cartethyia 2.5.4 → 2.5.7

package/src/transport/dispatch/attempt-loop.ts

@@ -1,530 +1,531 @@
-/**
- * Shared per-candidate attempt lifecycle for every dispatch route.
- *
- * Admission lease → pool slot → routing reservation, the route-specific
- * upstream operation, then one `completeAttempt` for the attempt's accounting.
- * Retry, backoff, pool-cooldown flagging, and reactive OAuth refresh are one
- * policy here, so the canonical proxy routes and the native Responses-compact
- * route cannot drift. Route-specific behavior arrives as the `prepare` and
- * `attempt` closures; this module never forks retry policy.
- */
-import { GatewayError } from "../../protocol/gateway-error";
-import type { ResolvedCredential } from "../../providers/provider-contracts";
-import {
-  classifyTerminalOutcome,
-  fallbackRetryDelayMs,
-  isRetryableFailure,
-  MIN_RETRY_BUDGET_MS,
-  sleep,
-} from "../failure-policy";
-import { resolveRouteMaxAttempts } from "../../config/timeouts";
-import type { OAuthTokenRefresher } from "../../providers/authentication/oauth-refresh-service";
-import type { OAuthRefreshService } from "../../providers/authentication/oauth-refresh-service";
-import type { ValidatedNetworkBindingFactory } from "../../network/pool/resolver";
-import type { ByokUpstreamHost } from "../../providers/operations/provider-catalog-service";
-import type { AdmissionLease } from "../../security/admission/contracts";
-import type { RouteCandidate, Reservation, RouteSnapshotService } from "../routing/route-model";
-import type { CartethyiaDatabase } from "../../persistence/db-handle";
-import { acquireAttemptLeases, releaseAttemptLeases } from "./leases";
-import type { AttemptLeaseSource, AttemptLeases } from "./leases";
-import type { AttemptLeaseOwnership } from "./leases";
-import { metrics } from "../../observability/metrics";
-import type { NetworkPoolSelector } from "../../network/pool/selector";
-import type { ProxyRequestState } from "../request/state";
-import type { TelemetryBatchBuffer } from "../../observability/telemetry-buffer";
-import { flagPoolBlocked, flagPoolCooldown } from "../../network/pool-health";
-import { completeAttempt, errorClientResponseBody, estimatedUsage, type ProviderExchangeCapture } from "./attempt-finalize";
-import { repriceUsage } from "../../providers/provider-usage";
-import { isFreeTierIpBlock, isOAuthCredentialInvalidated, shouldCooldownPool } from "./retry-policy";
-import { drainAbortReason } from "../shutdown-notice";
-
-/**
- * Identity of one combo member for the request-scoped failure flag.
- *
- * Account, model, and endpoint distinguish siblings that share a provider.
- * The flag lives only for this request: it stops a failed member from being
- * retried while another member remains, without writing a durable cooldown.
- */
-function comboCandidateKey(candidate: RouteCandidate): string {
-  return [
-    candidate.provider_id,
-    candidate.provider_account_id ?? "global",
-    candidate.model_id,
-    candidate.endpoint,
-  ].join(":");
-}
-
-/**
- * Which combo member a candidate belongs to.
- *
- * A combo's members are distinct models (usually one per provider), so the
- * model id is what separates them — the accounts of one member are siblings
- * that share it. Used only to give each member its own share of the attempt
- * budget; a candidate list that is not a combo collapses to one member, which
- * keeps the ordinary single-model path untouched.
- */
-function comboMemberKey(candidate: RouteCandidate): string {
-  return `${candidate.provider_id}:${candidate.model_id}:${candidate.endpoint}`;
-}
-
-/**
- * One member's ordinal, resolved by *identity* rather than by position.
- *
- * Deliberately not a run-length scan. `plan()` reorders the eligible list — its
- * cooling partition emits every healthy candidate before every cooling one — so
- * one member's accounts can be non-contiguous and even interleaved with another
- * member's. A run-length scan then reports more "members" than actually exist:
- * `[cb, wb, cb, wb]` counts four, which pushed each member's skip threshold to
- * `cap * 4` — well past the whole budget — and silently disabled the cap on
- * exactly the degraded lists it exists for.
- *
- * Keyed off the same string the member count is derived from, so the two can
- * never disagree.
- */
-function buildMemberOrdinals(
-  candidates: readonly RouteCandidate[],
-): ReadonlyMap<string, number> {
-  const ordinals = new Map<string, number>();
-  for (const candidate of candidates) {
-    const key = comboMemberKey(candidate);
-    if (!ordinals.has(key)) ordinals.set(key, ordinals.size);
-  }
-  return ordinals;
-}
-
-/** Route-specific preconditions resolved for one candidate before its leases are taken. */
-interface PreparedAttempt<TAdapter> {
-  readonly credential: ResolvedCredential;
-  readonly adapter: TAdapter;
-}
-
-/** One candidate attempt with its admission lease, pool slot, and reservation held. */
-interface AttemptContext<TAdapter> {
-  readonly candidate: RouteCandidate;
-  readonly credential: ResolvedCredential;
-  readonly adapter: TAdapter;
-  readonly leases: AttemptLeases;
-  /** Zero-based candidate index; also this attempt's failover count. */
-  readonly attemptIndex: number;
-  readonly providerCapture: ProviderExchangeCapture;
-  /**
-   * Ownership of this attempt's account/network-pool release, when the route
-   * acquired a pool slot. A provider session that outlives this request (a
-   * Cursor Run held open for the client's tool result) calls
-   * `transferToProviderSession()` and takes it over; otherwise the request
-   * releases it through `releaseByResponse()`.
-   */
-  readonly leaseOwnership?: AttemptLeaseOwnership | undefined;
-  /**
-   * Hands lease ownership to the value this attempt returns (a streaming
-   * response releases its own lease on stream completion/cancel). Without it
-   * the loop releases the leases in `finally`.
-   */
-  retainLeases(): void;
-}
-
-/** Collaborators shared by every dispatch route. */
-interface AttemptLoopDeps {
-  readonly db: CartethyiaDatabase;
-  readonly poolSelector?: NetworkPoolSelector;
-  readonly networkBindingFactory?: ValidatedNetworkBindingFactory;
-  readonly snapshotService?: RouteSnapshotService;
-  readonly telemetryBuffer?: TelemetryBatchBuffer;
-  readonly resolveOAuthRefresher?: (providerId: string) => Promise<OAuthTokenRefresher | undefined>;
-  readonly oauthRefreshService?: OAuthRefreshService;
-}
-
-interface AttemptLoopInput<TResult, TAdapter> {
-  readonly state: ProxyRequestState;
-  readonly deps: AttemptLoopDeps;
-  readonly leaseSource: AttemptLeaseSource;
-  readonly candidates: readonly RouteCandidate[];
-  readonly tenantId: string | null;
-  /** Strict pool selection rejects direct egress when every pool is at capacity. */
-  readonly strictPoolSelection: boolean;
-  /** BYOK host for a candidate, when this route validates network bindings. */
-  readonly resolveHost?: (candidate: RouteCandidate) => ByokUpstreamHost | undefined;
-  readonly prepare: (candidate: RouteCandidate) => Promise<PreparedAttempt<TAdapter>>;
-  readonly attempt: (context: AttemptContext<TAdapter>) => Promise<TResult>;
-  /** Raised when the candidate list is empty and no attempt ever ran. */
-  readonly exhaustedError: GatewayError;
-}
-
-/**
- * Shared per-candidate attempt lifecycle for the canonical proxy routes and
- * the native Responses-compact route: admission lease → pool slot → routing
- * reservation, the route-specific upstream operation, then one
- * `completeAttempt` for the attempt's accounting. Retry, backoff,
- * pool-cooldown flagging, and reactive OAuth refresh are one policy here, so
- * the two routes cannot drift.
- */
-export async function runAttemptLoop<TResult, TAdapter>(
-  input: AttemptLoopInput<TResult, TAdapter>,
-): Promise<TResult> {
-  const { state, deps, leaseSource } = input;
-  const candidates = [...input.candidates];
-  let lastError: unknown = input.exhaustedError;
-  const refreshedCandidates = new Set<string>();
-  const failedCandidates = new Set<string>();
-  // Every attempt owns a fresh capture sink, but only the terminal attempt's
-  // sink reaches the payload store. When the terminal attempt fails before it
-  // fires an outbound request (lease/acquire/prepare threw, or failover
-  // exhausted), its sink is empty and the drawer reads "—" for a request that
-  // DID reach an upstream on an earlier attempt. Keep the fired exchanges so
-  // the terminal completion can fall back to the last one that has a request.
-  const firedCaptures: ProviderExchangeCapture[] = [];
-  const maxAttempts = resolveRouteMaxAttempts();
-  // One combo member must not be able to spend the whole budget.
-  //
-  // A combo's member list is flattened into this candidate list, so the first
-  // member's accounts occupy its head. With one member holding more accounts
-  // than `maxAttempts` (default 8), every attempt was spent inside that member
-  // and a later, perfectly healthy member was never dialed — a `deepseek` combo
-  // of [cb, wb] stayed stuck on cb while wb sat idle, because cb's twelve
-  // accounts exhausted all eight attempts before the list reached wb.
-  //
-  // Each member gets an equal share of the budget, counted **per member** — not
-  // as a cumulative watermark over total attempts, or one member that over-runs
-  // silently steals the tail's share. The share is a ceiling, not a reservation:
-  // a member that used less than its share simply never reaches it, and the
-  // unused budget stays available to the members behind it.
-  const memberOrdinals = buildMemberOrdinals(candidates);
-  const memberCount = memberOrdinals.size;
-  const perMemberCap =
-    memberCount > 1 ? Math.max(1, Math.ceil(maxAttempts / memberCount)) : maxAttempts;
-  const attemptsByMember = new Map<string, number>();
-  // The budget itself is owned by failure-policy, so this loop, the
-  // empty-stream replay, and the Codex retry budget cannot drift apart.
-  let index = 0;
-  let attempts = 0;
-  while (index < candidates.length && attempts < maxAttempts) {
-    const candidate = candidates[index];
-    if (!candidate) break;
-    const memberKey = comboMemberKey(candidate);
-    // A member that has spent its share is skipped: the loop advances past the
-    // rest of its accounts to the next member instead of burning the budget
-    // they would have consumed.
-    if (
-      memberCount > 1 &&
-      (attemptsByMember.get(memberKey) ?? 0) >= perMemberCap
-    ) {
-      index += 1;
-      continue;
-    }
-    // Do not start an attempt that cannot finish. The measured pre-dispatch
-    // cost of the previous attempt (`resolveMs` ran 60s+ on a slow upstream)
-    // is what a follow-on attempt has to spend again before the upstream is
-    // even called, so a retry that begins with less than that left is a retry
-    // the deadline will kill mid-flight. Stopping here surfaces the first
-    // attempt's own error instead of a second one's timeout.
-    if (attempts > 0) {
-      const remainingMs = state.deadlineMs - Date.now();
-      if (remainingMs < MIN_RETRY_BUDGET_MS) {
-        if (lastError !== undefined && !state.completed) {
-          const terminal = classifyTerminalOutcome(lastError, state.abortController.signal);
-          const lastFired = [...firedCaptures]
-            .reverse()
-            .find((capture) => capture.request !== null);
-          await completeAttempt(state, {
-            status: terminal.status,
-            providerId: candidate.provider_id,
-            ...(candidate.provider_account_id
-              ? { accountId: candidate.provider_account_id }
-              : {}),
-            ...(candidate.provider_account_label
-              ? { accountLabel: candidate.provider_account_label }
-              : {}),
-            errorCategory: terminal.errorCategory,
-            errorOrigin: terminal.errorOrigin,
-            modelId: candidate.model_id,
-            error: lastError,
-            terminal: true,
-            tenantId: input.tenantId,
-            ingressBody: state.ingressBody,
-            responseBody: null,
-            ...(terminal.status !== "cancelled"
-              ? { clientResponseText: errorClientResponseBody(lastError) }
-              : {}),
-            // Same fallback the terminal catch uses: a request that did fire
-            // upstream keeps that exchange in Request Detail.
-            ...(lastFired ? { providerCapture: lastFired } : {}),
-            db: deps.db,
-            ...(deps.telemetryBuffer ? { telemetryBuffer: deps.telemetryBuffer } : {}),
-            ...(deps.snapshotService ? { snapshotService: deps.snapshotService } : {}),
-          });
-          metrics.proxy_requests_total.inc(1, {
-            status: terminal.status === "cancelled" ? "cancelled" : "failed",
-          });
-        }
-        throw lastError;
-      }
-    }
-    let lease: AdmissionLease | undefined;
-    let reservation: Reservation | undefined;
-    let proxySlot: { poolId: string; release: () => void } | undefined;
-    let networkPoolId: string | undefined;
-    let leaseOwnership: AttemptLeaseOwnership | undefined;
-    let bindingEstablished = false;
-    let leasesRetained = false;
-    let retryDelayMs = 0;
-    const providerCapture: ProviderExchangeCapture = {
-      request: null,
-      response: Promise.resolve(null),
-    };
-    try {
-      if (state.abortController.signal.aborted)
-        throw new GatewayError("transport_closed", 499, "request was cancelled");
-      const { credential, adapter } = await input.prepare(candidate);
-      const host = input.resolveHost?.(candidate);
-      // Single attempt-lease manager owns acquisition order (admission →
-      // pool slot → reservation) with reverse-order unwind. Assigned only on
-      // success, so a mid-acquire throw leaves these undefined for `finally`.
-      const attemptLeases = await acquireAttemptLeases(leaseSource, candidate, {
-        signal: state.abortController.signal,
-        ...(deps.poolSelector ? { poolSelector: deps.poolSelector } : {}),
-        ...(deps.networkBindingFactory ? { networkBindingFactory: deps.networkBindingFactory } : {}),
-        ...(host ? { host } : {}),
-        strictPoolSelection: input.strictPoolSelection,
-      });
-      lease = attemptLeases.lease;
-      reservation = attemptLeases.reservation;
-      proxySlot = attemptLeases.proxySlot;
-      networkPoolId = attemptLeases.networkPoolId;
-      leaseOwnership = attemptLeases.ownership;
-      // Downstream error paths gate `commitUsage(estimated)` on this flag so
-      // the estimate is only charged once the request truly consumed
-      // upstream resources; the lease and reservation are held from here on.
-      bindingEstablished = true;
-      state.startProviderFlight();
-      return await input.attempt({
-        candidate,
-        credential,
-        adapter,
-        leases: attemptLeases,
-        attemptIndex: index,
-        providerCapture,
-        ...(leaseOwnership === undefined ? {} : { leaseOwnership }),
-        retainLeases: () => {
-          leasesRetained = true;
-        },
-      });
-    } catch (error) {
-      // A drain aborts the controller, so a mid-flight attempt unwinds here as
-      // an abort. Surface the typed shutdown error (503 shutting_down /
-      // restart_for_update) instead of an opaque `transport_closed`, so the
-      // client's JSON error — not just the SSE path — names the real cause.
-      const drain = drainAbortReason(state.abortController.signal.reason);
-      if (drain !== undefined) error = drain;
-      lastError = error;
-      // A drain is a *server* close, not a client cancel: it must be recorded
-      // as a failure with the shutdown code, and it is never retryable.
-      // Status, category, and origin come from one classifier so they cannot
-      // disagree (a `cancelled` row used to carry `transport_unavailable`/502).
-      const terminal = classifyTerminalOutcome(error, state.abortController.signal);
-      const cancelled = terminal.status === "cancelled";
-      // A failure the *pool* caused retires that pool, not the candidate: the
-      // next attempt keeps the same provider/model/account and selects a
-      // different egress. That is a genuinely new attempt even though the
-      // candidate index does not move, so it must not be judged terminal by
-      // "this was the last candidate" — that rule is about candidates, and
-      // opencode has exactly one (it is credential-less, so the plan holds a
-      // single candidate carrying every pool). Judging it there is what made a
-      // refused egress end the request instead of failing over to a healthy
-      // one.
-      const rateLimitedPool = !cancelled && shouldCooldownPool(error, candidate.provider_id);
-      // A free-tier refusal is only ever this provider's: the guard is gated on
-      // `ipScopedRateLimit` in the provider metadata, so a 403 from any other
-      // upstream still means what it always did and never retires an egress.
-      const blockedPool = !cancelled && isFreeTierIpBlock(error, candidate.provider_id);
-      const poolScopedFailure =
-        (rateLimitedPool || blockedPool) &&
-        networkPoolId !== undefined &&
-        deps.poolSelector !== undefined;
-      // Flag before the decision below, so the retry that follows already sees
-      // the pool excluded. Both flag helpers write the local entry
-      // synchronously, so the ordering holds even though the Redis mirror is
-      // fire-and-forget.
-      if (networkPoolId && deps.poolSelector && error instanceof GatewayError) {
-        if (rateLimitedPool) {
-          flagPoolCooldown(deps.poolSelector, deps.db, networkPoolId, candidate.provider_id, error);
-        } else if (blockedPool) {
-          // The address itself is refused, not the request: retire the egress
-          // so the next attempt leaves by a different one instead of retrying
-          // a proxy the upstream has already flagged.
-          flagPoolBlocked(deps.poolSelector, deps.db, networkPoolId, candidate.provider_id, error);
-        }
-      }
-      const terminalAttempt =
-        cancelled ||
-        !isRetryableFailure(error) ||
-        (index === candidates.length - 1 && !poolScopedFailure);
-      if (providerCapture.request !== null) firedCaptures.push(providerCapture);
-      // The payload store only keeps the terminal attempt's exchange. When the
-      // terminal attempt never fired (failover walked past the attempt that
-      // did, or prepare/lease threw first), fall back to the last fired
-      // exchange so Request Detail still shows the translated provider
-      // request instead of "—" for a request that reached an upstream.
-      const terminalCapture =
-        providerCapture.request !== null
-          ? providerCapture
-          : [...firedCaptures].reverse().find((capture) => capture.request !== null) ??
-            providerCapture;
-      await completeAttempt(state, {
-        status: terminal.status,
-        providerId: candidate.provider_id,
-        ...(candidate.provider_account_id ? { accountId: candidate.provider_account_id } : {}),
-        ...(candidate.provider_account_label ? { accountLabel: candidate.provider_account_label } : {}),
-        errorCategory: terminal.errorCategory,
-        errorOrigin: terminal.errorOrigin,
-        modelId: candidate.model_id,
-        ...(networkPoolId ? { networkPoolId } : {}),
-        error,
-        ...(lease ? { lease } : {}),
-        // A failed attempt may have sent bytes upstream, so it is not free —
-        // but it must not be charged the full input estimate plus a full
-        // output reserve either, because a request that walks N candidates
-        // would then leave N estimates on the counters for one client request.
-        // Commit the input estimate only: nothing was generated, and the
-        // `finally` release then refunds the (uncommitted) output reserve.
-        ...(!cancelled && bindingEstablished
-          ? {
-              commitUsage: repriceUsage(
-                estimatedUsage(leaseSource.estimatedInputTokens, 0),
-                candidate.provider_id,
-                candidate.model_id,
-              ),
-            }
-          : {}),
-        terminal: terminalAttempt,
-        tenantId: input.tenantId,
-        ingressBody: state.ingressBody,
-        responseBody: null,
-        // The client still receives the public error envelope on the wire —
-        // store it so Request Detail's Client Response panel traces failures
-        // exactly as the client saw them. Cancelled requests have no client
-        // response (the caller is gone).
-        ...(terminalAttempt && !cancelled
-          ? { clientResponseText: errorClientResponseBody(error) }
-          : {}),
-        providerCapture: terminalCapture,
-        db: deps.db,
-        ...(deps.telemetryBuffer ? { telemetryBuffer: deps.telemetryBuffer } : {}),
-        ...(deps.snapshotService ? { snapshotService: deps.snapshotService } : {}),
-      });
-      // A 403 can be a deterministic provider policy rejection, not an
-      // invalid OAuth credential. Only the evidence-based classifier above
-      // may trigger a refresh.
-      const candidateKey = `${candidate.provider_id}:${candidate.provider_account_id ?? "global"}:${candidate.model_id}:${candidate.endpoint}`;
-      const authInvalidated = isOAuthCredentialInvalidated(error);
-      if (
-        !cancelled &&
-        !refreshedCandidates.has(candidateKey) &&
-        authInvalidated &&
-        candidate.provider_account_id &&
-        deps.oauthRefreshService &&
-        deps.resolveOAuthRefresher
-      ) {
-        const refresher = await deps.resolveOAuthRefresher(candidate.provider_id);
-        if (refresher) {
-          const fresh = await deps.oauthRefreshService.ensureFreshAccessToken(
-            candidate.provider_account_id,
-            refresher,
-            { force: true },
-          );
-          if (fresh) {
-            refreshedCandidates.add(candidateKey);
-            continue;
-          }
-        }
-      }
-      if (terminalAttempt) {
-        metrics.proxy_requests_total.inc(1, { status: cancelled ? "cancelled" : "failed" });
-        throw error;
-      }
-      // A failed combo member stays flagged for this request. Move it behind
-      // every candidate that has not failed yet, so the next attempt cannot
-      // stick to it while a sibling remains. It becomes eligible again only
-      // after that sibling list is exhausted.
-      //
-      // A pool-scoped failure is exempt: the candidate never got its turn, so
-      // marking it failed would spend its only remaining chance on the pool
-      // that just refused it. The candidate stays at the head, and the next
-      // attempt selects a different pool for it. The attempt budget still
-      // bounds the loop, and once every pool is retired the selection itself
-      // fails with `proxy_pool_blocked`, which is terminal.
-      const failureKey = comboCandidateKey(candidate);
-      if (poolScopedFailure) {
-        // Deliberately no index movement and no `failedCandidates` entry.
-      } else if (!failedCandidates.has(failureKey)) {
-        failedCandidates.add(failureKey);
-        candidates.push(candidates.splice(index, 1)[0]!);
-      } else {
-        index += 1;
-      }
-      retryDelayMs = fallbackRetryDelayMs(attempts);
-    } finally {
-      if (!leasesRetained)
-        await releaseAttemptLeases(
-          { lease, reservation, proxySlot, ...(leaseOwnership === undefined ? {} : { ownership: leaseOwnership }) },
-          leaseSource.routingEngine,
-        );
-    }
-    attempts += 1;
-    // Charged to the member that spent it, so one member's own attempts are
-    // what exhaust its share — not the running total across every member.
-    attemptsByMember.set(memberKey, (attemptsByMember.get(memberKey) ?? 0) + 1);
-    if (retryDelayMs > 0) {
-      try {
-        await sleep(retryDelayMs, state.abortController.signal);
-      } catch {
-        // Abort during backoff used to escape as the sleep's own error with
-        // no counter increment: attempts already recorded their telemetry,
-        // but `proxy_requests_total` missed the request. Fail with the last
-        // attempt's error so the client sees the real cause, counted once.
-        metrics.proxy_requests_total.inc(1, { status: "cancelled" });
-        throw lastError;
-      }
-    }
-  }
-  // The loop can fall out with a retryable error still in hand (every attempt
-  // was retryable, so the terminal branch in the catch never fired). That used
-  // to throw with no completion: no telemetry row, no payload capture, and
-  // `state.completed` left clear — so a request that genuinely exhausted its
-  // attempts was invisible in Request Detail. Record it as terminal here.
-  if (!state.completed) {
-    const terminal = classifyTerminalOutcome(lastError, state.abortController.signal);
-    const lastCandidate = candidates[Math.min(index, candidates.length - 1)];
-    const terminalCapture = [...firedCaptures]
-      .reverse()
-      .find((capture) => capture.request !== null);
-    await completeAttempt(state, {
-      status: terminal.status,
-      providerId: lastCandidate?.provider_id ?? "",
-      ...(lastCandidate?.provider_account_id
-        ? { accountId: lastCandidate.provider_account_id }
-        : {}),
-      ...(lastCandidate?.provider_account_label
-        ? { accountLabel: lastCandidate.provider_account_label }
-        : {}),
-      errorCategory: terminal.errorCategory,
-      errorOrigin: terminal.errorOrigin,
-      ...(lastCandidate?.model_id === undefined ? {} : { modelId: lastCandidate.model_id }),
-      error: lastError,
-      terminal: true,
-      tenantId: input.tenantId,
-      ingressBody: state.ingressBody,
-      responseBody: null,
-      // The client receives the public error envelope on the wire — store it
-      // so the Client Response panel traces the failure as the client saw it.
-      ...(terminal.status === "cancelled"
-        ? {}
-        : { clientResponseText: errorClientResponseBody(lastError) }),
-      ...(terminalCapture === undefined ? {} : { providerCapture: terminalCapture }),
-      db: deps.db,
-      ...(deps.telemetryBuffer ? { telemetryBuffer: deps.telemetryBuffer } : {}),
-      ...(deps.snapshotService ? { snapshotService: deps.snapshotService } : {}),
-    });
-  }
-  metrics.proxy_requests_total.inc(1, { status: "failed" });
-  throw lastError;
-}
+/**
+ * Shared per-candidate attempt lifecycle for every dispatch route.
+ *
+ * Admission lease → pool slot → routing reservation, the route-specific
+ * upstream operation, then one `completeAttempt` for the attempt's accounting.
+ * Retry, backoff, pool-cooldown flagging, and reactive OAuth refresh are one
+ * policy here, so the canonical proxy routes and the native Responses-compact
+ * route cannot drift. Route-specific behavior arrives as the `prepare` and
+ * `attempt` closures; this module never forks retry policy.
+ */
+import { GatewayError } from "../../protocol/gateway-error";
+import type { ResolvedCredential } from "../../providers/provider-contracts";
+import {
+  classifyTerminalOutcome,
+  fallbackRetryDelayMs,
+  isRetryableFailure,
+  MIN_RETRY_BUDGET_MS,
+  sleep,
+} from "../failure-policy";
+import { resolveRouteMaxAttempts } from "../../config/timeouts";
+import type { OAuthTokenRefresher } from "../../providers/authentication/oauth-refresh-service";
+import type { OAuthRefreshService } from "../../providers/authentication/oauth-refresh-service";
+import type { ValidatedNetworkBindingFactory } from "../../network/pool/resolver";
+import type { ByokUpstreamHost } from "../../providers/operations/provider-catalog-service";
+import type { AdmissionLease } from "../../security/admission/contracts";
+import type { RouteCandidate, Reservation, RouteSnapshotService } from "../routing/route-model";
+import type { CartethyiaDatabase } from "../../persistence/db-handle";
+import { acquireAttemptLeases, releaseAttemptLeases } from "./leases";
+import type { AttemptLeaseSource, AttemptLeases } from "./leases";
+import type { AttemptLeaseOwnership } from "./leases";
+import { metrics } from "../../observability/metrics";
+import type { NetworkPoolSelector } from "../../network/pool/selector";
+import type { ProxyRequestState } from "../request/state";
+import type { TelemetryBatchBuffer } from "../../observability/telemetry-buffer";
+import { flagPoolBlocked, flagPoolCooldown } from "../../network/pool-health";
+import { completeAttempt, errorClientResponseBody, estimatedUsage, type ProviderExchangeCapture } from "./attempt-finalize";
+import { repriceUsage } from "../../providers/provider-usage";
+import { isFreeTierIpBlock, isOAuthCredentialInvalidated, shouldCooldownPool } from "./retry-policy";
+import { drainAbortReason } from "../shutdown-notice";
+
+/**
+ * Identity of one combo member for the request-scoped failure flag.
+ *
+ * Account, model, and endpoint distinguish siblings that share a provider.
+ * The flag lives only for this request: it stops a failed member from being
+ * retried while another member remains, without writing a durable cooldown.
+ */
+function comboCandidateKey(candidate: RouteCandidate): string {
+  return [
+    candidate.provider_id,
+    candidate.provider_account_id ?? "global",
+    candidate.model_id,
+    candidate.endpoint,
+  ].join(":");
+}
+
+/**
+ * Which combo member a candidate belongs to.
+ *
+ * A combo's members are distinct models (usually one per provider), so the
+ * model id is what separates them — the accounts of one member are siblings
+ * that share it. Used only to give each member its own share of the attempt
+ * budget; a candidate list that is not a combo collapses to one member, which
+ * keeps the ordinary single-model path untouched.
+ */
+function comboMemberKey(candidate: RouteCandidate): string {
+  return `${candidate.provider_id}:${candidate.model_id}:${candidate.endpoint}`;
+}
+
+/**
+ * One member's ordinal, resolved by *identity* rather than by position.
+ *
+ * Deliberately not a run-length scan. `plan()` reorders the eligible list — its
+ * cooling partition emits every healthy candidate before every cooling one — so
+ * one member's accounts can be non-contiguous and even interleaved with another
+ * member's. A run-length scan then reports more "members" than actually exist:
+ * `[cb, wb, cb, wb]` counts four, which pushed each member's skip threshold to
+ * `cap * 4` — well past the whole budget — and silently disabled the cap on
+ * exactly the degraded lists it exists for.
+ *
+ * Keyed off the same string the member count is derived from, so the two can
+ * never disagree.
+ */
+function buildMemberOrdinals(
+  candidates: readonly RouteCandidate[],
+): ReadonlyMap<string, number> {
+  const ordinals = new Map<string, number>();
+  for (const candidate of candidates) {
+    const key = comboMemberKey(candidate);
+    if (!ordinals.has(key)) ordinals.set(key, ordinals.size);
+  }
+  return ordinals;
+}
+
+/** Route-specific preconditions resolved for one candidate before its leases are taken. */
+interface PreparedAttempt<TAdapter> {
+  readonly credential: ResolvedCredential;
+  readonly adapter: TAdapter;
+}
+
+/** One candidate attempt with its admission lease, pool slot, and reservation held. */
+interface AttemptContext<TAdapter> {
+  readonly candidate: RouteCandidate;
+  readonly credential: ResolvedCredential;
+  readonly adapter: TAdapter;
+  readonly leases: AttemptLeases;
+  /** Zero-based candidate index; also this attempt's failover count. */
+  readonly attemptIndex: number;
+  readonly providerCapture: ProviderExchangeCapture;
+  /**
+   * Ownership of this attempt's account/network-pool release, when the route
+   * acquired a pool slot. A provider session that outlives this request (a
+   * Cursor Run held open for the client's tool result) calls
+   * `transferToProviderSession()` and takes it over; otherwise the request
+   * releases it through `releaseByResponse()`.
+   */
+  readonly leaseOwnership?: AttemptLeaseOwnership | undefined;
+  /**
+   * Hands lease ownership to the value this attempt returns (a streaming
+   * response releases its own lease on stream completion/cancel). Without it
+   * the loop releases the leases in `finally`.
+   */
+  retainLeases(): void;
+}
+
+/** Collaborators shared by every dispatch route. */
+interface AttemptLoopDeps {
+  readonly db: CartethyiaDatabase;
+  readonly poolSelector?: NetworkPoolSelector;
+  readonly networkBindingFactory?: ValidatedNetworkBindingFactory;
+  readonly snapshotService?: RouteSnapshotService;
+  readonly telemetryBuffer?: TelemetryBatchBuffer;
+  readonly resolveOAuthRefresher?: (providerId: string) => Promise<OAuthTokenRefresher | undefined>;
+  readonly oauthRefreshService?: OAuthRefreshService;
+}
+
+interface AttemptLoopInput<TResult, TAdapter> {
+  readonly state: ProxyRequestState;
+  readonly deps: AttemptLoopDeps;
+  readonly leaseSource: AttemptLeaseSource;
+  readonly candidates: readonly RouteCandidate[];
+  readonly tenantId: string | null;
+  /** Strict pool selection rejects direct egress when every pool is at capacity. */
+  readonly strictPoolSelection: boolean;
+  /** BYOK host for a candidate, when this route validates network bindings. */
+  readonly resolveHost?: (candidate: RouteCandidate) => ByokUpstreamHost | undefined;
+  readonly prepare: (candidate: RouteCandidate) => Promise<PreparedAttempt<TAdapter>>;
+  readonly attempt: (context: AttemptContext<TAdapter>) => Promise<TResult>;
+  /** Raised when the candidate list is empty and no attempt ever ran. */
+  readonly exhaustedError: GatewayError;
+}
+
+/**
+ * Shared per-candidate attempt lifecycle for the canonical proxy routes and
+ * the native Responses-compact route: admission lease → pool slot → routing
+ * reservation, the route-specific upstream operation, then one
+ * `completeAttempt` for the attempt's accounting. Retry, backoff,
+ * pool-cooldown flagging, and reactive OAuth refresh are one policy here, so
+ * the two routes cannot drift.
+ */
+export async function runAttemptLoop<TResult, TAdapter>(
+  input: AttemptLoopInput<TResult, TAdapter>,
+): Promise<TResult> {
+  const { state, deps, leaseSource } = input;
+  const candidates = [...input.candidates];
+  let lastError: unknown = input.exhaustedError;
+  const refreshedCandidates = new Set<string>();
+  const failedCandidates = new Set<string>();
+  // Every attempt owns a fresh capture sink, but only the terminal attempt's
+  // sink reaches the payload store. When the terminal attempt fails before it
+  // fires an outbound request (lease/acquire/prepare threw, or failover
+  // exhausted), its sink is empty and the drawer reads "—" for a request that
+  // DID reach an upstream on an earlier attempt. Keep the fired exchanges so
+  // the terminal completion can fall back to the last one that has a request.
+  const firedCaptures: ProviderExchangeCapture[] = [];
+  const maxAttempts = resolveRouteMaxAttempts();
+  // One combo member must not be able to spend the whole budget.
+  //
+  // A combo's member list is flattened into this candidate list, so the first
+  // member's accounts occupy its head. With one member holding more accounts
+  // than `maxAttempts` (default 8), every attempt was spent inside that member
+  // and a later, perfectly healthy member was never dialed — a `deepseek` combo
+  // of [cb, wb] stayed stuck on cb while wb sat idle, because cb's twelve
+  // accounts exhausted all eight attempts before the list reached wb.
+  //
+  // Each member gets an equal share of the budget, counted **per member** — not
+  // as a cumulative watermark over total attempts, or one member that over-runs
+  // silently steals the tail's share. The share is a ceiling, not a reservation:
+  // a member that used less than its share simply never reaches it, and the
+  // unused budget stays available to the members behind it.
+  const memberOrdinals = buildMemberOrdinals(candidates);
+  const memberCount = memberOrdinals.size;
+  const perMemberCap =
+    memberCount > 1 ? Math.max(1, Math.ceil(maxAttempts / memberCount)) : maxAttempts;
+  const attemptsByMember = new Map<string, number>();
+  // The budget itself is owned by failure-policy, so this loop, the
+  // empty-stream replay, and the Codex retry budget cannot drift apart.
+  let index = 0;
+  let attempts = 0;
+  while (index < candidates.length && attempts < maxAttempts) {
+    const candidate = candidates[index];
+    if (!candidate) break;
+    const memberKey = comboMemberKey(candidate);
+    // A member that has spent its share is skipped: the loop advances past the
+    // rest of its accounts to the next member instead of burning the budget
+    // they would have consumed.
+    if (
+      memberCount > 1 &&
+      (attemptsByMember.get(memberKey) ?? 0) >= perMemberCap
+    ) {
+      index += 1;
+      continue;
+    }
+    // Do not start an attempt that cannot finish. The measured pre-dispatch
+    // cost of the previous attempt (`resolveMs` ran 60s+ on a slow upstream)
+    // is what a follow-on attempt has to spend again before the upstream is
+    // even called, so a retry that begins with less than that left is a retry
+    // the deadline will kill mid-flight. Stopping here surfaces the first
+    // attempt's own error instead of a second one's timeout.
+    if (attempts > 0) {
+      const remainingMs = state.deadlineMs - Date.now();
+      if (remainingMs < MIN_RETRY_BUDGET_MS) {
+        if (lastError !== undefined && !state.completed) {
+          const terminal = classifyTerminalOutcome(lastError, state.abortController.signal);
+          const lastFired = [...firedCaptures]
+            .reverse()
+            .find((capture) => capture.request !== null);
+          await completeAttempt(state, {
+            status: terminal.status,
+            providerId: candidate.provider_id,
+            ...(candidate.provider_account_id
+              ? { accountId: candidate.provider_account_id }
+              : {}),
+            ...(candidate.provider_account_label
+              ? { accountLabel: candidate.provider_account_label }
+              : {}),
+            errorCategory: terminal.errorCategory,
+            errorOrigin: terminal.errorOrigin,
+            modelId: candidate.model_id,
+            error: lastError,
+            terminal: true,
+            tenantId: input.tenantId,
+            ingressBody: state.ingressBody,
+            responseBody: null,
+            ...(terminal.status !== "cancelled"
+              ? { clientResponseText: errorClientResponseBody(lastError) }
+              : {}),
+            // Same fallback the terminal catch uses: a request that did fire
+            // upstream keeps that exchange in Request Detail.
+            ...(lastFired ? { providerCapture: lastFired } : {}),
+            db: deps.db,
+            ...(deps.telemetryBuffer ? { telemetryBuffer: deps.telemetryBuffer } : {}),
+            ...(deps.snapshotService ? { snapshotService: deps.snapshotService } : {}),
+          });
+          metrics.proxy_requests_total.inc(1, {
+            status: terminal.status === "cancelled" ? "cancelled" : "failed",
+          });
+        }
+        throw lastError;
+      }
+    }
+    let lease: AdmissionLease | undefined;
+    let reservation: Reservation | undefined;
+    let proxySlot: { poolId: string; release: () => void } | undefined;
+    let networkPoolId: string | undefined;
+    let leaseOwnership: AttemptLeaseOwnership | undefined;
+    let bindingEstablished = false;
+    let leasesRetained = false;
+    let retryDelayMs = 0;
+    const providerCapture: ProviderExchangeCapture = {
+      request: null,
+      response: Promise.resolve(null),
+    };
+    try {
+      if (state.abortController.signal.aborted)
+        throw new GatewayError("transport_closed", 499, "request was cancelled");
+      const { credential, adapter } = await input.prepare(candidate);
+      const host = input.resolveHost?.(candidate);
+      // Single attempt-lease manager owns acquisition order (admission →
+      // pool slot → reservation) with reverse-order unwind. Assigned only on
+      // success, so a mid-acquire throw leaves these undefined for `finally`.
+      const attemptLeases = await acquireAttemptLeases(leaseSource, candidate, {
+        signal: state.abortController.signal,
+        ...(deps.poolSelector ? { poolSelector: deps.poolSelector } : {}),
+        ...(deps.networkBindingFactory ? { networkBindingFactory: deps.networkBindingFactory } : {}),
+        ...(host ? { host } : {}),
+        strictPoolSelection: input.strictPoolSelection,
+      });
+      lease = attemptLeases.lease;
+      reservation = attemptLeases.reservation;
+      proxySlot = attemptLeases.proxySlot;
+      networkPoolId = attemptLeases.networkPoolId;
+      leaseOwnership = attemptLeases.ownership;
+      // Downstream error paths gate `commitUsage(estimated)` on this flag so
+      // the estimate is only charged once the request truly consumed
+      // upstream resources; the lease and reservation are held from here on.
+      bindingEstablished = true;
+      state.startProviderFlight();
+      return await input.attempt({
+        candidate,
+        credential,
+        adapter,
+        leases: attemptLeases,
+        attemptIndex: index,
+        providerCapture,
+        ...(leaseOwnership === undefined ? {} : { leaseOwnership }),
+        retainLeases: () => {
+          leasesRetained = true;
+        },
+      });
+    } catch (error) {
+      // A drain aborts the controller, so a mid-flight attempt unwinds here as
+      // an abort. Surface the typed shutdown error (503 shutting_down /
+      // restart_for_update) instead of an opaque `transport_closed`, so the
+      // client's JSON error — not just the SSE path — names the real cause.
+      const drain = drainAbortReason(state.abortController.signal.reason);
+      if (drain !== undefined) error = drain;
+      lastError = error;
+      // A drain is a *server* close, not a client cancel: it must be recorded
+      // as a failure with the shutdown code, and it is never retryable.
+      // Status, category, and origin come from one classifier so they cannot
+      // disagree and mark a `cancelled` row `transport_unavailable`/502.
+      const terminal = classifyTerminalOutcome(error, state.abortController.signal);
+      const cancelled = terminal.status === "cancelled";
+      // A failure the *pool* caused retires that pool, not the candidate: the
+      // next attempt keeps the same provider/model/account and selects a
+      // different egress. That is a genuinely new attempt even though the
+      // candidate index does not move, so it must not be judged terminal by
+      // "this was the last candidate" — that rule is about candidates, and
+      // opencode has exactly one (it is credential-less, so the plan holds a
+      // single candidate carrying every pool). Judging it there is what made a
+      // refused egress end the request instead of failing over to a healthy
+      // one.
+      const rateLimitedPool = !cancelled && shouldCooldownPool(error, candidate.provider_id);
+      // A free-tier refusal is only ever this provider's: the guard is gated on
+      // `ipScopedRateLimit` in the provider metadata, so a 403 from any other
+      // upstream still means what it always did and never retires an egress.
+      const blockedPool = !cancelled && isFreeTierIpBlock(error, candidate.provider_id);
+      const poolScopedFailure =
+        (rateLimitedPool || blockedPool) &&
+        networkPoolId !== undefined &&
+        deps.poolSelector !== undefined;
+      // Flag before the decision below, so the retry that follows already sees
+      // the pool excluded. Both flag helpers write the local entry
+      // synchronously, so the ordering holds even though the Redis mirror is
+      // fire-and-forget.
+      if (networkPoolId && deps.poolSelector && error instanceof GatewayError) {
+        if (rateLimitedPool) {
+          flagPoolCooldown(deps.poolSelector, deps.db, networkPoolId, candidate.provider_id, error);
+        } else if (blockedPool) {
+          // The address itself is refused, not the request: retire the egress
+          // so the next attempt leaves by a different one instead of retrying
+          // a proxy the upstream has already flagged.
+          flagPoolBlocked(deps.poolSelector, deps.db, networkPoolId, candidate.provider_id, error);
+        }
+      }
+      const terminalAttempt =
+        cancelled ||
+        !isRetryableFailure(error) ||
+        (index === candidates.length - 1 && !poolScopedFailure);
+      if (providerCapture.request !== null) firedCaptures.push(providerCapture);
+      // The payload store only keeps the terminal attempt's exchange. When the
+      // terminal attempt never fired (failover walked past the attempt that
+      // did, or prepare/lease threw first), fall back to the last fired
+      // exchange so Request Detail still shows the translated provider
+      // request instead of "—" for a request that reached an upstream.
+      const terminalCapture =
+        providerCapture.request !== null
+          ? providerCapture
+          : [...firedCaptures].reverse().find((capture) => capture.request !== null) ??
+            providerCapture;
+      await completeAttempt(state, {
+        status: terminal.status,
+        providerId: candidate.provider_id,
+        ...(candidate.provider_account_id ? { accountId: candidate.provider_account_id } : {}),
+        ...(candidate.provider_account_label ? { accountLabel: candidate.provider_account_label } : {}),
+        errorCategory: terminal.errorCategory,
+        errorOrigin: terminal.errorOrigin,
+        modelId: candidate.model_id,
+        ...(networkPoolId ? { networkPoolId } : {}),
+        error,
+        ...(lease ? { lease } : {}),
+        // A failed attempt may have sent bytes upstream, so it is not free —
+        // but it must not be charged the full input estimate plus a full
+        // output reserve either, because a request that walks N candidates
+        // would then leave N estimates on the counters for one client request.
+        // Commit the input estimate only: nothing was generated, and the
+        // `finally` release then refunds the (uncommitted) output reserve.
+        ...(!cancelled && bindingEstablished
+          ? {
+              commitUsage: repriceUsage(
+                estimatedUsage(leaseSource.estimatedInputTokens, 0),
+                candidate.provider_id,
+                candidate.model_id,
+              ),
+            }
+          : {}),
+        terminal: terminalAttempt,
+        tenantId: input.tenantId,
+        ingressBody: state.ingressBody,
+        responseBody: null,
+        // The client still receives the public error envelope on the wire —
+        // store it so Request Detail's Client Response panel traces failures
+        // exactly as the client saw them. Cancelled requests have no client
+        // response (the caller is gone).
+        ...(terminalAttempt && !cancelled
+          ? { clientResponseText: errorClientResponseBody(error) }
+          : {}),
+        providerCapture: terminalCapture,
+        db: deps.db,
+        ...(deps.telemetryBuffer ? { telemetryBuffer: deps.telemetryBuffer } : {}),
+        ...(deps.snapshotService ? { snapshotService: deps.snapshotService } : {}),
+      });
+      // A 403 can be a deterministic provider policy rejection, not an
+      // invalid OAuth credential. Only the evidence-based classifier above
+      // may trigger a refresh.
+      const candidateKey = `${candidate.provider_id}:${candidate.provider_account_id ?? "global"}:${candidate.model_id}:${candidate.endpoint}`;
+      const authInvalidated = isOAuthCredentialInvalidated(error);
+      if (
+        !cancelled &&
+        !refreshedCandidates.has(candidateKey) &&
+        authInvalidated &&
+        candidate.provider_account_id &&
+        deps.oauthRefreshService &&
+        deps.resolveOAuthRefresher
+      ) {
+        const refresher = await deps.resolveOAuthRefresher(candidate.provider_id);
+        if (refresher) {
+          const fresh = await deps.oauthRefreshService.ensureFreshAccessToken(
+            candidate.provider_account_id,
+            refresher,
+            { force: true },
+          );
+          if (fresh) {
+            refreshedCandidates.add(candidateKey);
+            continue;
+          }
+        }
+      }
+      if (terminalAttempt) {
+        metrics.proxy_requests_total.inc(1, { status: cancelled ? "cancelled" : "failed" });
+        throw error;
+      }
+      // A failed combo member stays flagged for this request. Move it behind
+      // every candidate that has not failed yet, so the next attempt cannot
+      // stick to it while a sibling remains. It becomes eligible again only
+      // after that sibling list is exhausted.
+      //
+      // A pool-scoped failure is exempt: the candidate never got its turn, so
+      // marking it failed would spend its only remaining chance on the pool
+      // that just refused it. The candidate stays at the head, and the next
+      // attempt selects a different pool for it. The attempt budget still
+      // bounds the loop, and once every pool is retired the selection itself
+      // fails with `proxy_pool_blocked`, which is terminal.
+      const failureKey = comboCandidateKey(candidate);
+      if (poolScopedFailure) {
+        // Deliberately no index movement and no `failedCandidates` entry.
+      } else if (!failedCandidates.has(failureKey)) {
+        failedCandidates.add(failureKey);
+        candidates.push(candidates.splice(index, 1)[0]!);
+      } else {
+        index += 1;
+      }
+      retryDelayMs = fallbackRetryDelayMs(attempts);
+    } finally {
+      if (!leasesRetained)
+        await releaseAttemptLeases(
+          { lease, reservation, proxySlot, ...(leaseOwnership === undefined ? {} : { ownership: leaseOwnership }) },
+          leaseSource.routingEngine,
+        );
+    }
+    attempts += 1;
+    // Charged to the member that spent it, so one member's own attempts are
+    // what exhaust its share — not the running total across every member.
+    attemptsByMember.set(memberKey, (attemptsByMember.get(memberKey) ?? 0) + 1);
+    if (retryDelayMs > 0) {
+      try {
+        await sleep(retryDelayMs, state.abortController.signal);
+      } catch {
+        // Fail with the last attempt's error so the client sees the real cause,
+        // counted once: attempts already recorded their telemetry, but escaping
+        // as the sleep's own error would leave `proxy_requests_total` missing
+        // the request.
+        metrics.proxy_requests_total.inc(1, { status: "cancelled" });
+        throw lastError;
+      }
+    }
+  }
+  // The loop can fall out with a retryable error still in hand (every attempt
+  // was retryable, so the terminal branch in the catch never fired). Throwing
+  // here instead of recording it would leave no telemetry row, no payload
+  // capture, and `state.completed` clear — so a request that genuinely
+  // exhausted its attempts would be invisible in Request Detail. Record it as
+  // terminal here.
+  if (!state.completed) {
+    const terminal = classifyTerminalOutcome(lastError, state.abortController.signal);
+    const lastCandidate = candidates[Math.min(index, candidates.length - 1)];
+    const terminalCapture = [...firedCaptures]
+      .reverse()
+      .find((capture) => capture.request !== null);
+    await completeAttempt(state, {
+      status: terminal.status,
+      providerId: lastCandidate?.provider_id ?? "",
+      ...(lastCandidate?.provider_account_id
+        ? { accountId: lastCandidate.provider_account_id }
+        : {}),
+      ...(lastCandidate?.provider_account_label
+        ? { accountLabel: lastCandidate.provider_account_label }
+        : {}),
+      errorCategory: terminal.errorCategory,
+      errorOrigin: terminal.errorOrigin,
+      ...(lastCandidate?.model_id === undefined ? {} : { modelId: lastCandidate.model_id }),
+      error: lastError,
+      terminal: true,
+      tenantId: input.tenantId,
+      ingressBody: state.ingressBody,
+      responseBody: null,
+      // The client receives the public error envelope on the wire — store it
+      // so the Client Response panel traces the failure as the client saw it.
+      ...(terminal.status === "cancelled"
+        ? {}
+        : { clientResponseText: errorClientResponseBody(lastError) }),
+      ...(terminalCapture === undefined ? {} : { providerCapture: terminalCapture }),
+      db: deps.db,
+      ...(deps.telemetryBuffer ? { telemetryBuffer: deps.telemetryBuffer } : {}),
+      ...(deps.snapshotService ? { snapshotService: deps.snapshotService } : {}),
+    });
+  }
+  metrics.proxy_requests_total.inc(1, { status: "failed" });
+  throw lastError;
+}