@effect-app/vue 4.0.0-beta.98 → 4.0.0

package/src/atomQuery.ts

@@ -0,0 +1,600 @@
+/* eslint-disable @typescript-eslint/no-explicit-any */
+/**
+ * Shared atom core for the query/mutation engine (used by query.ts + mutate.ts).
+ *
+ * Replaces the @tanstack/vue-query engine with Effect `Atom`, keeping the public
+ * `.query()/.suspense()/.mutate()` contract unchanged:
+ *   - cache identity   = the atom reference (one per [handler, input], via Atom.family)
+ *   - invalidation key = reactivity keys (= the app's existing namespace query keys)
+ *   - SWR + focus      = Atom.swr (+ windowFocusSignal)
+ *   - gcTime           = Atom.setIdleTTL / Atom.keepAlive
+ *   - retry            = Effect.retry inside the atom effect
+ *
+ * Built over the app's RPC client through the existing `RequestHandlerWithInput`
+ * abstraction (mirrors AtomRpc's recipe; see docs/atom-query-plan.md).
+ */
+import { defaultRegistry } from "@effect/atom-vue"
+import { DataDependencies, makeQueryKey } from "effect-app/client"
+import type { ClientForOptions } from "effect-app/client/clientFor"
+import { ServiceUnavailableError } from "effect-app/client/errors"
+import * as Effect from "effect-app/Effect"
+import * as Option from "effect-app/Option"
+import * as S from "effect-app/Schema"
+import * as Cause from "effect/Cause"
+import * as Duration from "effect/Duration"
+import * as Equal from "effect/Equal"
+import * as Hash from "effect/Hash"
+import { isHttpClientError } from "effect/http/HttpClientError"
+import type * as Layer from "effect/Layer"
+import * as AsyncResult from "effect/reactivity/AsyncResult"
+import * as Atom from "effect/reactivity/Atom"
+import * as AtomRegistry from "effect/reactivity/AtomRegistry"
+import * as Ref from "effect/Ref"
+import * as Stream from "effect/Stream"
+import type * as Tracer from "effect/Tracer"
+import { clearQueryReadDependencies, getQueryReadDependencies, type QueryInvalidationMode, registerQueryInvalidationMode, setQueryReadDependencies } from "./dependencyMetadata.ts"
+import { reportRuntimeError } from "./lib.ts"
+import { beginLiveQueryFetch, endLiveQueryFetch, type LiveQueryOptions, registerLiveQuery } from "./liveQueryInvalidation.ts"
+import { atomsToRefetch, observeQueryAtom } from "./queryLifetime.ts"
+
+/** All non-empty prefixes of a key, longest last. `[a,b,c]` -> `[[a],[a,b],[a,b,c]]`. */
+const prefixesOf = (key: ReadonlyArray<unknown>): ReadonlyArray<ReadonlyArray<unknown>> =>
+  key.map((_, i) => key.slice(0, i + 1))
+
+const uniqueKeys = (keys: ReadonlyArray<ReadonlyArray<unknown>>): ReadonlyArray<ReadonlyArray<unknown>> => {
+  const out: Array<ReadonlyArray<unknown>> = []
+  const seen = new Set<number>()
+  for (const key of keys) {
+    const hash = Hash.hash(key)
+    if (seen.has(hash)) continue
+    seen.add(hash)
+    out.push(key)
+  }
+  return out
+}
+
+// --- awaitable invalidation -------------------------------------------------------------------
+// keyHash -> query atoms registered under that key. A query atom stays in the map while it is
+// alive in the registry (mounted OR cached within idle-ttl). Invalidation *marks* every matching
+// atom stale; it only *refetches* atoms that currently have observers (TanStack refetchType:
+// "active"). Idle cache entries refetch on the next mount.
+const keyAtoms = new Map<number, Set<Atom.Atom<AsyncResult.AsyncResult<any, any>>>>()
+
+const trackByKeys =
+  (keys: ReadonlyArray<ReadonlyArray<unknown>>) =>
+  <A, E>(atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>): Atom.Atom<AsyncResult.AsyncResult<A, E>> =>
+    Atom.transform(atom, (get) => {
+      for (const key of keys) {
+        const h = Hash.hash(key)
+        let set = keyAtoms.get(h)
+        if (!set) keyAtoms.set(h, set = new Set())
+        set.add(atom)
+        get.addFinalizer(() => {
+          set.delete(atom)
+          if (set.size === 0) keyAtoms.delete(h)
+        })
+      }
+      return get(atom)
+    }, { initialValueTarget: atom })
+
+const trackWritableByKeys =
+  (keys: ReadonlyArray<ReadonlyArray<unknown>>) =>
+  <A, E, W>(atom: Atom.Writable<AsyncResult.AsyncResult<A, E>, W>): Atom.Writable<AsyncResult.AsyncResult<A, E>, W> => {
+    const tracked = trackByKeys(keys)(atom)
+    return Atom.writable(
+      (get) => get(tracked),
+      (ctx, value) => ctx.set(atom, value),
+      (refresh) => refresh(tracked)
+    )
+  }
+
+/**
+ * Keep a query's recorded read-dependencies registered for as long as it is alive, and drop them when
+ * its atom is GC'd (mirrors `trackByKeys`).
+ *
+ * `recordReads` only stores the deps on an actual fetch. A query that unmounts and later remounts from
+ * cache (within idle-ttl) does NOT re-run the handler, so its registry entry — cleared by the previous
+ * teardown's finalizer — would never be restored, and a subsequent mutation could not derive it as an
+ * invalidation target. Re-asserting the last-known reads on every (re)subscribe closes that gap, so the
+ * registry mirrors the live query even across cache-hit remounts.
+ */
+const trackReadDependencies =
+  (key: ReadonlyArray<unknown>, getReads: () => DataDependencies.DataDependencies) =>
+  <A, E>(atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>): Atom.Atom<AsyncResult.AsyncResult<A, E>> =>
+    Atom.transform(atom, (get) => {
+      const reads = getReads()
+      if (DataDependencies.isNonEmpty(reads)) setQueryReadDependencies(key, reads)
+      get.addFinalizer(() => clearQueryReadDependencies(key))
+      return get(atom)
+    }, { initialValueTarget: atom })
+
+const atomsForKeys = (keys: ReadonlyArray<unknown>): ReadonlyArray<Atom.Atom<AsyncResult.AsyncResult<any, any>>> => {
+  const atoms = new Set<Atom.Atom<AsyncResult.AsyncResult<any, any>>>()
+  for (const key of keys) {
+    const set = keyAtoms.get(Hash.hash(key))
+    if (set) { for (const a of set) atoms.add(a) }
+  }
+  return [...atoms]
+}
+
+/** Refresh observed query atoms without making the triggering mutation await them. */
+export const invalidateSoft = (keys: ReadonlyArray<unknown>): Effect.Effect<void> =>
+  Effect
+    .gen(function*() {
+      const atoms = atomsToRefetch(atomsForKeys(keys))
+      yield* Effect.forEach(atoms, captureAtomQueryParentSpan, { discard: true, concurrency: "unbounded" })
+      if (atoms.length === 0) return
+      yield* Effect.forEach(atoms, (atom) => Effect.sync(() => defaultRegistry.refresh(atom)), {
+        discard: true
+      })
+    })
+    .pipe(Effect.orDie)
+
+/**
+ * Invalidate the given keys and AWAIT the result. `keyAtoms` resolves all matching hierarchical
+ * keys to a deduplicated set of query atoms. Refresh that set directly: sending the whole key set
+ * through `Reactivity.invalidate` invokes one atom's registered callback once per matching key,
+ * repeatedly superseding the same fetch when a mutation carries many row/prefix keys.
+ *
+ * Resolves once every **observed** matching query has settled, so a mutation can
+ * `yield*` this and know the on-screen data is fresh. Idle (observers === 0)
+ * matches are only marked stale — they do not block the mutation. Soft
+ * invalidation (`invalidateSoft`) still fires without awaiting.
+ */
+export const invalidateAndAwait = (keys: ReadonlyArray<unknown>): Effect.Effect<void> =>
+  Effect
+    .gen(function*() {
+      const atoms = atomsToRefetch(atomsForKeys(keys))
+      yield* Effect.forEach(atoms, captureAtomQueryParentSpan, { discard: true, concurrency: "unbounded" })
+      if (atoms.length === 0) return
+      yield* Effect.forEach(atoms, (atom) => Effect.sync(() => defaultRegistry.refresh(atom)), {
+        discard: true
+      })
+      // Contract: observers > 0 ⇒ wait for this refresh to finish.
+      yield* Effect.forEach(atoms, (a) => awaitAtomResult(defaultRegistry, a).pipe(Effect.exit))
+    })
+    .pipe(Effect.orDie)
+
+const isPlainObject = (o: unknown): o is Record<string, unknown> => {
+  if (typeof o !== "object" || o === null) return false
+  const proto = Object.getPrototypeOf(o)
+  return proto === Object.prototype || proto === null
+}
+
+/**
+ * Structural sharing (tanstack `replaceEqualDeep`): walk `next` against `prev` and reuse `prev`'s
+ * reference for any unchanged array/object sub-tree, so unchanged data keeps referential identity
+ * (Vue skips re-rendering it). Leaves — including decoded Schema CLASS INSTANCES — are compared with
+ * Effect `Equal.equals` (structural), which reuses equal instances that tanstack's `===` could not.
+ */
+export const replaceEqualDeep = (prev: any, next: any): any => {
+  if (prev === next) return prev
+  const bothArrays = Array.isArray(prev) && Array.isArray(next)
+  if (bothArrays || (isPlainObject(prev) && isPlainObject(next))) {
+    const a: Record<PropertyKey, any> = prev
+    const b: Record<PropertyKey, any> = next
+    const copy: Record<PropertyKey, any> = bothArrays ? [] : {}
+    const nextKeys: Array<PropertyKey> = bothArrays ? (b as Array<any>).map((_, i) => i) : Object.keys(b)
+    const nextSize = nextKeys.length
+    const prevSize = bothArrays ? (a as Array<any>).length : Object.keys(a).length
+    let equalItems = 0
+    for (let i = 0; i < nextSize; i++) {
+      const key = nextKeys[i]!
+      copy[key] = replaceEqualDeep(a[key], b[key])
+      if (copy[key] === a[key] && a[key] !== undefined) equalItems++
+    }
+    return prevSize === nextSize && equalItems === prevSize ? prev : copy
+  }
+  return Equal.equals(prev, next) ? prev : next
+}
+
+/** Atom combinator: share each new `Success` value structurally against the previous one. */
+const structuralShare = <A, E>(
+  self: Atom.Atom<AsyncResult.AsyncResult<A, E>>
+): Atom.Atom<AsyncResult.AsyncResult<A, E>> =>
+  Atom.transform(self, (get) => {
+    const next = get(self)
+    if (next._tag !== "Success") return next
+    const prev = Option.flatMap(get.self<AsyncResult.AsyncResult<A, E>>(), AsyncResult.value)
+    if (Option.isNone(prev)) return next
+    const shared = replaceEqualDeep(prev.value, next.value)
+    return shared === next.value
+      ? next
+      : AsyncResult.success(shared, { waiting: next.waiting, timestamp: next.timestamp })
+  }, { initialValueTarget: self })
+
+export interface AtomClientRuntime {
+  readonly runtime: Atom.AtomRuntime<any, never>
+  readonly factory: Atom.RuntimeFactory
+}
+
+/**
+ * Build one AtomRuntime (and its factory) from an already-built app context.
+ * Shares the ManagedRuntime's `memoMap` so layers are not built twice, and so
+ * query-registration and mutation-invalidation resolve the SAME `Reactivity`.
+ */
+export const makeAtomClientRuntime = <R>(
+  getContext: () => Layer.Layer<R, never, never>,
+  memoMap: Layer.MemoMap
+): AtomClientRuntime => {
+  const factory = Atom.context({ memoMap })
+  const runtime = factory((_get) => getContext())
+  return { runtime, factory }
+}
+
+const isRetryable = (e: unknown): boolean =>
+  isHttpClientError(e)
+  || S.is(ServiceUnavailableError)(e)
+  || (typeof e === "object" && e !== null && (e as any)._tag === "RpcClientError")
+
+export interface AtomQueryOptions {
+  /** background-refresh threshold (TanStack staleTime; default 5s) */
+  readonly staleTime?: Duration.Input
+  /** dispose-when-idle (TanStack gcTime; default 5min). "infinity" => keepAlive */
+  readonly gcTime?: Duration.Input | "infinity"
+  readonly invalidation?: QueryInvalidationMode
+  /**
+   * Revalidate a stale query on window focus AND on network reconnect (default on, matching
+   * tanstack refetchOnWindowFocus + refetchOnReconnect).
+   */
+  readonly revalidateOnFocus?: boolean
+  /**
+   * Reuse references of unchanged sub-trees across refetches (default on, matching tanstack
+   * structuralSharing). Uses Effect `Equal` so decoded Schema instances share too — more effective
+   * than tanstack's `===`, but a deep compare per refetch (O(rows·fields)). Set `false` for very
+   * large or mostly-changing result sets where the compare costs more than the saved re-renders.
+   */
+  readonly structuralSharing?: boolean
+  /** poll: re-fetch every N ms (tanstack refetchInterval). */
+  readonly refetchInterval?: number
+  /** Refresh when writes intersect the dependencies recorded by this query. */
+  readonly live?: boolean | LiveQueryOptions
+}
+
+const defaults = { staleTime: Duration.seconds(5), gcTime: Duration.minutes(5) }
+
+export interface AtomQueryMetadata {
+  readonly staleTimeMs: number
+}
+
+const atomQueryMetadata = new WeakMap<Atom.Atom<AsyncResult.AsyncResult<any, any>>, AtomQueryMetadata>()
+const atomQueryParentSpans = new WeakMap<Atom.Atom<AsyncResult.AsyncResult<any, any>>, Tracer.AnySpan>()
+const atomQueryKeys = new WeakMap<Atom.Atom<AsyncResult.AsyncResult<any, any>>, ReadonlyArray<unknown>>()
+
+export const queryKeyForAtom = <A, E>(
+  atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>
+): ReadonlyArray<unknown> | undefined => atomQueryKeys.get(atom)
+
+const atomSpanTarget = <A, E>(atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>) => {
+  let target = atom
+  while (target.initialValueTarget !== undefined) {
+    target = target.initialValueTarget
+  }
+  return target
+}
+
+const takeAtomQueryParentSpan = <A, E>(atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>) => {
+  const target = atomSpanTarget(atom)
+  const span = atomQueryParentSpans.get(target)
+  if (span !== undefined) atomQueryParentSpans.delete(target)
+  return span
+}
+
+const setAtomQueryParentSpan = <A, E>(
+  atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>,
+  span: Tracer.AnySpan
+) => {
+  atomQueryParentSpans.set(atomSpanTarget(atom), span)
+}
+
+export const captureAtomQueryParentSpan = <A, E>(
+  atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>
+): Effect.Effect<void> =>
+  Effect.currentParentSpan.pipe(
+    Effect.tap((span) => Effect.sync(() => setAtomQueryParentSpan(atom, span))),
+    Effect.ignore
+  )
+
+export const refreshAtomWithCurrentSpan = <A, E>(
+  registry: AtomRegistry.AtomRegistry,
+  atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>
+): Effect.Effect<void> =>
+  captureAtomQueryParentSpan(atom).pipe(
+    Effect.andThen(Effect.sync(() => registry.refresh(atom)))
+  )
+
+const setAtomQueryMetadata = <A, E>(
+  atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>,
+  opts: AtomQueryOptions = {}
+) => {
+  const staleTimeMs = staleTimeMsOf(opts)
+  const previous = atomQueryMetadata.get(atom)
+  atomQueryMetadata.set(atom, {
+    staleTimeMs: previous === undefined ? staleTimeMs : Math.min(previous.staleTimeMs, staleTimeMs)
+  })
+  return atom
+}
+
+export const getAtomQueryMetadata = <A, E>(
+  atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>
+): AtomQueryMetadata | undefined => atomQueryMetadata.get(atom)
+
+/** Exported so the vue hook can do refetch-on-mount-per-observer with the same rule as swr. */
+export const isStaleResult = (r: AsyncResult.AsyncResult<any, any>, staleTimeMs: number): boolean => {
+  if (r.waiting) return false
+  const ts = r._tag === "Success"
+    ? r.timestamp
+    : r._tag === "Failure"
+    ? Option.getOrUndefined(Option.map(r.previousSuccess, (s) => s.timestamp))
+    : undefined
+  if (ts === undefined) return r._tag !== "Initial"
+  return Date.now() - ts >= staleTimeMs
+}
+
+export const staleTimeMsOf = (opts: AtomQueryOptions): number =>
+  Duration.toMillis(Duration.fromInputUnsafe(opts.staleTime ?? defaults.staleTime))
+
+// A query fetch that is interrupted (a subscriber lost interest / a refresh superseded it / the
+// component navigated) leaves the family atom `waiting=true` with NO fiber running — effect-core
+// `makeEffect` removes its result-observer before interrupting, so the interrupt is never written
+// back; it stays `waitingFrom(previous)`. `swr`/`isStaleResult` both short-circuit
+// `if (waiting) return false`, so that stuck `waiting` is neither completing nor considered stale —
+// mount/focus/staleness all skip it. Left alone the interrupt is TERMINAL: any current or future
+// subscriber inherits the `waiting` forever (cold → `latestDefined` throws → white screen; warm →
+// the Mako Bauhaus PickList stale wedge).
+//
+// The invariant we restore in this layer (no effect-core change): an interrupt must never be a
+// terminal state. We do NOT auto-retry the interrupted fetch — an interrupt is intentional (that
+// caller lost interest), and the first observer recovers through its normal refresh/staleness rules.
+// We only guarantee that a *genuine* new/parallel/future observer is never stuck on a departed
+// observer's interrupt: per family atom we count live computes (`inFlight`), so "stuck" = a `waiting`
+// result with `inFlight === 0` (nothing is actually fetching). On mount a stuck atom triggers one
+// fresh fetch instead of adopting the dangling `waiting`; a genuinely in-flight fetch (`inFlight > 0`)
+// is joined, not superseded (dedup / re-entrancy preserved). `recovering` dedupes concurrent mounts
+// so exactly one recovery fetch is issued until a fetch is running again.
+export interface QueryFetchState {
+  inFlight: number
+  recovering: boolean
+}
+export const queryFetchStates = new WeakMap<Atom.Atom<any>, QueryFetchState>()
+
+const recoverStuckWaitingOnMount =
+  <A, E>(familyAtom: Atom.Atom<AsyncResult.AsyncResult<A, E>>) =>
+  (wrapped: Atom.Atom<AsyncResult.AsyncResult<A, E>>): Atom.Atom<AsyncResult.AsyncResult<A, E>> =>
+    Atom.transform(wrapped, (get) => {
+      const current = get.once(wrapped)
+      get.subscribe(wrapped, (value) => get.setSelf(value))
+      const state = queryFetchStates.get(familyAtom)
+      const waiting = current?.waiting === true
+      // Stuck: the result says `waiting` yet nothing is in-flight — an interrupt was hidden and left
+      // `waiting` dangling. Treat it as not-yet-fetched and refetch; a live fetch (`inFlight > 0`) is
+      // joined, not superseded. `recovering` prevents concurrent mounts from issuing more than one
+      // recovery fetch (cleared once a fetch is actually running again).
+      if (state && waiting && state.inFlight === 0 && !state.recovering) {
+        state.recovering = true
+        get.refresh(familyAtom)
+      }
+      return current
+    }, { initialValueTarget: wrapped })
+
+export const withQueryOptions = <A, E>(
+  self: Atom.Atom<AsyncResult.AsyncResult<A, E>>,
+  opts: AtomQueryOptions = {},
+  liveKey?: ReadonlyArray<unknown>
+): Atom.Atom<AsyncResult.AsyncResult<A, E>> => {
+  setAtomQueryMetadata(self, opts)
+  const staleTime: Duration.Input = opts.staleTime ?? defaults.staleTime
+  let atom = self
+  if (liveKey !== undefined) {
+    atom = Atom.transform(atom, (get) => {
+      const unregister = registerQueryInvalidationMode(liveKey, opts.invalidation ?? "await")
+      get.addFinalizer(unregister)
+      return get(self)
+    }, { initialValueTarget: self })
+  }
+  if (opts.live && liveKey !== undefined) {
+    const liveOptions = opts.live === true ? {} : opts.live
+    atom = Atom.transform(atom, (get) => {
+      const unregister = registerLiveQuery(
+        liveKey,
+        () => getQueryReadDependencies(liveKey),
+        liveOptions
+      )
+      get.addFinalizer(unregister)
+      return get(self)
+    }, { initialValueTarget: self })
+  }
+  const revalidateOnFocus = opts.revalidateOnFocus ?? true
+  atom = Atom.swr({
+    staleTime,
+    revalidateOnFocus,
+    focusSignal: revalidateOnFocus ? focusOrReconnectSignal : undefined
+  })(atom)
+  atom = recoverStuckWaitingOnMount(self)(atom)
+  if (opts.refetchInterval) atom = Atom.withRefresh(Duration.millis(opts.refetchInterval))(atom)
+  if (opts.structuralSharing ?? true) atom = structuralShare(atom)
+  return atom
+}
+
+/** Constant atom for disabled / `mode:"optional"`-None queries: stays Initial, never fetches. */
+export const disabledQueryAtom: Atom.Atom<AsyncResult.AsyncResult<any, any>> = Atom.readable(() =>
+  AsyncResult.initial(false)
+)
+
+/**
+ * Bumps when the browser regains connectivity (the `online` event) — the tanstack
+ * `refetchOnReconnect` trigger. One shared listener (module-level). SSR-guarded.
+ */
+const onlineSignal: Atom.Atom<number> = Atom.readable((get) => {
+  let count = 0
+  if (typeof window === "undefined") return count
+  const update = () => {
+    if (navigator.onLine) get.setSelf(++count)
+  }
+  window.addEventListener("online", update)
+  get.addFinalizer(() => window.removeEventListener("online", update))
+  return count
+})
+
+/**
+ * Focus OR reconnect, as a single signal for `swr` — both should stale-revalidate a query.
+ * swr takes one `focusSignal`, so we fold window-focus + reconnect into one derived atom;
+ * a bump from either triggers swr's stale check.
+ */
+const focusOrReconnectSignal: Atom.Atom<number> = Atom.make((get) => get(Atom.windowFocusSignal) + get(onlineSignal))
+
+/**
+ * Build the per-input atom family for a request handler — the query CACHE IDENTITY.
+ *
+ * This is the TanStack `queryKey = [handler, input]` equivalent and the piece that makes
+ * caching cross-component: `Atom.family` memoizes one atom per structurally-distinct input
+ * (v4 hashes the input via Hash/Equal), so every component querying the same handler+input
+ * reads the SAME atom instance => one fetch, one shared result in the global registry,
+ * ref-counted and GC'd on idle ttl. (The registry + ttl give lifetime; reactivity keys give
+ * invalidation; the family gives identity/sharing — all three are needed.)
+ *
+ * The family is created once per handler (see query.ts's per-handler cache), so it is shared
+ * process-wide via the registry.
+ *
+ * Invalidation is hierarchical: each atom registers under EVERY prefix of its full key
+ * `[...makeQueryKey(self), input]`. Since reactivity matches keys by exact hash, registering
+ * all prefixes means `invalidate(P)` refreshes every atom whose key starts with `P` — e.g.
+ * `["$X"]` refreshes all inputs, `["$X","$List",input]` only that input. (`makeQueryKey`'s
+ * collapsed form `getQueryKey` — what mutations invalidate by default — is one of the prefixes.)
+ */
+export const buildQueryFamily = <I, A, E>(
+  rt: AtomClientRuntime,
+  self: {
+    readonly id: string
+    readonly handler: (i: I) => Effect.Effect<A, E, any>
+    readonly options?: ClientForOptions
+    readonly queryKeyProjectionHash?: string
+  }
+) => {
+  const baseKey = makeQueryKey(self) // hierarchical, input-independent
+
+  return Atom.family((input: I) => {
+    const fullKey = [...baseKey, input]
+    // Record the repository/server read-dependencies seen while fetching, keyed by `fullKey`, so a
+    // later mutation whose writes intersect them can derive this query as an invalidation target.
+    // The last recorded reads are retained on the (memoized) family atom so `trackReadDependencies`
+    // can re-assert them on a cache-hit remount that never re-runs the handler.
+    let lastReads: DataDependencies.DataDependencies = DataDependencies.empty()
+    // Fetch bookkeeping read by `recoverStuckWaitingOnMount`. `inFlight` counts live computes, so a
+    // `waiting` result with `inFlight === 0` is a stuck/hidden interrupt. `recovering` guards the
+    // recovery refresh against concurrent mounts; a running fetch clears it.
+    const fetchState: QueryFetchState = { inFlight: 0, recovering: false }
+    let atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>
+    atom = rt.runtime.atom(
+      Effect.suspend(() => {
+        // A fetch is now running: the atom is not stuck, and any pending recovery is fulfilled.
+        fetchState.recovering = false
+        fetchState.inFlight++
+        const recordReads = Effect
+          .gen(function*() {
+            const readsRef = yield* Ref.make(DataDependencies.empty())
+            const writesRef = yield* Ref.make(DataDependencies.empty())
+            const recorder = DataDependencies.makeDataDependencyRecorder(readsRef, writesRef)
+            const result = yield* self
+              .handler(input)
+              .pipe(Effect.provideService(DataDependencies.DataDependencyRecorder, recorder))
+            lastReads = yield* Ref.get(readsRef)
+            setQueryReadDependencies(fullKey, lastReads)
+            return result
+          })
+        let liveFetch = false
+        const effect = Effect
+          .gen(function*() {
+            liveFetch = yield* Effect.promise(() => beginLiveQueryFetch(fullKey))
+            return yield* recordReads.pipe(Effect.retry({ times: 5, while: isRetryable }))
+          })
+          .pipe(
+            Effect.ensuring(Effect.sync(() => endLiveQueryFetch(liveFetch))),
+            Effect.tapCauseIf(Cause.hasDies, (cause) => reportRuntimeError(cause)),
+            // On exit, the compute is no longer in-flight. An interrupt (subscriber lost interest / a
+            // superseding refresh) may leave the result at `waiting`; with `inFlight` back at 0 that
+            // reads as "stuck", so the next mount recovers it (`recoverStuckWaitingOnMount`). We do not
+            // re-fire here — the interrupt was intentional; recovery is driven by a genuine (re)mount.
+            Effect.onExit(() =>
+              Effect.sync(() => {
+                fetchState.inFlight = Math.max(0, fetchState.inFlight - 1)
+              })
+            ),
+            Effect.withSpan(`query ${self.id}`, {}, { captureStackTrace: false })
+          )
+        const parentSpan = takeAtomQueryParentSpan(atom)
+        return parentSpan === undefined
+          ? effect
+          : effect.pipe(Effect.withParentSpan(parentSpan, { captureStackTrace: false }))
+      })
+    )
+    // Register under every prefix of the full key => hierarchical (prefix) invalidation. Two roles:
+    //   - withReactivity: `Reactivity.invalidate(key)` refreshes this atom (the actual refetch).
+    //   - trackByKeys:    records the atom in `keyAtoms` so the mutation can AWAIT its settle.
+    const projectedFullKey = self.queryKeyProjectionHash === undefined
+      ? fullKey
+      : [...baseKey, self.queryKeyProjectionHash, input]
+    const reactivityKeys = uniqueKeys([...prefixesOf(fullKey), ...prefixesOf(projectedFullKey)])
+    atom = rt.factory.withReactivity(reactivityKeys)(atom)
+    atom = trackByKeys(reactivityKeys)(atom)
+    atom = trackReadDependencies(fullKey, () => lastReads)(atom)
+    // gcTime LAST so the whole chain (incl. the registration + tracking) stays alive through the
+    // idle window. Invalidation still finds the cached atom and marks it stale; it does not
+    // refetch until an observer remounts (TanStack refetchType: "active").
+    atom = Atom.setIdleTTL(atom, defaults.gcTime)
+    const registered = setAtomQueryMetadata(Atom.withLabel(`query:${self.id}`)(atom))
+    const writable = Atom.writable(
+      (get) => {
+        const current = get.once(registered)
+        get.subscribe(registered, (value) => get.setSelf(value))
+        return current
+      },
+      (ctx, value: AsyncResult.AsyncResult<A, E>) => ctx.setSelf(value),
+      (refresh) => refresh(registered)
+    )
+    const writableWithTarget = Object.assign(writable, { initialValueTarget: registered })
+    const result = setAtomQueryMetadata(Atom.withLabel(`query-cache:${self.id}`)(writableWithTarget))
+    // Key the fetch state by the atom `withQueryOptions` receives, so its mount hook can find it.
+    const observed = observeQueryAtom(result)
+    queryFetchStates.set(result, fetchState)
+    queryFetchStates.set(observed, fetchState)
+    atomQueryKeys.set(result, fullKey)
+    atomQueryKeys.set(observed, fullKey)
+    return observed
+  })
+}
+
+export const buildStreamQueryFamily = <I, A, E>(
+  rt: AtomClientRuntime,
+  self: {
+    readonly id: string
+    readonly handler: (i: I) => Stream.Stream<A, E, any>
+    readonly options?: ClientForOptions
+    readonly queryKeyProjectionHash?: string
+  }
+) => {
+  const baseKey = makeQueryKey(self)
+
+  return Atom.family((input: I) => {
+    let atom = rt.runtime.pull(
+      self.handler(input).pipe(
+        Stream.tapCause((cause) => Cause.hasDies(cause) ? reportRuntimeError(cause) : Effect.void)
+      )
+    )
+    const fullKey = [...baseKey, input]
+    const projectedFullKey = self.queryKeyProjectionHash === undefined
+      ? fullKey
+      : [...baseKey, self.queryKeyProjectionHash, input]
+    const reactivityKeys = uniqueKeys([...prefixesOf(fullKey), ...prefixesOf(projectedFullKey)])
+    atom = rt.factory.withReactivity(reactivityKeys)(atom)
+    atom = trackWritableByKeys(reactivityKeys)(atom)
+    atom = Atom.setIdleTTL(atom, defaults.gcTime)
+    return observeQueryAtom(setAtomQueryMetadata(Atom.withLabel(`stream-query:${self.id}`)(atom)))
+  })
+}
+
+/** Await the first resolved (non-Waiting) result of an atom. Failing query results fail the Effect. */
+export const awaitAtomResult = <A, E>(
+  registry: AtomRegistry.AtomRegistry,
+  atom: Atom.Atom<AsyncResult.AsyncResult<A, E>>
+) => AtomRegistry.getResult(registry, atom, { suspendOnWaiting: true })