getUser({ id: '1' })
+// error === null, data === null
+
+// 4.0.0
+const { data, error } = await client.getUser({ id: '1' })
+// error.kind === 'parse', error.status === 200, error.body === '{}'
+```
+
+`error.body` carries the raw response text, which is the only useful answer to
+"then what did the server send?".
+
+**GraphQL errors are unchanged.** A `{"data": null, "errors": [...]}` response —
+the legitimate shape for a field error — still reports `kind: 'http'` with the
+errors in `error.body` and any partial result in `error.partialData`, exactly as
+before. Only a response reporting *no* errors and *no* data is affected, which
+the GraphQL over HTTP spec does not permit.
+
+### 3. The one-time empty-body warning is gone
+
+3.1.0's `console.warn` has served its purpose and is removed. Nothing replaces
+it — the condition it warned about is now reported as an error through the
+normal `Result`.
+
+### Nothing to do if…
+
+- every endpoint that returns no body already declares `responseType: 'none'`, or
+- you never saw 3.1.0's empty-body warning, or
+- your GraphQL server always answers with `data` or `errors` (i.e. it is
+  spec-compliant).
+
+In those cases 4.0.0 is a drop-in upgrade.
+
+---
+
+## Upgrading to 3.1.0
+
+3.1.0 is additive — no existing behaviour changed, and no action is required
+to upgrade. It adds one new `responseType` option and a diagnostic warning;
+read on if either applies to you.
+
+### 1. New: `responseType: 'none'`
+
+Declares that an endpoint returns no body on success — the accurate
+declaration for a `204`, or a `200` with an empty body, most commonly a
+`DELETE`. Before 3.1.0, that shape only had `responseType: 'json'` (the
+default) to reach for, which parses an empty body to `data: null` at runtime
+while `TResponse` claims otherwise:
+
+```ts
+// Before — TResponse widened to admit the null the endpoint actually returns
+const deleteUserBefore = new Request<{ id: string }, { deleted: boolean } | null>({
+  method: 'DELETE',
+  path: '/users/:id',
+})
+const before = await api.deleteUserBefore({ id: '42' })
+if (before.error) return
+if (before.data) console.log(before.data.deleted) // null-check required even though error was null
+
+// After — responseType: 'none' says exactly what happens: no body, ever
+const deleteUserAfter = new Request<{ id: string }, undefined>({
+  method: 'DELETE',
+  path: '/users/:id',
+  responseType: 'none',
+})
+const after = await api.deleteUserAfter({ id: '42' })
+if (after.error) return
+console.log(after.data) // undefined -- no null-check needed, and none is possible
+```
+
+Declare `TResponse` as `undefined` when using `responseType: 'none'` — this
+is a convention, not a compile-time guarantee, and `new Request<P,
+User>({ responseType: 'none' })` compiles without error. A non-2xx response
+is unaffected: its body is still read and parsed as JSON for `error.body`,
+since an error body (a message, a code) is worth reading even when the
+caller wants nothing back on success.
+
+**Known limitation:** the compiler does not check the `responseType: 'none'`
+/ `TResponse` pairing. Mismatch them and `data` is `undefined` at runtime
+behind whatever type you declared, with no compile error to catch it.
+
+### 2. You may see a one-time console warning
+
+If any existing endpoint declares `responseType: 'json'` (the default) and
+the server answers with an empty body, 3.1.0 now logs this once per request
+name, per `createApi` instance:
+
+```
+[apify] deleteUser: server returned an empty body for responseType 'json'. This yields data: null today and will be an error in 4.0.0. Declare responseType: 'none' if the endpoint returns no content.
+```
+
+This is a diagnostic, not a behaviour change — the call still resolves as a
+success with `data: null`, exactly as it always has. It means the named
+request is one of the empty-body endpoints described above, and declaring
+`responseType: 'none'` on it both makes `TResponse` accurate and silences the
+warning.
+
+### 3. 4.0.0: the empty-body rule (shipped)
+
+This shipped. See [Upgrading to 4.0.0](#upgrading-to-400) — an empty body under
+`responseType: 'json'` is now a `kind: 'parse'` error, and declaring
+`responseType: 'none'` makes that upgrade a no-op.
+
+## Upgrading to 3.0.0
+
+3.0.0 tightens contracts the library always implied but never enforced. Most
+of it is types catching up to behaviour that was already there; the error
+*classification* changes (parse failures, aborts) change what a `Result`
+actually contains for a narrow set of cases. Read the "Nothing to do if…"
+section at the end first — it covers the common case.
+
+### 1. `Result<T>` is a discriminated union, not an interface
+
+`data` and `error` used to be independent nullable fields, so `if (error)
+return` never narrowed `data` — every consumer had to write `data!` or a
+redundant null check. `Result<T>` is now `SuccessResult<T> | ErrorResult<T>`
+(both exported), and checking `error` narrows `data` for real:
+
+```ts
+const { data, error } = await api.getUser({ id: '42' })
+if (error) return
+// 2.x → data: User | null, so data!.name (or a redundant `if (!data) return`)
+// 3.0.0 → data: User, so data.name — the `!` and the redundant check can go
+console.log(data.name)
+```
+
+**If you have custom middleware that synthesises a *success* `Result`**
+(short-circuits with `{ data, error: null, response, retry }` rather than
+calling `next()`), it must now supply a non-null `Response` — the type no
+longer allows `response: null` on the success branch. There is no runtime
+change here: the library's own factories (`createSuccessResult` and friends)
+already only ever produced exactly these shapes, so this is a compile-time
+tightening, not a behaviour change, for any middleware that was already
+well-formed.
+
+### 2. `Request` generics are no longer interchangeable
+
+`Request<TParams, TResponse>` never referenced its own generics in the class
+body, so every instantiation was structurally identical to TypeScript and
+`Request<{ id: string }, User>` silently accepted a `Request<{ slug: string
+}, Post>` wherever one was expected. Phantom fields now make the generics
+load-bearing:
+
+```ts
+const getUser = new Request<{ id: string }, User>({ method: 'GET', path: '/users/:id' })
+const getPost = new Request<{ slug: string }, Post>({ method: 'GET', path: '/posts/:slug' })
+
+function useRequest(r: Request<{ id: string }, User>) { /* ... */ }
+
+useRequest(getUser)  // fine, always was
+useRequest(getPost)
+// 2.x   → compiled (both Requests looked identical to the type system)
+// 3.0.0 → compile error: Request<{ slug: string }, Post> is not assignable
+//         to Request<{ id: string }, User>
+```
+
+**What to do:** if this fires after upgrading, the assignment was already
+wrong — the two `Request`s describe different endpoints and were never
+actually interchangeable at runtime. Fix the annotation to match the real
+endpoint.
+
+### 3. `ApiError.kind` is required, and gained `'middleware'`
+
+`kind` shipped optional in 2.2.0; every construction site inside the library
+already set it, so this is the type catching up. `ApiErrorKind` is now
+`'http' | 'network' | 'abort' | 'timeout' | 'parse' | 'middleware'`.
+
+```ts
+// Custom middleware constructing its own ApiError (e.g. to short-circuit
+// with a validation failure) must now supply `kind`:
+new ApiError({
+  status: 422,
+  kind: 'http',            // 3.0.0: required — omitting this is a type error
+  statusText: 'Unprocessable Entity',
+  body: { message: 'invalid payload' },
+  headers: new Headers(),
+  request: { method: 'POST', url, params },
+})
+```
+
+**If you have an exhaustive `switch (error.kind)`** (or a mapped type keyed on
+`ApiErrorKind`), it needs a new `'middleware'` arm — see #5 below for what
+produces it.
+
+### 4. Parse failures on a 2xx response changed shape
+
+**Scope: 2xx responses only.** A 2xx response whose body failed to parse
+according to `responseType` previously reported `status: 0`, `response:
+null`, `kind: 'network'` — indistinguishable from being offline, and the
+`Response` (status, headers) was discarded even though the server did
+respond. It now reports the real `status`, a non-null `response`, and `kind:
+'parse'`:
+
+```ts
+const { error } = await api.getUser({ id: '42' })  // server sent 200 with an unparseable body
+// 2.x   → error.status === 0, error.response === null, error.kind === 'network'
+// 3.0.0 → error.status === 200, error.response !== null, error.kind === 'parse'
+```
+
+**Non-2xx responses are unaffected.** `!response.ok` is checked before the
+body is parsed, so a 5xx with an unparseable body already reported (and still
+reports) `kind: 'http'` with the real status — **`retryMiddleware`'s default
+5xx retry behaviour has not changed.** Do not treat this as "parse errors are
+now retried differently"; only the 2xx case moved.
+
+**What to do:** code that branched on `status === 0` (or `kind === 'network'`)
+to mean "the user is offline" now needs to also handle `kind: 'parse'`
+explicitly if it wants to keep distinguishing "offline" from "the server
+responded with something we couldn't read." Code that only checked `if
+(error)` and logged generically needs no changes.
+
+### 5. A throwing middleware returns a `Result` instead of rejecting
+
+`composeMiddleware`'s dispatch has no guard, so an `async` middleware that
+threw used to reject the caller's promise — breaking the "never throws"
+contract for the one path most likely to have a bug (your own middleware). It
+now produces an ordinary `Result` with `kind: 'middleware'`:
+
+```ts
+const buggyMiddleware: Middleware = async (ctx, next) => {
+  throw new Error('oops')
+}
+
+// 2.x
+try {
+  const result = await api.getUser({ id: '42' })
+} catch (err) {
+  // had to catch here — a middleware bug rejected the call
+}
+
+// 3.0.0 — no try/catch needed; remove it
+const { error } = await api.getUser({ id: '42' })
+if (error?.kind === 'middleware') {
+  // error.body is the Error the middleware threw
+}
+```
+
+**What to do:** delete any `try`/`catch` you placed around an API call
+specifically to catch a middleware's throw. It's dead code now — the call
+never rejects — and the failure is available as `error.kind === 'middleware'`
+instead.
+
+### 6. `onError` no longer fires for `error.kind === 'abort'`
+
+Every `dedupe` supersede and every caller-initiated cancellation used to reach
+`onError` — i.e. your Sentry — as a reported error, even though the library
+caused it deliberately. `'timeout'` is unaffected and still fires: a deadline
+you missed is a real failure, unlike a cancellation you caused yourself.
+
+```ts
+const controller = new AbortController()
+const promise = api.getUser({ id: '42' }, { signal: controller.signal })
+controller.abort()
+await promise
+// 2.x   → onError(error) fires with error.kind === 'abort'
+// 3.0.0 → onError does not fire; the caller still gets the abort Result back
+```
+
+**What to do:** delete any hand-rolled filtering you added to your `onError`
+handler to ignore `AbortError`/cancellations (e.g. `if (error.body?.name ===
+'AbortError') return`) — the library now does this for you, unconditionally,
+for every abort it produces.
+
+### 7. Aborts are classified by provenance, not by the reason's name
+
+This is the change most likely to surface silently, because it changes
+`kind` for shapes that used to look like something else entirely.
+Previously, the library guessed a cancellation by sniffing the *thrown
+value's* `.name` (`'AbortError'` / `'TimeoutError'`). Now it asks a different
+question: **did the signal that actually governs this request abort?** If
+yes, the failure is `'abort'`/`'timeout'` regardless of what was thrown or
+what it's named; if no, sniffing the thrown value's shape is the fallback.
+Four consequences:
+
+- **A caller's own custom abort reason is now silent, not reported.**
+  `controller.abort(new Error('unmounted'))` or `controller.abort('cancelled')`
+  used to fail the old name-based sniff (the reason isn't named
+  `AbortError`), so it fell through to `kind: 'network'` and reached
+  `onError`. It's now `kind: 'abort'` — correctly classified as *your*
+  cancellation — and, per #6, `onError` doesn't fire for it.
+
+  ```ts
+  controller.abort(new Error('component unmounted'))
+  // 2.x   → error.kind === 'network', reported to onError
+  // 3.0.0 → error.kind === 'abort', not reported
+  ```
+
+  **What to do:** if you were branching on `error.kind === 'network'` (or
+  `error.status === 0`) to mean "offline," and relying on a custom abort
+  reason to *also* hit that branch, it no longer will. Branch on `kind ===
+  'abort'` explicitly if you still want to observe your own cancellations.
+
+- **A middleware that propagates the library's own abort reason now returns a
+  silent `kind: 'abort'` `Result`, instead of rejecting the caller's promise
+  outright.** This covers both re-throwing the exact reason (`throw
+  ctx.request.signal.reason`) and the shape `node:timers/promises` and most
+  abortable helpers actually produce — a fresh `AbortError` whose `.cause` is
+  the signal's reason:
+
+  ```ts
+  import { setTimeout as delay } from 'node:timers/promises'
+
+  const backoff: Middleware = async (ctx, next) => {
+    await delay(100, undefined, { signal: ctx.request.signal })
+    return next()
+  }
+  // if ctx.request.signal aborts during the delay, `delay` rejects with an
+  // AbortError whose `.cause` is ctx.request.signal.reason
+  // 2.x   → non-shared: composeMiddleware's chain had no rejection handler at
+  //         all, so the caller's own promise rejects with the raw AbortError
+  //         — no Result, "kind" does not apply, onError never runs.
+  //         Under `share: true` specifically, this *was* already converted to
+  //         a Result (the share site's own rejection handler, present since
+  //         2.2.0), classified `kind: 'abort'` by sniffing the thrown value's
+  //         `.name` — the same name-based sniff #7's intro paragraph
+  //         describes, so it could not tell this genuine propagation apart
+  //         from bullet 3's unrelated-failure case below. Reported either way
+  //         (no abort suppression existed yet).
+  // 3.0.0 → error.kind === 'abort' (recognised as propagating our own signal,
+  //         not merely name-matched), returned as an ordinary Result on every
+  //         path, not reported
+  ```
+
+  **What to do:** delete any `try`/`catch` you placed around this kind of
+  call for the same reason as #5. Code was already correct if it treated this
+  as `kind: 'abort'` — it just could not have relied on that being *reliable*
+  (see bullet 3, which used to collide with this one under the old
+  name-based sniff). Middleware authors: see the "worth knowing" note below
+  about not attaching `signal.reason` as `.cause` to your *own* unrelated
+  failures — doing so makes them indistinguishable from this propagation
+  case.
+
+- **A middleware throwing its own, unrelated `AbortError`-named failure now
+  reaches `onError` classified as `'middleware'`, as an ordinary `Result`,
+  instead of rejecting the caller's promise outright.** An IndexedDB quota
+  abort, say, rethrown by a caching middleware, has nothing to do with this
+  request's own signal:
+
+  ```ts
+  // 2.x   → non-shared: composeMiddleware's chain had no rejection handler at
+  //         all, so the caller's own promise rejects with the raw AbortError
+  //         — no Result, "kind" does not apply, onError never runs.
+  //         Under `share: true`, this was already converted to a Result, but
+  //         misclassified `kind: 'abort'` by the same name-based sniff as
+  //         bullet 2 above — 'middleware' did not exist as a kind at all
+  //         before this release, on either path — and was reported (no
+  //         suppression existed for 'abort' yet either).
+  // 3.0.0 → error.kind === 'middleware' (not a propagation of our signal),
+  //         returned as an ordinary Result on every path, reported
+  ```
+
+  **What to do:** nothing to change in your code, but expect to *start*
+  seeing these correctly labelled `kind: 'middleware'` instead of either an
+  unhandled rejection (non-shared) or a misleading `kind: 'abort'` (shared) —
+  if you have middleware that can throw an `AbortError`-named failure
+  unrelated to request cancellation. See #5 above: this is the same
+  "throwing middleware now returns a Result" change, just for a failure that
+  happens to be named like an abort.
+
+- **Any abort of the exact signal handed to `fetch` — including one installed
+  by middleware — is now `'abort'`/`'timeout'` and silent**, not classified by
+  what was thrown. A middleware that replaces `ctx.request.signal` (e.g. a
+  per-attempt timeout) and whose replacement signal aborts now gets the same
+  cancellation treatment as any other abort on the operative signal.
+
+**What to do, generally:** grep your `onError` handler and any code branching
+on `error.kind` or `error.status === 0` for logic that assumed "not named
+`AbortError`" meant "not a cancellation," or that assumed `kind ===
+'middleware'` was reserved for genuine middleware bugs. Both assumptions are
+now wrong in the specific ways above.
+
+### 8. Cancelling during the response body download is now classified as the cancellation
+
+Aborting after headers arrive but while the body is still downloading — a
+component unmounting mid-fetch, a deadline firing mid-download — used to be
+misclassified for a **non-2xx** response specifically. The **2xx** case
+already produced the right `kind`/`status`/`response` shape in 2.2.1; only
+whether it was *reported* changes there, which is entirely #6 (`onError` no
+longer fires for `'abort'`), not a distinct shape change:
+
+```ts
+// A slow response body, aborted partway through download:
+// 2xx response (e.g. a slow success payload):
+//   2.x   → kind: 'abort' (or 'network', if the thrown value wasn't
+//           name-recognisable as AbortError/TimeoutError), status: 0,
+//           response: null — reported (2.2.1 had no abort suppression at all)
+//   3.0.0 → kind: 'abort' (or 'timeout'), status: 0, response: null — not
+//           reported (same shape, see #6 for why reporting stops)
+// non-2xx response (e.g. a slow 502 gateway page) — this is the real shape change:
+//   2.x   → kind: 'http', the real status, response present, body: null —
+//           reported, regardless of whether the cancellation was ours
+//   3.0.0 → kind: 'abort' (or 'timeout'), status: 0, response: null — not reported
+```
+
+Two concrete hazards to check for, both specific to the **non-2xx** case:
+
+- **Code branching on `error.status` for a cancellation that lands while a
+  non-2xx body downloads** now sees `status: 0` instead of the real status —
+  the same "was this reported?" question as #6/#7 applies.
+- **Code reading `result.response!.headers` (or any non-null assertion on
+  `response`) for that same non-2xx-cancellation case** must now handle
+  `response === null` — it previously had a real `response` (with `body:
+  null`), even though the request never actually finished.
+
+**Also:** `retryMiddleware`'s default `retryOn` (and any custom `retryOn`
+keyed on `status >= 500`) no longer retries a cancellation caught in this
+window, because `status` is now `0`, not the real (often 5xx-shaped) status.
+This is strictly correct — retrying a cancellation you caused is never
+useful — but it means **fewer requests** for code that happened to rely on
+the old misclassification triggering a retry.
+
+### 9. GraphQL partial data is preserved (additive)
+
+A GraphQL response with both `data` and `errors` (partial success — a
+nullable field errored while the rest of the query resolved) used to discard
+`data` entirely. It's now available as `error.partialData`:
+
+```ts
+const { error } = await graphql.getCategory({ id: '123' })
+if (error) {
+  console.log(error.body)          // GraphQLError[]
+  console.log(error.partialData)   // the data the server sent alongside the errors, or undefined
+}
+```
+
+This lives on `error.partialData`, not `result.data` — putting it on `data`
+would break the `Result` union's narrowing from #1: a non-null `error` means
+`data` is null, and a null `error` means the call succeeded (see the
+empty-body caveat below). Nothing to change unless you want to start using
+it.
+
+### Worth knowing, no action needed
+
+- **A genuinely malformed body arriving while the signal *happens* to already
+  be aborted is classified as the cancellation, and so dropped from
+  `onError`**, even when the abort didn't actually cause the parse failure.
+  The guard asks "is the signal aborted right now," not "did the abort cause
+  this" — narrowing that further would require `parseResponse` to
+  distinguish its own read failure from a parse failure across all five
+  response types, which it doesn't attempt. In practice this only matters if
+  you're relying on `onError` to catch every malformed-body case with
+  certainty; it's a narrow, pre-existing edge case, not a new hazard to
+  design around.
+- **Middleware authors: do not attach `signal.reason` as `.cause` to your
+  own, unrelated failure.** `throw new Error('cache write failed', { cause:
+  ctx.request.signal.reason })` is read as *relaying* the library's own
+  cancellation (see #7's `.cause`-unwrapping case) and silently dropped as
+  `'abort'`, even though your error is about something else entirely (a cache
+  write, not a cancellation) that merely happened to occur while the signal
+  was aborted. Use a different field to carry that context —
+  `{ cause: new Error('disk full') }`, or a custom property — and reserve
+  `.cause` for genuine propagation of the signal's own reason.
+- **`SuccessResult.data` is typed `TResponse`, not `TResponse | null` — but an
+  empty body still parses to `null` at runtime.** A `DELETE` that answers
+  `200` with no body (or a `204`) is one of the most common REST shapes, and
+  `parseResponse` returns `null` for it exactly as it always has (see the
+  `responseType` reference for the empty-body case). 3.0.0 removes the `| null`
+  from the *type*, not from the *behaviour*: `error` is still `null` on that
+  response, so `data` narrows to `TResponse` and compiles, but the value you
+  get is `null` anyway.
+
+  ```ts
+  const deleteUser = new Request<{ id: string }, { deleted: boolean }>({
+    method: 'DELETE', path: '/users/:id',
+  })
+  const { data, error } = await api.deleteUser({ id: '42' })
+  if (error) return
+  console.log(data.deleted) // throws: data is null at runtime for a 200/204 empty body
+  ```
+
+  If an endpoint can answer 204 or an empty 200, say so accurately in its own
+  `TResponse`. **This guidance changed in 3.1.0:** at the time of the 3.0.0
+  release, the only option was widening to `Request<{ id: string }, { deleted:
+  boolean } | null>` and handling the `null` case explicitly; as of 3.1.0, use
+  `responseType: 'none'` instead (see [Upgrading to
+  3.1.0](#upgrading-to-310) above) — it's more accurate (declares "no body,"
+  not "body or null") and it silences the empty-body warning 3.1.0 added. This
+  is not a new *behaviour* (2.2.1 had the identical runtime `null`); what
+  changed in 3.0.0 is that the type system no longer forces you to handle it,
+  and what changed in 3.1.0 is the recommended way to handle it anyway.
+
+### Nothing to do if…
+
+…you only call API methods and check `error`, **and every endpoint's
+`TResponse` accounts for its own empty-body responses** (see the bullet
+above — as of 3.1.0, a `DELETE`/204/empty-200 endpoint should declare
+`responseType: 'none'` to stay accurate; everything else needs no change):
+
+```ts
+const { data, error } = await api.getUser({ id: '42' })
+if (error) {
+  console.error(error.status, error.body)
+  return
+}
+console.log(data.name)
+```
+
+The only change visible here is a good one: `data!.name` becomes `data.name`
+(the `!` is now unnecessary and can be deleted, but leaving it is harmless —
+a non-null assertion on an already-non-null value is a no-op). No behavior
+changes for this pattern.
+
+## Upgrading to 2.2.0
+
+2.2.0 is additive — no API changed shape, and there is nothing you need to do.
+
+### Worth knowing, no action needed
+
+- **Abort reasons now survive `dedupe`.** Previously, a `dedupe: true`
+  request's merged signal always aborted with a generic, reason-less
+  `AbortError`, regardless of what actually caused the abort:
+
+  ```ts
+  const withTimeout: Middleware = async (ctx, next) => {
+    ctx.request.signal = AbortSignal.timeout(20)
+    return next()
+  }
+
+  const api = createApi({
+    baseUrl: '/api',
+    requests: { getUser }, // getUser has dedupe: true
+    middleware: [withTimeout],
+  })
+
+  const { error } = await api.getUser({ id: '42' })
+  // 2.1.0 → error.body.name === 'AbortError'   -- the real cause (a timeout) was lost
+  // 2.2.0 → error.body.name === 'TimeoutError' -- the actual cause survives
+  ```
+
+  This only differs when `dedupe: true` is combined with a caller-supplied
+  `signal` or a signal-setting middleware — plain `dedupe: true` with no
+  external signal involved is unaffected. `error.body.name` (and any custom
+  reason you pass to your own `AbortController.abort(reason)`) is a
+  **pre-2.2.0 surface** — existing code reading it does not need to touch
+  anything new to notice this, since it never had to opt into `kind` to read
+  `.name` in the first place. Going forward, prefer branching on the new
+  `error.kind` (`'timeout'` vs `'abort'` vs `'network'`) instead of
+  `error.body.name` — it's the field the library commits to maintaining.
+
+- **Retries now back off instead of firing instantly.** `retryMiddleware(3)`
+  previously made all four attempts in the same tick, with no delay between
+  them. It now waits out a real backoff (exponential by default, with jitter)
+  between attempts, so a retrying request takes measurably longer in
+  wall-clock terms. Nothing breaks, but a test asserting on elapsed time
+  around a retrying call may need its tolerance revisited — or pass
+  `retryMiddleware({ baseDelay: 0, jitter: false })` to keep the old, instant
+  timing.
+
+## Upgrading to 2.1.0
+
+2.1.0 is a non-breaking release, but **two changes can surface as new errors** in
+code that was already subtly wrong. Neither requires an API change on your side.
+
+### The Node floor moved from 18 to 20
+
+Node 18 reached end-of-life on 2025-04-30, and CI never tested it — the "Node 18+"
+claim in the README was untested, which is worse than not making it. `package.json`
+now declares `"engines": { "node": ">=20" }`.
+
+**What to do:** if you are on Node 18, nothing breaks today — the library uses no
+API that Node 18 lacks. But the version is unsupported and untested, so treat this
+as notice rather than a guarantee.
+
+### Path parameter mismatches now fail loudly
+
+Previously, a mismatch between a request's params and its path template shipped a
+malformed URL and said nothing:
+
+```ts
+const getUser = new Request<{ id: string }, User>({
+  method: 'GET',
+  path: '/users/:userId',   // note: :userId, but the param is `id`
+})
+
+await api.getUser({ id: '42' })
+// 2.0.0 → GET /api/users/:userId?id=42   ← literal token, value duplicated as a query param
+// 2.1.0 → Result with an error; no request is made
+```
+
+The 2.1.0 result is an ordinary error `Result`, not a thrown exception — the
+never-throws contract is intact:
+
+```ts
+const { error } = await api.getUser({ id: '42' })
+// error.status === 0
+// String(error.body) === 'TypeError: Unresolved path parameter :userId in path "/users/:userId". …'
+```
+
+**What to do:** if this fires after upgrading, the endpoint was making a malformed
+request all along. Align the param name with the path token (or vice versa). The
+error message names the offending token and the path.
+
+**Related:** `baseUrl` and `path` now join with exactly one slash, so a `baseUrl`
+ending in `/` no longer produces `//`. If a server was tolerating (or redirecting)
+the double slash, requests will now go to the correct URL — worth checking if you
+have path-sensitive routing or logging.
+
+### Worth knowing, no action needed
+
+- `ctx.request.signal` is now readable and writable from middleware, which makes
+  timeouts and cancel-on-condition writable in userland for the first time:
+
+  ```ts
+  const timeout = (ms: number): Middleware => async (ctx, next) => {
+    ctx.request.signal = AbortSignal.timeout(ms)
+    return next()
+  }
+  ```
+
+  It composes with `dedupe: true` — dedupe merges whatever signal is current
+  rather than discarding it.
+
+- The published package is roughly half its former size (240.6 kB → 112.7 kB
+  unpacked). Source maps were dropped: they referenced `../src/*.ts`, which was
+  never published, and carried no `sourcesContent`, so they resolved to nothing
+  in every consumer. JSDoc still ships in the `.d.ts` files, so editor hover
+  documentation is unchanged.