@remnic/core 10.0.25 → 10.0.27

package/dist/chunk-3VRMUNWL.js

@@ -0,0 +1,289 @@
+// src/local-llm-rate-limit.ts
+var RATE_LIMIT_STATUSES = /* @__PURE__ */ new Set([408, 429, 503]);
+var RATE_LIMIT_BODY_PATTERN = /no deployments available|cooldown_list|all deployments.*cooldown/i;
+var TRY_AGAIN_PATTERN = /try again in (\d+(?:\.\d+)?) ?(seconds?|secs?|s|milliseconds?|ms|minutes?|mins?)\b/i;
+var MAX_RECORDED_BACKOFF_MS = 10 * 6e4;
+var DEFAULT_RATE_LIMIT_RETRY_COUNT = 3;
+var DEFAULT_RATE_LIMIT_BACKOFF_BASE_MS = 1e3;
+var DEFAULT_RATE_LIMIT_BACKOFF_MAX_MS = 6e4;
+var DEFAULT_MAX_CONCURRENCY = 4;
+function nonNegativeInt(value, fallback) {
+  if (value === void 0 || !Number.isFinite(value) || value < 0) return fallback;
+  return Math.floor(value);
+}
+function resolveRateLimitSettings(config) {
+  const backoffBaseMs = Math.max(
+    1,
+    nonNegativeInt(config.localLlmRateLimitBackoffBaseMs, DEFAULT_RATE_LIMIT_BACKOFF_BASE_MS)
+  );
+  const backoffMaxMs = Math.max(
+    backoffBaseMs,
+    nonNegativeInt(config.localLlmRateLimitBackoffMaxMs, DEFAULT_RATE_LIMIT_BACKOFF_MAX_MS)
+  );
+  return {
+    retryCount: nonNegativeInt(config.localLlmRateLimitRetryCount, DEFAULT_RATE_LIMIT_RETRY_COUNT),
+    backoffBaseMs,
+    backoffMaxMs,
+    maxConcurrency: nonNegativeInt(config.localLlmMaxConcurrency, DEFAULT_MAX_CONCURRENCY)
+  };
+}
+function isRateLimitedResponse(status, bodyText) {
+  if (RATE_LIMIT_STATUSES.has(status)) return true;
+  return status >= 500 && RATE_LIMIT_BODY_PATTERN.test(bodyText);
+}
+function parseRetryAfterMs(value, nowMs = Date.now()) {
+  if (value === null || value === void 0) return null;
+  const trimmed = value.trim();
+  if (!trimmed) return null;
+  if (/^\d+(\.\d+)?$/.test(trimmed)) {
+    return Math.min(MAX_RECORDED_BACKOFF_MS, Math.round(Number(trimmed) * 1e3));
+  }
+  const dateMs = Date.parse(trimmed);
+  if (Number.isNaN(dateMs)) return null;
+  return Math.min(MAX_RECORDED_BACKOFF_MS, Math.max(0, dateMs - nowMs));
+}
+function parseTryAgainHintMs(bodyText) {
+  const match = bodyText.match(TRY_AGAIN_PATTERN);
+  if (!match) return null;
+  const amount = Number(match[1]);
+  if (!Number.isFinite(amount)) return null;
+  const unit = match[2].toLowerCase();
+  let ms;
+  if (unit.startsWith("ms") || unit.startsWith("milli")) ms = amount;
+  else if (unit.startsWith("min")) ms = amount * 6e4;
+  else ms = amount * 1e3;
+  return Math.min(MAX_RECORDED_BACKOFF_MS, Math.round(ms));
+}
+function serverRetryHintMs(retryAfterHeader, bodyText, nowMs = Date.now()) {
+  return parseRetryAfterMs(retryAfterHeader, nowMs) ?? parseTryAgainHintMs(bodyText);
+}
+function computeFullJitterBackoffMs(attempt, baseMs, maxMs, random = Math.random) {
+  const exponent = Math.max(0, Math.min(30, attempt - 1));
+  const ceiling = Math.min(maxMs, baseMs * 2 ** exponent);
+  return Math.floor(random() * ceiling);
+}
+function computeRateLimitDelayMs(attempt, settings, serverHintMs, random = Math.random) {
+  const jittered = computeFullJitterBackoffMs(attempt, settings.backoffBaseMs, settings.backoffMaxMs, random);
+  return serverHintMs === null ? jittered : Math.max(serverHintMs, jittered);
+}
+var LocalLlmConcurrencyLimiter = class {
+  active = 0;
+  max;
+  waiters = [];
+  constructor(max) {
+    this.max = max;
+  }
+  setMax(max) {
+    this.max = max;
+    this.pump();
+  }
+  get inFlight() {
+    return this.active;
+  }
+  get queued() {
+    return this.waiters.length;
+  }
+  /**
+   * Wait for a slot. Resolves to a release function, or null when the
+   * signal aborts first. `max <= 0` means unlimited.
+   */
+  acquire(options = {}) {
+    const { signal, timeoutMs } = options;
+    if (signal?.aborted) return Promise.resolve(null);
+    if (this.max <= 0 || this.active < this.max && this.waiters.length === 0) {
+      this.active += 1;
+      return Promise.resolve(this.releaser());
+    }
+    if (timeoutMs !== void 0 && timeoutMs <= 0) return Promise.resolve(null);
+    return new Promise((resolve) => {
+      let timer;
+      const waiter = {
+        highPriority: options.highPriority === true,
+        grant: () => {
+          signal?.removeEventListener("abort", onAbort);
+          if (timer) clearTimeout(timer);
+          resolve(this.releaser());
+        }
+      };
+      const onAbort = () => {
+        const index = this.waiters.indexOf(waiter);
+        if (index >= 0) this.waiters.splice(index, 1);
+        signal?.removeEventListener("abort", onAbort);
+        if (timer) clearTimeout(timer);
+        resolve(null);
+      };
+      if (waiter.highPriority) {
+        const firstLow = this.waiters.findIndex((w) => !w.highPriority);
+        if (firstLow < 0) this.waiters.push(waiter);
+        else this.waiters.splice(firstLow, 0, waiter);
+      } else {
+        this.waiters.push(waiter);
+      }
+      signal?.addEventListener("abort", onAbort, { once: true });
+      if (timeoutMs !== void 0) timer = setTimeout(onAbort, timeoutMs);
+    });
+  }
+  releaser() {
+    let released = false;
+    return () => {
+      if (released) return;
+      released = true;
+      this.active -= 1;
+      this.pump();
+    };
+  }
+  pump() {
+    while (this.waiters.length > 0 && (this.max <= 0 || this.active < this.max)) {
+      const next = this.waiters.shift();
+      if (!next) break;
+      this.active += 1;
+      next.grant();
+    }
+  }
+};
+function globalState() {
+  globalThis.__remnicLocalLlmRateLimitState ??= {
+    limiter: new LocalLlmConcurrencyLimiter(DEFAULT_MAX_CONCURRENCY),
+    endpoints: /* @__PURE__ */ new Map()
+  };
+  return globalThis.__remnicLocalLlmRateLimitState;
+}
+function getLocalLlmConcurrencyLimiter(maxConcurrency) {
+  const { limiter } = globalState();
+  limiter.setMax(maxConcurrency);
+  return limiter;
+}
+function endpointBackoffRemainingMs(endpoint, nowMs = Date.now()) {
+  const entry = globalState().endpoints.get(endpoint);
+  return entry ? Math.max(0, entry.untilMs - nowMs) : 0;
+}
+function noteEndpointRateLimited(endpoint) {
+  const { endpoints } = globalState();
+  const entry = endpoints.get(endpoint) ?? { untilMs: 0, consecutiveRateLimits: 0 };
+  entry.consecutiveRateLimits += 1;
+  endpoints.set(endpoint, entry);
+  return entry.consecutiveRateLimits;
+}
+function extendEndpointBackoff(endpoint, delayMs, nowMs = Date.now()) {
+  const { endpoints } = globalState();
+  const entry = endpoints.get(endpoint) ?? { untilMs: 0, consecutiveRateLimits: 0 };
+  const bounded = Math.min(MAX_RECORDED_BACKOFF_MS, Math.max(0, delayMs));
+  entry.untilMs = Math.max(entry.untilMs, nowMs + bounded);
+  endpoints.set(endpoint, entry);
+}
+function noteEndpointRecovered(endpoint) {
+  const entry = globalState().endpoints.get(endpoint);
+  if (entry) entry.consecutiveRateLimits = 0;
+}
+function resetLocalLlmRateLimitStateForTests() {
+  const state = globalState();
+  state.endpoints.clear();
+  state.limiter = new LocalLlmConcurrencyLimiter(DEFAULT_MAX_CONCURRENCY);
+}
+var LocalLlmRateLimitSession = class {
+  /**
+   * @param waitBudgetMs the caller's per-attempt time budget. Waiting for the
+   * endpoint backoff or a concurrency slot never exceeds it, and a backoff
+   * longer than it skips the request instead (fail open to the caller).
+   */
+  constructor(endpoint, settings, highPriority, waitBudgetMs = settings.backoffMaxMs) {
+    this.endpoint = endpoint;
+    this.highPriority = highPriority;
+    this.waitBudgetMs = waitBudgetMs;
+    this.settings = settings;
+    this.limiter = getLocalLlmConcurrencyLimiter(settings.maxConcurrency);
+  }
+  endpoint;
+  highPriority;
+  waitBudgetMs;
+  retries = 0;
+  limiter;
+  settings;
+  get retryCount() {
+    return this.retries;
+  }
+  async beforeAttempt(signal) {
+    const deadlineMs = Date.now() + Math.max(0, this.waitBudgetMs);
+    for (; ; ) {
+      const remainingMs = endpointBackoffRemainingMs(this.endpoint);
+      if (remainingMs > 0) {
+        if (remainingMs > Math.min(this.settings.backoffMaxMs, deadlineMs - Date.now())) {
+          return { kind: "skip", reason: "backoff", remainingMs };
+        }
+        if (!await waitWithSignal(remainingMs, signal)) return { kind: "aborted" };
+        continue;
+      }
+      const release = await this.limiter.acquire({
+        highPriority: this.highPriority,
+        signal,
+        timeoutMs: deadlineMs - Date.now()
+      });
+      if (!release) return signal?.aborted ? { kind: "aborted" } : { kind: "skip", reason: "queue", remainingMs: 0 };
+      if (endpointBackoffRemainingMs(this.endpoint) > 0) {
+        release();
+        continue;
+      }
+      return { kind: "go", release };
+    }
+  }
+  onSuccess() {
+    noteEndpointRecovered(this.endpoint);
+  }
+  onResponse(status, retryAfterHeader, bodyText) {
+    if (!isRateLimitedResponse(status, bodyText)) {
+      noteEndpointRecovered(this.endpoint);
+      return { kind: "not-rate-limited" };
+    }
+    const consecutive = noteEndpointRateLimited(this.endpoint);
+    const hintMs = serverRetryHintMs(retryAfterHeader, bodyText);
+    const delayMs = computeRateLimitDelayMs(consecutive, this.settings, hintMs);
+    extendEndpointBackoff(this.endpoint, delayMs);
+    if (this.retries >= this.settings.retryCount || delayMs > Math.min(this.settings.backoffMaxMs, this.waitBudgetMs)) {
+      return { kind: "give-up", delayMs, retries: this.retries };
+    }
+    this.retries += 1;
+    return { kind: "retry", delayMs, hintMs, consecutive, retry: this.retries };
+  }
+};
+function waitWithSignal(ms, signal) {
+  return new Promise((resolve) => {
+    if (signal?.aborted) {
+      resolve(false);
+      return;
+    }
+    const onAbort = () => {
+      clearTimeout(timer);
+      resolve(false);
+    };
+    const timer = setTimeout(() => {
+      signal?.removeEventListener("abort", onAbort);
+      resolve(true);
+    }, ms);
+    signal?.addEventListener("abort", onAbort, { once: true });
+  });
+}
+
+export {
+  RATE_LIMIT_STATUSES,
+  MAX_RECORDED_BACKOFF_MS,
+  DEFAULT_RATE_LIMIT_RETRY_COUNT,
+  DEFAULT_RATE_LIMIT_BACKOFF_BASE_MS,
+  DEFAULT_RATE_LIMIT_BACKOFF_MAX_MS,
+  DEFAULT_MAX_CONCURRENCY,
+  resolveRateLimitSettings,
+  isRateLimitedResponse,
+  parseRetryAfterMs,
+  parseTryAgainHintMs,
+  serverRetryHintMs,
+  computeFullJitterBackoffMs,
+  computeRateLimitDelayMs,
+  LocalLlmConcurrencyLimiter,
+  getLocalLlmConcurrencyLimiter,
+  endpointBackoffRemainingMs,
+  noteEndpointRateLimited,
+  extendEndpointBackoff,
+  noteEndpointRecovered,
+  resetLocalLlmRateLimitStateForTests,
+  LocalLlmRateLimitSession
+};
+//# sourceMappingURL=chunk-3VRMUNWL.js.map

package/dist/chunk-3VRMUNWL.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/local-llm-rate-limit.ts"],"sourcesContent":["/**\n * Process-wide rate-limit handling for the local LLM client (COS-447).\n *\n * A LiteLLM proxy in front of the local models answers `429 No deployments\n * available` (and `503`/`408`) when every deployment is busy or in cooldown.\n * Before this module the client returned on the first 429 and the caller\n * loop immediately issued the next request, so one daemon could hammer the\n * proxy with thousands of calls per hour while it was saturated.\n *\n * This module provides three pieces that the client composes:\n *\n * - classification of rate-limit responses and of the retry hint\n *   (`Retry-After` header as delta-seconds or HTTP-date, falling back to\n *   LiteLLM's `Try again in N seconds` body text);\n * - exponential backoff with full jitter;\n * - process-wide state shared by every `LocalLlmClient` instance: a\n *   concurrency semaphore for in-flight HTTP attempts and a per-endpoint\n *   \"do not call before\" deadline, so one client's 429 slows every client\n *   pointed at the same endpoint.\n *\n * State lives on `globalThis` so duplicated module copies in bundled output\n * still share one limiter.\n */\n\nexport const RATE_LIMIT_STATUSES: ReadonlySet<number> = new Set([408, 429, 503]);\n\nconst RATE_LIMIT_BODY_PATTERN = /no deployments available|cooldown_list|all deployments.*cooldown/i;\nconst TRY_AGAIN_PATTERN = /try again in (\\d+(?:\\.\\d+)?) ?(seconds?|secs?|s|milliseconds?|ms|minutes?|mins?)\\b/i;\n\n/** Upper bound on any server-provided wait we will record (sanity clamp). */\nexport const MAX_RECORDED_BACKOFF_MS = 10 * 60_000;\n\nexport const DEFAULT_RATE_LIMIT_RETRY_COUNT = 3;\nexport const DEFAULT_RATE_LIMIT_BACKOFF_BASE_MS = 1_000;\nexport const DEFAULT_RATE_LIMIT_BACKOFF_MAX_MS = 60_000;\nexport const DEFAULT_MAX_CONCURRENCY = 4;\n\nexport interface LocalLlmRateLimitSettings {\n  retryCount: number;\n  backoffBaseMs: number;\n  backoffMaxMs: number;\n  /** 0 disables the process-wide concurrency cap. */\n  maxConcurrency: number;\n}\n\nfunction nonNegativeInt(value: number | undefined, fallback: number): number {\n  if (value === undefined || !Number.isFinite(value) || value < 0) return fallback;\n  return Math.floor(value);\n}\n\nexport function resolveRateLimitSettings(config: {\n  localLlmRateLimitRetryCount?: number;\n  localLlmRateLimitBackoffBaseMs?: number;\n  localLlmRateLimitBackoffMaxMs?: number;\n  localLlmMaxConcurrency?: number;\n}): LocalLlmRateLimitSettings {\n  const backoffBaseMs = Math.max(\n    1,\n    nonNegativeInt(config.localLlmRateLimitBackoffBaseMs, DEFAULT_RATE_LIMIT_BACKOFF_BASE_MS)\n  );\n  const backoffMaxMs = Math.max(\n    backoffBaseMs,\n    nonNegativeInt(config.localLlmRateLimitBackoffMaxMs, DEFAULT_RATE_LIMIT_BACKOFF_MAX_MS)\n  );\n  return {\n    retryCount: nonNegativeInt(config.localLlmRateLimitRetryCount, DEFAULT_RATE_LIMIT_RETRY_COUNT),\n    backoffBaseMs,\n    backoffMaxMs,\n    maxConcurrency: nonNegativeInt(config.localLlmMaxConcurrency, DEFAULT_MAX_CONCURRENCY),\n  };\n}\n\n/**\n * True for responses that mean \"the endpoint is saturated, slow down\":\n * 408/429/503, or a 5xx whose body carries LiteLLM's cooldown wording.\n * 400s are never rate limits (context-window errors mention fallbacks).\n */\nexport function isRateLimitedResponse(status: number, bodyText: string): boolean {\n  if (RATE_LIMIT_STATUSES.has(status)) return true;\n  return status >= 500 && RATE_LIMIT_BODY_PATTERN.test(bodyText);\n}\n\n/**\n * Parse an HTTP `Retry-After` value (RFC 9110: delta-seconds or HTTP-date).\n * Returns milliseconds to wait, or null when absent/invalid.\n */\nexport function parseRetryAfterMs(value: string | null | undefined, nowMs: number = Date.now()): number | null {\n  if (value === null || value === undefined) return null;\n  const trimmed = value.trim();\n  if (!trimmed) return null;\n  if (/^\\d+(\\.\\d+)?$/.test(trimmed)) {\n    return Math.min(MAX_RECORDED_BACKOFF_MS, Math.round(Number(trimmed) * 1000));\n  }\n  const dateMs = Date.parse(trimmed);\n  if (Number.isNaN(dateMs)) return null;\n  return Math.min(MAX_RECORDED_BACKOFF_MS, Math.max(0, dateMs - nowMs));\n}\n\n/** Parse LiteLLM's `Try again in N seconds` hint from an error body. */\nexport function parseTryAgainHintMs(bodyText: string): number | null {\n  const match = bodyText.match(TRY_AGAIN_PATTERN);\n  if (!match) return null;\n  const amount = Number(match[1]);\n  if (!Number.isFinite(amount)) return null;\n  const unit = match[2].toLowerCase();\n  let ms: number;\n  if (unit.startsWith(\"ms\") || unit.startsWith(\"milli\")) ms = amount;\n  else if (unit.startsWith(\"min\")) ms = amount * 60_000;\n  else ms = amount * 1000;\n  return Math.min(MAX_RECORDED_BACKOFF_MS, Math.round(ms));\n}\n\n/** Server-provided wait: `Retry-After` header first, then the body hint. */\nexport function serverRetryHintMs(\n  retryAfterHeader: string | null | undefined,\n  bodyText: string,\n  nowMs: number = Date.now()\n): number | null {\n  return parseRetryAfterMs(retryAfterHeader, nowMs) ?? parseTryAgainHintMs(bodyText);\n}\n\n/**\n * Exponential backoff with full jitter: uniform in [0, min(max, base * 2^(n-1))].\n * `attempt` is 1-based (the number of consecutive rate-limit responses).\n */\nexport function computeFullJitterBackoffMs(\n  attempt: number,\n  baseMs: number,\n  maxMs: number,\n  random: () => number = Math.random\n): number {\n  const exponent = Math.max(0, Math.min(30, attempt - 1));\n  const ceiling = Math.min(maxMs, baseMs * 2 ** exponent);\n  return Math.floor(random() * ceiling);\n}\n\n/**\n * Delay before the next attempt after a rate-limit response. A server hint\n * is a floor (never call earlier than asked); jitter is added on top of the\n * exponential term so synchronized callers spread out.\n */\nexport function computeRateLimitDelayMs(\n  attempt: number,\n  settings: Pick<LocalLlmRateLimitSettings, \"backoffBaseMs\" | \"backoffMaxMs\">,\n  serverHintMs: number | null,\n  random: () => number = Math.random\n): number {\n  const jittered = computeFullJitterBackoffMs(attempt, settings.backoffBaseMs, settings.backoffMaxMs, random);\n  return serverHintMs === null ? jittered : Math.max(serverHintMs, jittered);\n}\n\ntype Waiter = { highPriority: boolean; grant: () => void };\n\n/** FIFO counting semaphore with a high-priority lane (recall-critical first). */\nexport class LocalLlmConcurrencyLimiter {\n  private active = 0;\n  private max: number;\n  private readonly waiters: Waiter[] = [];\n\n  constructor(max: number) {\n    this.max = max;\n  }\n\n  setMax(max: number): void {\n    this.max = max;\n    this.pump();\n  }\n\n  get inFlight(): number {\n    return this.active;\n  }\n\n  get queued(): number {\n    return this.waiters.length;\n  }\n\n  /**\n   * Wait for a slot. Resolves to a release function, or null when the\n   * signal aborts first. `max <= 0` means unlimited.\n   */\n  acquire(\n    options: { highPriority?: boolean; signal?: AbortSignal; timeoutMs?: number } = {}\n  ): Promise<(() => void) | null> {\n    const { signal, timeoutMs } = options;\n    if (signal?.aborted) return Promise.resolve(null);\n    if (this.max <= 0 || (this.active < this.max && this.waiters.length === 0)) {\n      this.active += 1;\n      return Promise.resolve(this.releaser());\n    }\n    if (timeoutMs !== undefined && timeoutMs <= 0) return Promise.resolve(null);\n    return new Promise((resolve) => {\n      let timer: NodeJS.Timeout | undefined;\n      const waiter: Waiter = {\n        highPriority: options.highPriority === true,\n        grant: () => {\n          signal?.removeEventListener(\"abort\", onAbort);\n          if (timer) clearTimeout(timer);\n          resolve(this.releaser());\n        },\n      };\n      const onAbort = (): void => {\n        const index = this.waiters.indexOf(waiter);\n        if (index >= 0) this.waiters.splice(index, 1);\n        signal?.removeEventListener(\"abort\", onAbort);\n        if (timer) clearTimeout(timer);\n        resolve(null);\n      };\n      if (waiter.highPriority) {\n        const firstLow = this.waiters.findIndex((w) => !w.highPriority);\n        if (firstLow < 0) this.waiters.push(waiter);\n        else this.waiters.splice(firstLow, 0, waiter);\n      } else {\n        this.waiters.push(waiter);\n      }\n      signal?.addEventListener(\"abort\", onAbort, { once: true });\n      if (timeoutMs !== undefined) timer = setTimeout(onAbort, timeoutMs);\n    });\n  }\n\n  private releaser(): () => void {\n    let released = false;\n    return () => {\n      if (released) return;\n      released = true;\n      this.active -= 1;\n      this.pump();\n    };\n  }\n\n  private pump(): void {\n    while (this.waiters.length > 0 && (this.max <= 0 || this.active < this.max)) {\n      const next = this.waiters.shift();\n      if (!next) break;\n      this.active += 1;\n      next.grant();\n    }\n  }\n}\n\ninterface EndpointBackoffState {\n  untilMs: number;\n  consecutiveRateLimits: number;\n}\n\ninterface RateLimitGlobalState {\n  limiter: LocalLlmConcurrencyLimiter;\n  endpoints: Map<string, EndpointBackoffState>;\n}\n\ndeclare global {\n  // Shared across duplicated module copies in bundled output.\n  var __remnicLocalLlmRateLimitState: RateLimitGlobalState | undefined;\n}\n\nfunction globalState(): RateLimitGlobalState {\n  globalThis.__remnicLocalLlmRateLimitState ??= {\n    limiter: new LocalLlmConcurrencyLimiter(DEFAULT_MAX_CONCURRENCY),\n    endpoints: new Map(),\n  };\n  return globalThis.__remnicLocalLlmRateLimitState;\n}\n\n/** Process-wide limiter, resized to the caller's configured cap. */\nexport function getLocalLlmConcurrencyLimiter(maxConcurrency: number): LocalLlmConcurrencyLimiter {\n  const { limiter } = globalState();\n  limiter.setMax(maxConcurrency);\n  return limiter;\n}\n\n/** Milliseconds until `endpoint` may be called again (0 when clear). */\nexport function endpointBackoffRemainingMs(endpoint: string, nowMs: number = Date.now()): number {\n  const entry = globalState().endpoints.get(endpoint);\n  return entry ? Math.max(0, entry.untilMs - nowMs) : 0;\n}\n\n/**\n * Record a rate-limit response for `endpoint`. Returns the consecutive\n * rate-limit count (1-based) to feed the backoff exponent.\n */\nexport function noteEndpointRateLimited(endpoint: string): number {\n  const { endpoints } = globalState();\n  const entry = endpoints.get(endpoint) ?? { untilMs: 0, consecutiveRateLimits: 0 };\n  entry.consecutiveRateLimits += 1;\n  endpoints.set(endpoint, entry);\n  return entry.consecutiveRateLimits;\n}\n\n/** Push the endpoint's \"do not call before\" deadline out to at least `nowMs + delayMs`. */\nexport function extendEndpointBackoff(endpoint: string, delayMs: number, nowMs: number = Date.now()): void {\n  const { endpoints } = globalState();\n  const entry = endpoints.get(endpoint) ?? { untilMs: 0, consecutiveRateLimits: 0 };\n  const bounded = Math.min(MAX_RECORDED_BACKOFF_MS, Math.max(0, delayMs));\n  entry.untilMs = Math.max(entry.untilMs, nowMs + bounded);\n  endpoints.set(endpoint, entry);\n}\n\n/**\n * Reset the backoff exponent after a non-rate-limited response. An already\n * recorded deadline is left to expire so concurrent callers that saw the\n * 429 still wait it out.\n */\nexport function noteEndpointRecovered(endpoint: string): void {\n  const entry = globalState().endpoints.get(endpoint);\n  if (entry) entry.consecutiveRateLimits = 0;\n}\n\n/** Test seam: drop all process-wide state. */\nexport function resetLocalLlmRateLimitStateForTests(): void {\n  const state = globalState();\n  state.endpoints.clear();\n  state.limiter = new LocalLlmConcurrencyLimiter(DEFAULT_MAX_CONCURRENCY);\n}\n\nexport type RateLimitAttemptGate =\n  | { kind: \"go\"; release: () => void }\n  | { kind: \"aborted\" }\n  | { kind: \"skip\"; reason: \"backoff\" | \"queue\"; remainingMs: number };\n\nexport type RateLimitDecision =\n  | { kind: \"not-rate-limited\" }\n  | { kind: \"retry\"; delayMs: number; hintMs: number | null; consecutive: number; retry: number }\n  | { kind: \"give-up\"; delayMs: number; retries: number };\n\n/**\n * Per-request driver for the rate-limit policy. The client calls\n * `beforeAttempt` ahead of every HTTP attempt (waits out the shared endpoint\n * deadline, then takes a concurrency slot) and `onResponse` after every\n * non-ok response.\n */\nexport class LocalLlmRateLimitSession {\n  private retries = 0;\n  private readonly limiter: LocalLlmConcurrencyLimiter;\n  readonly settings: LocalLlmRateLimitSettings;\n\n  /**\n   * @param waitBudgetMs the caller's per-attempt time budget. Waiting for the\n   * endpoint backoff or a concurrency slot never exceeds it, and a backoff\n   * longer than it skips the request instead (fail open to the caller).\n   */\n  constructor(\n    private readonly endpoint: string,\n    settings: LocalLlmRateLimitSettings,\n    private readonly highPriority: boolean,\n    private readonly waitBudgetMs: number = settings.backoffMaxMs\n  ) {\n    this.settings = settings;\n    this.limiter = getLocalLlmConcurrencyLimiter(settings.maxConcurrency);\n  }\n\n  get retryCount(): number {\n    return this.retries;\n  }\n\n  async beforeAttempt(signal?: AbortSignal): Promise<RateLimitAttemptGate> {\n    const deadlineMs = Date.now() + Math.max(0, this.waitBudgetMs);\n    // Recheck the shared deadline after every async wait: another attempt may\n    // have recorded or extended a cooldown while this one slept or queued.\n    for (;;) {\n      const remainingMs = endpointBackoffRemainingMs(this.endpoint);\n      if (remainingMs > 0) {\n        if (remainingMs > Math.min(this.settings.backoffMaxMs, deadlineMs - Date.now())) {\n          return { kind: \"skip\", reason: \"backoff\", remainingMs };\n        }\n        if (!(await waitWithSignal(remainingMs, signal))) return { kind: \"aborted\" };\n        continue;\n      }\n      const release = await this.limiter.acquire({\n        highPriority: this.highPriority,\n        signal,\n        timeoutMs: deadlineMs - Date.now(),\n      });\n      if (!release) return signal?.aborted ? { kind: \"aborted\" } : { kind: \"skip\", reason: \"queue\", remainingMs: 0 };\n      if (endpointBackoffRemainingMs(this.endpoint) > 0) {\n        release();\n        continue;\n      }\n      return { kind: \"go\", release };\n    }\n  }\n\n  onSuccess(): void {\n    noteEndpointRecovered(this.endpoint);\n  }\n\n  onResponse(status: number, retryAfterHeader: string | null | undefined, bodyText: string): RateLimitDecision {\n    if (!isRateLimitedResponse(status, bodyText)) {\n      noteEndpointRecovered(this.endpoint);\n      return { kind: \"not-rate-limited\" };\n    }\n    const consecutive = noteEndpointRateLimited(this.endpoint);\n    const hintMs = serverRetryHintMs(retryAfterHeader, bodyText);\n    const delayMs = computeRateLimitDelayMs(consecutive, this.settings, hintMs);\n    extendEndpointBackoff(this.endpoint, delayMs);\n    if (this.retries >= this.settings.retryCount || delayMs > Math.min(this.settings.backoffMaxMs, this.waitBudgetMs)) {\n      return { kind: \"give-up\", delayMs, retries: this.retries };\n    }\n    this.retries += 1;\n    return { kind: \"retry\", delayMs, hintMs, consecutive, retry: this.retries };\n  }\n}\n\nfunction waitWithSignal(ms: number, signal?: AbortSignal): Promise<boolean> {\n  return new Promise((resolve) => {\n    if (signal?.aborted) {\n      resolve(false);\n      return;\n    }\n    const onAbort = (): void => {\n      clearTimeout(timer);\n      resolve(false);\n    };\n    const timer = setTimeout(() => {\n      signal?.removeEventListener(\"abort\", onAbort);\n      resolve(true);\n    }, ms);\n    signal?.addEventListener(\"abort\", onAbort, { once: true });\n  });\n}\n"],"mappings":";AAwBO,IAAM,sBAA2C,oBAAI,IAAI,CAAC,KAAK,KAAK,GAAG,CAAC;AAE/E,IAAM,0BAA0B;AAChC,IAAM,oBAAoB;AAGnB,IAAM,0BAA0B,KAAK;AAErC,IAAM,iCAAiC;AACvC,IAAM,qCAAqC;AAC3C,IAAM,oCAAoC;AAC1C,IAAM,0BAA0B;AAUvC,SAAS,eAAe,OAA2B,UAA0B;AAC3E,MAAI,UAAU,UAAa,CAAC,OAAO,SAAS,KAAK,KAAK,QAAQ,EAAG,QAAO;AACxE,SAAO,KAAK,MAAM,KAAK;AACzB;AAEO,SAAS,yBAAyB,QAKX;AAC5B,QAAM,gBAAgB,KAAK;AAAA,IACzB;AAAA,IACA,eAAe,OAAO,gCAAgC,kCAAkC;AAAA,EAC1F;AACA,QAAM,eAAe,KAAK;AAAA,IACxB;AAAA,IACA,eAAe,OAAO,+BAA+B,iCAAiC;AAAA,EACxF;AACA,SAAO;AAAA,IACL,YAAY,eAAe,OAAO,6BAA6B,8BAA8B;AAAA,IAC7F;AAAA,IACA;AAAA,IACA,gBAAgB,eAAe,OAAO,wBAAwB,uBAAuB;AAAA,EACvF;AACF;AAOO,SAAS,sBAAsB,QAAgB,UAA2B;AAC/E,MAAI,oBAAoB,IAAI,MAAM,EAAG,QAAO;AAC5C,SAAO,UAAU,OAAO,wBAAwB,KAAK,QAAQ;AAC/D;AAMO,SAAS,kBAAkB,OAAkC,QAAgB,KAAK,IAAI,GAAkB;AAC7G,MAAI,UAAU,QAAQ,UAAU,OAAW,QAAO;AAClD,QAAM,UAAU,MAAM,KAAK;AAC3B,MAAI,CAAC,QAAS,QAAO;AACrB,MAAI,gBAAgB,KAAK,OAAO,GAAG;AACjC,WAAO,KAAK,IAAI,yBAAyB,KAAK,MAAM,OAAO,OAAO,IAAI,GAAI,CAAC;AAAA,EAC7E;AACA,QAAM,SAAS,KAAK,MAAM,OAAO;AACjC,MAAI,OAAO,MAAM,MAAM,EAAG,QAAO;AACjC,SAAO,KAAK,IAAI,yBAAyB,KAAK,IAAI,GAAG,SAAS,KAAK,CAAC;AACtE;AAGO,SAAS,oBAAoB,UAAiC;AACnE,QAAM,QAAQ,SAAS,MAAM,iBAAiB;AAC9C,MAAI,CAAC,MAAO,QAAO;AACnB,QAAM,SAAS,OAAO,MAAM,CAAC,CAAC;AAC9B,MAAI,CAAC,OAAO,SAAS,MAAM,EAAG,QAAO;AACrC,QAAM,OAAO,MAAM,CAAC,EAAE,YAAY;AAClC,MAAI;AACJ,MAAI,KAAK,WAAW,IAAI,KAAK,KAAK,WAAW,OAAO,EAAG,MAAK;AAAA,WACnD,KAAK,WAAW,KAAK,EAAG,MAAK,SAAS;AAAA,MAC1C,MAAK,SAAS;AACnB,SAAO,KAAK,IAAI,yBAAyB,KAAK,MAAM,EAAE,CAAC;AACzD;AAGO,SAAS,kBACd,kBACA,UACA,QAAgB,KAAK,IAAI,GACV;AACf,SAAO,kBAAkB,kBAAkB,KAAK,KAAK,oBAAoB,QAAQ;AACnF;AAMO,SAAS,2BACd,SACA,QACA,OACA,SAAuB,KAAK,QACpB;AACR,QAAM,WAAW,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,UAAU,CAAC,CAAC;AACtD,QAAM,UAAU,KAAK,IAAI,OAAO,SAAS,KAAK,QAAQ;AACtD,SAAO,KAAK,MAAM,OAAO,IAAI,OAAO;AACtC;AAOO,SAAS,wBACd,SACA,UACA,cACA,SAAuB,KAAK,QACpB;AACR,QAAM,WAAW,2BAA2B,SAAS,SAAS,eAAe,SAAS,cAAc,MAAM;AAC1G,SAAO,iBAAiB,OAAO,WAAW,KAAK,IAAI,cAAc,QAAQ;AAC3E;AAKO,IAAM,6BAAN,MAAiC;AAAA,EAC9B,SAAS;AAAA,EACT;AAAA,EACS,UAAoB,CAAC;AAAA,EAEtC,YAAY,KAAa;AACvB,SAAK,MAAM;AAAA,EACb;AAAA,EAEA,OAAO,KAAmB;AACxB,SAAK,MAAM;AACX,SAAK,KAAK;AAAA,EACZ;AAAA,EAEA,IAAI,WAAmB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,IAAI,SAAiB;AACnB,WAAO,KAAK,QAAQ;AAAA,EACtB;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,QACE,UAAgF,CAAC,GACnD;AAC9B,UAAM,EAAE,QAAQ,UAAU,IAAI;AAC9B,QAAI,QAAQ,QAAS,QAAO,QAAQ,QAAQ,IAAI;AAChD,QAAI,KAAK,OAAO,KAAM,KAAK,SAAS,KAAK,OAAO,KAAK,QAAQ,WAAW,GAAI;AAC1E,WAAK,UAAU;AACf,aAAO,QAAQ,QAAQ,KAAK,SAAS,CAAC;AAAA,IACxC;AACA,QAAI,cAAc,UAAa,aAAa,EAAG,QAAO,QAAQ,QAAQ,IAAI;AAC1E,WAAO,IAAI,QAAQ,CAAC,YAAY;AAC9B,UAAI;AACJ,YAAM,SAAiB;AAAA,QACrB,cAAc,QAAQ,iBAAiB;AAAA,QACvC,OAAO,MAAM;AACX,kBAAQ,oBAAoB,SAAS,OAAO;AAC5C,cAAI,MAAO,cAAa,KAAK;AAC7B,kBAAQ,KAAK,SAAS,CAAC;AAAA,QACzB;AAAA,MACF;AACA,YAAM,UAAU,MAAY;AAC1B,cAAM,QAAQ,KAAK,QAAQ,QAAQ,MAAM;AACzC,YAAI,SAAS,EAAG,MAAK,QAAQ,OAAO,OAAO,CAAC;AAC5C,gBAAQ,oBAAoB,SAAS,OAAO;AAC5C,YAAI,MAAO,cAAa,KAAK;AAC7B,gBAAQ,IAAI;AAAA,MACd;AACA,UAAI,OAAO,cAAc;AACvB,cAAM,WAAW,KAAK,QAAQ,UAAU,CAAC,MAAM,CAAC,EAAE,YAAY;AAC9D,YAAI,WAAW,EAAG,MAAK,QAAQ,KAAK,MAAM;AAAA,YACrC,MAAK,QAAQ,OAAO,UAAU,GAAG,MAAM;AAAA,MAC9C,OAAO;AACL,aAAK,QAAQ,KAAK,MAAM;AAAA,MAC1B;AACA,cAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AACzD,UAAI,cAAc,OAAW,SAAQ,WAAW,SAAS,SAAS;AAAA,IACpE,CAAC;AAAA,EACH;AAAA,EAEQ,WAAuB;AAC7B,QAAI,WAAW;AACf,WAAO,MAAM;AACX,UAAI,SAAU;AACd,iBAAW;AACX,WAAK,UAAU;AACf,WAAK,KAAK;AAAA,IACZ;AAAA,EACF;AAAA,EAEQ,OAAa;AACnB,WAAO,KAAK,QAAQ,SAAS,MAAM,KAAK,OAAO,KAAK,KAAK,SAAS,KAAK,MAAM;AAC3E,YAAM,OAAO,KAAK,QAAQ,MAAM;AAChC,UAAI,CAAC,KAAM;AACX,WAAK,UAAU;AACf,WAAK,MAAM;AAAA,IACb;AAAA,EACF;AACF;AAiBA,SAAS,cAAoC;AAC3C,aAAW,mCAAmC;AAAA,IAC5C,SAAS,IAAI,2BAA2B,uBAAuB;AAAA,IAC/D,WAAW,oBAAI,IAAI;AAAA,EACrB;AACA,SAAO,WAAW;AACpB;AAGO,SAAS,8BAA8B,gBAAoD;AAChG,QAAM,EAAE,QAAQ,IAAI,YAAY;AAChC,UAAQ,OAAO,cAAc;AAC7B,SAAO;AACT;AAGO,SAAS,2BAA2B,UAAkB,QAAgB,KAAK,IAAI,GAAW;AAC/F,QAAM,QAAQ,YAAY,EAAE,UAAU,IAAI,QAAQ;AAClD,SAAO,QAAQ,KAAK,IAAI,GAAG,MAAM,UAAU,KAAK,IAAI;AACtD;AAMO,SAAS,wBAAwB,UAA0B;AAChE,QAAM,EAAE,UAAU,IAAI,YAAY;AAClC,QAAM,QAAQ,UAAU,IAAI,QAAQ,KAAK,EAAE,SAAS,GAAG,uBAAuB,EAAE;AAChF,QAAM,yBAAyB;AAC/B,YAAU,IAAI,UAAU,KAAK;AAC7B,SAAO,MAAM;AACf;AAGO,SAAS,sBAAsB,UAAkB,SAAiB,QAAgB,KAAK,IAAI,GAAS;AACzG,QAAM,EAAE,UAAU,IAAI,YAAY;AAClC,QAAM,QAAQ,UAAU,IAAI,QAAQ,KAAK,EAAE,SAAS,GAAG,uBAAuB,EAAE;AAChF,QAAM,UAAU,KAAK,IAAI,yBAAyB,KAAK,IAAI,GAAG,OAAO,CAAC;AACtE,QAAM,UAAU,KAAK,IAAI,MAAM,SAAS,QAAQ,OAAO;AACvD,YAAU,IAAI,UAAU,KAAK;AAC/B;AAOO,SAAS,sBAAsB,UAAwB;AAC5D,QAAM,QAAQ,YAAY,EAAE,UAAU,IAAI,QAAQ;AAClD,MAAI,MAAO,OAAM,wBAAwB;AAC3C;AAGO,SAAS,sCAA4C;AAC1D,QAAM,QAAQ,YAAY;AAC1B,QAAM,UAAU,MAAM;AACtB,QAAM,UAAU,IAAI,2BAA2B,uBAAuB;AACxE;AAkBO,IAAM,2BAAN,MAA+B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUpC,YACmB,UACjB,UACiB,cACA,eAAuB,SAAS,cACjD;AAJiB;AAEA;AACA;AAEjB,SAAK,WAAW;AAChB,SAAK,UAAU,8BAA8B,SAAS,cAAc;AAAA,EACtE;AAAA,EAPmB;AAAA,EAEA;AAAA,EACA;AAAA,EAbX,UAAU;AAAA,EACD;AAAA,EACR;AAAA,EAiBT,IAAI,aAAqB;AACvB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,cAAc,QAAqD;AACvE,UAAM,aAAa,KAAK,IAAI,IAAI,KAAK,IAAI,GAAG,KAAK,YAAY;AAG7D,eAAS;AACP,YAAM,cAAc,2BAA2B,KAAK,QAAQ;AAC5D,UAAI,cAAc,GAAG;AACnB,YAAI,cAAc,KAAK,IAAI,KAAK,SAAS,cAAc,aAAa,KAAK,IAAI,CAAC,GAAG;AAC/E,iBAAO,EAAE,MAAM,QAAQ,QAAQ,WAAW,YAAY;AAAA,QACxD;AACA,YAAI,CAAE,MAAM,eAAe,aAAa,MAAM,EAAI,QAAO,EAAE,MAAM,UAAU;AAC3E;AAAA,MACF;AACA,YAAM,UAAU,MAAM,KAAK,QAAQ,QAAQ;AAAA,QACzC,cAAc,KAAK;AAAA,QACnB;AAAA,QACA,WAAW,aAAa,KAAK,IAAI;AAAA,MACnC,CAAC;AACD,UAAI,CAAC,QAAS,QAAO,QAAQ,UAAU,EAAE,MAAM,UAAU,IAAI,EAAE,MAAM,QAAQ,QAAQ,SAAS,aAAa,EAAE;AAC7G,UAAI,2BAA2B,KAAK,QAAQ,IAAI,GAAG;AACjD,gBAAQ;AACR;AAAA,MACF;AACA,aAAO,EAAE,MAAM,MAAM,QAAQ;AAAA,IAC/B;AAAA,EACF;AAAA,EAEA,YAAkB;AAChB,0BAAsB,KAAK,QAAQ;AAAA,EACrC;AAAA,EAEA,WAAW,QAAgB,kBAA6C,UAAqC;AAC3G,QAAI,CAAC,sBAAsB,QAAQ,QAAQ,GAAG;AAC5C,4BAAsB,KAAK,QAAQ;AACnC,aAAO,EAAE,MAAM,mBAAmB;AAAA,IACpC;AACA,UAAM,cAAc,wBAAwB,KAAK,QAAQ;AACzD,UAAM,SAAS,kBAAkB,kBAAkB,QAAQ;AAC3D,UAAM,UAAU,wBAAwB,aAAa,KAAK,UAAU,MAAM;AAC1E,0BAAsB,KAAK,UAAU,OAAO;AAC5C,QAAI,KAAK,WAAW,KAAK,SAAS,cAAc,UAAU,KAAK,IAAI,KAAK,SAAS,cAAc,KAAK,YAAY,GAAG;AACjH,aAAO,EAAE,MAAM,WAAW,SAAS,SAAS,KAAK,QAAQ;AAAA,IAC3D;AACA,SAAK,WAAW;AAChB,WAAO,EAAE,MAAM,SAAS,SAAS,QAAQ,aAAa,OAAO,KAAK,QAAQ;AAAA,EAC5E;AACF;AAEA,SAAS,eAAe,IAAY,QAAwC;AAC1E,SAAO,IAAI,QAAQ,CAAC,YAAY;AAC9B,QAAI,QAAQ,SAAS;AACnB,cAAQ,KAAK;AACb;AAAA,IACF;AACA,UAAM,UAAU,MAAY;AAC1B,mBAAa,KAAK;AAClB,cAAQ,KAAK;AAAA,IACf;AACA,UAAM,QAAQ,WAAW,MAAM;AAC7B,cAAQ,oBAAoB,SAAS,OAAO;AAC5C,cAAQ,IAAI;AAAA,IACd,GAAG,EAAE;AACL,YAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AAAA,EAC3D,CAAC;AACH;","names":[]}

package/dist/{chunk-466EGMFQ.js → chunk-42KSEX6C.js}

@@ -19,4 +19,4 @@ export {
   confidenceTier,
   SPECULATIVE_TTL_DAYS
 };
-//# sourceMappingURL=chunk-466EGMFQ.js.map
+//# sourceMappingURL=chunk-42KSEX6C.js.map