@rangojs/router 0.0.0-experimental.14 → 0.0.0-experimental.141

package/src/rsc/shell-capture.ts

@@ -0,0 +1,897 @@
+/**
+ * PPR shell capture orchestration (Axis 2, see docs/design/ppr-shell-resume.md).
+ *
+ * Capture does NOT flow through the HTTP middleware pipeline. The integrated PPR
+ * serve path (rsc-rendering.ts + shell-serve.ts) builds a ShellCaptureDescriptor
+ * from the route's `ppr` path option after the served response is built and calls
+ * scheduleShellCapture. The capture then runs as a background task that re-derives
+ * the page via `ctx.router.match()` under its OWN derived request context — fresh
+ * handle store, `_shellCaptureRun: true` so loaders mask (loader-mask.ts) and every
+ * loading() subtree postpones. The render is MIXED-CHAIN: cache()'d segments replay
+ * from ring 3, uncached segments execute their handlers fresh. It drives the static
+ * prerender to a quiescent shell, aborts to freeze the prelude + postponed state,
+ * and stores the pair via putShell. Because it uses match() rather than the HTTP
+ * pipeline, the middleware chain (auth, logging) never re-runs — it already ran for
+ * the triggering request, and the derived context inherits its post-middleware
+ * state (variables, cache store). Guarding is serve-time.
+ */
+
+import React from "react";
+import { bufferToBase64 } from "../cache/cf/cf-base64.js";
+import { reportCacheError } from "../cache/cache-error.js";
+import { runBackground } from "../cache/background-task.js";
+import { observePhase, PHASES } from "../router/instrument.js";
+import {
+  runWithRequestContext,
+  setRequestContextParams,
+  UNTRACKED_BACKGROUND_TASK,
+  type RequestContext,
+} from "../server/request-context.js";
+import { createHandleStore, type HandleStore } from "../server/handle-store.js";
+import type {
+  ShellCacheEntry,
+  SegmentCacheStore,
+  ShellSnapshotRecord,
+} from "../cache/types.js";
+import {
+  RecordingShellStore,
+  getRecordingStore,
+} from "../cache/shell-snapshot.js";
+import type { HandlerContext } from "./handler-context.js";
+import type { RscPayload, SSRModule } from "./types.js";
+import { buildFullPayload } from "./full-payload.js";
+import { resolveDeferredHandleValues } from "../handles/deferred-resolution.js";
+
+/**
+ * Task-quantized quiesce: the number of consecutive macrotask hops with zero new
+ * Flight bytes that marks the shell "quiet". This replaces the old 50ms
+ * wall-clock debounce.
+ *
+ * The capture Flight render is a REGULAR renderToReadableStream (not a static
+ * prerender), so React schedules both its retries and its byte-flush on
+ * setTimeout(0) MACROTASKS (verified against the vendored edge production
+ * react-server-dom build: pingTask uses scheduleMicrotask only when
+ * request.type === PRERENDER, otherwise setTimeout; enqueueFlush is always
+ * setTimeout). Masked loaders are the live lane — their rows never emit — so once
+ * the shell rows finish flushing the stream goes permanently byte-silent, and K
+ * consecutive quiet macrotask hops after the last observed byte declare quiesce.
+ *
+ * K=2 gives a race window of ~two event-loop turns: shell work still producing
+ * bytes keeps resetting the counter; anything not producing bytes within the
+ * window (the masked loaders, and any genuinely pending I/O) becomes a hole. The
+ * only residual is raw per-request I/O rendered directly in shell (not via a
+ * loader) that resolves inside the window — a documented shell anti-pattern; put
+ * per-request data in loaders. See docs/design/ppr-shell-resume.md.
+ */
+const FLIGHT_QUIET_HOPS = 2;
+
+/** Default upper bound on the capture prerender wait before forcing the abort. */
+const SHELL_CAPTURE_MAX_WAIT_MS = 5000;
+
+/**
+ * Upper bound on waiting for the capture's DEFERRED cache writes to settle before
+ * draining the snapshot. Cache writes run under waitUntil (fire-and-forget on
+ * Node, executionContext on workerd), so a MISS-at-capture value's setItem/set —
+ * hence its snapshot record — can land after the shell has quiesced. We collect
+ * those write promises and await them here so the written value is pinned. Kept
+ * short: a pathological slow write must never stall the background capture; a key
+ * that does not settle in time is simply left unpinned (it drifts, the
+ * pre-snapshot behavior) rather than hanging. Reads that HIT are recorded
+ * synchronously during the render and do not depend on this.
+ */
+const SHELL_SNAPSHOT_WRITE_SETTLE_MS = 1000;
+
+/**
+ * Upper bound on the pre-render WRITE BARRIER: before the capture's match/render,
+ * settle the background tasks the FOREGROUND request already scheduled — its
+ * deferred ring-3 cacheRoute and ring-1 setItem writes all go through
+ * reqCtx.waitUntil, and every one of them is scheduled BEFORE scheduleShellCapture
+ * runs (the response, and its onResponse callbacks, are committed first). Draining
+ * them turns the capture's cache reads from a RACE into an ORDERING EDGE: the
+ * capture deterministically observes the foreground's cache generation, replays it
+ * (handler skipped, module-level side effects untouched), and records THAT
+ * generation into the snapshot — so prelude, snapshot, and ring-3 all agree on the
+ * foreground's generation and the capture can never clobber a foreground-produced
+ * entry with a re-render of its own. Scar tissue: without this, the capture's
+ * ring-3 lookup could land between the foreground write chain's serialization and
+ * its store.set, MISS, re-execute the route handler (bumping module-level
+ * counters), and — via the synthetic onResponse fire below — overwrite the
+ * foreground's entry (the mini shell-manifest regression). Bounded: a slow
+ * consumer waitUntil task must never stall the background capture; on timeout the
+ * capture proceeds with the pre-barrier (racy) behavior.
+ */
+const SHELL_CAPTURE_WRITE_BARRIER_MS = 1500;
+
+/**
+ * Settle the tracked background tasks on `reqCtx._pendingBackgroundTasks`,
+ * ITERATIVELY: a settled task can have scheduled a nested one (cache-store's
+ * cacheRoute outer task schedules the actual store.set in a second waitUntil), so
+ * each awaited batch may append more. Loop until no new tasks appear or the
+ * deadline passes. The capture's own task never enters the list
+ * (UNTRACKED_BACKGROUND_TASK), so the loop terminates.
+ */
+async function settleTrackedBackgroundTasks(
+  reqCtx: RequestContext<any>,
+  timeoutMs: number,
+): Promise<void> {
+  const tasks = reqCtx._pendingBackgroundTasks;
+  if (!tasks) return;
+  const deadline = Date.now() + timeoutMs;
+  let seen = 0;
+  while (tasks.length > seen) {
+    const remaining = deadline - Date.now();
+    if (remaining <= 0) return;
+    const batch = tasks.slice(seen);
+    seen = tasks.length;
+    let timer: ReturnType<typeof setTimeout> | undefined;
+    const guard = new Promise<void>((resolve) => {
+      timer = setTimeout(resolve, remaining);
+      (timer as { unref?: () => void }).unref?.();
+    });
+    await Promise.race([Promise.allSettled(batch).then(() => {}), guard]);
+    if (timer) clearTimeout(timer);
+  }
+}
+
+/**
+ * Delay before the in-place retry of a capture that produced no usable shell.
+ *
+ * The dominant reason a first capture comes back with a trivial prelude is a
+ * COLD render: in dev the module transform graph (route modules, the SSR/Flight
+ * transforms) is being built lazily and outlasts the task-quantized quiesce, so
+ * the shell has not finished rendering when we freeze it; on a cold worker the
+ * first invocation pays the same one-time cost. The first attempt WARMS that
+ * graph, so a second attempt a short beat later usually completes the shell in
+ * the SAME background task — no extra HTTP request needed. Short enough to feel
+ * instant, long enough for the module graph to settle. See
+ * docs/design/ppr-shell-resume.md ("Capture retry-in-place").
+ */
+const SHELL_CAPTURE_RETRY_DELAY_MS = 400;
+
+/** Sleep `ms`, unref'd so a Node dev process is never kept alive by the timer. */
+function delay(ms: number): Promise<void> {
+  return new Promise((resolve) => {
+    const t = setTimeout(resolve, ms);
+    (t as { unref?: () => void }).unref?.();
+  });
+}
+
+/**
+ * Module-level in-flight key set: the stampede guard for background captures, and
+ * its single owner. One capture runs per key per isolate; concurrent MISS/stale
+ * requests for the same key coalesce onto the first (the rest see the key present
+ * in scheduleShellCapture and skip). Added when a capture is scheduled and cleared
+ * in the task's finally once it settles, so a later request can recapture when TTL
+ * rolls. Living here (not split across the middleware) keeps the add/clear
+ * lifecycle in one layer.
+ */
+const inFlightCaptures = new Set<string>();
+
+/**
+ * Refused-capture backoff bounds. The window is EXPONENTIAL in the consecutive
+ * failure count: `min(BASE * 2^(failures-1), MAX)` — 1s, 2s, 4s, … capped at 60s.
+ *
+ * Why exponential and not a flat 60s: a flat long window conflates two very
+ * different failures. A STRUCTURALLY ineligible route (no loading(), a cookie
+ * reader) fails forever and wants the long 60s cap. But a cold-but-ELIGIBLE route
+ * can also fail the in-place retry under a truly cold graph (dev module transform,
+ * or a cold worker under parallel load) — and it must recover FAST, on the next
+ * request or two, not be frozen for 60s (that would re-break the very cold-start DX
+ * the retry fixes; it bit the cloudflare dev e2e). Escalating from 1s means the
+ * eligible route re-probes almost immediately (warm now → HIT and clear), while the
+ * doomed route ramps to the 60s cap within a handful of failures. Either way an
+ * app-wide mount never re-renders a doomed route on EVERY request.
+ */
+const REFUSED_CAPTURE_BASE_MS = 1_000;
+const REFUSED_CAPTURE_MAX_MS = 60_000;
+
+/**
+ * Refused-capture backoff: key -> { consecutive failure count, epoch ms until which
+ * the key is not re-probed }. A key enters backoff only after runShellCapture's
+ * in-place retry ALSO failed (or a genuine error). A successful capture clears the
+ * entry outright (failure count resets). Module-level (same lifetime as
+ * inFlightCaptures) so the whole lifecycle lives in one layer.
+ */
+const refusedCaptures = new Map<string, { failures: number; until: number }>();
+
+/** True iff `key` is still inside its (exponential) backoff window. */
+function isCaptureBackedOff(key: string): boolean {
+  const entry = refusedCaptures.get(key);
+  if (entry === undefined) return false;
+  // Window elapsed: allow a re-probe. Keep the entry (its failure count drives the
+  // NEXT window's escalation if the re-probe also fails); a success clears it.
+  return Date.now() < entry.until;
+}
+
+/** Record a refused/failed capture, escalating the backoff window exponentially. */
+function markCaptureBackoff(key: string): void {
+  const failures = (refusedCaptures.get(key)?.failures ?? 0) + 1;
+  const window = Math.min(
+    REFUSED_CAPTURE_BASE_MS * 2 ** (failures - 1),
+    REFUSED_CAPTURE_MAX_MS,
+  );
+  refusedCaptures.set(key, { failures, until: Date.now() + window });
+}
+
+/** Clear any backoff for a key that just captured successfully. */
+function clearCaptureBackoff(key: string): void {
+  refusedCaptures.delete(key);
+}
+
+/**
+ * Keys already warned about a refused (null) capture, so the eternal-MISS shape
+ * logs once per key per isolate instead of on every request.
+ */
+const warnedNullCaptures = new Set<string>();
+
+/**
+ * Warn once per key that a capture produced no usable shell EVEN AFTER the
+ * in-place retry (runShellCapture attempt 2). Naming both causes with the
+ * distinguishing signal — does the route ever flip to HIT — is the whole point:
+ * the pre-retry version blamed "a loader route without loading()" unconditionally
+ * and misled users whose route DID have loading() and was merely cold. Because the
+ * retry already absorbs the cold-start case, by the time this fires cold-start has
+ * usually healed, so a firing warning leans toward the structural cause — but we
+ * still name both so a cold-start straggler is not misdiagnosed.
+ *
+ * The pointer is shipped-path-safe (a05c8251 convention): the /ppr skill ships in
+ * the npm tarball, but docs/design/ is repo-only, so link it by absolute GitHub URL
+ * rather than a relative path that dead-ends for consumers.
+ */
+function warnNullCaptureOnce(key: string): void {
+  if (warnedNullCaptures.has(key)) return;
+  warnedNullCaptures.add(key);
+  console.warn(
+    `[rango] Shell capture for "${key}" produced no usable shell after an in-place ` +
+      "retry; nothing was stored, so this request stays on MISS. Two things cause this, " +
+      "told apart by whether the route ever flips to HIT:\n" +
+      "  1. Cold-start warmup (dev module transform, or a cold worker): the capture raced " +
+      "an unfinished shell render. This SELF-HEALS — the route flips to HIT once a later " +
+      "request warms the modules. Usually nothing to do.\n" +
+      "  2. A loader route WITHOUT a route-level loading() boundary: its loader data is " +
+      "awaited at tree-build, so under capture's masked loaders no shell exists above " +
+      "<body>, and the route NEVER flips to HIT. Add loading() to the loader route (and " +
+      "keep shell material in a layout) to make it PPR-capturable.\n" +
+      'See the /ppr skill (node_modules/@rangojs/router/skills/ppr/SKILL.md), "The hole ' +
+      'contract", or the design doc: ' +
+      "https://github.com/ivogt/vite-rsc/blob/main/packages/rangojs-router/docs/design/ppr-shell-resume.md",
+  );
+}
+
+export interface FlightCaptureGate {
+  /** Identity passthrough of the source stream; feed this to captureShellHTML. */
+  stream: ReadableStream<Uint8Array>;
+  /**
+   * Resolves once the source has been byte-quiet for FLIGHT_QUIET_HOPS macrotask
+   * hops (or has closed — the DATA variant). At that instant the gate FREEZES:
+   * no further source byte reaches the fizz side, and the readable is left open
+   * (never closed / errored) so fizz postpones the still-pending references
+   * instead of seeing "Connection closed".
+   */
+  quiesce: Promise<void>;
+  /**
+   * Stop the internal macrotask-hop loop. captureShellHTML's maxWaitMs bounds the
+   * overall wait; dispose() is the clean shutdown for the pathological case where
+   * the source never goes byte-quiet (quiesce never fires), so the hop loop would
+   * otherwise keep rescheduling after captureShellHTML has already aborted and
+   * returned.
+   */
+  dispose(): void;
+}
+
+/**
+ * Wrap the capture Flight stream so the fizz shell prerender reads a stream that
+ * (a) forwards the shell rows unchanged, (b) resolves `quiesce` after the rows go
+ * byte-silent for FLIGHT_QUIET_HOPS macrotask hops, and (c) FREEZES at that
+ * instant — dropping any later byte without closing or erroring the readable, so
+ * the pending masked-loader references stay pending and fizz postpones them (the
+ * "unclosing stream" property, here for free because the masked rows never emit).
+ * Freezing also guarantees no post-quiesce byte — including an error row from any
+ * later abort/cancel of the underlying render — can corrupt the frozen prelude.
+ *
+ * Quiet is measured in TASKS, not wall-clock: after the first byte a macrotask
+ * hop loop compares a byte counter each turn and fires after K quiet turns. The
+ * hop timers are unref'd so they never keep a Node process alive, and the source
+ * closing (no holes) fires quiesce immediately for the DATA variant — the
+ * TransformStream then closes the readable, so fizz completes with postponed null.
+ *
+ * `holdUntil` keeps the gate from FREEZING before shell material with real latency
+ * has emitted. The hole doctrine bakes TOP-LEVEL pushed handle promises into the
+ * shell (resolvedHandleStream awaits them before the handles row emits), but a
+ * pushed promise that takes longer than the quiet window would otherwise be frozen
+ * out — the handles row would never reach fizz and the prelude would come back
+ * trivial. While `holdUntil` is pending, byte-quiet detection keeps running but the
+ * gate neither fires nor freezes; once it resolves, the quiet counter restarts so a
+ * burst of rows unblocked by it (the resolved handles row) is still captured. It
+ * never delays a HOLE from postponing: holes are pending promises that emit no
+ * bytes, so holding the gate open longer only ever admits shell rows. Bounded by
+ * captureShellHTML's maxWaitMs like every other quiesce input.
+ */
+export function gateFlightForCapture(
+  source: ReadableStream<Uint8Array>,
+  quietHops: number = FLIGHT_QUIET_HOPS,
+  holdUntil?: Promise<unknown>,
+): FlightCaptureGate {
+  let resolveQuiet!: () => void;
+  const quiesce = new Promise<void>((resolve) => {
+    resolveQuiet = resolve;
+  });
+
+  let bytesSeen = 0;
+  let armed = false;
+  let settled = false;
+  let disposed = false;
+  let frozen = false;
+  let held = holdUntil !== undefined;
+  let heldFirePending = false;
+
+  if (holdUntil !== undefined) {
+    const release = (): void => {
+      held = false;
+      if (heldFirePending && !settled && !disposed) {
+        // Quiet elapsed while held: restart the quiet count instead of firing
+        // immediately, so rows unblocked by the hold (the baked handles row)
+        // still flow before the freeze.
+        heldFirePending = false;
+        armed = false;
+        arm();
+      }
+    };
+    // Resolve OR reject releases the hold (a rejected handle value is dropped by
+    // resolveDeferredHandleValues; the capture must not hang on it).
+    holdUntil.then(release, release);
+  }
+
+  const fire = (): void => {
+    if (settled) return;
+    if (held) {
+      heldFirePending = true;
+      return;
+    }
+    settled = true;
+    frozen = true;
+    resolveQuiet();
+  };
+
+  const scheduleHop = (fn: () => void): void => {
+    const t = setTimeout(fn, 0);
+    // Never let the quiet-detection hop alone keep a Node process alive
+    // (no-op on workerd).
+    (t as { unref?: () => void }).unref?.();
+  };
+
+  // The hop loop starts only after the first byte, so it can never declare
+  // quiesce before fizz has begun pulling rows through the transform.
+  const arm = (): void => {
+    if (armed || settled || disposed) return;
+    armed = true;
+    let lastSeen = bytesSeen;
+    let quiet = 0;
+    const hop = (): void => {
+      if (settled || disposed) return;
+      if (bytesSeen === lastSeen) {
+        quiet += 1;
+        if (quiet >= quietHops) {
+          fire();
+          return;
+        }
+      } else {
+        lastSeen = bytesSeen;
+        quiet = 0;
+      }
+      scheduleHop(hop);
+    };
+    scheduleHop(hop);
+  };
+
+  const monitor = new TransformStream<Uint8Array, Uint8Array>({
+    transform(chunk, controller) {
+      // Post-quiesce: drop the byte. Do NOT enqueue and do NOT close/error — the
+      // frozen fizz input must stay a fixed byte set behind an open (unclosing)
+      // readable so still-pending references postpone.
+      if (frozen) return;
+      bytesSeen += chunk.length;
+      arm();
+      controller.enqueue(chunk);
+    },
+    flush() {
+      // Source closed with no freeze => DATA variant (no holes): quiet
+      // immediately. The TransformStream then closes the readable, so fizz
+      // completes and postponed comes back null.
+      fire();
+    },
+  });
+
+  return {
+    stream: source.pipeThrough(monitor),
+    quiesce,
+    dispose(): void {
+      disposed = true;
+    },
+  };
+}
+
+/**
+ * The background shell-capture descriptor: everything the capture task needs to
+ * store the shell. Built by the integrated PPR serve path (rsc-rendering.ts) from
+ * the route's `ppr` path option (`PartialPrerenderProps`) and the app-level cache
+ * store, and passed to scheduleShellCapture directly — it is NOT threaded through
+ * the request context. `tags` carries the route's OPERATIONAL `ppr.tags`; the
+ * capture UNIONS them with the shell's own auto-collected (non-loader) request
+ * tags from its derived render (the collected set stays authoritative). `store`
+ * is the same store the serve path resolved for its getShell read
+ * (requestCtx._cacheStore), so the capture writes where the serve reads.
+ */
+export interface ShellCaptureDescriptor {
+  key: string;
+  ttl?: number;
+  swr?: number;
+  tags?: string[];
+  store?: SegmentCacheStore<any>;
+  /** Gates the concise per-attempt capture breadcrumbs (INTERNAL_RANGO_DEBUG). */
+  debug?: boolean;
+}
+
+/**
+ * Schedule the background shell capture for a served document. Stampede-guarded:
+ * one capture per key per isolate. Runs via runBackground (waitUntil on workerd,
+ * fire-and-forget in Node dev), so the served response is never blocked on it. Any
+ * error is routed through reportCacheError — capture is best-effort; a failure just
+ * means the next request recaptures.
+ *
+ * Eligibility (nonce/allReady/partial/status/strategy) is decided by the caller
+ * (rsc-rendering.ts maybeScheduleShellCapture); this function only owns the
+ * stampede guard and the background dispatch.
+ */
+export function scheduleShellCapture(
+  ctx: HandlerContext<any>,
+  request: Request,
+  env: any,
+  url: URL,
+  reqCtx: RequestContext<any>,
+  ssrModule: SSRModule,
+  descriptor: ShellCaptureDescriptor,
+): void {
+  const key = descriptor.key;
+  if (inFlightCaptures.has(key)) return;
+  // Refused/failed within the window → skip the doomed re-render (one probe per
+  // key per window per isolate). Expired entries self-evict inside the check.
+  if (isCaptureBackedOff(key)) return;
+  inFlightCaptures.add(key);
+  const captureTask = async () => {
+    try {
+      const outcome = await runShellCapture(
+        ctx,
+        request,
+        env,
+        url,
+        reqCtx,
+        ssrModule,
+        descriptor,
+      );
+      // Update the negative cache off the terminal outcome. A stored shell clears
+      // any prior backoff; a `no-shell` (after the in-place retry) backs the key
+      // off so the next requests don't re-probe it. A `redirect` has no shell but
+      // is not a doomed render — leave the backoff untouched.
+      if (outcome === "stored") clearCaptureBackoff(key);
+      else if (outcome === "no-shell") markCaptureBackoff(key);
+    } catch (error) {
+      // Detached background task — pass reqCtx so onError still fires when the ALS
+      // context is gone. A genuine failure recurs, so back it off too (re-probe
+      // once per window, not every request) and report it once.
+      markCaptureBackoff(key);
+      reportCacheError(error, "cache-write", "[ShellCache] capture", reqCtx);
+    } finally {
+      inFlightCaptures.delete(key);
+    }
+  };
+  // The capture's own task must NOT enter reqCtx._pendingBackgroundTasks: the
+  // capture drains that list before rendering (the write-barrier ordering edge),
+  // and awaiting its own still-running promise would burn the whole barrier
+  // deadline on every capture.
+  (captureTask as { [UNTRACKED_BACKGROUND_TASK]?: boolean })[
+    UNTRACKED_BACKGROUND_TASK
+  ] = true;
+  runBackground(reqCtx, captureTask);
+}
+
+/**
+ * The outcome of one capture attempt.
+ * - `stored`: a usable shell was captured (and a putShell was attempted; a store
+ *   I/O failure is reported separately and does NOT make the attempt retryable —
+ *   the capture itself worked).
+ * - `redirect`: the matched route redirects, so there is no shell to capture.
+ * - `no-shell`: the prelude came back trivial (no <body>) OR captureShellHTML
+ *   rejected with our own abort. This is the only RETRYABLE outcome.
+ */
+type CaptureAttemptOutcome = "stored" | "redirect" | "no-shell";
+
+/**
+ * Run the shell capture with a single in-place retry, then store the result.
+ *
+ * Each attempt re-derives EVERYTHING (fresh context, fresh router.match, fresh
+ * Flight render) via {@link attemptCapture} — a capture consumes its handle store,
+ * its request-tag set, and its one-shot Flight stream, so none of them are
+ * reusable across attempts. A first attempt that comes back `no-shell` is almost
+ * always a cold render (dev module transform / cold worker) that had not finished
+ * when we froze the shell; the attempt itself warmed the module graph, so a second
+ * attempt a short beat later usually completes the shell in the SAME background
+ * task. That kills the old multi-request warmup where the caller had to re-issue
+ * several HTTP requests before a capture stuck. We retry ONLY on `no-shell` (and a
+ * defensively-caught abort); a genuine render error is NOT retried — it propagates
+ * to scheduleShellCapture's reportCacheError. See docs/design/ppr-shell-resume.md.
+ *
+ * `retryDelayMs` is a parameter (defaulting to the module const) so unit tests can
+ * drive the retry without a real 400ms wall-clock wait.
+ */
+async function runShellCapture(
+  ctx: HandlerContext<any>,
+  request: Request,
+  env: any,
+  url: URL,
+  reqCtx: RequestContext<any>,
+  ssrModule: SSRModule,
+  descriptor: ShellCaptureDescriptor,
+  retryDelayMs: number = SHELL_CAPTURE_RETRY_DELAY_MS,
+): Promise<CaptureAttemptOutcome> {
+  const log = descriptor.debug
+    ? (message: string) => console.log(message)
+    : () => {};
+
+  const first = await attemptCapture(
+    ctx,
+    request,
+    env,
+    url,
+    reqCtx,
+    ssrModule,
+    descriptor,
+  );
+  // "stored" (success) or "redirect" (no shell exists): nothing to retry.
+  if (first !== "no-shell") return first;
+
+  // Attempt 1 produced no usable shell. Retry ONCE in place — the first attempt
+  // warmed the dev transform graph / cold worker, so attempt 2 typically completes
+  // the shell without another HTTP request. The concise line is gated on the
+  // middleware's debug flag (threaded via the descriptor) so it replaces the old
+  // full DOMException dump with one readable breadcrumb.
+  log(
+    `[ShellCache] capture attempt 1/2 for ${descriptor.key} aborted before shell completed (cold modules?) — retrying`,
+  );
+  await delay(retryDelayMs);
+  const second = await attemptCapture(
+    ctx,
+    request,
+    env,
+    url,
+    reqCtx,
+    ssrModule,
+    descriptor,
+  );
+  if (second !== "no-shell") return second;
+
+  // Both attempts came back with no usable shell. Cold-start would have healed by
+  // now, so the eternal-MISS structural shape (a loader route without loading()) is
+  // the likely cause — warn once per key. Ordering matters: because the retry
+  // absorbs cold-start, cold-start routes almost never reach this warning. The
+  // caller (scheduleShellCapture) reads this `no-shell` return to back the key off.
+  log(
+    `[ShellCache] capture attempt 2/2 for ${descriptor.key} aborted — giving up until next request`,
+  );
+  warnNullCaptureOnce(descriptor.key);
+  return "no-shell";
+}
+
+/**
+ * One capture attempt in a DERIVED request context.
+ *
+ * The derived context is `Object.create(reqCtx)` so it inherits the foreground's
+ * post-middleware state (variables, cache store, env/request/url, waitUntil) while
+ * overriding the render-scoped accumulators as own properties:
+ *   - _handleStore: a fresh store. The foreground store is already drained to
+ *     completion (its stream() flipped `completed` on settle) and would throw
+ *     LateHandlePushError on any re-push. Every downstream reader resolves the
+ *     store off the ambient context (setupLoaderAccess captures
+ *     _getRequestContext()._handleStore; trackHandler reads it), so the fresh
+ *     store on the derived context is what the capture match() writes handles to.
+ *   - _requestTags: a fresh Set. The capture collects its OWN shell tags here —
+ *     non-loader tags only, since loaders are masked — which is exactly the tag
+ *     set a shell entry should be invalidatable by (loader tags belong to holes).
+ *   - _transitionWhen: a fresh [] so the capture's transition gating is its own.
+ *   - _shellCaptureRun: true — the switch loaders/cookies/headers guards read.
+ *   - _metricsStore: undefined so the capture never appends to the foreground's
+ *     (already-finalized) metrics.
+ *
+ * The capture is MIXED-CHAIN: its match() behaves like a normal render with
+ * respect to the segment cache — cache()'d segments replay from ring 3, UNCACHED
+ * segments execute their handlers fresh (which is why the cookies()/headers()
+ * capture guard is load-bearing). Middleware is NOT re-run: it already ran for the
+ * triggering request, and the derived context inherits its post-middleware state
+ * (guarding is serve-time; the shell is never served without the full chain).
+ *
+ * A FRESH context (and match/render) per attempt is what makes the retry sound:
+ * the second attempt is a clean capture, not a resumption of the first.
+ */
+async function attemptCapture(
+  ctx: HandlerContext<any>,
+  request: Request,
+  env: any,
+  url: URL,
+  reqCtx: RequestContext<any>,
+  ssrModule: SSRModule,
+  descriptor: ShellCaptureDescriptor,
+): Promise<CaptureAttemptOutcome> {
+  // WRITE BARRIER (ordering edge, not a narrower race): settle the foreground's
+  // already-scheduled background tasks — its deferred ring-3/ring-1 cache writes —
+  // BEFORE this attempt's match/render, so the capture's cache reads observe the
+  // foreground's generation deterministically. Contract: a capture must never
+  // clobber a ring-3 entry the foreground produced; with the barrier, the
+  // capture's ring-3 lookup HITs the foreground's entry and REPLAYS it (handler
+  // skipped, cache-store middleware's write path gated off by state.cacheHit), so
+  // prelude, snapshot, and ring-3 agree on the foreground's generation. Runs per
+  // attempt (the retry re-checks; already-settled promises are free).
+  await settleTrackedBackgroundTasks(reqCtx, SHELL_CAPTURE_WRITE_BARRIER_MS);
+
+  const freshHandleStore = createHandleStore();
+  freshHandleStore.onError = reqCtx._handleStore.onError;
+
+  const derivedCtx: RequestContext = Object.create(reqCtx);
+  derivedCtx._handleStore = freshHandleStore;
+  derivedCtx._requestTags = new Set<string>();
+  derivedCtx._transitionWhen = [];
+  derivedCtx._shellCaptureRun = true;
+  derivedCtx._metricsStore = undefined;
+  // Own onResponse list so the capture's match-middleware callbacks (the ring-3
+  // segment cache write registers here) are ISOLATED from the foreground's shared
+  // array AND can be fired by captureAndStoreShell. The segment write is gated
+  // behind onResponse, which the capture never triggers (it builds no Response) —
+  // without firing it, a ring-3 cache() MISS at capture renders fresh into the
+  // prelude but is never written, so it is never recorded and drifts on a HIT.
+  derivedCtx._onResponseCallbacks = [];
+
+  // Capture data snapshot: read every cache-store hit/write through a recording
+  // wrapper on the DERIVED context's store (own property, so the shared
+  // reqCtx._cacheStore is untouched — the snapshot is per-capture). Its records
+  // ride inside the ShellCacheEntry so a HIT can reproduce the shell's cached
+  // content byte-identically. See cache/shell-snapshot.ts and the design doc.
+  //
+  // Cache writes are deferred (waitUntil): a MISS-at-capture value's setItem/set
+  // — hence its record — would otherwise land after the shell quiesces. Override
+  // the derived context's waitUntil to COLLECT those write promises (still
+  // forwarding to the parent so the write persists and the worker stays alive),
+  // then captureAndStoreShell awaits them before draining. Reads that HIT are
+  // recorded synchronously during the render and need none of this.
+  if (reqCtx._cacheStore) {
+    const recordingStore = new RecordingShellStore(reqCtx._cacheStore);
+    derivedCtx._cacheStore = recordingStore;
+    derivedCtx.waitUntil = (fn: () => Promise<void>): void => {
+      const p = Promise.resolve().then(fn);
+      recordingStore.trackWrite(p);
+      reqCtx.waitUntil(() => p);
+    };
+  }
+
+  return runWithRequestContext(derivedCtx, async () => {
+    const match = await ctx.router.match(request, { env });
+    // A route that redirects has no shell to capture — bail (no store write, no
+    // retry: a redirect is deterministic).
+    if (match.redirect) return "redirect";
+
+    setRequestContextParams(match.params, match.routeName);
+
+    const payload = buildFullPayload(
+      match,
+      ctx,
+      url,
+      derivedCtx,
+      freshHandleStore,
+    );
+    const rscStream = ctx.renderToReadableStream<RscPayload>(payload, {
+      onError: (error: unknown) => {
+        ctx.callOnError(error, "rendering", { request, url, env });
+      },
+    });
+
+    // Shell tags = the non-loader request tags the capture render recorded on its
+    // own fresh _requestTags (loaders are masked, so loader cache tags — which
+    // belong to the holes, not the shell — are correctly excluded), UNIONED with
+    // the middleware's operational `tags` option (descriptor.tags). The collected
+    // set is authoritative; the option only adds tags the render cannot know.
+    const collected = [...derivedCtx._requestTags];
+    const union = new Set<string>([...(descriptor.tags ?? []), ...collected]);
+    const tags = union.size > 0 ? [...union] : undefined;
+
+    return captureAndStoreShell(
+      ssrModule,
+      rscStream,
+      freshHandleStore,
+      derivedCtx,
+      {
+        ...descriptor,
+        tags,
+      },
+    );
+  });
+}
+
+/**
+ * Seal handles, derive the quiesce signal, prerender + abort via the SSR module's
+ * captureShellHTML, and store the result. Returns the attempt outcome (the caller
+ * owns retry/warn decisions — this function no longer warns). Never throws out of
+ * the store write: a failed putShell is routed through reportCacheError so the
+ * background task stays best-effort, and the attempt still counts as `stored` (the
+ * capture worked; only the store I/O failed). `ssrModule.captureShellHTML` MUST be
+ * present (eligibility is checked before scheduling).
+ *
+ * A `no-shell` result (trivial prelude, or a defensively-caught abort) is the only
+ * retryable outcome; a genuine (non-abort) captureShellHTML error propagates so it
+ * reaches reportCacheError and is NOT retried.
+ */
+async function captureAndStoreShell(
+  ssrModule: SSRModule,
+  rscStream: ReadableStream<Uint8Array>,
+  handleStore: HandleStore,
+  reqCtx: RequestContext<any>,
+  capture: ShellCaptureDescriptor,
+): Promise<Exclude<CaptureAttemptOutcome, "redirect">> {
+  const captureShellHTML = ssrModule.captureShellHTML!;
+
+  // Seal the handle store so the payload's handles generator (resolvedHandleStream
+  // -> handleStore.stream()) converges and completes even though masked loaders
+  // never resolve. handleStore.settled gates ONLY on tracked HANDLER promises
+  // (handleStore.track, via trackHandler) — NOT on deferred handle VALUES pushed
+  // through ctx.use(Handle).defer(), which are plain pushed promises. So seal()
+  // does not reject or hang on outstanding defers: settled resolves once the
+  // handlers settle, and each deferred slot resolves on its own createDeferred
+  // timeout (defer.ts, default 10s) or when its resolver fires. A defer whose
+  // resolver depends on a masked loader can never fire, so it stays pending until
+  // that 10s timeout — longer than maxWaitMs (5s). At the abort the handles
+  // generator has not yielded, SsrRoot suspends at the root (consumeAsyncGenerator
+  // sits above every boundary), the prelude comes back trivial, and
+  // captureShellHTML's sanity gate returns null: the designed fail-safe no-op, not
+  // an error. This mirrors the __prerender_collect seal+settled regime, which also
+  // excludes loaders. See docs/design/ppr-shell-resume.md ("Loaders and handles").
+  handleStore.seal();
+
+  // Handles contract, shell half ("nesting = liveness"): TOP-LEVEL pushed handle
+  // promises are BAKED into the shell — resolvedHandleStream awaits them before
+  // the payload's handles row emits. A pushed promise with real latency would lose
+  // the byte-quiet race (the pending handles row emits no bytes, the gate freezes,
+  // the row is dropped, SsrRoot suspends at the root), so the gate is HELD open
+  // until the same await completes: handlesBaked mirrors resolvedHandleStream's
+  // resolution (getData waits the tracked-handler barrier; resolveDeferredHandleValues
+  // awaits the top-level thenables). NESTED promises inside pushed containers are
+  // shallow-skipped by isThenable and never hold the gate — they stay holes.
+  // Bounded by maxWaitMs like every quiesce input (a defer hanging on a masked
+  // loader still ends in the sanity-gate refusal).
+  const handlesBaked = handleStore.getData().then(resolveDeferredHandleValues);
+  const gate = gateFlightForCapture(rscStream, undefined, handlesBaked);
+  // Quiesce = handles settled AND the Flight shell rows went task-quiet. Either
+  // half stalling is bounded by captureShellHTML's maxWaitMs.
+  const quiesce = Promise.all([handleStore.settled, gate.quiesce]).then(
+    () => {},
+  );
+
+  try {
+    // captureShellHTML CONSUMES the (gated) stream — it is not also SSR'd.
+    let result: Awaited<ReturnType<typeof captureShellHTML>>;
+    try {
+      result = await observePhase(PHASES.ssr, () =>
+        captureShellHTML(gate.stream, {
+          quiesce,
+          maxWaitMs: SHELL_CAPTURE_MAX_WAIT_MS,
+        }),
+      );
+    } catch (error) {
+      // captureShellHTML normally converts its OWN deliberate abort to a null
+      // return (index.tsx). This catch is defensive: if an AbortError still escapes
+      // (a runtime where the abort surfaces as a stream rejection outside its
+      // guard), treat it as the same retryable "no usable shell" degradation rather
+      // than a failure — do NOT report it as an error. A genuine (non-abort) render
+      // error is a real failure: rethrow so it reaches reportCacheError (no retry).
+      if ((error as { name?: string } | null)?.name === "AbortError") {
+        return "no-shell";
+      }
+      throw error;
+    }
+
+    // null = sanity gate refused (trivial/empty prelude, no <body>). Store nothing
+    // and report `no-shell` so the caller (runShellCapture) can retry once and, if
+    // that also fails, warn once per key. On a cold render this is the shell not
+    // yet finished; on a loader route WITHOUT a route-level loading() boundary it is
+    // the structural eternal-MISS shape (the masked loader pins the tree above
+    // <body> at tree-build). The caller's warning names both.
+    if (result === null) {
+      return "no-shell";
+    }
+
+    // Store per the flag's key/ttl/swr/tags, into the flag's store: the middleware
+    // threads the SAME store it resolved for its getShell read (options.store ??
+    // _cacheStore), so a store-attached middleware writes captures where it reads
+    // them. The _cacheStore fallback covers a flag armed without a store (tests).
+    // reactVersion is read from the same React.version import the middleware
+    // validates reads against, so capture and serve always agree.
+    // Fire the capture's isolated onResponse callbacks with a synthetic 200 so
+    // the ring-3 segment cache write (cacheScope.cacheRoute, registered via
+    // onResponse by the cache-store match-middleware and gated on a 200) runs
+    // DURING capture, routed through the recording store. The foreground path
+    // never fires for the capture — it builds no Response — so without this a
+    // cache() SEGMENT that MISSED at capture would be rendered fresh into the
+    // prelude yet never written, hence never recorded, and would drift on a HIT
+    // (an item-family "use cache" write already runs inline during the render, so
+    // it needs none of this; only segment writes are onResponse-gated). The
+    // derived context's own _onResponseCallbacks holds only capture match-
+    // middleware callbacks (HTTP middleware never runs for a capture), so firing
+    // them is safe. Best-effort: a throwing callback must not fail the capture.
+    const responseCallbacks = reqCtx._onResponseCallbacks;
+    if (responseCallbacks && responseCallbacks.length > 0) {
+      const synthetic = new Response(null, { status: 200 });
+      for (const cb of responseCallbacks) {
+        try {
+          cb(synthetic);
+        } catch {
+          // A capture-time cache write that throws is degradation, not failure.
+        }
+      }
+    }
+
+    // Drain the capture data snapshot from the recording store on the derived
+    // context. Await the deferred cache writes first so a MISS-at-capture value
+    // (setItem/set scheduled under waitUntil, including the segment write just
+    // fired) is pinned, not just read-hits. When no recording store is installed
+    // (unit tests that call this directly), there is simply no snapshot.
+    const recording = getRecordingStore(reqCtx._cacheStore);
+    let snapshot: ShellSnapshotRecord[] | undefined;
+    if (recording) {
+      await recording.settleWrites(SHELL_SNAPSHOT_WRITE_SETTLE_MS);
+      snapshot = recording.drainSnapshot();
+    }
+
+    const store = capture.store ?? reqCtx._cacheStore;
+    if (store?.putShell) {
+      try {
+        const entry: ShellCacheEntry = {
+          // slice() copies just this view's bytes into a fresh ArrayBuffer, so a
+          // prelude that is a subarray of a larger backing buffer encodes only its
+          // own region — bufferToBase64 reads the whole ArrayBuffer it is handed.
+          prelude: bufferToBase64(result.prelude.slice().buffer as ArrayBuffer),
+          postponed: result.postponed,
+          reactVersion: React.version,
+          // The theme this capture's payload was built with (buildFullPayload
+          // reads reqCtx.theme off the derived context). The serve tail replays
+          // it so the resume tree matches the frozen prelude — see
+          // ShellCacheEntry.initialTheme.
+          initialTheme: reqCtx.theme,
+          snapshot,
+          createdAt: Date.now(),
+        };
+        await store.putShell(
+          capture.key,
+          entry,
+          capture.ttl,
+          capture.swr,
+          capture.tags,
+        );
+      } catch (error) {
+        // Best-effort: a failed put must never throw out of the background task.
+        reportCacheError(
+          error,
+          "cache-write",
+          "[ShellCache] capture put",
+          reqCtx,
+        );
+      }
+    }
+    // A shell was captured (the store I/O may have failed, but that is reported,
+    // not retried) — so this attempt is `stored` and the caller does not retry.
+    return "stored";
+  } finally {
+    // Stop the hop loop for the pathological never-quiets path (quiesce never
+    // fired, capture returned via maxWaitMs). On the normal path the loop already
+    // stopped when it fired quiesce; dispose() is then a no-op.
+    gate.dispose();
+  }
+}
+
+// Exported for unit tests that drive the capture core directly.
+export { runShellCapture, captureAndStoreShell };