dsh-plugin-subscriptions 0.5.2 → 0.6.0

package/lib/providers/pool-family.d.ts

@@ -0,0 +1,56 @@
+/**
+ * Same-subscription account routing: the picker is the union of every
+ * account's catalog (deduped by wire id). A model listed by two or more
+ * accounts failovers between them; a model listed by only one account is
+ * sent to that account. No extra pool identity, no cross-provider
+ * aggregation.
+ */
+import type { LlmModelInfo } from '@deepseek-ai/dsh-llm';
+import type { ProviderId } from '../auth/store.js';
+/** One pool member: an exact provider/account/model route. */
+export interface PoolMemberRef {
+    provider: ProviderId;
+    /** Account key; omitted in configured members to mean "the default account". */
+    account?: string;
+    model: string;
+}
+/** A member with its account resolved (no config indirection left). */
+export type ConcretePoolMember = PoolMemberRef & {
+    account: string;
+};
+/** One pool: its members plus display metadata for the picker. */
+export interface PoolDefinition {
+    members: PoolMemberRef[];
+    /** Display name of the catalog entry (account pools) or the pool id (extras). */
+    name?: string;
+    /** Description of the catalog entry, when the pool borrowed one. */
+    description?: string;
+    /**
+     * When true, the pool is an extra picker entry (a configured tier). Account
+     * pools leave this unset so they reuse the provider's existing catalog row.
+     */
+    extra?: boolean;
+}
+/** One account's catalog as seen through that account's credentials. */
+export interface AccountCatalog {
+    account: string;
+    models: readonly LlmModelInfo[];
+}
+/** One provider's contribution to account-pool aggregation. */
+export interface ProviderPoolSource {
+    /** Per-account catalogs, default first. A model pools only the accounts that list it. */
+    catalogs: readonly AccountCatalog[];
+}
+/** Map key for one provider's pool of one model (ids collide across providers). */
+export declare function poolKey(provider: string, model: string): string;
+/**
+ * Build per-provider account routes. Each model id becomes a definition of
+ * the accounts that list it: two or more fail over; one is pinned to that
+ * account (so a Max-only model is never sent to a Plus login). The picker
+ * unions these catalogs; a logout that drops a model to one account keeps
+ * the same id and pins it to whoever remains.
+ * @param sources - per-account catalogs (providers with no accounts list
+ *   nothing and simply never join a pool).
+ * @returns `provider/model` → pool definition (not listed as an extra entry).
+ */
+export declare function buildAccountPools(sources: Partial<Record<ProviderId, ProviderPoolSource>>): Map<string, PoolDefinition>;

package/lib/providers/pool-family.js

@@ -0,0 +1,45 @@
+/**
+ * Same-subscription account routing: the picker is the union of every
+ * account's catalog (deduped by wire id). A model listed by two or more
+ * accounts failovers between them; a model listed by only one account is
+ * sent to that account. No extra pool identity, no cross-provider
+ * aggregation.
+ */
+/** Map key for one provider's pool of one model (ids collide across providers). */
+export function poolKey(provider, model) {
+    return `${provider}/${model}`;
+}
+/**
+ * Build per-provider account routes. Each model id becomes a definition of
+ * the accounts that list it: two or more fail over; one is pinned to that
+ * account (so a Max-only model is never sent to a Plus login). The picker
+ * unions these catalogs; a logout that drops a model to one account keeps
+ * the same id and pins it to whoever remains.
+ * @param sources - per-account catalogs (providers with no accounts list
+ *   nothing and simply never join a pool).
+ * @returns `provider/model` → pool definition (not listed as an extra entry).
+ */
+export function buildAccountPools(sources) {
+    const pools = new Map();
+    for (const [provider, source] of Object.entries(sources)) {
+        const byModel = new Map();
+        for (const catalog of source.catalogs) {
+            for (const model of catalog.models) {
+                let entry = byModel.get(model.id);
+                if (entry === undefined) {
+                    entry = { members: [], info: model };
+                    byModel.set(model.id, entry);
+                }
+                entry.members.push({ provider, account: catalog.account, model: model.id });
+            }
+        }
+        for (const [id, { members, info }] of byModel) {
+            pools.set(poolKey(provider, id), {
+                members,
+                ...info.name === undefined || info.name === id ? {} : { name: info.name },
+                ...info.description === undefined ? {} : { description: info.description },
+            });
+        }
+    }
+    return pools;
+}

package/lib/providers/pool-health.d.ts

@@ -0,0 +1,74 @@
+/**
+ * Health bookkeeping for pool members: which `(provider, account, model)`
+ * member is cooling down after a failure, and for how long. Purely in-memory
+ * — a restart re-probes members naturally, so nothing here is persisted.
+ *
+ * The failure classifier maps the adapters' stable `LlmError` codes (see
+ * `httpLlmError`/`mapFetchFailure` in `common.ts`) to one of three actions:
+ * switch to another member with a cooldown, switch without recording a
+ * cooldown (transport blips say nothing about the account), or rethrow
+ * (the request itself is at fault and another account would fail alike).
+ */
+import type { ProviderId } from '../auth/store.js';
+/** Registry key for one pool member. */
+export declare function memberKey(provider: ProviderId, account: string, model: string): string;
+/** Registry key parking EVERY member of one account (account-level failures). */
+export declare function accountKey(provider: ProviderId, account: string): string;
+/** Default cooldown when a quota/rate failure carries no `retry-after`. */
+export declare const DEFAULT_QUOTA_COOLDOWN_MS: number;
+/** Auth failures recheck after a day; a re-login clears the record immediately. */
+export declare const AUTH_COOLDOWN_MS: number;
+/** Transient server-side failures cool down briefly. */
+export declare const TRANSIENT_COOLDOWN_MS = 60000;
+/** Whether a failure parks one member or the account's whole quota. */
+export type PoolFailureScope = 'member' | 'account';
+/** What the pool should do with a member that just failed. */
+export type PoolFailureAction = {
+    action: 'switch';
+    cooldownMs: number;
+    reason: string;
+    scope: PoolFailureScope;
+} | {
+    action: 'switch';
+} | {
+    action: 'throw';
+};
+/**
+ * Classify a member failure. Quota and rate-limit failures cool down (using
+ * the provider's own `retry-after` when sent, which is more accurate than
+ * any fixed guess) — account-wide for account-metered providers, per-member
+ * for model-scoped ones; auth failures park the account until re-login
+ * (credentials are account-level); server/timeout failures get a short
+ * per-member cooldown; transport failures switch without a record;
+ * everything else — most importantly CONTEXT_WINDOW_EXCEEDED and ABORTED —
+ * is the request's own fault and is rethrown untouched.
+ * @param error - the failure thrown by a member adapter's stream.
+ * @param provider - the failing member's provider (decides the quota scope).
+ * @returns the action the pool should take.
+ */
+export declare function classifyPoolFailure(error: unknown, provider: ProviderId): PoolFailureAction;
+/**
+ * Cooldown registry keyed by {@link memberKey}. A member whose cooldown has
+ * expired is simply available again — recovery is proven by the next real
+ * request, not by a background probe.
+ */
+export declare class PoolHealthRegistry {
+    private readonly records;
+    /** Whether a member may serve: neither it nor its whole account is cooling. */
+    isMemberAvailable(provider: ProviderId, account: string, model: string, now?: number): boolean;
+    /** Whether one registry key is clear right now. */
+    isAvailable(key: string, now?: number): boolean;
+    /** Park a member for `cooldownMs`; a longer existing cooldown wins. */
+    markUnavailable(key: string, cooldownMs: number, reason: string, now?: number): void;
+    /**
+     * Epoch ms at which the earliest cooling record among `keys` recovers;
+     * `undefined` when none of them is cooling. The registry is shared by
+     * every pool, so the caller passes the keys of ITS members (member and
+     * account keys alike) — an unrelated pool's cooldown must not shape this
+     * pool's retry hint. Feeds the pool-exhausted error's
+     * `providerRetryAfterMs`.
+     */
+    earliestRecovery(keys: ReadonlySet<string>, now?: number): number | undefined;
+    /** Drop records of one provider, or of a single account when given (auth changes). */
+    clear(provider: ProviderId, account?: string): void;
+}

package/lib/providers/pool-health.js

@@ -0,0 +1,148 @@
+/**
+ * Health bookkeeping for pool members: which `(provider, account, model)`
+ * member is cooling down after a failure, and for how long. Purely in-memory
+ * — a restart re-probes members naturally, so nothing here is persisted.
+ *
+ * The failure classifier maps the adapters' stable `LlmError` codes (see
+ * `httpLlmError`/`mapFetchFailure` in `common.ts`) to one of three actions:
+ * switch to another member with a cooldown, switch without recording a
+ * cooldown (transport blips say nothing about the account), or rethrow
+ * (the request itself is at fault and another account would fail alike).
+ */
+import { CONTEXT_WINDOW_EXCEEDED_CODE, LlmError, QUOTA_EXCEEDED_CODE } from '@deepseek-ai/dsh-llm';
+/** Registry key for one pool member. */
+export function memberKey(provider, account, model) {
+    return `${provider}/${account}/${model}`;
+}
+/** Registry key parking EVERY member of one account (account-level failures). */
+export function accountKey(provider, account) {
+    return `${provider}/${account}/*`;
+}
+/** Default cooldown when a quota/rate failure carries no `retry-after`. */
+export const DEFAULT_QUOTA_COOLDOWN_MS = 5 * 60_000;
+/** Auth failures recheck after a day; a re-login clears the record immediately. */
+export const AUTH_COOLDOWN_MS = 24 * 60 * 60_000;
+/** Transient server-side failures cool down briefly. */
+export const TRANSIENT_COOLDOWN_MS = 60_000;
+/**
+ * Providers whose quota windows are model-scoped, so a quota failure on one
+ * model says nothing about its siblings (Claude's Opus/Sonnet lanes). Every
+ * other provider meters the account as a whole: one member hitting the wall
+ * means its siblings on the SAME account would too, so the cooldown parks
+ * the account (other accounts of the provider are unaffected).
+ */
+const MODEL_SCOPED_QUOTA_PROVIDERS = new Set(['claude']);
+/** The `retry-after` an adapter propagated through `httpLlmError`, when any. */
+function retryAfterMs(error) {
+    return error.failure.providerRetryAfterMs;
+}
+/**
+ * Classify a member failure. Quota and rate-limit failures cool down (using
+ * the provider's own `retry-after` when sent, which is more accurate than
+ * any fixed guess) — account-wide for account-metered providers, per-member
+ * for model-scoped ones; auth failures park the account until re-login
+ * (credentials are account-level); server/timeout failures get a short
+ * per-member cooldown; transport failures switch without a record;
+ * everything else — most importantly CONTEXT_WINDOW_EXCEEDED and ABORTED —
+ * is the request's own fault and is rethrown untouched.
+ * @param error - the failure thrown by a member adapter's stream.
+ * @param provider - the failing member's provider (decides the quota scope).
+ * @returns the action the pool should take.
+ */
+export function classifyPoolFailure(error, provider) {
+    if (!(error instanceof LlmError))
+        return { action: 'throw' };
+    switch (error.code) {
+        case QUOTA_EXCEEDED_CODE:
+        case 'RATE_LIMIT':
+            return {
+                action: 'switch',
+                cooldownMs: retryAfterMs(error) ?? DEFAULT_QUOTA_COOLDOWN_MS,
+                reason: error.code,
+                scope: MODEL_SCOPED_QUOTA_PROVIDERS.has(provider) ? 'member' : 'account',
+            };
+        case 'AUTH':
+        case 'INVALID_CREDENTIAL':
+        case 'MISSING_CREDENTIAL':
+            return { action: 'switch', cooldownMs: AUTH_COOLDOWN_MS, reason: error.code, scope: 'account' };
+        case 'SERVER':
+        case 'TIMEOUT':
+        case 'EMPTY_RESPONSE':
+            return { action: 'switch', cooldownMs: TRANSIENT_COOLDOWN_MS, reason: error.code, scope: 'member' };
+        case 'TRANSPORT':
+            return { action: 'switch' };
+        case 'HTTP_402':
+        case 'HTTP_404':
+            // Plan/model availability is account-shaped: another account of the
+            // same subscription may still serve. A 400 that is not a context-window
+            // error stays HTTP_400 and throws (the request itself is at fault).
+            return { action: 'switch', cooldownMs: TRANSIENT_COOLDOWN_MS, reason: error.code, scope: 'member' };
+        case CONTEXT_WINDOW_EXCEEDED_CODE:
+        case 'ABORTED':
+        default:
+            return { action: 'throw' };
+    }
+}
+/**
+ * Cooldown registry keyed by {@link memberKey}. A member whose cooldown has
+ * expired is simply available again — recovery is proven by the next real
+ * request, not by a background probe.
+ */
+export class PoolHealthRegistry {
+    records = new Map();
+    /** Whether a member may serve: neither it nor its whole account is cooling. */
+    isMemberAvailable(provider, account, model, now = Date.now()) {
+        return this.isAvailable(accountKey(provider, account), now)
+            && this.isAvailable(memberKey(provider, account, model), now);
+    }
+    /** Whether one registry key is clear right now. */
+    isAvailable(key, now = Date.now()) {
+        const record = this.records.get(key);
+        if (record === undefined)
+            return true;
+        if (record.unavailableUntil <= now) {
+            this.records.delete(key);
+            return true;
+        }
+        return false;
+    }
+    /** Park a member for `cooldownMs`; a longer existing cooldown wins. */
+    markUnavailable(key, cooldownMs, reason, now = Date.now()) {
+        const until = now + cooldownMs;
+        const existing = this.records.get(key);
+        if (existing !== undefined && existing.unavailableUntil > until)
+            return;
+        this.records.set(key, { unavailableUntil: until, reason });
+    }
+    /**
+     * Epoch ms at which the earliest cooling record among `keys` recovers;
+     * `undefined` when none of them is cooling. The registry is shared by
+     * every pool, so the caller passes the keys of ITS members (member and
+     * account keys alike) — an unrelated pool's cooldown must not shape this
+     * pool's retry hint. Feeds the pool-exhausted error's
+     * `providerRetryAfterMs`.
+     */
+    earliestRecovery(keys, now = Date.now()) {
+        let earliest;
+        for (const [key, record] of this.records) {
+            if (record.unavailableUntil <= now) {
+                this.records.delete(key);
+                continue;
+            }
+            if (!keys.has(key))
+                continue;
+            if (earliest === undefined || record.unavailableUntil < earliest) {
+                earliest = record.unavailableUntil;
+            }
+        }
+        return earliest;
+    }
+    /** Drop records of one provider, or of a single account when given (auth changes). */
+    clear(provider, account) {
+        const prefix = account === undefined ? `${provider}/` : `${provider}/${account}/`;
+        for (const key of [...this.records.keys()]) {
+            if (key.startsWith(prefix))
+                this.records.delete(key);
+        }
+    }
+}

package/lib/providers/pool-usage.d.ts

@@ -0,0 +1,78 @@
+/**
+ * Quota tracking for pool members: polls the providers' usage endpoints
+ * (the same normalized `ProviderUsage` shape the Settings page consumes) and
+ * turns the windows into a scheduling score.
+ *
+ * The score is a REQUIRED BURN RATE: the fraction of the window that must be
+ * consumed per millisecond for the quota to be exactly used up at reset time
+ * (`remaining / timeUntilReset`). Subscription quota does not roll over, so a
+ * window about to reset with plenty left is the most urgent to spend — the
+ * `quota_aware` strategy therefore prefers the highest-urgency member, which
+ * over time converges on every window hitting zero right at its reset.
+ */
+import type { ProviderUsage } from './common.js';
+import type { ProviderId } from '../auth/store.js';
+import type { ConcretePoolMember } from './pool-family.js';
+/** A member is taken out of rotation once any window crosses this fill level. */
+export declare const QUOTA_FULL_PERCENT = 95;
+/** How long a usage snapshot is trusted before a background refresh. */
+export declare const USAGE_TTL_MS: number;
+/** The scheduling view of one member's quota. */
+export interface MemberQuota {
+    /** False when a window is effectively full or the login is gone. */
+    available: boolean;
+    /** Required burn rate (fraction of window per ms); 0 when unknown. */
+    urgency: number;
+    /** Epoch ms of the snapshot this was computed from; 0 when none. */
+    fetchedAt: number;
+}
+/**
+ * Per-ACCOUNT usage snapshots with in-flight dedupe and
+ * stale-while-revalidate refresh. Providers without a usage endpoint
+ * (copilot) resolve no fetcher and score a constant zero urgency — which
+ * naturally ranks them behind every measured member. Fetchers are resolved
+ * lazily per (provider, account) so accounts added after startup join
+ * tracking on their first score.
+ */
+export declare class PoolUsageTracker {
+    private readonly fetcherFor;
+    private readonly ttlMs;
+    private readonly entries;
+    private readonly inflight;
+    constructor(fetcherFor: (provider: ProviderId, account: string) => (() => Promise<ProviderUsage>) | undefined, ttlMs?: number);
+    /**
+     * The quota view of one member. A cold cache awaits the first fetch; a
+     * stale one answers immediately while the refresh serves the NEXT call
+     * (member selection must never block on the network mid-conversation). A
+     * failure still cooling down degrades immediately with no network call.
+     * @param member - the pool member to score (account resolved).
+     * @returns availability plus the urgency score.
+     */
+    quotaFor(member: ConcretePoolMember): Promise<MemberQuota>;
+    /**
+     * Same cache as {@link quotaFor}, for direct display (the Settings page):
+     * the raw snapshot, or the original fetch error, instead of a routing
+     * score.
+     * @param provider - the account's provider.
+     * @param account - the account key.
+     * @param force - bypass a fresh cached SNAPSHOT for an honest re-check (the
+     *   manual Refresh button). A live failure cooldown is never bypassed —
+     *   retrying through it is exactly what turns a 429 into a permanent
+     *   lockout, so even a forced call still answers from the negative cache.
+     * @returns `{ supported: false }` when the provider has no usage fetcher.
+     */
+    snapshotFor(provider: ProviderId, account: string, force?: boolean): Promise<ProviderUsage>;
+    /** Drop cached snapshots: one account, or a whole provider when `account` is omitted. */
+    invalidate(provider: ProviderId, account?: string): void;
+    /**
+     * Run (or join) the single in-flight fetch for one account key, caching
+     * either outcome. A missing/invalid credential is deliberately NOT
+     * negative-cached: it costs no network round trip (the session lookup
+     * fails before the request goes out) and re-checking live means the
+     * member rejoins routing the instant its login is fixed, rather than
+     * waiting out a stale cooldown.
+     */
+    private refresh;
+    /** Score one member against a snapshot's windows. */
+    private score;
+}

package/lib/providers/pool-usage.js

@@ -0,0 +1,185 @@
+/**
+ * Quota tracking for pool members: polls the providers' usage endpoints
+ * (the same normalized `ProviderUsage` shape the Settings page consumes) and
+ * turns the windows into a scheduling score.
+ *
+ * The score is a REQUIRED BURN RATE: the fraction of the window that must be
+ * consumed per millisecond for the quota to be exactly used up at reset time
+ * (`remaining / timeUntilReset`). Subscription quota does not roll over, so a
+ * window about to reset with plenty left is the most urgent to spend — the
+ * `quota_aware` strategy therefore prefers the highest-urgency member, which
+ * over time converges on every window hitting zero right at its reset.
+ */
+import { isMissingOrInvalidCredential, OAuthEndpointError } from './common.js';
+/** A member is taken out of rotation once any window crosses this fill level. */
+export const QUOTA_FULL_PERCENT = 95;
+/** How long a usage snapshot is trusted before a background refresh. */
+export const USAGE_TTL_MS = 5 * 60_000;
+/** Assumed window length when the provider discloses no `resetsAt`. */
+const FALLBACK_HORIZON_MS = {
+    session: 5 * 60 * 60_000,
+    weekly: 7 * 24 * 60 * 60_000,
+    other: 30 * 24 * 60 * 60_000,
+};
+/**
+ * Per-ACCOUNT usage snapshots with in-flight dedupe and
+ * stale-while-revalidate refresh. Providers without a usage endpoint
+ * (copilot) resolve no fetcher and score a constant zero urgency — which
+ * naturally ranks them behind every measured member. Fetchers are resolved
+ * lazily per (provider, account) so accounts added after startup join
+ * tracking on their first score.
+ */
+export class PoolUsageTracker {
+    fetcherFor;
+    ttlMs;
+    entries = new Map();
+    inflight = new Map();
+    constructor(fetcherFor, ttlMs = USAGE_TTL_MS) {
+        this.fetcherFor = fetcherFor;
+        this.ttlMs = ttlMs;
+    }
+    /**
+     * The quota view of one member. A cold cache awaits the first fetch; a
+     * stale one answers immediately while the refresh serves the NEXT call
+     * (member selection must never block on the network mid-conversation). A
+     * failure still cooling down degrades immediately with no network call.
+     * @param member - the pool member to score (account resolved).
+     * @returns availability plus the urgency score.
+     */
+    async quotaFor(member) {
+        const key = `${member.provider}/${member.account}`;
+        const fetcher = this.fetcherFor(member.provider, member.account);
+        if (fetcher === undefined)
+            return { available: true, urgency: 0, fetchedAt: 0 };
+        const entry = this.entries.get(key);
+        if (entry !== undefined) {
+            const fresh = Date.now() - entry.at < (entry.cooldownMs ?? this.ttlMs);
+            if (entry.snapshot !== undefined) {
+                if (!fresh)
+                    void this.refresh(key, fetcher).catch(() => undefined);
+                return this.score(member, entry);
+            }
+            if (fresh)
+                return degradedQuota(entry.error);
+            // The cooldown expired: fall through to a fresh, blocking attempt.
+        }
+        try {
+            const snapshot = await this.refresh(key, fetcher);
+            return this.score(member, { snapshot, at: Date.now() });
+        }
+        catch (error) {
+            return degradedQuota(error);
+        }
+    }
+    /**
+     * Same cache as {@link quotaFor}, for direct display (the Settings page):
+     * the raw snapshot, or the original fetch error, instead of a routing
+     * score.
+     * @param provider - the account's provider.
+     * @param account - the account key.
+     * @param force - bypass a fresh cached SNAPSHOT for an honest re-check (the
+     *   manual Refresh button). A live failure cooldown is never bypassed —
+     *   retrying through it is exactly what turns a 429 into a permanent
+     *   lockout, so even a forced call still answers from the negative cache.
+     * @returns `{ supported: false }` when the provider has no usage fetcher.
+     */
+    async snapshotFor(provider, account, force = false) {
+        const fetcher = this.fetcherFor(provider, account);
+        if (fetcher === undefined)
+            return { supported: false };
+        const key = `${provider}/${account}`;
+        const entry = this.entries.get(key);
+        if (entry !== undefined && Date.now() - entry.at < (entry.cooldownMs ?? this.ttlMs)) {
+            if (entry.snapshot !== undefined) {
+                if (!force)
+                    return entry.snapshot;
+            }
+            else {
+                throw entry.error;
+            }
+        }
+        return this.refresh(key, fetcher);
+    }
+    /** Drop cached snapshots: one account, or a whole provider when `account` is omitted. */
+    invalidate(provider, account) {
+        if (account !== undefined) {
+            this.entries.delete(`${provider}/${account}`);
+            return;
+        }
+        for (const key of [...this.entries.keys()]) {
+            if (key.startsWith(`${provider}/`))
+                this.entries.delete(key);
+        }
+    }
+    /**
+     * Run (or join) the single in-flight fetch for one account key, caching
+     * either outcome. A missing/invalid credential is deliberately NOT
+     * negative-cached: it costs no network round trip (the session lookup
+     * fails before the request goes out) and re-checking live means the
+     * member rejoins routing the instant its login is fixed, rather than
+     * waiting out a stale cooldown.
+     */
+    refresh(key, fetcher) {
+        let pending = this.inflight.get(key);
+        if (pending === undefined) {
+            pending = fetcher().then((snapshot) => {
+                this.entries.set(key, { snapshot, at: Date.now() });
+                return snapshot;
+            }, (error) => {
+                if (!isMissingOrInvalidCredential(error)) {
+                    this.entries.set(key, { error, at: Date.now(), cooldownMs: cooldownFor(error, this.ttlMs) });
+                }
+                throw error;
+            }).finally(() => {
+                this.inflight.delete(key);
+            });
+            this.inflight.set(key, pending);
+        }
+        return pending;
+    }
+    /** Score one member against a snapshot's windows. */
+    score(member, entry) {
+        const windows = (entry.snapshot.windows ?? []).filter(window => windowApplies(window, member.model));
+        let available = true;
+        let urgency = 0;
+        for (const window of windows) {
+            if (window.usedPercent >= QUOTA_FULL_PERCENT)
+                available = false;
+            urgency = Math.max(urgency, windowUrgency(window));
+        }
+        return { available, urgency, fetchedAt: entry.at };
+    }
+}
+/**
+ * The routing view of a fetch failure. Logged out: the member cannot serve
+ * at all. Any other failure (network, endpoint rate limit) must not block
+ * routing — the member stays available with a zero score, degrading the
+ * strategy to plain priority order for it.
+ */
+function degradedQuota(error) {
+    return isMissingOrInvalidCredential(error)
+        ? { available: false, urgency: 0, fetchedAt: 0 }
+        : { available: true, urgency: 0, fetchedAt: 0 };
+}
+/** How long to hold a failure in the negative cache: the endpoint's own `retry-after`, or the default TTL. */
+function cooldownFor(error, defaultTtlMs) {
+    return error instanceof OAuthEndpointError && error.retryAfterMs !== undefined ? error.retryAfterMs : defaultTtlMs;
+}
+/**
+ * Whether a window constrains this model: unscoped windows always do; a
+ * model-scoped window (Claude's Opus/Sonnet lanes) applies when its scope
+ * names the model family.
+ */
+function windowApplies(window, model) {
+    if (window.scope === undefined)
+        return true;
+    return model.toLowerCase().includes(window.scope.toLowerCase());
+}
+/** The required burn rate of one window (fraction per ms). */
+function windowUrgency(window, now = Date.now()) {
+    const remaining = Math.max(0, 1 - window.usedPercent / 100);
+    const horizon = window.resetsAt !== undefined
+        ? Math.max(window.resetsAt - now, 1)
+        : FALLBACK_HORIZON_MS[window.kind];
+    return remaining / horizon;
+}