@almadar/llm 2.31.0 → 2.33.0

package/dist/{chunk-SXSP6M24.js → chunk-CZX533SU.js}

@@ -246,13 +246,20 @@ var TokenTracker = class {
     };
     /** Sum of provider-reported authoritative costs (e.g. OpenRouter `usage.cost`). */
     this.authoritativeCostUSD = 0;
-    /** Token buckets for calls WITHOUT an authoritative cost — priced cache-aware. */
-    this.computed = { promptTokens: 0, completionTokens: 0, cachedPromptTokens: 0, cacheWriteTokens: 0 };
+    /** Token buckets for calls WITHOUT an authoritative cost — priced cache-aware.
+     *  Keyed by the model that MADE the calls: `setModel` flips the tracker
+     *  between models mid-process (coordinator on pro, subagents on flash), and
+     *  pricing the aggregate at whichever model is current when someone reads
+     *  the estimate mispriced the whole history by up to the models' rate ratio
+     *  (snapshot/diff consumers saw phantom multi-dollar deltas — or $0 after a
+     *  negative-clamp — per battery spec). Per-model buckets keep the
+     *  retroactive late-pricing property without cross-model contamination. */
+    this.computedByModel = /* @__PURE__ */ new Map();
     this.model = model;
   }
   /** Cache-aware cost for one (or an aggregate of) call(s), in USD. */
-  costFor(promptTokens, completionTokens, cached, written) {
-    const costs = getCostForModel(this.model);
+  costFor(model, promptTokens, completionTokens, cached, written) {
+    const costs = getCostForModel(model);
     const cacheReadRate = costs.cacheReadCostPer1K ?? costs.promptCostPer1K;
     const cacheWriteRate = costs.cacheWriteCostPer1K ?? costs.promptCostPer1K;
     const uncached = Math.max(0, promptTokens - cached - written);
@@ -280,11 +287,18 @@ var TokenTracker = class {
       this.authoritativeCostUSD += authoritative;
       estimatedCost = authoritative;
     } else {
-      this.computed.promptTokens += promptTokens;
-      this.computed.completionTokens += completionTokens;
-      this.computed.cachedPromptTokens += cached;
-      this.computed.cacheWriteTokens += written;
-      estimatedCost = this.costFor(promptTokens, completionTokens, cached, written);
+      const bucket = this.computedByModel.get(this.model) ?? {
+        promptTokens: 0,
+        completionTokens: 0,
+        cachedPromptTokens: 0,
+        cacheWriteTokens: 0
+      };
+      bucket.promptTokens += promptTokens;
+      bucket.completionTokens += completionTokens;
+      bucket.cachedPromptTokens += cached;
+      bucket.cacheWriteTokens += written;
+      this.computedByModel.set(this.model, bucket);
+      estimatedCost = this.costFor(this.model, promptTokens, completionTokens, cached, written);
     }
     const entry = {
       timestamp: (/* @__PURE__ */ new Date()).toISOString(),
@@ -309,12 +323,17 @@ var TokenTracker = class {
     return { ...this.usage };
   }
   getEstimatedCost() {
-    return this.authoritativeCostUSD + this.costFor(
-      this.computed.promptTokens,
-      this.computed.completionTokens,
-      this.computed.cachedPromptTokens,
-      this.computed.cacheWriteTokens
-    );
+    let computed = 0;
+    for (const [model, b] of this.computedByModel) {
+      computed += this.costFor(
+        model,
+        b.promptTokens,
+        b.completionTokens,
+        b.cachedPromptTokens,
+        b.cacheWriteTokens
+      );
+    }
+    return this.authoritativeCostUSD + computed;
   }
   getFormattedCost() {
     const cost = this.getEstimatedCost();
@@ -343,7 +362,7 @@ var TokenTracker = class {
       cacheWriteTokens: 0
     };
     this.authoritativeCostUSD = 0;
-    this.computed = { promptTokens: 0, completionTokens: 0, cachedPromptTokens: 0, cacheWriteTokens: 0 };
+    this.computedByModel.clear();
   }
   setModel(model) {
     this.model = model;
@@ -370,4 +389,4 @@ export {
   getGlobalTokenTracker,
   resetGlobalTokenTracker
 };
-//# sourceMappingURL=chunk-SXSP6M24.js.map
+//# sourceMappingURL=chunk-CZX533SU.js.map

package/dist/chunk-CZX533SU.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/rate-limiter.ts","../src/token-tracker.ts"],"sourcesContent":["/**\n * Rate Limiter for LLM API Calls\n *\n * Implements token bucket algorithm with:\n * - Configurable requests per minute/second\n * - Automatic backoff on 429 errors\n * - Queue for pending requests\n *\n * @packageDocumentation\n */\n\nexport interface RateLimiterOptions {\n  /** Maximum requests per minute (default: 60) */\n  requestsPerMinute?: number;\n  /** Maximum requests per second (default: 3) */\n  requestsPerSecond?: number;\n  /** Maximum concurrent requests (default: 5) */\n  maxConcurrent?: number;\n  /** Base delay for exponential backoff in ms (default: 1000) */\n  baseBackoffMs?: number;\n  /** Maximum backoff delay in ms (default: 60000) */\n  maxBackoffMs?: number;\n}\n\ninterface QueuedRequest<T> {\n  execute: () => Promise<T>;\n  resolve: (value: T) => void;\n  reject: (error: Error) => void;\n  retryCount: number;\n}\n\n/**\n * Rate limiter for LLM API calls using token bucket algorithm.\n *\n * @example\n * ```typescript\n * const limiter = new RateLimiter({ requestsPerMinute: 30 });\n * const result = await limiter.execute(() => llm.invoke(messages));\n * ```\n */\nexport class RateLimiter {\n  private requestsPerMinute: number;\n  private requestsPerSecond: number;\n  private maxConcurrent: number;\n  private baseBackoffMs: number;\n  private maxBackoffMs: number;\n\n  private minuteTokens: number;\n  private secondTokens: number;\n  private activeRequests: number = 0;\n  private queue: QueuedRequest<unknown>[] = [];\n  private lastMinuteReset: number = Date.now();\n  private lastSecondReset: number = Date.now();\n  private processing: boolean = false;\n  private currentBackoffMs: number = 0;\n\n  constructor(options: RateLimiterOptions = {}) {\n    this.requestsPerMinute = options.requestsPerMinute ?? 60;\n    this.requestsPerSecond = options.requestsPerSecond ?? 3;\n    this.maxConcurrent = options.maxConcurrent ?? 5;\n    this.baseBackoffMs = options.baseBackoffMs ?? 1000;\n    this.maxBackoffMs = options.maxBackoffMs ?? 60000;\n\n    this.minuteTokens = this.requestsPerMinute;\n    this.secondTokens = this.requestsPerSecond;\n  }\n\n  async execute<T>(fn: () => Promise<T>, _maxRetries: number = 3): Promise<T> {\n    return new Promise<T>((resolve, reject) => {\n      this.queue.push({\n        execute: fn as () => Promise<unknown>,\n        resolve: resolve as (value: unknown) => void,\n        reject,\n        retryCount: 0,\n      });\n      this.processQueue();\n    });\n  }\n\n  getStatus(): {\n    queueLength: number;\n    activeRequests: number;\n    minuteTokens: number;\n    secondTokens: number;\n    backoffMs: number;\n  } {\n    return {\n      queueLength: this.queue.length,\n      activeRequests: this.activeRequests,\n      minuteTokens: this.minuteTokens,\n      secondTokens: this.secondTokens,\n      backoffMs: this.currentBackoffMs,\n    };\n  }\n\n  reset(): void {\n    this.minuteTokens = this.requestsPerMinute;\n    this.secondTokens = this.requestsPerSecond;\n    this.activeRequests = 0;\n    this.queue = [];\n    this.currentBackoffMs = 0;\n    this.lastMinuteReset = Date.now();\n    this.lastSecondReset = Date.now();\n  }\n\n  private async processQueue(): Promise<void> {\n    if (this.processing) return;\n    this.processing = true;\n\n    while (this.queue.length > 0) {\n      this.refillTokens();\n\n      if (!this.canMakeRequest()) {\n        const waitTime = this.getWaitTime();\n        await this.sleep(waitTime);\n        continue;\n      }\n\n      if (this.currentBackoffMs > 0) {\n        await this.sleep(this.currentBackoffMs);\n        this.currentBackoffMs = 0;\n      }\n\n      const request = this.queue.shift();\n      if (!request) continue;\n\n      this.consumeTokens();\n      this.activeRequests++;\n\n      try {\n        const result = await request.execute();\n        request.resolve(result);\n        this.currentBackoffMs = 0;\n      } catch (error) {\n        const err = error instanceof Error ? error : new Error(String(error));\n\n        if (this.isRateLimitError(err)) {\n          this.currentBackoffMs = Math.min(\n            this.baseBackoffMs * Math.pow(2, request.retryCount),\n            this.maxBackoffMs,\n          );\n\n          console.warn(\n            `[RateLimiter] Rate limited. Backing off for ${this.currentBackoffMs}ms ` +\n              `(retry ${request.retryCount + 1})`,\n          );\n\n          if (request.retryCount < 3) {\n            this.queue.unshift({\n              ...request,\n              retryCount: request.retryCount + 1,\n            });\n          } else {\n            request.reject(\n              new Error(\n                `Rate limit exceeded after ${request.retryCount + 1} retries: ${err.message}`,\n              ),\n            );\n          }\n        } else {\n          request.reject(err);\n        }\n      } finally {\n        this.activeRequests--;\n      }\n    }\n\n    this.processing = false;\n  }\n\n  private refillTokens(): void {\n    const now = Date.now();\n    if (now - this.lastMinuteReset >= 60000) {\n      this.minuteTokens = this.requestsPerMinute;\n      this.lastMinuteReset = now;\n    }\n    if (now - this.lastSecondReset >= 1000) {\n      this.secondTokens = this.requestsPerSecond;\n      this.lastSecondReset = now;\n    }\n  }\n\n  private canMakeRequest(): boolean {\n    return (\n      this.minuteTokens > 0 &&\n      this.secondTokens > 0 &&\n      this.activeRequests < this.maxConcurrent\n    );\n  }\n\n  private consumeTokens(): void {\n    this.minuteTokens--;\n    this.secondTokens--;\n  }\n\n  private getWaitTime(): number {\n    const now = Date.now();\n    if (this.secondTokens <= 0) {\n      return Math.max(0, 1000 - (now - this.lastSecondReset));\n    }\n    if (this.minuteTokens <= 0) {\n      return Math.max(0, 60000 - (now - this.lastMinuteReset));\n    }\n    return 100;\n  }\n\n  private isRateLimitError(error: Error): boolean {\n    const message = error.message.toLowerCase();\n    return (\n      message.includes('429') ||\n      message.includes('rate limit') ||\n      message.includes('too many requests') ||\n      message.includes('quota exceeded')\n    );\n  }\n\n  private sleep(ms: number): Promise<void> {\n    return new Promise((resolve) => setTimeout(resolve, ms));\n  }\n}\n\n// Singleton instance\nlet globalRateLimiter: RateLimiter | null = null;\n\nexport function getGlobalRateLimiter(\n  options?: RateLimiterOptions,\n): RateLimiter {\n  if (!globalRateLimiter) {\n    globalRateLimiter = new RateLimiter(options);\n  }\n  return globalRateLimiter;\n}\n\nexport function resetGlobalRateLimiter(): void {\n  globalRateLimiter?.reset();\n  globalRateLimiter = null;\n}\n","/**\n * Token Tracker for LLM Usage\n *\n * Tracks token usage across multiple LLM calls for:\n * - Cost estimation (pricing fetched from OpenRouter models API)\n * - Usage monitoring\n * - Quota management\n * - Per-call JSONL logging\n *\n * @packageDocumentation\n */\n\nimport { appendFileSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';\nimport { dirname, join } from 'node:path';\n\nexport interface TokenUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  callCount: number;\n  /** Subset of `promptTokens` that were cache reads (billed at the cache-read rate). */\n  cachedPromptTokens: number;\n  /** Subset of `promptTokens` written to cache (Anthropic; billed at the cache-write rate). */\n  cacheWriteTokens: number;\n}\n\nexport interface TokenCost {\n  promptCostPer1K: number;\n  completionCostPer1K: number;\n  /** Per-1K rate for cache-read (cache-hit) prompt tokens. Falls back to prompt rate when absent. */\n  cacheReadCostPer1K?: number;\n  /** Per-1K rate for cache-write (cache-creation) prompt tokens. Falls back to prompt rate when absent. */\n  cacheWriteCostPer1K?: number;\n}\n\nexport interface CallLogEntry {\n  timestamp: string;\n  provider: string;\n  model: string;\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  /** Cache-read (cache-hit) subset of promptTokens, billed at the discounted rate. */\n  cachedPromptTokens?: number;\n  /** Cache-write subset of promptTokens (Anthropic). */\n  cacheWriteTokens?: number;\n  estimatedCost: number;\n  durationMs?: number;\n  source: 'local-log';\n}\n\n// ---------------------------------------------------------------------------\n// Pricing: fetched from OpenRouter /api/v1/models, cached to disk for 24h\n// ---------------------------------------------------------------------------\n\nconst ALMADAR_ROOT = process.env['ALMADAR_ROOT'] ?? process.cwd();\nconst PRICING_CACHE_PATH = join(ALMADAR_ROOT, '.llm-pricing-cache.json');\nconst CALL_LOG_PATH = join(ALMADAR_ROOT, '.llm-call-log.jsonl');\nconst CACHE_TTL_MS = 24 * 60 * 60 * 1000; // 24 hours\n// Bump when the cached TokenCost shape changes so stale on-disk caches are\n// invalidated on upgrade. v2 added cacheReadCostPer1K / cacheWriteCostPer1K.\nconst PRICING_CACHE_VERSION = 2;\n\n/** Map from our local model name to OpenRouter model ID */\nconst MODEL_ID_MAP: Record<string, string> = {\n  // Anthropic\n  'claude-opus-4-5-20250929': 'anthropic/claude-opus-4.5',\n  'claude-sonnet-4-5-20250929': 'anthropic/claude-sonnet-4.5',\n  'claude-sonnet-4-20250514': 'anthropic/claude-sonnet-4',\n  'claude-3-5-haiku-20241022': 'anthropic/claude-3.5-haiku',\n  // DeepSeek — map to current versions on OpenRouter\n  'deepseek-chat': 'deepseek/deepseek-v4-flash',\n  'deepseek-coder': 'deepseek/deepseek-v4-flash',\n  'deepseek-reasoner': 'deepseek/deepseek-v4-flash',\n  'deepseek-v4-pro': 'deepseek/deepseek-v4-pro',\n  'deepseek-v4-flash': 'deepseek/deepseek-v4-flash',\n  // Kimi\n  'kimi-k2.5': 'moonshotai/kimi-k2.5',\n};\n\n// Fallback: zero cost — forces OpenRouter fetch for real pricing\nconst FALLBACK_COSTS: Record<string, TokenCost> = {};\n\ninterface PricingCache {\n  version?: number;\n  fetchedAt: number;\n  models: Record<string, TokenCost>;\n}\n\nlet pricingCache: PricingCache | null = null;\n\nfunction loadCachedPricing(): PricingCache | null {\n  try {\n    const raw = readFileSync(PRICING_CACHE_PATH, 'utf-8');\n    const parsed = JSON.parse(raw) as PricingCache;\n    if (parsed.version === PRICING_CACHE_VERSION && Date.now() - parsed.fetchedAt < CACHE_TTL_MS) {\n      return parsed;\n    }\n  } catch {\n    // No cache or expired\n  }\n  return null;\n}\n\nasync function fetchPricingFromOpenRouter(): Promise<Record<string, TokenCost>> {\n  const res = await fetch('https://openrouter.ai/api/v1/models');\n  if (!res.ok) throw new Error(`OpenRouter models API: HTTP ${res.status}`);\n  const json = await res.json() as {\n    data?: Array<{\n      id: string;\n      pricing?: {\n        prompt?: string;\n        completion?: string;\n        input_cache_read?: string;\n        input_cache_write?: string;\n      };\n    }>;\n  };\n  const models: Record<string, TokenCost> = {};\n  for (const m of json.data ?? []) {\n    const promptPerToken = parseFloat(m.pricing?.prompt ?? '0');\n    const completionPerToken = parseFloat(m.pricing?.completion ?? '0');\n    const cacheReadPerToken = parseFloat(m.pricing?.input_cache_read ?? '0');\n    const cacheWritePerToken = parseFloat(m.pricing?.input_cache_write ?? '0');\n    if (promptPerToken > 0 || completionPerToken > 0) {\n      models[m.id] = {\n        promptCostPer1K: promptPerToken * 1000,\n        completionCostPer1K: completionPerToken * 1000,\n        // 0 (field absent) → leave undefined so cost math falls back to the prompt rate.\n        ...(cacheReadPerToken > 0 ? { cacheReadCostPer1K: cacheReadPerToken * 1000 } : {}),\n        ...(cacheWritePerToken > 0 ? { cacheWriteCostPer1K: cacheWritePerToken * 1000 } : {}),\n      };\n    }\n  }\n  return models;\n}\n\n/**\n * Get pricing for all models. Uses 24h disk cache, fetches from OpenRouter on miss.\n * Non-blocking: returns cached/fallback immediately, refreshes in background if stale.\n */\nfunction getPricing(): Record<string, TokenCost> {\n  if (pricingCache) return pricingCache.models;\n\n  const diskCache = loadCachedPricing();\n  if (diskCache) {\n    pricingCache = diskCache;\n    return diskCache.models;\n  }\n\n  // Trigger background fetch, return fallback for now\n  refreshPricingCache();\n  return FALLBACK_COSTS;\n}\n\nfunction refreshPricingCache(): void {\n  fetchPricingFromOpenRouter()\n    .then((models) => {\n      pricingCache = { version: PRICING_CACHE_VERSION, fetchedAt: Date.now(), models };\n      try {\n        mkdirSync(dirname(PRICING_CACHE_PATH), { recursive: true });\n        writeFileSync(PRICING_CACHE_PATH, JSON.stringify(pricingCache));\n      } catch {\n        // Non-critical\n      }\n    })\n    .catch(() => {\n      // Silently fail, use fallback\n    });\n}\n\nfunction getCostForModel(model: string): TokenCost {\n  const pricing = getPricing();\n  // Try direct match on OpenRouter ID\n  const orId = MODEL_ID_MAP[model];\n  if (orId && pricing[orId]) return pricing[orId];\n  // Try direct key match (e.g., user passed \"openai/gpt-4o\")\n  if (pricing[model]) return pricing[model];\n  // Fuzzy: find first key containing the model name\n  for (const [key, cost] of Object.entries(pricing)) {\n    if (key.includes(model) || model.includes(key.split('/')[1] ?? '')) return cost;\n  }\n  // No pricing available — return zero (OpenRouter fetch pending or model not listed)\n  return { promptCostPer1K: 0, completionCostPer1K: 0 };\n}\n\n// ---------------------------------------------------------------------------\n// TokenTracker\n// ---------------------------------------------------------------------------\n\nexport class TokenTracker {\n  private model: string;\n  private usage: TokenUsage = {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n    callCount: 0,\n    cachedPromptTokens: 0,\n    cacheWriteTokens: 0,\n  };\n\n  /** Sum of provider-reported authoritative costs (e.g. OpenRouter `usage.cost`). */\n  private authoritativeCostUSD = 0;\n  /** Token buckets for calls WITHOUT an authoritative cost — priced cache-aware.\n   *  Keyed by the model that MADE the calls: `setModel` flips the tracker\n   *  between models mid-process (coordinator on pro, subagents on flash), and\n   *  pricing the aggregate at whichever model is current when someone reads\n   *  the estimate mispriced the whole history by up to the models' rate ratio\n   *  (snapshot/diff consumers saw phantom multi-dollar deltas — or $0 after a\n   *  negative-clamp — per battery spec). Per-model buckets keep the\n   *  retroactive late-pricing property without cross-model contamination. */\n  private computedByModel = new Map<\n    string,\n    { promptTokens: number; completionTokens: number; cachedPromptTokens: number; cacheWriteTokens: number }\n  >();\n\n  constructor(model: string = 'claude-sonnet-4-5-20250929') {\n    this.model = model;\n  }\n\n  /** Cache-aware cost for one (or an aggregate of) call(s), in USD. */\n  private costFor(model: string, promptTokens: number, completionTokens: number, cached: number, written: number): number {\n    const costs = getCostForModel(model);\n    const cacheReadRate = costs.cacheReadCostPer1K ?? costs.promptCostPer1K;\n    const cacheWriteRate = costs.cacheWriteCostPer1K ?? costs.promptCostPer1K;\n    const uncached = Math.max(0, promptTokens - cached - written);\n    return (\n      (uncached / 1000) * costs.promptCostPer1K +\n      (cached / 1000) * cacheReadRate +\n      (written / 1000) * cacheWriteRate +\n      (completionTokens / 1000) * costs.completionCostPer1K\n    );\n  }\n\n  /**\n   * Record one LLM call's usage. `promptTokens` is the TOTAL input count\n   * (cache reads + cache writes + uncached); `cachedPromptTokens` and\n   * `cacheWriteTokens` are subsets of it, priced at their own (cheaper /\n   * pricier) rates. Providers that don't report cache detail pass 0, which\n   * reduces to the previous flat-rate behaviour.\n   */\n  addUsage(\n    promptTokens: number,\n    completionTokens: number,\n    options?: {\n      provider?: string;\n      durationMs?: number;\n      cachedPromptTokens?: number;\n      cacheWriteTokens?: number;\n      /** Provider-reported authoritative cost (e.g. OpenRouter `usage.cost`). When set, used verbatim. */\n      costUSD?: number;\n    },\n  ): void {\n    const cached = Math.min(promptTokens, Math.max(0, options?.cachedPromptTokens ?? 0));\n    const written = Math.min(promptTokens - cached, Math.max(0, options?.cacheWriteTokens ?? 0));\n\n    this.usage.promptTokens += promptTokens;\n    this.usage.completionTokens += completionTokens;\n    this.usage.totalTokens += promptTokens + completionTokens;\n    this.usage.cachedPromptTokens += cached;\n    this.usage.cacheWriteTokens += written;\n    this.usage.callCount++;\n\n    // Prefer the provider's authoritative cost (already cache- and routing-\n    // adjusted). Otherwise bucket the tokens and price them cache-aware so\n    // late-arriving pricing still applies retroactively to the estimate.\n    const authoritative = options?.costUSD;\n    let estimatedCost: number;\n    if (authoritative != null && Number.isFinite(authoritative)) {\n      this.authoritativeCostUSD += authoritative;\n      estimatedCost = authoritative;\n    } else {\n      const bucket = this.computedByModel.get(this.model) ?? {\n        promptTokens: 0,\n        completionTokens: 0,\n        cachedPromptTokens: 0,\n        cacheWriteTokens: 0,\n      };\n      bucket.promptTokens += promptTokens;\n      bucket.completionTokens += completionTokens;\n      bucket.cachedPromptTokens += cached;\n      bucket.cacheWriteTokens += written;\n      this.computedByModel.set(this.model, bucket);\n      estimatedCost = this.costFor(this.model, promptTokens, completionTokens, cached, written);\n    }\n\n    const entry: CallLogEntry = {\n      timestamp: new Date().toISOString(),\n      provider: options?.provider ?? 'unknown',\n      model: this.model,\n      promptTokens,\n      completionTokens,\n      totalTokens: promptTokens + completionTokens,\n      cachedPromptTokens: cached,\n      cacheWriteTokens: written,\n      estimatedCost,\n      durationMs: options?.durationMs,\n      source: 'local-log',\n    };\n\n    try {\n      mkdirSync(dirname(CALL_LOG_PATH), { recursive: true });\n      appendFileSync(CALL_LOG_PATH, JSON.stringify(entry) + '\\n');\n    } catch {\n      // Non-critical: don't break LLM calls if logging fails\n    }\n  }\n\n  getSummary(): TokenUsage {\n    return { ...this.usage };\n  }\n\n  getEstimatedCost(): number {\n    let computed = 0;\n    for (const [model, b] of this.computedByModel) {\n      computed += this.costFor(\n        model,\n        b.promptTokens,\n        b.completionTokens,\n        b.cachedPromptTokens,\n        b.cacheWriteTokens,\n      );\n    }\n    return this.authoritativeCostUSD + computed;\n  }\n\n  getFormattedCost(): string {\n    const cost = this.getEstimatedCost();\n    return `$${cost.toFixed(4)}`;\n  }\n\n  getReport(): string {\n    const summary = this.getSummary();\n    const cost = this.getEstimatedCost();\n    return [\n      `Token Usage Report (${this.model})`,\n      `─────────────────────────────`,\n      `Calls:              ${summary.callCount}`,\n      `Prompt Tokens:      ${summary.promptTokens.toLocaleString()}`,\n      `Completion Tokens:  ${summary.completionTokens.toLocaleString()}`,\n      `Total Tokens:       ${summary.totalTokens.toLocaleString()}`,\n      `Estimated Cost:     $${cost.toFixed(4)}`,\n    ].join('\\n');\n  }\n\n  reset(): void {\n    this.usage = {\n      promptTokens: 0,\n      completionTokens: 0,\n      totalTokens: 0,\n      callCount: 0,\n      cachedPromptTokens: 0,\n      cacheWriteTokens: 0,\n    };\n    this.authoritativeCostUSD = 0;\n    this.computedByModel.clear();\n  }\n\n  setModel(model: string): void {\n    this.model = model;\n  }\n}\n\n// Global tracker instance\nlet globalTracker: TokenTracker | null = null;\n\nexport function getGlobalTokenTracker(model?: string): TokenTracker {\n  if (!globalTracker) {\n    globalTracker = new TokenTracker(model);\n  } else if (model) {\n    globalTracker.setModel(model);\n  }\n  return globalTracker;\n}\n\nexport function resetGlobalTokenTracker(): void {\n  globalTracker?.reset();\n}\n\nexport function getCallLogPath(): string {\n  return CALL_LOG_PATH;\n}\n\n/** Force-refresh the pricing cache from OpenRouter. */\nexport async function refreshPricing(): Promise<void> {\n  const models = await fetchPricingFromOpenRouter();\n  pricingCache = { version: PRICING_CACHE_VERSION, fetchedAt: Date.now(), models };\n  mkdirSync(dirname(PRICING_CACHE_PATH), { recursive: true });\n  writeFileSync(PRICING_CACHE_PATH, JSON.stringify(pricingCache));\n}\n"],"mappings":";AAwCO,IAAM,cAAN,MAAkB;AAAA,EAgBvB,YAAY,UAA8B,CAAC,GAAG;AAP9C,SAAQ,iBAAyB;AACjC,SAAQ,QAAkC,CAAC;AAC3C,SAAQ,kBAA0B,KAAK,IAAI;AAC3C,SAAQ,kBAA0B,KAAK,IAAI;AAC3C,SAAQ,aAAsB;AAC9B,SAAQ,mBAA2B;AAGjC,SAAK,oBAAoB,QAAQ,qBAAqB;AACtD,SAAK,oBAAoB,QAAQ,qBAAqB;AACtD,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,eAAe,QAAQ,gBAAgB;AAE5C,SAAK,eAAe,KAAK;AACzB,SAAK,eAAe,KAAK;AAAA,EAC3B;AAAA,EAEA,MAAM,QAAW,IAAsB,cAAsB,GAAe;AAC1E,WAAO,IAAI,QAAW,CAAC,SAAS,WAAW;AACzC,WAAK,MAAM,KAAK;AAAA,QACd,SAAS;AAAA,QACT;AAAA,QACA;AAAA,QACA,YAAY;AAAA,MACd,CAAC;AACD,WAAK,aAAa;AAAA,IACpB,CAAC;AAAA,EACH;AAAA,EAEA,YAME;AACA,WAAO;AAAA,MACL,aAAa,KAAK,MAAM;AAAA,MACxB,gBAAgB,KAAK;AAAA,MACrB,cAAc,KAAK;AAAA,MACnB,cAAc,KAAK;AAAA,MACnB,WAAW,KAAK;AAAA,IAClB;AAAA,EACF;AAAA,EAEA,QAAc;AACZ,SAAK,eAAe,KAAK;AACzB,SAAK,eAAe,KAAK;AACzB,SAAK,iBAAiB;AACtB,SAAK,QAAQ,CAAC;AACd,SAAK,mBAAmB;AACxB,SAAK,kBAAkB,KAAK,IAAI;AAChC,SAAK,kBAAkB,KAAK,IAAI;AAAA,EAClC;AAAA,EAEA,MAAc,eAA8B;AAC1C,QAAI,KAAK,WAAY;AACrB,SAAK,aAAa;AAElB,WAAO,KAAK,MAAM,SAAS,GAAG;AAC5B,WAAK,aAAa;AAElB,UAAI,CAAC,KAAK,eAAe,GAAG;AAC1B,cAAM,WAAW,KAAK,YAAY;AAClC,cAAM,KAAK,MAAM,QAAQ;AACzB;AAAA,MACF;AAEA,UAAI,KAAK,mBAAmB,GAAG;AAC7B,cAAM,KAAK,MAAM,KAAK,gBAAgB;AACtC,aAAK,mBAAmB;AAAA,MAC1B;AAEA,YAAM,UAAU,KAAK,MAAM,MAAM;AACjC,UAAI,CAAC,QAAS;AAEd,WAAK,cAAc;AACnB,WAAK;AAEL,UAAI;AACF,cAAM,SAAS,MAAM,QAAQ,QAAQ;AACrC,gBAAQ,QAAQ,MAAM;AACtB,aAAK,mBAAmB;AAAA,MAC1B,SAAS,OAAO;AACd,cAAM,MAAM,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;AAEpE,YAAI,KAAK,iBAAiB,GAAG,GAAG;AAC9B,eAAK,mBAAmB,KAAK;AAAA,YAC3B,KAAK,gBAAgB,KAAK,IAAI,GAAG,QAAQ,UAAU;AAAA,YACnD,KAAK;AAAA,UACP;AAEA,kBAAQ;AAAA,YACN,+CAA+C,KAAK,gBAAgB,aACxD,QAAQ,aAAa,CAAC;AAAA,UACpC;AAEA,cAAI,QAAQ,aAAa,GAAG;AAC1B,iBAAK,MAAM,QAAQ;AAAA,cACjB,GAAG;AAAA,cACH,YAAY,QAAQ,aAAa;AAAA,YACnC,CAAC;AAAA,UACH,OAAO;AACL,oBAAQ;AAAA,cACN,IAAI;AAAA,gBACF,6BAA6B,QAAQ,aAAa,CAAC,aAAa,IAAI,OAAO;AAAA,cAC7E;AAAA,YACF;AAAA,UACF;AAAA,QACF,OAAO;AACL,kBAAQ,OAAO,GAAG;AAAA,QACpB;AAAA,MACF,UAAE;AACA,aAAK;AAAA,MACP;AAAA,IACF;AAEA,SAAK,aAAa;AAAA,EACpB;AAAA,EAEQ,eAAqB;AAC3B,UAAM,MAAM,KAAK,IAAI;AACrB,QAAI,MAAM,KAAK,mBAAmB,KAAO;AACvC,WAAK,eAAe,KAAK;AACzB,WAAK,kBAAkB;AAAA,IACzB;AACA,QAAI,MAAM,KAAK,mBAAmB,KAAM;AACtC,WAAK,eAAe,KAAK;AACzB,WAAK,kBAAkB;AAAA,IACzB;AAAA,EACF;AAAA,EAEQ,iBAA0B;AAChC,WACE,KAAK,eAAe,KACpB,KAAK,eAAe,KACpB,KAAK,iBAAiB,KAAK;AAAA,EAE/B;AAAA,EAEQ,gBAAsB;AAC5B,SAAK;AACL,SAAK;AAAA,EACP;AAAA,EAEQ,cAAsB;AAC5B,UAAM,MAAM,KAAK,IAAI;AACrB,QAAI,KAAK,gBAAgB,GAAG;AAC1B,aAAO,KAAK,IAAI,GAAG,OAAQ,MAAM,KAAK,gBAAgB;AAAA,IACxD;AACA,QAAI,KAAK,gBAAgB,GAAG;AAC1B,aAAO,KAAK,IAAI,GAAG,OAAS,MAAM,KAAK,gBAAgB;AAAA,IACzD;AACA,WAAO;AAAA,EACT;AAAA,EAEQ,iBAAiB,OAAuB;AAC9C,UAAM,UAAU,MAAM,QAAQ,YAAY;AAC1C,WACE,QAAQ,SAAS,KAAK,KACtB,QAAQ,SAAS,YAAY,KAC7B,QAAQ,SAAS,mBAAmB,KACpC,QAAQ,SAAS,gBAAgB;AAAA,EAErC;AAAA,EAEQ,MAAM,IAA2B;AACvC,WAAO,IAAI,QAAQ,CAAC,YAAY,WAAW,SAAS,EAAE,CAAC;AAAA,EACzD;AACF;AAGA,IAAI,oBAAwC;AAErC,SAAS,qBACd,SACa;AACb,MAAI,CAAC,mBAAmB;AACtB,wBAAoB,IAAI,YAAY,OAAO;AAAA,EAC7C;AACA,SAAO;AACT;AAEO,SAAS,yBAA+B;AAC7C,qBAAmB,MAAM;AACzB,sBAAoB;AACtB;;;AChOA,SAAS,gBAAgB,WAAW,cAAc,qBAAqB;AACvE,SAAS,SAAS,YAAY;AA0C9B,IAAM,eAAe,QAAQ,IAAI,cAAc,KAAK,QAAQ,IAAI;AAChE,IAAM,qBAAqB,KAAK,cAAc,yBAAyB;AACvE,IAAM,gBAAgB,KAAK,cAAc,qBAAqB;AAC9D,IAAM,eAAe,KAAK,KAAK,KAAK;AAGpC,IAAM,wBAAwB;AAG9B,IAAM,eAAuC;AAAA;AAAA,EAE3C,4BAA4B;AAAA,EAC5B,8BAA8B;AAAA,EAC9B,4BAA4B;AAAA,EAC5B,6BAA6B;AAAA;AAAA,EAE7B,iBAAiB;AAAA,EACjB,kBAAkB;AAAA,EAClB,qBAAqB;AAAA,EACrB,mBAAmB;AAAA,EACnB,qBAAqB;AAAA;AAAA,EAErB,aAAa;AACf;AAGA,IAAM,iBAA4C,CAAC;AAQnD,IAAI,eAAoC;AAExC,SAAS,oBAAyC;AAChD,MAAI;AACF,UAAM,MAAM,aAAa,oBAAoB,OAAO;AACpD,UAAM,SAAS,KAAK,MAAM,GAAG;AAC7B,QAAI,OAAO,YAAY,yBAAyB,KAAK,IAAI,IAAI,OAAO,YAAY,cAAc;AAC5F,aAAO;AAAA,IACT;AAAA,EACF,QAAQ;AAAA,EAER;AACA,SAAO;AACT;AAEA,eAAe,6BAAiE;AAC9E,QAAM,MAAM,MAAM,MAAM,qCAAqC;AAC7D,MAAI,CAAC,IAAI,GAAI,OAAM,IAAI,MAAM,+BAA+B,IAAI,MAAM,EAAE;AACxE,QAAM,OAAO,MAAM,IAAI,KAAK;AAW5B,QAAM,SAAoC,CAAC;AAC3C,aAAW,KAAK,KAAK,QAAQ,CAAC,GAAG;AAC/B,UAAM,iBAAiB,WAAW,EAAE,SAAS,UAAU,GAAG;AAC1D,UAAM,qBAAqB,WAAW,EAAE,SAAS,cAAc,GAAG;AAClE,UAAM,oBAAoB,WAAW,EAAE,SAAS,oBAAoB,GAAG;AACvE,UAAM,qBAAqB,WAAW,EAAE,SAAS,qBAAqB,GAAG;AACzE,QAAI,iBAAiB,KAAK,qBAAqB,GAAG;AAChD,aAAO,EAAE,EAAE,IAAI;AAAA,QACb,iBAAiB,iBAAiB;AAAA,QAClC,qBAAqB,qBAAqB;AAAA;AAAA,QAE1C,GAAI,oBAAoB,IAAI,EAAE,oBAAoB,oBAAoB,IAAK,IAAI,CAAC;AAAA,QAChF,GAAI,qBAAqB,IAAI,EAAE,qBAAqB,qBAAqB,IAAK,IAAI,CAAC;AAAA,MACrF;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAMA,SAAS,aAAwC;AAC/C,MAAI,aAAc,QAAO,aAAa;AAEtC,QAAM,YAAY,kBAAkB;AACpC,MAAI,WAAW;AACb,mBAAe;AACf,WAAO,UAAU;AAAA,EACnB;AAGA,sBAAoB;AACpB,SAAO;AACT;AAEA,SAAS,sBAA4B;AACnC,6BAA2B,EACxB,KAAK,CAAC,WAAW;AAChB,mBAAe,EAAE,SAAS,uBAAuB,WAAW,KAAK,IAAI,GAAG,OAAO;AAC/E,QAAI;AACF,gBAAU,QAAQ,kBAAkB,GAAG,EAAE,WAAW,KAAK,CAAC;AAC1D,oBAAc,oBAAoB,KAAK,UAAU,YAAY,CAAC;AAAA,IAChE,QAAQ;AAAA,IAER;AAAA,EACF,CAAC,EACA,MAAM,MAAM;AAAA,EAEb,CAAC;AACL;AAEA,SAAS,gBAAgB,OAA0B;AACjD,QAAM,UAAU,WAAW;AAE3B,QAAM,OAAO,aAAa,KAAK;AAC/B,MAAI,QAAQ,QAAQ,IAAI,EAAG,QAAO,QAAQ,IAAI;AAE9C,MAAI,QAAQ,KAAK,EAAG,QAAO,QAAQ,KAAK;AAExC,aAAW,CAAC,KAAK,IAAI,KAAK,OAAO,QAAQ,OAAO,GAAG;AACjD,QAAI,IAAI,SAAS,KAAK,KAAK,MAAM,SAAS,IAAI,MAAM,GAAG,EAAE,CAAC,KAAK,EAAE,EAAG,QAAO;AAAA,EAC7E;AAEA,SAAO,EAAE,iBAAiB,GAAG,qBAAqB,EAAE;AACtD;AAMO,IAAM,eAAN,MAAmB;AAAA,EA0BxB,YAAY,QAAgB,8BAA8B;AAxB1D,SAAQ,QAAoB;AAAA,MAC1B,cAAc;AAAA,MACd,kBAAkB;AAAA,MAClB,aAAa;AAAA,MACb,WAAW;AAAA,MACX,oBAAoB;AAAA,MACpB,kBAAkB;AAAA,IACpB;AAGA;AAAA,SAAQ,uBAAuB;AAS/B;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,SAAQ,kBAAkB,oBAAI,IAG5B;AAGA,SAAK,QAAQ;AAAA,EACf;AAAA;AAAA,EAGQ,QAAQ,OAAe,cAAsB,kBAA0B,QAAgB,SAAyB;AACtH,UAAM,QAAQ,gBAAgB,KAAK;AACnC,UAAM,gBAAgB,MAAM,sBAAsB,MAAM;AACxD,UAAM,iBAAiB,MAAM,uBAAuB,MAAM;AAC1D,UAAM,WAAW,KAAK,IAAI,GAAG,eAAe,SAAS,OAAO;AAC5D,WACG,WAAW,MAAQ,MAAM,kBACzB,SAAS,MAAQ,gBACjB,UAAU,MAAQ,iBAClB,mBAAmB,MAAQ,MAAM;AAAA,EAEtC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,SACE,cACA,kBACA,SAQM;AACN,UAAM,SAAS,KAAK,IAAI,cAAc,KAAK,IAAI,GAAG,SAAS,sBAAsB,CAAC,CAAC;AACnF,UAAM,UAAU,KAAK,IAAI,eAAe,QAAQ,KAAK,IAAI,GAAG,SAAS,oBAAoB,CAAC,CAAC;AAE3F,SAAK,MAAM,gBAAgB;AAC3B,SAAK,MAAM,oBAAoB;AAC/B,SAAK,MAAM,eAAe,eAAe;AACzC,SAAK,MAAM,sBAAsB;AACjC,SAAK,MAAM,oBAAoB;AAC/B,SAAK,MAAM;AAKX,UAAM,gBAAgB,SAAS;AAC/B,QAAI;AACJ,QAAI,iBAAiB,QAAQ,OAAO,SAAS,aAAa,GAAG;AAC3D,WAAK,wBAAwB;AAC7B,sBAAgB;AAAA,IAClB,OAAO;AACL,YAAM,SAAS,KAAK,gBAAgB,IAAI,KAAK,KAAK,KAAK;AAAA,QACrD,cAAc;AAAA,QACd,kBAAkB;AAAA,QAClB,oBAAoB;AAAA,QACpB,kBAAkB;AAAA,MACpB;AACA,aAAO,gBAAgB;AACvB,aAAO,oBAAoB;AAC3B,aAAO,sBAAsB;AAC7B,aAAO,oBAAoB;AAC3B,WAAK,gBAAgB,IAAI,KAAK,OAAO,MAAM;AAC3C,sBAAgB,KAAK,QAAQ,KAAK,OAAO,cAAc,kBAAkB,QAAQ,OAAO;AAAA,IAC1F;AAEA,UAAM,QAAsB;AAAA,MAC1B,YAAW,oBAAI,KAAK,GAAE,YAAY;AAAA,MAClC,UAAU,SAAS,YAAY;AAAA,MAC/B,OAAO,KAAK;AAAA,MACZ;AAAA,MACA;AAAA,MACA,aAAa,eAAe;AAAA,MAC5B,oBAAoB;AAAA,MACpB,kBAAkB;AAAA,MAClB;AAAA,MACA,YAAY,SAAS;AAAA,MACrB,QAAQ;AAAA,IACV;AAEA,QAAI;AACF,gBAAU,QAAQ,aAAa,GAAG,EAAE,WAAW,KAAK,CAAC;AACrD,qBAAe,eAAe,KAAK,UAAU,KAAK,IAAI,IAAI;AAAA,IAC5D,QAAQ;AAAA,IAER;AAAA,EACF;AAAA,EAEA,aAAyB;AACvB,WAAO,EAAE,GAAG,KAAK,MAAM;AAAA,EACzB;AAAA,EAEA,mBAA2B;AACzB,QAAI,WAAW;AACf,eAAW,CAAC,OAAO,CAAC,KAAK,KAAK,iBAAiB;AAC7C,kBAAY,KAAK;AAAA,QACf;AAAA,QACA,EAAE;AAAA,QACF,EAAE;AAAA,QACF,EAAE;AAAA,QACF,EAAE;AAAA,MACJ;AAAA,IACF;AACA,WAAO,KAAK,uBAAuB;AAAA,EACrC;AAAA,EAEA,mBAA2B;AACzB,UAAM,OAAO,KAAK,iBAAiB;AACnC,WAAO,IAAI,KAAK,QAAQ,CAAC,CAAC;AAAA,EAC5B;AAAA,EAEA,YAAoB;AAClB,UAAM,UAAU,KAAK,WAAW;AAChC,UAAM,OAAO,KAAK,iBAAiB;AACnC,WAAO;AAAA,MACL,uBAAuB,KAAK,KAAK;AAAA,MACjC;AAAA,MACA,uBAAuB,QAAQ,SAAS;AAAA,MACxC,uBAAuB,QAAQ,aAAa,eAAe,CAAC;AAAA,MAC5D,uBAAuB,QAAQ,iBAAiB,eAAe,CAAC;AAAA,MAChE,uBAAuB,QAAQ,YAAY,eAAe,CAAC;AAAA,MAC3D,wBAAwB,KAAK,QAAQ,CAAC,CAAC;AAAA,IACzC,EAAE,KAAK,IAAI;AAAA,EACb;AAAA,EAEA,QAAc;AACZ,SAAK,QAAQ;AAAA,MACX,cAAc;AAAA,MACd,kBAAkB;AAAA,MAClB,aAAa;AAAA,MACb,WAAW;AAAA,MACX,oBAAoB;AAAA,MACpB,kBAAkB;AAAA,IACpB;AACA,SAAK,uBAAuB;AAC5B,SAAK,gBAAgB,MAAM;AAAA,EAC7B;AAAA,EAEA,SAAS,OAAqB;AAC5B,SAAK,QAAQ;AAAA,EACf;AACF;AAGA,IAAI,gBAAqC;AAElC,SAAS,sBAAsB,OAA8B;AAClE,MAAI,CAAC,eAAe;AAClB,oBAAgB,IAAI,aAAa,KAAK;AAAA,EACxC,WAAW,OAAO;AAChB,kBAAc,SAAS,KAAK;AAAA,EAC9B;AACA,SAAO;AACT;AAEO,SAAS,0BAAgC;AAC9C,iBAAe,MAAM;AACvB;","names":[]}

package/dist/{chunk-QOJXMUWL.js → chunk-EHZE3HBY.js}

@@ -5,7 +5,7 @@ import {
   RateLimiter,
   getGlobalRateLimiter,
   getGlobalTokenTracker
-} from "./chunk-SXSP6M24.js";
+} from "./chunk-CZX533SU.js";
 
 // src/client.ts
 import { ChatOpenAI } from "@langchain/openai";
@@ -1051,4 +1051,4 @@ export {
   createOpenRouterClient,
   createZhipuClient
 };
-//# sourceMappingURL=chunk-QOJXMUWL.js.map
+//# sourceMappingURL=chunk-EHZE3HBY.js.map

package/dist/{chunk-IDXSWM57.js → chunk-S6KQIMQQ.js}

@@ -2,7 +2,7 @@ import {
   RateLimiter,
   getGlobalRateLimiter,
   getGlobalTokenTracker
-} from "./chunk-SXSP6M24.js";
+} from "./chunk-CZX533SU.js";
 
 // src/structured-output.ts
 import OpenAI from "openai";
@@ -175,4 +175,4 @@ export {
   resetStructuredOutputClient,
   isStructuredOutputAvailable
 };
-//# sourceMappingURL=chunk-IDXSWM57.js.map
+//# sourceMappingURL=chunk-S6KQIMQQ.js.map

package/dist/{client-Bfd-fT35.d.ts → client-DWcDgr3a.d.ts}

@@ -1,4 +1,4 @@
-import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-7-oOHrtX.js';
+import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-C12cctao.js';
 import { ChatOpenAI } from '@langchain/openai';
 import { ChatAnthropic } from '@langchain/anthropic';
 import { z } from 'zod';

package/dist/client.d.ts

@@ -1,6 +1,6 @@
-import './rate-limiter-7-oOHrtX.js';
+import './rate-limiter-C12cctao.js';
 import '@langchain/openai';
 import '@langchain/anthropic';
 import 'zod';
-export { A as ANTHROPIC_MODELS, C as CacheAwareLLMCallOptions, b as CacheableBlock, D as DEEPSEEK_MODELS, K as KIMI_MODELS, j as LLMCallOptions, a as LLMClient, k as LLMClientOptions, L as LLMFinishReason, l as LLMProvider, m as LLMResponse, n as LLMStreamChunk, o as LLMStreamOptions, p as LLMUsage, O as OPENAI_MODELS, q as OPENROUTER_MODELS, P as ProviderConfig, r as createAnthropicClient, s as createCreativeClient, t as createDeepSeekClient, u as createFixClient, v as createKimiClient, w as createOpenAIClient, x as createOpenRouterClient, y as createRequirementsClient, z as createZhipuClient, B as getAvailableProvider, E as getSharedLLMClient, F as isProviderAvailable, H as resetSharedLLMClient } from './client-Bfd-fT35.js';
+export { A as ANTHROPIC_MODELS, C as CacheAwareLLMCallOptions, b as CacheableBlock, D as DEEPSEEK_MODELS, K as KIMI_MODELS, j as LLMCallOptions, a as LLMClient, k as LLMClientOptions, L as LLMFinishReason, l as LLMProvider, m as LLMResponse, n as LLMStreamChunk, o as LLMStreamOptions, p as LLMUsage, O as OPENAI_MODELS, q as OPENROUTER_MODELS, P as ProviderConfig, r as createAnthropicClient, s as createCreativeClient, t as createDeepSeekClient, u as createFixClient, v as createKimiClient, w as createOpenAIClient, x as createOpenRouterClient, y as createRequirementsClient, z as createZhipuClient, B as getAvailableProvider, E as getSharedLLMClient, F as isProviderAvailable, H as resetSharedLLMClient } from './client-DWcDgr3a.js';
 import '@almadar/core';

package/dist/client.js

@@ -18,9 +18,9 @@ import {
   getSharedLLMClient,
   isProviderAvailable,
   resetSharedLLMClient
-} from "./chunk-QOJXMUWL.js";
+} from "./chunk-EHZE3HBY.js";
 import "./chunk-P4VCT25B.js";
-import "./chunk-SXSP6M24.js";
+import "./chunk-CZX533SU.js";
 export {
   ANTHROPIC_MODELS,
   DEEPSEEK_MODELS,

package/dist/index.d.ts

@@ -1,6 +1,6 @@
-import { L as LLMFinishReason, a as LLMClient } from './client-Bfd-fT35.js';
-export { A as ANTHROPIC_MODELS, C as CacheAwareLLMCallOptions, b as CacheableBlock, c as ChatCompletionChoice, d as ChatCompletionMessage, e as ChatCompletionResponse, f as ChatCompletionRole, g as ChatCompletionToolCall, h as ChatCompletionToolDef, i as ChatCompletionUsage, D as DEEPSEEK_MODELS, K as KIMI_MODELS, j as LLMCallOptions, k as LLMClientOptions, l as LLMProvider, m as LLMResponse, n as LLMStreamChunk, o as LLMStreamOptions, p as LLMUsage, O as OPENAI_MODELS, q as OPENROUTER_MODELS, P as ProviderConfig, r as createAnthropicClient, s as createCreativeClient, t as createDeepSeekClient, u as createFixClient, v as createKimiClient, w as createOpenAIClient, x as createOpenRouterClient, y as createRequirementsClient, z as createZhipuClient, B as getAvailableProvider, E as getSharedLLMClient, F as isProviderAvailable, G as parseChatCompletionResponse, H as resetSharedLLMClient } from './client-Bfd-fT35.js';
-export { a as RateLimiter, R as RateLimiterOptions, b as TokenTracker, T as TokenUsage, g as getGlobalRateLimiter, c as getGlobalTokenTracker, r as resetGlobalRateLimiter, d as resetGlobalTokenTracker } from './rate-limiter-7-oOHrtX.js';
+import { L as LLMFinishReason, a as LLMClient } from './client-DWcDgr3a.js';
+export { A as ANTHROPIC_MODELS, C as CacheAwareLLMCallOptions, b as CacheableBlock, c as ChatCompletionChoice, d as ChatCompletionMessage, e as ChatCompletionResponse, f as ChatCompletionRole, g as ChatCompletionToolCall, h as ChatCompletionToolDef, i as ChatCompletionUsage, D as DEEPSEEK_MODELS, K as KIMI_MODELS, j as LLMCallOptions, k as LLMClientOptions, l as LLMProvider, m as LLMResponse, n as LLMStreamChunk, o as LLMStreamOptions, p as LLMUsage, O as OPENAI_MODELS, q as OPENROUTER_MODELS, P as ProviderConfig, r as createAnthropicClient, s as createCreativeClient, t as createDeepSeekClient, u as createFixClient, v as createKimiClient, w as createOpenAIClient, x as createOpenRouterClient, y as createRequirementsClient, z as createZhipuClient, B as getAvailableProvider, E as getSharedLLMClient, F as isProviderAvailable, G as parseChatCompletionResponse, H as resetSharedLLMClient } from './client-DWcDgr3a.js';
+export { a as RateLimiter, R as RateLimiterOptions, b as TokenTracker, T as TokenUsage, g as getGlobalRateLimiter, c as getGlobalTokenTracker, r as resetGlobalRateLimiter, d as resetGlobalTokenTracker } from './rate-limiter-C12cctao.js';
 export { autoCloseJson, extractJsonFromText, isValidJson, parseJsonResponse, safeParseJson } from './json-parser.js';
 import { z } from 'zod';
 export { JsonSchema, STRUCTURED_OUTPUT_MODELS, StructuredGenerationOptions, StructuredGenerationResult, StructuredOutputClient, StructuredOutputOptions, getStructuredOutputClient, isStructuredOutputAvailable, resetStructuredOutputClient } from './structured-output.js';
@@ -65,6 +65,7 @@ declare class EmbeddingClient {
      * (≈190 entries) rather than per-entry requests.
      */
     embedBatch(texts: ReadonlyArray<string>): Promise<EmbeddingBatchResult>;
+    private embedBatchOnce;
     private resolveApiKey;
 }
 /**

package/dist/index.js

@@ -19,7 +19,7 @@ import {
   isProviderAvailable,
   parseChatCompletionResponse,
   resetSharedLLMClient
-} from "./chunk-QOJXMUWL.js";
+} from "./chunk-EHZE3HBY.js";
 import {
   autoCloseJson,
   extractJsonFromText,
@@ -33,7 +33,7 @@ import {
   getStructuredOutputClient,
   isStructuredOutputAvailable,
   resetStructuredOutputClient
-} from "./chunk-IDXSWM57.js";
+} from "./chunk-S6KQIMQQ.js";
 import {
   RateLimiter,
   TokenTracker,
@@ -41,7 +41,7 @@ import {
   getGlobalTokenTracker,
   resetGlobalRateLimiter,
   resetGlobalTokenTracker
-} from "./chunk-SXSP6M24.js";
+} from "./chunk-CZX533SU.js";
 import {
   MasarError,
   MasarProvider,
@@ -50,6 +50,9 @@ import {
 } from "./chunk-MUTXGY6D.js";
 
 // src/embedding-client.ts
+var EMBED_MAX_ATTEMPTS = 3;
+var EMBED_RETRY_BACKOFF_MS = [500, 2e3];
+var TOKEN_OVERFLOW_RE = /context length|maximum input length|reduce the length of the input|input_tokens/i;
 var DEFAULT_MODELS = {
   openai: "text-embedding-3-small",
   openrouter: "baai/bge-base-en-v1.5"
@@ -93,7 +96,26 @@ var EmbeddingClient = class {
     if (texts.length === 0) {
       return { embeddings: [], usage: { promptTokens: 0, totalTokens: 0 } };
     }
-    const trimmed = texts.map((t) => t.length > 2e3 ? t.slice(0, 2e3) : t);
+    let charCap = 2e3;
+    let lastError = new Error("EmbeddingClient: unreachable");
+    for (let attempt = 1; attempt <= EMBED_MAX_ATTEMPTS; attempt++) {
+      const trimmed = texts.map((t) => t.length > charCap ? t.slice(0, charCap) : t);
+      try {
+        return await this.embedBatchOnce(texts.length, trimmed);
+      } catch (e) {
+        lastError = e instanceof Error ? e : new Error(String(e));
+        if (TOKEN_OVERFLOW_RE.test(lastError.message) && charCap > 500) {
+          charCap = Math.floor(charCap / 2);
+          continue;
+        }
+        const backoff = EMBED_RETRY_BACKOFF_MS[attempt - 1];
+        if (backoff === void 0) break;
+        await new Promise((r) => setTimeout(r, backoff));
+      }
+    }
+    throw lastError;
+  }
+  async embedBatchOnce(inputCount, trimmed) {
     const response = await fetch(`${this.baseUrl}/embeddings`, {
       method: "POST",
       headers: {
@@ -108,14 +130,25 @@ var EmbeddingClient = class {
         `EmbeddingClient: ${this.provider} returned ${response.status} ${response.statusText}: ${errorText}`
       );
     }
-    const body = await response.json();
+    const rawBody = await response.text();
+    let body;
+    try {
+      body = JSON.parse(rawBody);
+    } catch {
+      throw new Error(
+        `EmbeddingClient: ${this.provider} returned unparseable body: ${rawBody}`
+      );
+    }
     if (!Array.isArray(body.data)) {
-      throw new Error("EmbeddingClient: provider response missing `data` array");
+      const detail = body.error?.message ?? rawBody;
+      throw new Error(
+        `EmbeddingClient: ${this.provider} response has no \`data\` array: ${detail}`
+      );
     }
     const ordered = [...body.data].sort((a, b) => a.index - b.index);
-    if (ordered.length !== texts.length) {
+    if (ordered.length !== inputCount) {
       throw new Error(
-        `EmbeddingClient: provider returned ${ordered.length} embeddings for ${texts.length} inputs`
+        `EmbeddingClient: provider returned ${ordered.length} embeddings for ${inputCount} inputs`
       );
     }
     return {

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/embedding-client.ts","../src/truncation-detector.ts","../src/continuation.ts"],"sourcesContent":["/**\n * Embedding Client\n *\n * Single-purpose client for text embeddings. Used by `@almadar-io/agent`'s\n * cosine-similarity catalog retrieval (rank organisms + atoms against the\n * user's request before rendering Stage A's prompt) and by\n * `@almadar/std`'s publish-time embedding bake step.\n *\n * Providers:\n *   - `openai` (default model `text-embedding-3-small`, 1536-d) — requires\n *     `OPENAI_API_KEY`.\n *   - `openrouter` (default model `baai/bge-base-en-v1.5`, 768-d) — requires\n *     `OPEN_ROUTER_API_KEY`. Same OpenAI-compatible request shape, just a\n *     different base URL.\n *\n * Both providers return the same response envelope (`{data:[{embedding,index}]}`),\n * so the request/response code is shared.\n *\n * @packageDocumentation\n */\n\nexport type EmbeddingProvider = 'openai' | 'openrouter';\n\nexport interface EmbeddingClientOptions {\n  provider: EmbeddingProvider;\n  /** Defaults: openai → text-embedding-3-small, openrouter → baai/bge-base-en-v1.5. */\n  model?: string;\n  /** Override API key. Defaults to provider-specific env var. */\n  apiKey?: string;\n  /** Override base URL. Defaults to provider canonical endpoint. */\n  baseUrl?: string;\n}\n\nexport interface EmbeddingUsage {\n  promptTokens: number;\n  totalTokens: number;\n}\n\nexport interface EmbeddingResult {\n  embedding: readonly number[];\n  usage: EmbeddingUsage;\n}\n\nexport interface EmbeddingBatchResult {\n  embeddings: readonly (readonly number[])[];\n  usage: EmbeddingUsage;\n}\n\ninterface OpenAIEmbeddingApiResponse {\n  data: Array<{ embedding: number[]; index: number }>;\n  usage?: { prompt_tokens?: number; total_tokens?: number };\n}\n\nconst DEFAULT_MODELS: Record<EmbeddingProvider, string> = {\n  openai: 'text-embedding-3-small',\n  openrouter: 'baai/bge-base-en-v1.5',\n};\n\nconst DEFAULT_BASE_URLS: Record<EmbeddingProvider, string> = {\n  openai: 'https://api.openai.com/v1',\n  openrouter: 'https://openrouter.ai/api/v1',\n};\n\nconst API_KEY_ENV_VARS: Record<EmbeddingProvider, string> = {\n  openai: 'OPENAI_API_KEY',\n  openrouter: 'OPEN_ROUTER_API_KEY',\n};\n\nexport class EmbeddingClient {\n  private readonly provider: EmbeddingProvider;\n  private readonly model: string;\n  private readonly apiKey: string;\n  private readonly baseUrl: string;\n\n  constructor(opts: EmbeddingClientOptions) {\n    this.provider = opts.provider;\n    this.model = opts.model ?? DEFAULT_MODELS[opts.provider];\n    this.baseUrl = (opts.baseUrl ?? DEFAULT_BASE_URLS[opts.provider]).replace(/\\/$/, '');\n\n    const apiKey = opts.apiKey ?? this.resolveApiKey();\n    if (!apiKey) {\n      throw new Error(\n        `EmbeddingClient: API key for provider '${this.provider}' not found. ` +\n          `Set ${API_KEY_ENV_VARS[this.provider]} in your environment or pass options.apiKey.`,\n      );\n    }\n    this.apiKey = apiKey;\n  }\n\n  /** Embed a single text. */\n  async embed(text: string): Promise<EmbeddingResult> {\n    const batch = await this.embedBatch([text]);\n    const embedding = batch.embeddings[0];\n    if (!embedding) {\n      throw new Error('EmbeddingClient.embed: provider returned no embedding');\n    }\n    return { embedding, usage: batch.usage };\n  }\n\n  /**\n   * Embed a batch of texts. Used by `@almadar/std`'s publish-time\n   * `build-embeddings.ts` script — one batch call for the full catalog\n   * (≈190 entries) rather than per-entry requests.\n   */\n  async embedBatch(texts: ReadonlyArray<string>): Promise<EmbeddingBatchResult> {\n    if (texts.length === 0) {\n      return { embeddings: [], usage: { promptTokens: 0, totalTokens: 0 } };\n    }\n    // bge-base-en-v1.5 has a 512-token context limit; trim each input so\n    // providers (notably OpenRouter) don't return a malformed response body.\n    // 2000 chars ≈ 500 tokens for typical English text — same cap used by\n    // the runtime's tryEmbedRequest helper.\n    const trimmed = texts.map((t) => (t.length > 2000 ? t.slice(0, 2000) : t));\n    const response = await fetch(`${this.baseUrl}/embeddings`, {\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        Authorization: `Bearer ${this.apiKey}`,\n      },\n      body: JSON.stringify({ model: this.model, input: trimmed }),\n    });\n    if (!response.ok) {\n      const errorText = await response.text().catch(() => '<no body>');\n      throw new Error(\n        `EmbeddingClient: ${this.provider} returned ${response.status} ${response.statusText}: ${errorText}`,\n      );\n    }\n    const body = (await response.json()) as OpenAIEmbeddingApiResponse;\n    if (!Array.isArray(body.data)) {\n      throw new Error('EmbeddingClient: provider response missing `data` array');\n    }\n    const ordered = [...body.data].sort((a, b) => a.index - b.index);\n    if (ordered.length !== texts.length) {\n      throw new Error(\n        `EmbeddingClient: provider returned ${ordered.length} embeddings for ${texts.length} inputs`,\n      );\n    }\n    return {\n      embeddings: ordered.map((d) => d.embedding),\n      usage: {\n        promptTokens: body.usage?.prompt_tokens ?? 0,\n        totalTokens: body.usage?.total_tokens ?? 0,\n      },\n    };\n  }\n\n  private resolveApiKey(): string | undefined {\n    return process.env[API_KEY_ENV_VARS[this.provider]];\n  }\n}\n\n/**\n * Cosine similarity between two equal-length vectors. Returns a value in\n * [-1, 1]; higher means more similar. Used by `@almadar-io/agent`'s\n * catalog retrieval to rank organism/atom descriptions against the user\n * request embedding.\n *\n * Defensive: returns 0 when either vector is the zero vector (would\n * otherwise divide by zero).\n */\nexport function cosineSimilarity(a: ReadonlyArray<number>, b: ReadonlyArray<number>): number {\n  if (a.length !== b.length) {\n    throw new Error(`cosineSimilarity: vector length mismatch (${a.length} vs ${b.length})`);\n  }\n  let dot = 0;\n  let normA = 0;\n  let normB = 0;\n  for (let i = 0; i < a.length; i++) {\n    const av = a[i];\n    const bv = b[i];\n    dot += av * bv;\n    normA += av * av;\n    normB += bv * bv;\n  }\n  if (normA === 0 || normB === 0) return 0;\n  return dot / (Math.sqrt(normA) * Math.sqrt(normB));\n}\n","/**\n * Truncation Detector\n *\n * Utilities for detecting when LLM output has been truncated and\n * extracting usable content from partial responses.\n *\n * @packageDocumentation\n */\n\nimport type { LLMFinishReason } from './client.js';\nimport { autoCloseJson } from './json-parser.js';\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport type TruncationReason =\n  | 'finish_reason'\n  | 'json_incomplete'\n  | 'bracket_mismatch'\n  | 'none';\n\nexport interface TruncationResult {\n  isTruncated: boolean;\n  reason: TruncationReason;\n  partialContent?: string;\n  lastCompleteElement?: unknown;\n  missingCloseBrackets?: number;\n  missingCloseBraces?: number;\n}\n\n// ============================================================================\n// Main Detection Function\n// ============================================================================\n\n/**\n * Detect if an LLM response has been truncated.\n *\n * Analyzes the response content and finish reason to determine if\n * the output was cut off due to token limits or other issues.\n *\n * @param {string} response - The LLM response text\n * @param {LLMFinishReason} finishReason - The finish reason from the LLM\n * @returns {TruncationResult} Detection result with truncation details\n */\nexport function detectTruncation(\n  response: string,\n  finishReason: LLMFinishReason,\n): TruncationResult {\n  if (finishReason === 'length') {\n    const bracketInfo = countBrackets(response);\n    return {\n      isTruncated: true,\n      reason: 'finish_reason',\n      partialContent: response,\n      lastCompleteElement: findLastCompleteElement(response),\n      missingCloseBrackets: bracketInfo.missingCloseBrackets,\n      missingCloseBraces: bracketInfo.missingCloseBraces,\n    };\n  }\n\n  try {\n    JSON.parse(response);\n    return { isTruncated: false, reason: 'none' };\n  } catch {\n    // JSON is invalid, check if due to truncation\n  }\n\n  if (finishReason === 'stop' || finishReason === null) {\n    const trimmed = response.trim();\n\n    const isMidContent =\n      trimmed.endsWith(',') ||\n      trimmed.endsWith(':') ||\n      trimmed.endsWith('\": ') ||\n      /:\\s*$/.test(trimmed) ||\n      /,\\s*$/.test(trimmed);\n\n    if (isMidContent) {\n      const bracketInfo = countBrackets(response);\n      return {\n        isTruncated: true,\n        reason: 'json_incomplete',\n        partialContent: response,\n        lastCompleteElement: findLastCompleteElement(response),\n        missingCloseBrackets: bracketInfo.missingCloseBrackets,\n        missingCloseBraces: bracketInfo.missingCloseBraces,\n      };\n    }\n\n    try {\n      const closed = autoCloseJson(trimmed);\n      JSON.parse(closed);\n      return { isTruncated: false, reason: 'none' };\n    } catch {\n      return { isTruncated: false, reason: 'none' };\n    }\n  }\n\n  const bracketInfo = countBrackets(response);\n  if (\n    bracketInfo.missingCloseBrackets > 0 ||\n    bracketInfo.missingCloseBraces > 0\n  ) {\n    return {\n      isTruncated: true,\n      reason: 'bracket_mismatch',\n      partialContent: response,\n      lastCompleteElement: findLastCompleteElement(response),\n      missingCloseBrackets: bracketInfo.missingCloseBrackets,\n      missingCloseBraces: bracketInfo.missingCloseBraces,\n    };\n  }\n\n  return { isTruncated: false, reason: 'none' };\n}\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\nfunction countBrackets(json: string): {\n  openBrackets: number;\n  closeBrackets: number;\n  openBraces: number;\n  closeBraces: number;\n  missingCloseBrackets: number;\n  missingCloseBraces: number;\n} {\n  let inString = false;\n  let escaped = false;\n  let openBrackets = 0;\n  let closeBrackets = 0;\n  let openBraces = 0;\n  let closeBraces = 0;\n\n  for (const char of json) {\n    if (escaped) {\n      escaped = false;\n      continue;\n    }\n    if (char === '\\\\' && inString) {\n      escaped = true;\n      continue;\n    }\n    if (char === '\"') {\n      inString = !inString;\n      continue;\n    }\n    if (inString) continue;\n\n    switch (char) {\n      case '[':\n        openBrackets++;\n        break;\n      case ']':\n        closeBrackets++;\n        break;\n      case '{':\n        openBraces++;\n        break;\n      case '}':\n        closeBraces++;\n        break;\n    }\n  }\n\n  return {\n    openBrackets,\n    closeBrackets,\n    openBraces,\n    closeBraces,\n    missingCloseBrackets: Math.max(0, openBrackets - closeBrackets),\n    missingCloseBraces: Math.max(0, openBraces - closeBraces),\n  };\n}\n\nexport function findLastCompleteElement(json: string): unknown | null {\n  const autoClosed = autoCloseJson(json);\n  try {\n    return JSON.parse(autoClosed);\n  } catch {\n    // Auto-close didn't work\n  }\n\n  const trimmed = json.trim();\n\n  if (trimmed.startsWith('[')) {\n    const lastCompleteIndex = findLastCompleteArrayElement(trimmed);\n    if (lastCompleteIndex > 0) {\n      const subset = trimmed.substring(0, lastCompleteIndex) + ']';\n      try {\n        return JSON.parse(subset);\n      } catch {\n        // Continue\n      }\n    }\n  }\n\n  if (trimmed.startsWith('{')) {\n    const closed = autoCloseJson(trimmed);\n    try {\n      return JSON.parse(closed);\n    } catch {\n      const lastCompleteIndex = findLastCompleteObjectProperty(trimmed);\n      if (lastCompleteIndex > 0) {\n        const subset = trimmed.substring(0, lastCompleteIndex) + '}';\n        try {\n          return JSON.parse(subset);\n        } catch {\n          // Give up\n        }\n      }\n    }\n  }\n\n  return null;\n}\n\nfunction findLastCompleteArrayElement(json: string): number {\n  let depth = 0;\n  let inString = false;\n  let escaped = false;\n  let lastCompleteElementEnd = -1;\n\n  for (let i = 0; i < json.length; i++) {\n    const char = json[i];\n\n    if (escaped) {\n      escaped = false;\n      continue;\n    }\n    if (char === '\\\\' && inString) {\n      escaped = true;\n      continue;\n    }\n    if (char === '\"') {\n      inString = !inString;\n      continue;\n    }\n    if (inString) continue;\n\n    if (char === '[' || char === '{') {\n      depth++;\n    } else if (char === ']' || char === '}') {\n      depth--;\n      if (depth === 1) {\n        lastCompleteElementEnd = i + 1;\n      }\n    } else if (char === ',' && depth === 1) {\n      lastCompleteElementEnd = i;\n    }\n  }\n\n  return lastCompleteElementEnd > 0 ? lastCompleteElementEnd : -1;\n}\n\nfunction findLastCompleteObjectProperty(json: string): number {\n  let depth = 0;\n  let inString = false;\n  let escaped = false;\n  let lastCommaIndex = -1;\n\n  for (let i = 0; i < json.length; i++) {\n    const char = json[i];\n\n    if (escaped) {\n      escaped = false;\n      continue;\n    }\n    if (char === '\\\\' && inString) {\n      escaped = true;\n      continue;\n    }\n    if (char === '\"') {\n      inString = !inString;\n      continue;\n    }\n    if (inString) continue;\n\n    if (char === '[' || char === '{') {\n      depth++;\n    } else if (char === ']' || char === '}') {\n      depth--;\n    } else if (char === ',' && depth === 1) {\n      lastCommaIndex = i;\n    }\n  }\n\n  return lastCommaIndex > 0 ? lastCommaIndex : -1;\n}\n\nexport function isLikelyTruncated(content: string): boolean {\n  const trimmed = content.trim();\n  if (!trimmed) return false;\n\n  const brackets = countBrackets(trimmed);\n  if (\n    brackets.missingCloseBrackets > 0 ||\n    brackets.missingCloseBraces > 0\n  ) {\n    return true;\n  }\n\n  const abruptEndings = [\n    /,\\s*$/,\n    /:\\s*$/,\n    /\"\\s*:\\s*$/,\n    /\\[\\s*$/,\n    /{\\s*$/,\n  ];\n\n  for (const pattern of abruptEndings) {\n    if (pattern.test(trimmed)) return true;\n  }\n\n  return false;\n}\n","/**\n * LLM Continuation Utility\n *\n * Handles truncated LLM responses with automatic continuation.\n * - Detects truncation via finish_reason and JSON structure\n * - Automatically continues with full context\n * - Merges partial and continuation responses\n * - Salvages partial data if max continuations reached\n *\n * @packageDocumentation\n */\n\nimport { z } from 'zod';\nimport { LLMClient, type LLMFinishReason } from './client.js';\nimport { detectTruncation } from './truncation-detector.js';\nimport { extractJsonFromText, autoCloseJson, isValidJson } from './json-parser.js';\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface ContinuationOptions<T> {\n  client: LLMClient;\n  systemPrompt: string;\n  userPrompt: string;\n  schema?: z.ZodSchema<T>;\n  maxTokens?: number;\n  maxContinuations?: number;\n  maxRetries?: number;\n  buildContinuationPrompt: (\n    partialResponse: string,\n    attempt: number,\n  ) => string;\n  continuationSystemPrompt?: string;\n}\n\nexport interface ContinuationResult<T> {\n  data: T;\n  raw: string;\n  continuationCount: number;\n  warnings: string[];\n  wasSalvaged: boolean;\n}\n\n// ============================================================================\n// Constants\n// ============================================================================\n\nconst DEFAULT_MAX_TOKENS = 8192;\nconst DEFAULT_MAX_CONTINUATIONS = 3;\n\n/**\n * Default continuation system prompt.\n * Used when no custom continuationSystemPrompt is provided.\n */\nconst DEFAULT_CONTINUATION_SYSTEM_PROMPT = `You are a JSON continuation assistant. Your ONLY job is to continue generating JSON from where the previous response was truncated.\n\nRules:\n1. Continue from EXACTLY where the previous output stopped\n2. Do NOT repeat any content already generated\n3. Complete the JSON structure properly with all closing brackets\n4. Do NOT wrap in markdown code blocks\n5. Output ONLY the continuation JSON, nothing else`;\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\n/**\n * Merge a previous partial response with a continuation.\n *\n * Handles JSON structure merging, removing markdown code blocks,\n * and ensuring proper comma separation between merged parts.\n *\n * @param {string} previous - The previous partial response\n * @param {string} continuation - The continuation to merge\n * @returns {string} The merged response\n */\nexport function mergeResponses(\n  previous: string,\n  continuation: string,\n): string {\n  const trimmedPrev = previous.trimEnd();\n  const trimmedCont = continuation.trimStart();\n\n  let cleanedCont = trimmedCont\n    .replace(/^```json?\\s*/i, '')\n    .replace(/```\\s*$/i, '')\n    .trim();\n\n  if (cleanedCont.startsWith('{')) {\n    try {\n      const contParsed = JSON.parse(autoCloseJson(cleanedCont));\n      const keys = Object.keys(contParsed);\n      if (keys.length === 1 && Array.isArray(contParsed[keys[0]])) {\n        cleanedCont = contParsed[keys[0]]\n          .map((item: unknown) => JSON.stringify(item))\n          .join(',\\n');\n      }\n    } catch {\n      // Continue with original cleaning\n    }\n  }\n\n  if (cleanedCont.startsWith('}') || cleanedCont.startsWith(']')) {\n    return trimmedPrev + cleanedCont;\n  }\n\n  const prevEndsWithValue = /[\\}\\]\\\"\\d]$/.test(trimmedPrev);\n  const contStartsWithValue = /^[\\{\\[\\\"]/.test(cleanedCont);\n\n  if (prevEndsWithValue && contStartsWithValue) {\n    return trimmedPrev + ',\\n' + cleanedCont;\n  }\n\n  return trimmedPrev + cleanedCont;\n}\n\nexport function salvagePartialResponse<T>(rawResponse: string): T | null {\n  console.warn('[Continuation] Attempting to salvage partial response');\n\n  try {\n    const cleanedResponse = extractJsonFromText(rawResponse) || rawResponse;\n    const closed = autoCloseJson(cleanedResponse);\n    const parsed = JSON.parse(closed) as T;\n    console.log('[Continuation] Successfully salvaged partial response');\n    return parsed;\n  } catch (error) {\n    console.error('[Continuation] Could not salvage response:', error);\n  }\n\n  return null;\n}\n\n// ============================================================================\n// Main Function\n// ============================================================================\n\n/**\n * Call an LLM with automatic continuation handling.\n *\n * Manages token limits by detecting truncation and making continuation\n * calls until the complete response is received or max continuations reached.\n *\n * @template T - Expected response type\n * @param {ContinuationOptions<T>} options - Continuation call options\n * @returns {Promise<ContinuationResult<T>>} Result with data and continuation metadata\n */\nexport async function callWithContinuation<T>(\n  options: ContinuationOptions<T>,\n): Promise<ContinuationResult<T>> {\n  const {\n    client,\n    systemPrompt,\n    userPrompt,\n    schema,\n    maxTokens = DEFAULT_MAX_TOKENS,\n    maxContinuations = DEFAULT_MAX_CONTINUATIONS,\n    buildContinuationPrompt,\n    continuationSystemPrompt = DEFAULT_CONTINUATION_SYSTEM_PROMPT,\n  } = options;\n\n  let rawResponse = '';\n  let continuationCount = 0;\n  const warnings: string[] = [];\n  let wasSalvaged = false;\n\n  console.log('[Continuation] Starting LLM call with continuation support');\n  console.log(\n    `[Continuation] Max tokens: ${maxTokens}, Max continuations: ${maxContinuations}`,\n  );\n\n  try {\n    const response = await client.callRawWithMetadata({\n      systemPrompt,\n      userPrompt,\n      maxTokens,\n    });\n\n    rawResponse = extractJsonFromText(response.raw) || response.raw;\n\n    console.log(\n      `[Continuation] Initial response: ${rawResponse.length} chars, finish_reason: ${response.finishReason}`,\n    );\n\n    let truncation = detectTruncation(rawResponse, response.finishReason);\n\n    while (truncation.isTruncated && continuationCount < maxContinuations) {\n      continuationCount++;\n      const warningMsg = `Response truncated (${truncation.reason}), continuing (attempt ${continuationCount}/${maxContinuations})`;\n      console.log(`[Continuation] ${warningMsg}`);\n      warnings.push(warningMsg);\n\n      const contPrompt = buildContinuationPrompt(\n        rawResponse,\n        continuationCount,\n      );\n\n      const contResponse = await client.callRawWithMetadata({\n        systemPrompt: continuationSystemPrompt,\n        userPrompt: contPrompt,\n        maxTokens,\n      });\n\n      console.log(\n        `[Continuation] Continuation response: ${contResponse.raw.length} chars, finish_reason: ${contResponse.finishReason}`,\n      );\n\n      const cleanedContResponse =\n        extractJsonFromText(contResponse.raw) || contResponse.raw;\n      rawResponse = mergeResponses(rawResponse, cleanedContResponse);\n\n      truncation = detectTruncation(rawResponse, contResponse.finishReason);\n    }\n\n    if (\n      continuationCount >= maxContinuations &&\n      truncation.isTruncated\n    ) {\n      console.warn(\n        `[Continuation] Reached max continuations (${maxContinuations}), attempting to salvage...`,\n      );\n      warnings.push(\n        `Reached max continuations - some content may be incomplete`,\n      );\n      wasSalvaged = true;\n    }\n\n    const cleanedResponse =\n      extractJsonFromText(rawResponse) || rawResponse;\n    let data: T;\n\n    try {\n      if (isValidJson(cleanedResponse)) {\n        data = JSON.parse(cleanedResponse) as T;\n      } else {\n        const closed = autoCloseJson(cleanedResponse);\n        data = JSON.parse(closed) as T;\n        if (!wasSalvaged) {\n          warnings.push('Response required auto-closing of JSON brackets');\n        }\n      }\n    } catch (parseError) {\n      const salvaged = salvagePartialResponse<T>(cleanedResponse);\n      if (salvaged) {\n        data = salvaged;\n        wasSalvaged = true;\n        warnings.push('Response was salvaged from partial data');\n      } else {\n        throw new Error(\n          `Failed to parse response after ${continuationCount} continuations: ${parseError}`,\n        );\n      }\n    }\n\n    if (schema) {\n      try {\n        data = schema.parse(data);\n      } catch (validationError) {\n        console.warn(\n          '[Continuation] Schema validation failed:',\n          validationError,\n        );\n        warnings.push(`Schema validation issue: ${validationError}`);\n      }\n    }\n\n    console.log(\n      `[Continuation] Complete. Continuations: ${continuationCount}, Warnings: ${warnings.length}`,\n    );\n\n    return {\n      data,\n      raw: rawResponse,\n      continuationCount,\n      warnings,\n      wasSalvaged,\n    };\n  } catch (error) {\n    console.error('[Continuation] Error during LLM call:', error);\n    throw error;\n  }\n}\n\n/**\n * Build a generic continuation prompt.\n *\n * Creates a prompt that instructs the LLM to continue from where\n * a previous truncated response left off.\n *\n * @param {string} context - Original context/prompt\n * @param {string} partialResponse - The partial response generated so far\n * @param {number} attempt - Current continuation attempt number\n * @param {number} [maxAttempts] - Maximum allowed continuation attempts\n * @returns {string} The formatted continuation prompt\n */\nexport function buildGenericContinuationPrompt(\n  context: string,\n  partialResponse: string,\n  attempt: number,\n  maxAttempts: number = DEFAULT_MAX_CONTINUATIONS,\n): string {\n  return `## CONTINUATION REQUEST (Attempt ${attempt}/${maxAttempts})\n\nYour previous response was truncated. Continue generating from where you left off.\n\n### ORIGINAL CONTEXT\n${context}\n\n### WHAT YOU GENERATED SO FAR\n\\`\\`\\`json\n${partialResponse}\n\\`\\`\\`\n\n### INSTRUCTIONS\n1. Continue from EXACTLY where the response was cut off\n2. Do NOT repeat any content already generated\n3. Complete the JSON structure properly\n4. Do NOT wrap your response in markdown code blocks\n\nContinue generating now:`;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqDA,IAAM,iBAAoD;AAAA,EACxD,QAAQ;AAAA,EACR,YAAY;AACd;AAEA,IAAM,oBAAuD;AAAA,EAC3D,QAAQ;AAAA,EACR,YAAY;AACd;AAEA,IAAM,mBAAsD;AAAA,EAC1D,QAAQ;AAAA,EACR,YAAY;AACd;AAEO,IAAM,kBAAN,MAAsB;AAAA,EAM3B,YAAY,MAA8B;AACxC,SAAK,WAAW,KAAK;AACrB,SAAK,QAAQ,KAAK,SAAS,eAAe,KAAK,QAAQ;AACvD,SAAK,WAAW,KAAK,WAAW,kBAAkB,KAAK,QAAQ,GAAG,QAAQ,OAAO,EAAE;AAEnF,UAAM,SAAS,KAAK,UAAU,KAAK,cAAc;AACjD,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR,0CAA0C,KAAK,QAAQ,oBAC9C,iBAAiB,KAAK,QAAQ,CAAC;AAAA,MAC1C;AAAA,IACF;AACA,SAAK,SAAS;AAAA,EAChB;AAAA;AAAA,EAGA,MAAM,MAAM,MAAwC;AAClD,UAAM,QAAQ,MAAM,KAAK,WAAW,CAAC,IAAI,CAAC;AAC1C,UAAM,YAAY,MAAM,WAAW,CAAC;AACpC,QAAI,CAAC,WAAW;AACd,YAAM,IAAI,MAAM,uDAAuD;AAAA,IACzE;AACA,WAAO,EAAE,WAAW,OAAO,MAAM,MAAM;AAAA,EACzC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,WAAW,OAA6D;AAC5E,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,EAAE,YAAY,CAAC,GAAG,OAAO,EAAE,cAAc,GAAG,aAAa,EAAE,EAAE;AAAA,IACtE;AAKA,UAAM,UAAU,MAAM,IAAI,CAAC,MAAO,EAAE,SAAS,MAAO,EAAE,MAAM,GAAG,GAAI,IAAI,CAAE;AACzE,UAAM,WAAW,MAAM,MAAM,GAAG,KAAK,OAAO,eAAe;AAAA,MACzD,QAAQ;AAAA,MACR,SAAS;AAAA,QACP,gBAAgB;AAAA,QAChB,eAAe,UAAU,KAAK,MAAM;AAAA,MACtC;AAAA,MACA,MAAM,KAAK,UAAU,EAAE,OAAO,KAAK,OAAO,OAAO,QAAQ,CAAC;AAAA,IAC5D,CAAC;AACD,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,YAAY,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,WAAW;AAC/D,YAAM,IAAI;AAAA,QACR,oBAAoB,KAAK,QAAQ,aAAa,SAAS,MAAM,IAAI,SAAS,UAAU,KAAK,SAAS;AAAA,MACpG;AAAA,IACF;AACA,UAAM,OAAQ,MAAM,SAAS,KAAK;AAClC,QAAI,CAAC,MAAM,QAAQ,KAAK,IAAI,GAAG;AAC7B,YAAM,IAAI,MAAM,yDAAyD;AAAA,IAC3E;AACA,UAAM,UAAU,CAAC,GAAG,KAAK,IAAI,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAC/D,QAAI,QAAQ,WAAW,MAAM,QAAQ;AACnC,YAAM,IAAI;AAAA,QACR,sCAAsC,QAAQ,MAAM,mBAAmB,MAAM,MAAM;AAAA,MACrF;AAAA,IACF;AACA,WAAO;AAAA,MACL,YAAY,QAAQ,IAAI,CAAC,MAAM,EAAE,SAAS;AAAA,MAC1C,OAAO;AAAA,QACL,cAAc,KAAK,OAAO,iBAAiB;AAAA,QAC3C,aAAa,KAAK,OAAO,gBAAgB;AAAA,MAC3C;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,gBAAoC;AAC1C,WAAO,QAAQ,IAAI,iBAAiB,KAAK,QAAQ,CAAC;AAAA,EACpD;AACF;AAWO,SAAS,iBAAiB,GAA0B,GAAkC;AAC3F,MAAI,EAAE,WAAW,EAAE,QAAQ;AACzB,UAAM,IAAI,MAAM,6CAA6C,EAAE,MAAM,OAAO,EAAE,MAAM,GAAG;AAAA,EACzF;AACA,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC;AACd,UAAM,KAAK,EAAE,CAAC;AACd,WAAO,KAAK;AACZ,aAAS,KAAK;AACd,aAAS,KAAK;AAAA,EAChB;AACA,MAAI,UAAU,KAAK,UAAU,EAAG,QAAO;AACvC,SAAO,OAAO,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,KAAK;AAClD;;;ACnIO,SAAS,iBACd,UACA,cACkB;AAClB,MAAI,iBAAiB,UAAU;AAC7B,UAAMA,eAAc,cAAc,QAAQ;AAC1C,WAAO;AAAA,MACL,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,gBAAgB;AAAA,MAChB,qBAAqB,wBAAwB,QAAQ;AAAA,MACrD,sBAAsBA,aAAY;AAAA,MAClC,oBAAoBA,aAAY;AAAA,IAClC;AAAA,EACF;AAEA,MAAI;AACF,SAAK,MAAM,QAAQ;AACnB,WAAO,EAAE,aAAa,OAAO,QAAQ,OAAO;AAAA,EAC9C,QAAQ;AAAA,EAER;AAEA,MAAI,iBAAiB,UAAU,iBAAiB,MAAM;AACpD,UAAM,UAAU,SAAS,KAAK;AAE9B,UAAM,eACJ,QAAQ,SAAS,GAAG,KACpB,QAAQ,SAAS,GAAG,KACpB,QAAQ,SAAS,KAAK,KACtB,QAAQ,KAAK,OAAO,KACpB,QAAQ,KAAK,OAAO;AAEtB,QAAI,cAAc;AAChB,YAAMA,eAAc,cAAc,QAAQ;AAC1C,aAAO;AAAA,QACL,aAAa;AAAA,QACb,QAAQ;AAAA,QACR,gBAAgB;AAAA,QAChB,qBAAqB,wBAAwB,QAAQ;AAAA,QACrD,sBAAsBA,aAAY;AAAA,QAClC,oBAAoBA,aAAY;AAAA,MAClC;AAAA,IACF;AAEA,QAAI;AACF,YAAM,SAAS,cAAc,OAAO;AACpC,WAAK,MAAM,MAAM;AACjB,aAAO,EAAE,aAAa,OAAO,QAAQ,OAAO;AAAA,IAC9C,QAAQ;AACN,aAAO,EAAE,aAAa,OAAO,QAAQ,OAAO;AAAA,IAC9C;AAAA,EACF;AAEA,QAAM,cAAc,cAAc,QAAQ;AAC1C,MACE,YAAY,uBAAuB,KACnC,YAAY,qBAAqB,GACjC;AACA,WAAO;AAAA,MACL,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,gBAAgB;AAAA,MAChB,qBAAqB,wBAAwB,QAAQ;AAAA,MACrD,sBAAsB,YAAY;AAAA,MAClC,oBAAoB,YAAY;AAAA,IAClC;AAAA,EACF;AAEA,SAAO,EAAE,aAAa,OAAO,QAAQ,OAAO;AAC9C;AAMA,SAAS,cAAc,MAOrB;AACA,MAAI,WAAW;AACf,MAAI,UAAU;AACd,MAAI,eAAe;AACnB,MAAI,gBAAgB;AACpB,MAAI,aAAa;AACjB,MAAI,cAAc;AAElB,aAAW,QAAQ,MAAM;AACvB,QAAI,SAAS;AACX,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,QAAQ,UAAU;AAC7B,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,KAAK;AAChB,iBAAW,CAAC;AACZ;AAAA,IACF;AACA,QAAI,SAAU;AAEd,YAAQ,MAAM;AAAA,MACZ,KAAK;AACH;AACA;AAAA,MACF,KAAK;AACH;AACA;AAAA,MACF,KAAK;AACH;AACA;AAAA,MACF,KAAK;AACH;AACA;AAAA,IACJ;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,sBAAsB,KAAK,IAAI,GAAG,eAAe,aAAa;AAAA,IAC9D,oBAAoB,KAAK,IAAI,GAAG,aAAa,WAAW;AAAA,EAC1D;AACF;AAEO,SAAS,wBAAwB,MAA8B;AACpE,QAAM,aAAa,cAAc,IAAI;AACrC,MAAI;AACF,WAAO,KAAK,MAAM,UAAU;AAAA,EAC9B,QAAQ;AAAA,EAER;AAEA,QAAM,UAAU,KAAK,KAAK;AAE1B,MAAI,QAAQ,WAAW,GAAG,GAAG;AAC3B,UAAM,oBAAoB,6BAA6B,OAAO;AAC9D,QAAI,oBAAoB,GAAG;AACzB,YAAM,SAAS,QAAQ,UAAU,GAAG,iBAAiB,IAAI;AACzD,UAAI;AACF,eAAO,KAAK,MAAM,MAAM;AAAA,MAC1B,QAAQ;AAAA,MAER;AAAA,IACF;AAAA,EACF;AAEA,MAAI,QAAQ,WAAW,GAAG,GAAG;AAC3B,UAAM,SAAS,cAAc,OAAO;AACpC,QAAI;AACF,aAAO,KAAK,MAAM,MAAM;AAAA,IAC1B,QAAQ;AACN,YAAM,oBAAoB,+BAA+B,OAAO;AAChE,UAAI,oBAAoB,GAAG;AACzB,cAAM,SAAS,QAAQ,UAAU,GAAG,iBAAiB,IAAI;AACzD,YAAI;AACF,iBAAO,KAAK,MAAM,MAAM;AAAA,QAC1B,QAAQ;AAAA,QAER;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,6BAA6B,MAAsB;AAC1D,MAAI,QAAQ;AACZ,MAAI,WAAW;AACf,MAAI,UAAU;AACd,MAAI,yBAAyB;AAE7B,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,OAAO,KAAK,CAAC;AAEnB,QAAI,SAAS;AACX,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,QAAQ,UAAU;AAC7B,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,KAAK;AAChB,iBAAW,CAAC;AACZ;AAAA,IACF;AACA,QAAI,SAAU;AAEd,QAAI,SAAS,OAAO,SAAS,KAAK;AAChC;AAAA,IACF,WAAW,SAAS,OAAO,SAAS,KAAK;AACvC;AACA,UAAI,UAAU,GAAG;AACf,iCAAyB,IAAI;AAAA,MAC/B;AAAA,IACF,WAAW,SAAS,OAAO,UAAU,GAAG;AACtC,+BAAyB;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO,yBAAyB,IAAI,yBAAyB;AAC/D;AAEA,SAAS,+BAA+B,MAAsB;AAC5D,MAAI,QAAQ;AACZ,MAAI,WAAW;AACf,MAAI,UAAU;AACd,MAAI,iBAAiB;AAErB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,OAAO,KAAK,CAAC;AAEnB,QAAI,SAAS;AACX,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,QAAQ,UAAU;AAC7B,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,KAAK;AAChB,iBAAW,CAAC;AACZ;AAAA,IACF;AACA,QAAI,SAAU;AAEd,QAAI,SAAS,OAAO,SAAS,KAAK;AAChC;AAAA,IACF,WAAW,SAAS,OAAO,SAAS,KAAK;AACvC;AAAA,IACF,WAAW,SAAS,OAAO,UAAU,GAAG;AACtC,uBAAiB;AAAA,IACnB;AAAA,EACF;AAEA,SAAO,iBAAiB,IAAI,iBAAiB;AAC/C;AAEO,SAAS,kBAAkB,SAA0B;AAC1D,QAAM,UAAU,QAAQ,KAAK;AAC7B,MAAI,CAAC,QAAS,QAAO;AAErB,QAAM,WAAW,cAAc,OAAO;AACtC,MACE,SAAS,uBAAuB,KAChC,SAAS,qBAAqB,GAC9B;AACA,WAAO;AAAA,EACT;AAEA,QAAM,gBAAgB;AAAA,IACpB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,aAAW,WAAW,eAAe;AACnC,QAAI,QAAQ,KAAK,OAAO,EAAG,QAAO;AAAA,EACpC;AAEA,SAAO;AACT;;;AC7QA,IAAM,qBAAqB;AAC3B,IAAM,4BAA4B;AAMlC,IAAM,qCAAqC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAuBpC,SAAS,eACd,UACA,cACQ;AACR,QAAM,cAAc,SAAS,QAAQ;AACrC,QAAM,cAAc,aAAa,UAAU;AAE3C,MAAI,cAAc,YACf,QAAQ,iBAAiB,EAAE,EAC3B,QAAQ,YAAY,EAAE,EACtB,KAAK;AAER,MAAI,YAAY,WAAW,GAAG,GAAG;AAC/B,QAAI;AACF,YAAM,aAAa,KAAK,MAAM,cAAc,WAAW,CAAC;AACxD,YAAM,OAAO,OAAO,KAAK,UAAU;AACnC,UAAI,KAAK,WAAW,KAAK,MAAM,QAAQ,WAAW,KAAK,CAAC,CAAC,CAAC,GAAG;AAC3D,sBAAc,WAAW,KAAK,CAAC,CAAC,EAC7B,IAAI,CAAC,SAAkB,KAAK,UAAU,IAAI,CAAC,EAC3C,KAAK,KAAK;AAAA,MACf;AAAA,IACF,QAAQ;AAAA,IAER;AAAA,EACF;AAEA,MAAI,YAAY,WAAW,GAAG,KAAK,YAAY,WAAW,GAAG,GAAG;AAC9D,WAAO,cAAc;AAAA,EACvB;AAEA,QAAM,oBAAoB,cAAc,KAAK,WAAW;AACxD,QAAM,sBAAsB,YAAY,KAAK,WAAW;AAExD,MAAI,qBAAqB,qBAAqB;AAC5C,WAAO,cAAc,QAAQ;AAAA,EAC/B;AAEA,SAAO,cAAc;AACvB;AAEO,SAAS,uBAA0B,aAA+B;AACvE,UAAQ,KAAK,uDAAuD;AAEpE,MAAI;AACF,UAAM,kBAAkB,oBAAoB,WAAW,KAAK;AAC5D,UAAM,SAAS,cAAc,eAAe;AAC5C,UAAM,SAAS,KAAK,MAAM,MAAM;AAChC,YAAQ,IAAI,uDAAuD;AACnE,WAAO;AAAA,EACT,SAAS,OAAO;AACd,YAAQ,MAAM,8CAA8C,KAAK;AAAA,EACnE;AAEA,SAAO;AACT;AAgBA,eAAsB,qBACpB,SACgC;AAChC,QAAM;AAAA,IACJ;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,YAAY;AAAA,IACZ,mBAAmB;AAAA,IACnB;AAAA,IACA,2BAA2B;AAAA,EAC7B,IAAI;AAEJ,MAAI,cAAc;AAClB,MAAI,oBAAoB;AACxB,QAAM,WAAqB,CAAC;AAC5B,MAAI,cAAc;AAElB,UAAQ,IAAI,4DAA4D;AACxE,UAAQ;AAAA,IACN,8BAA8B,SAAS,wBAAwB,gBAAgB;AAAA,EACjF;AAEA,MAAI;AACF,UAAM,WAAW,MAAM,OAAO,oBAAoB;AAAA,MAChD;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAED,kBAAc,oBAAoB,SAAS,GAAG,KAAK,SAAS;AAE5D,YAAQ;AAAA,MACN,oCAAoC,YAAY,MAAM,0BAA0B,SAAS,YAAY;AAAA,IACvG;AAEA,QAAI,aAAa,iBAAiB,aAAa,SAAS,YAAY;AAEpE,WAAO,WAAW,eAAe,oBAAoB,kBAAkB;AACrE;AACA,YAAM,aAAa,uBAAuB,WAAW,MAAM,0BAA0B,iBAAiB,IAAI,gBAAgB;AAC1H,cAAQ,IAAI,kBAAkB,UAAU,EAAE;AAC1C,eAAS,KAAK,UAAU;AAExB,YAAM,aAAa;AAAA,QACjB;AAAA,QACA;AAAA,MACF;AAEA,YAAM,eAAe,MAAM,OAAO,oBAAoB;AAAA,QACpD,cAAc;AAAA,QACd,YAAY;AAAA,QACZ;AAAA,MACF,CAAC;AAED,cAAQ;AAAA,QACN,yCAAyC,aAAa,IAAI,MAAM,0BAA0B,aAAa,YAAY;AAAA,MACrH;AAEA,YAAM,sBACJ,oBAAoB,aAAa,GAAG,KAAK,aAAa;AACxD,oBAAc,eAAe,aAAa,mBAAmB;AAE7D,mBAAa,iBAAiB,aAAa,aAAa,YAAY;AAAA,IACtE;AAEA,QACE,qBAAqB,oBACrB,WAAW,aACX;AACA,cAAQ;AAAA,QACN,6CAA6C,gBAAgB;AAAA,MAC/D;AACA,eAAS;AAAA,QACP;AAAA,MACF;AACA,oBAAc;AAAA,IAChB;AAEA,UAAM,kBACJ,oBAAoB,WAAW,KAAK;AACtC,QAAI;AAEJ,QAAI;AACF,UAAI,YAAY,eAAe,GAAG;AAChC,eAAO,KAAK,MAAM,eAAe;AAAA,MACnC,OAAO;AACL,cAAM,SAAS,cAAc,eAAe;AAC5C,eAAO,KAAK,MAAM,MAAM;AACxB,YAAI,CAAC,aAAa;AAChB,mBAAS,KAAK,iDAAiD;AAAA,QACjE;AAAA,MACF;AAAA,IACF,SAAS,YAAY;AACnB,YAAM,WAAW,uBAA0B,eAAe;AAC1D,UAAI,UAAU;AACZ,eAAO;AACP,sBAAc;AACd,iBAAS,KAAK,yCAAyC;AAAA,MACzD,OAAO;AACL,cAAM,IAAI;AAAA,UACR,kCAAkC,iBAAiB,mBAAmB,UAAU;AAAA,QAClF;AAAA,MACF;AAAA,IACF;AAEA,QAAI,QAAQ;AACV,UAAI;AACF,eAAO,OAAO,MAAM,IAAI;AAAA,MAC1B,SAAS,iBAAiB;AACxB,gBAAQ;AAAA,UACN;AAAA,UACA;AAAA,QACF;AACA,iBAAS,KAAK,4BAA4B,eAAe,EAAE;AAAA,MAC7D;AAAA,IACF;AAEA,YAAQ;AAAA,MACN,2CAA2C,iBAAiB,eAAe,SAAS,MAAM;AAAA,IAC5F;AAEA,WAAO;AAAA,MACL;AAAA,MACA,KAAK;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF,SAAS,OAAO;AACd,YAAQ,MAAM,yCAAyC,KAAK;AAC5D,UAAM;AAAA,EACR;AACF;AAcO,SAAS,+BACd,SACA,iBACA,SACA,cAAsB,2BACd;AACR,SAAO,oCAAoC,OAAO,IAAI,WAAW;AAAA;AAAA;AAAA;AAAA;AAAA,EAKjE,OAAO;AAAA;AAAA;AAAA;AAAA,EAIP,eAAe;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAUjB;","names":["bracketInfo"]}
+{"version":3,"sources":["../src/embedding-client.ts","../src/truncation-detector.ts","../src/continuation.ts"],"sourcesContent":["/**\n * Embedding Client\n *\n * Single-purpose client for text embeddings. Used by `@almadar-io/agent`'s\n * cosine-similarity catalog retrieval (rank organisms + atoms against the\n * user's request before rendering Stage A's prompt) and by\n * `@almadar/std`'s publish-time embedding bake step.\n *\n * Providers:\n *   - `openai` (default model `text-embedding-3-small`, 1536-d) — requires\n *     `OPENAI_API_KEY`.\n *   - `openrouter` (default model `baai/bge-base-en-v1.5`, 768-d) — requires\n *     `OPEN_ROUTER_API_KEY`. Same OpenAI-compatible request shape, just a\n *     different base URL.\n *\n * Both providers return the same response envelope (`{data:[{embedding,index}]}`),\n * so the request/response code is shared.\n *\n * @packageDocumentation\n */\n\nexport type EmbeddingProvider = 'openai' | 'openrouter';\n\nexport interface EmbeddingClientOptions {\n  provider: EmbeddingProvider;\n  /** Defaults: openai → text-embedding-3-small, openrouter → baai/bge-base-en-v1.5. */\n  model?: string;\n  /** Override API key. Defaults to provider-specific env var. */\n  apiKey?: string;\n  /** Override base URL. Defaults to provider canonical endpoint. */\n  baseUrl?: string;\n}\n\nexport interface EmbeddingUsage {\n  promptTokens: number;\n  totalTokens: number;\n}\n\nexport interface EmbeddingResult {\n  embedding: readonly number[];\n  usage: EmbeddingUsage;\n}\n\nexport interface EmbeddingBatchResult {\n  embeddings: readonly (readonly number[])[];\n  usage: EmbeddingUsage;\n}\n\ninterface OpenAIEmbeddingApiResponse {\n  data: Array<{ embedding: number[]; index: number }>;\n  usage?: { prompt_tokens?: number; total_tokens?: number };\n  /** OpenRouter reports provider hiccups as HTTP 200 with an error envelope and no `data`. */\n  error?: { message?: string; code?: number | string };\n}\n\n/** Transient-failure retry policy: attempts and backoff (ms) between them. */\nconst EMBED_MAX_ATTEMPTS = 3;\nconst EMBED_RETRY_BACKOFF_MS = [500, 2000];\n\n/** Provider rejects an input that exceeds the model's context window (bge: 512 tokens). */\nconst TOKEN_OVERFLOW_RE =\n  /context length|maximum input length|reduce the length of the input|input_tokens/i;\n\nconst DEFAULT_MODELS: Record<EmbeddingProvider, string> = {\n  openai: 'text-embedding-3-small',\n  openrouter: 'baai/bge-base-en-v1.5',\n};\n\nconst DEFAULT_BASE_URLS: Record<EmbeddingProvider, string> = {\n  openai: 'https://api.openai.com/v1',\n  openrouter: 'https://openrouter.ai/api/v1',\n};\n\nconst API_KEY_ENV_VARS: Record<EmbeddingProvider, string> = {\n  openai: 'OPENAI_API_KEY',\n  openrouter: 'OPEN_ROUTER_API_KEY',\n};\n\nexport class EmbeddingClient {\n  private readonly provider: EmbeddingProvider;\n  private readonly model: string;\n  private readonly apiKey: string;\n  private readonly baseUrl: string;\n\n  constructor(opts: EmbeddingClientOptions) {\n    this.provider = opts.provider;\n    this.model = opts.model ?? DEFAULT_MODELS[opts.provider];\n    this.baseUrl = (opts.baseUrl ?? DEFAULT_BASE_URLS[opts.provider]).replace(/\\/$/, '');\n\n    const apiKey = opts.apiKey ?? this.resolveApiKey();\n    if (!apiKey) {\n      throw new Error(\n        `EmbeddingClient: API key for provider '${this.provider}' not found. ` +\n          `Set ${API_KEY_ENV_VARS[this.provider]} in your environment or pass options.apiKey.`,\n      );\n    }\n    this.apiKey = apiKey;\n  }\n\n  /** Embed a single text. */\n  async embed(text: string): Promise<EmbeddingResult> {\n    const batch = await this.embedBatch([text]);\n    const embedding = batch.embeddings[0];\n    if (!embedding) {\n      throw new Error('EmbeddingClient.embed: provider returned no embedding');\n    }\n    return { embedding, usage: batch.usage };\n  }\n\n  /**\n   * Embed a batch of texts. Used by `@almadar/std`'s publish-time\n   * `build-embeddings.ts` script — one batch call for the full catalog\n   * (≈190 entries) rather than per-entry requests.\n   */\n  async embedBatch(texts: ReadonlyArray<string>): Promise<EmbeddingBatchResult> {\n    if (texts.length === 0) {\n      return { embeddings: [], usage: { promptTokens: 0, totalTokens: 0 } };\n    }\n    // bge-base-en-v1.5 has a 512-token context limit. Chars-per-token varies\n    // with density (~4 for typical English, ~2.5-3 for dense technical/code\n    // text), so no fixed char cap can guarantee a fit. Start at 2000 chars and,\n    // when the provider rejects an input for exceeding the context window,\n    // halve the cap and retry — the provider's own token count is the source\n    // of truth, not a chars-per-token guess.\n    let charCap = 2000;\n    let lastError: Error = new Error('EmbeddingClient: unreachable');\n    for (let attempt = 1; attempt <= EMBED_MAX_ATTEMPTS; attempt++) {\n      const trimmed = texts.map((t) => (t.length > charCap ? t.slice(0, charCap) : t));\n      try {\n        return await this.embedBatchOnce(texts.length, trimmed);\n      } catch (e) {\n        lastError = e instanceof Error ? e : new Error(String(e));\n        if (TOKEN_OVERFLOW_RE.test(lastError.message) && charCap > 500) {\n          charCap = Math.floor(charCap / 2);\n          continue;\n        }\n        const backoff = EMBED_RETRY_BACKOFF_MS[attempt - 1];\n        if (backoff === undefined) break;\n        await new Promise((r) => setTimeout(r, backoff));\n      }\n    }\n    throw lastError;\n  }\n\n  private async embedBatchOnce(\n    inputCount: number,\n    trimmed: ReadonlyArray<string>,\n  ): Promise<EmbeddingBatchResult> {\n    const response = await fetch(`${this.baseUrl}/embeddings`, {\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        Authorization: `Bearer ${this.apiKey}`,\n      },\n      body: JSON.stringify({ model: this.model, input: trimmed }),\n    });\n    if (!response.ok) {\n      const errorText = await response.text().catch(() => '<no body>');\n      throw new Error(\n        `EmbeddingClient: ${this.provider} returned ${response.status} ${response.statusText}: ${errorText}`,\n      );\n    }\n    const rawBody = await response.text();\n    let body: OpenAIEmbeddingApiResponse;\n    try {\n      body = JSON.parse(rawBody) as OpenAIEmbeddingApiResponse;\n    } catch {\n      throw new Error(\n        `EmbeddingClient: ${this.provider} returned unparseable body: ${rawBody}`,\n      );\n    }\n    if (!Array.isArray(body.data)) {\n      // OpenRouter reports provider hiccups as 200 + `{error}` with no data —\n      // surface the real body instead of a blind \"missing data\" (G9: this\n      // error was undiagnosable for weeks because the body was discarded).\n      const detail = body.error?.message ?? rawBody;\n      throw new Error(\n        `EmbeddingClient: ${this.provider} response has no \\`data\\` array: ${detail}`,\n      );\n    }\n    const ordered = [...body.data].sort((a, b) => a.index - b.index);\n    if (ordered.length !== inputCount) {\n      throw new Error(\n        `EmbeddingClient: provider returned ${ordered.length} embeddings for ${inputCount} inputs`,\n      );\n    }\n    return {\n      embeddings: ordered.map((d) => d.embedding),\n      usage: {\n        promptTokens: body.usage?.prompt_tokens ?? 0,\n        totalTokens: body.usage?.total_tokens ?? 0,\n      },\n    };\n  }\n\n  private resolveApiKey(): string | undefined {\n    return process.env[API_KEY_ENV_VARS[this.provider]];\n  }\n}\n\n/**\n * Cosine similarity between two equal-length vectors. Returns a value in\n * [-1, 1]; higher means more similar. Used by `@almadar-io/agent`'s\n * catalog retrieval to rank organism/atom descriptions against the user\n * request embedding.\n *\n * Defensive: returns 0 when either vector is the zero vector (would\n * otherwise divide by zero).\n */\nexport function cosineSimilarity(a: ReadonlyArray<number>, b: ReadonlyArray<number>): number {\n  if (a.length !== b.length) {\n    throw new Error(`cosineSimilarity: vector length mismatch (${a.length} vs ${b.length})`);\n  }\n  let dot = 0;\n  let normA = 0;\n  let normB = 0;\n  for (let i = 0; i < a.length; i++) {\n    const av = a[i];\n    const bv = b[i];\n    dot += av * bv;\n    normA += av * av;\n    normB += bv * bv;\n  }\n  if (normA === 0 || normB === 0) return 0;\n  return dot / (Math.sqrt(normA) * Math.sqrt(normB));\n}\n","/**\n * Truncation Detector\n *\n * Utilities for detecting when LLM output has been truncated and\n * extracting usable content from partial responses.\n *\n * @packageDocumentation\n */\n\nimport type { LLMFinishReason } from './client.js';\nimport { autoCloseJson } from './json-parser.js';\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport type TruncationReason =\n  | 'finish_reason'\n  | 'json_incomplete'\n  | 'bracket_mismatch'\n  | 'none';\n\nexport interface TruncationResult {\n  isTruncated: boolean;\n  reason: TruncationReason;\n  partialContent?: string;\n  lastCompleteElement?: unknown;\n  missingCloseBrackets?: number;\n  missingCloseBraces?: number;\n}\n\n// ============================================================================\n// Main Detection Function\n// ============================================================================\n\n/**\n * Detect if an LLM response has been truncated.\n *\n * Analyzes the response content and finish reason to determine if\n * the output was cut off due to token limits or other issues.\n *\n * @param {string} response - The LLM response text\n * @param {LLMFinishReason} finishReason - The finish reason from the LLM\n * @returns {TruncationResult} Detection result with truncation details\n */\nexport function detectTruncation(\n  response: string,\n  finishReason: LLMFinishReason,\n): TruncationResult {\n  if (finishReason === 'length') {\n    const bracketInfo = countBrackets(response);\n    return {\n      isTruncated: true,\n      reason: 'finish_reason',\n      partialContent: response,\n      lastCompleteElement: findLastCompleteElement(response),\n      missingCloseBrackets: bracketInfo.missingCloseBrackets,\n      missingCloseBraces: bracketInfo.missingCloseBraces,\n    };\n  }\n\n  try {\n    JSON.parse(response);\n    return { isTruncated: false, reason: 'none' };\n  } catch {\n    // JSON is invalid, check if due to truncation\n  }\n\n  if (finishReason === 'stop' || finishReason === null) {\n    const trimmed = response.trim();\n\n    const isMidContent =\n      trimmed.endsWith(',') ||\n      trimmed.endsWith(':') ||\n      trimmed.endsWith('\": ') ||\n      /:\\s*$/.test(trimmed) ||\n      /,\\s*$/.test(trimmed);\n\n    if (isMidContent) {\n      const bracketInfo = countBrackets(response);\n      return {\n        isTruncated: true,\n        reason: 'json_incomplete',\n        partialContent: response,\n        lastCompleteElement: findLastCompleteElement(response),\n        missingCloseBrackets: bracketInfo.missingCloseBrackets,\n        missingCloseBraces: bracketInfo.missingCloseBraces,\n      };\n    }\n\n    try {\n      const closed = autoCloseJson(trimmed);\n      JSON.parse(closed);\n      return { isTruncated: false, reason: 'none' };\n    } catch {\n      return { isTruncated: false, reason: 'none' };\n    }\n  }\n\n  const bracketInfo = countBrackets(response);\n  if (\n    bracketInfo.missingCloseBrackets > 0 ||\n    bracketInfo.missingCloseBraces > 0\n  ) {\n    return {\n      isTruncated: true,\n      reason: 'bracket_mismatch',\n      partialContent: response,\n      lastCompleteElement: findLastCompleteElement(response),\n      missingCloseBrackets: bracketInfo.missingCloseBrackets,\n      missingCloseBraces: bracketInfo.missingCloseBraces,\n    };\n  }\n\n  return { isTruncated: false, reason: 'none' };\n}\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\nfunction countBrackets(json: string): {\n  openBrackets: number;\n  closeBrackets: number;\n  openBraces: number;\n  closeBraces: number;\n  missingCloseBrackets: number;\n  missingCloseBraces: number;\n} {\n  let inString = false;\n  let escaped = false;\n  let openBrackets = 0;\n  let closeBrackets = 0;\n  let openBraces = 0;\n  let closeBraces = 0;\n\n  for (const char of json) {\n    if (escaped) {\n      escaped = false;\n      continue;\n    }\n    if (char === '\\\\' && inString) {\n      escaped = true;\n      continue;\n    }\n    if (char === '\"') {\n      inString = !inString;\n      continue;\n    }\n    if (inString) continue;\n\n    switch (char) {\n      case '[':\n        openBrackets++;\n        break;\n      case ']':\n        closeBrackets++;\n        break;\n      case '{':\n        openBraces++;\n        break;\n      case '}':\n        closeBraces++;\n        break;\n    }\n  }\n\n  return {\n    openBrackets,\n    closeBrackets,\n    openBraces,\n    closeBraces,\n    missingCloseBrackets: Math.max(0, openBrackets - closeBrackets),\n    missingCloseBraces: Math.max(0, openBraces - closeBraces),\n  };\n}\n\nexport function findLastCompleteElement(json: string): unknown | null {\n  const autoClosed = autoCloseJson(json);\n  try {\n    return JSON.parse(autoClosed);\n  } catch {\n    // Auto-close didn't work\n  }\n\n  const trimmed = json.trim();\n\n  if (trimmed.startsWith('[')) {\n    const lastCompleteIndex = findLastCompleteArrayElement(trimmed);\n    if (lastCompleteIndex > 0) {\n      const subset = trimmed.substring(0, lastCompleteIndex) + ']';\n      try {\n        return JSON.parse(subset);\n      } catch {\n        // Continue\n      }\n    }\n  }\n\n  if (trimmed.startsWith('{')) {\n    const closed = autoCloseJson(trimmed);\n    try {\n      return JSON.parse(closed);\n    } catch {\n      const lastCompleteIndex = findLastCompleteObjectProperty(trimmed);\n      if (lastCompleteIndex > 0) {\n        const subset = trimmed.substring(0, lastCompleteIndex) + '}';\n        try {\n          return JSON.parse(subset);\n        } catch {\n          // Give up\n        }\n      }\n    }\n  }\n\n  return null;\n}\n\nfunction findLastCompleteArrayElement(json: string): number {\n  let depth = 0;\n  let inString = false;\n  let escaped = false;\n  let lastCompleteElementEnd = -1;\n\n  for (let i = 0; i < json.length; i++) {\n    const char = json[i];\n\n    if (escaped) {\n      escaped = false;\n      continue;\n    }\n    if (char === '\\\\' && inString) {\n      escaped = true;\n      continue;\n    }\n    if (char === '\"') {\n      inString = !inString;\n      continue;\n    }\n    if (inString) continue;\n\n    if (char === '[' || char === '{') {\n      depth++;\n    } else if (char === ']' || char === '}') {\n      depth--;\n      if (depth === 1) {\n        lastCompleteElementEnd = i + 1;\n      }\n    } else if (char === ',' && depth === 1) {\n      lastCompleteElementEnd = i;\n    }\n  }\n\n  return lastCompleteElementEnd > 0 ? lastCompleteElementEnd : -1;\n}\n\nfunction findLastCompleteObjectProperty(json: string): number {\n  let depth = 0;\n  let inString = false;\n  let escaped = false;\n  let lastCommaIndex = -1;\n\n  for (let i = 0; i < json.length; i++) {\n    const char = json[i];\n\n    if (escaped) {\n      escaped = false;\n      continue;\n    }\n    if (char === '\\\\' && inString) {\n      escaped = true;\n      continue;\n    }\n    if (char === '\"') {\n      inString = !inString;\n      continue;\n    }\n    if (inString) continue;\n\n    if (char === '[' || char === '{') {\n      depth++;\n    } else if (char === ']' || char === '}') {\n      depth--;\n    } else if (char === ',' && depth === 1) {\n      lastCommaIndex = i;\n    }\n  }\n\n  return lastCommaIndex > 0 ? lastCommaIndex : -1;\n}\n\nexport function isLikelyTruncated(content: string): boolean {\n  const trimmed = content.trim();\n  if (!trimmed) return false;\n\n  const brackets = countBrackets(trimmed);\n  if (\n    brackets.missingCloseBrackets > 0 ||\n    brackets.missingCloseBraces > 0\n  ) {\n    return true;\n  }\n\n  const abruptEndings = [\n    /,\\s*$/,\n    /:\\s*$/,\n    /\"\\s*:\\s*$/,\n    /\\[\\s*$/,\n    /{\\s*$/,\n  ];\n\n  for (const pattern of abruptEndings) {\n    if (pattern.test(trimmed)) return true;\n  }\n\n  return false;\n}\n","/**\n * LLM Continuation Utility\n *\n * Handles truncated LLM responses with automatic continuation.\n * - Detects truncation via finish_reason and JSON structure\n * - Automatically continues with full context\n * - Merges partial and continuation responses\n * - Salvages partial data if max continuations reached\n *\n * @packageDocumentation\n */\n\nimport { z } from 'zod';\nimport { LLMClient, type LLMFinishReason } from './client.js';\nimport { detectTruncation } from './truncation-detector.js';\nimport { extractJsonFromText, autoCloseJson, isValidJson } from './json-parser.js';\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface ContinuationOptions<T> {\n  client: LLMClient;\n  systemPrompt: string;\n  userPrompt: string;\n  schema?: z.ZodSchema<T>;\n  maxTokens?: number;\n  maxContinuations?: number;\n  maxRetries?: number;\n  buildContinuationPrompt: (\n    partialResponse: string,\n    attempt: number,\n  ) => string;\n  continuationSystemPrompt?: string;\n}\n\nexport interface ContinuationResult<T> {\n  data: T;\n  raw: string;\n  continuationCount: number;\n  warnings: string[];\n  wasSalvaged: boolean;\n}\n\n// ============================================================================\n// Constants\n// ============================================================================\n\nconst DEFAULT_MAX_TOKENS = 8192;\nconst DEFAULT_MAX_CONTINUATIONS = 3;\n\n/**\n * Default continuation system prompt.\n * Used when no custom continuationSystemPrompt is provided.\n */\nconst DEFAULT_CONTINUATION_SYSTEM_PROMPT = `You are a JSON continuation assistant. Your ONLY job is to continue generating JSON from where the previous response was truncated.\n\nRules:\n1. Continue from EXACTLY where the previous output stopped\n2. Do NOT repeat any content already generated\n3. Complete the JSON structure properly with all closing brackets\n4. Do NOT wrap in markdown code blocks\n5. Output ONLY the continuation JSON, nothing else`;\n\n// ============================================================================\n// Helper Functions\n// ============================================================================\n\n/**\n * Merge a previous partial response with a continuation.\n *\n * Handles JSON structure merging, removing markdown code blocks,\n * and ensuring proper comma separation between merged parts.\n *\n * @param {string} previous - The previous partial response\n * @param {string} continuation - The continuation to merge\n * @returns {string} The merged response\n */\nexport function mergeResponses(\n  previous: string,\n  continuation: string,\n): string {\n  const trimmedPrev = previous.trimEnd();\n  const trimmedCont = continuation.trimStart();\n\n  let cleanedCont = trimmedCont\n    .replace(/^```json?\\s*/i, '')\n    .replace(/```\\s*$/i, '')\n    .trim();\n\n  if (cleanedCont.startsWith('{')) {\n    try {\n      const contParsed = JSON.parse(autoCloseJson(cleanedCont));\n      const keys = Object.keys(contParsed);\n      if (keys.length === 1 && Array.isArray(contParsed[keys[0]])) {\n        cleanedCont = contParsed[keys[0]]\n          .map((item: unknown) => JSON.stringify(item))\n          .join(',\\n');\n      }\n    } catch {\n      // Continue with original cleaning\n    }\n  }\n\n  if (cleanedCont.startsWith('}') || cleanedCont.startsWith(']')) {\n    return trimmedPrev + cleanedCont;\n  }\n\n  const prevEndsWithValue = /[\\}\\]\\\"\\d]$/.test(trimmedPrev);\n  const contStartsWithValue = /^[\\{\\[\\\"]/.test(cleanedCont);\n\n  if (prevEndsWithValue && contStartsWithValue) {\n    return trimmedPrev + ',\\n' + cleanedCont;\n  }\n\n  return trimmedPrev + cleanedCont;\n}\n\nexport function salvagePartialResponse<T>(rawResponse: string): T | null {\n  console.warn('[Continuation] Attempting to salvage partial response');\n\n  try {\n    const cleanedResponse = extractJsonFromText(rawResponse) || rawResponse;\n    const closed = autoCloseJson(cleanedResponse);\n    const parsed = JSON.parse(closed) as T;\n    console.log('[Continuation] Successfully salvaged partial response');\n    return parsed;\n  } catch (error) {\n    console.error('[Continuation] Could not salvage response:', error);\n  }\n\n  return null;\n}\n\n// ============================================================================\n// Main Function\n// ============================================================================\n\n/**\n * Call an LLM with automatic continuation handling.\n *\n * Manages token limits by detecting truncation and making continuation\n * calls until the complete response is received or max continuations reached.\n *\n * @template T - Expected response type\n * @param {ContinuationOptions<T>} options - Continuation call options\n * @returns {Promise<ContinuationResult<T>>} Result with data and continuation metadata\n */\nexport async function callWithContinuation<T>(\n  options: ContinuationOptions<T>,\n): Promise<ContinuationResult<T>> {\n  const {\n    client,\n    systemPrompt,\n    userPrompt,\n    schema,\n    maxTokens = DEFAULT_MAX_TOKENS,\n    maxContinuations = DEFAULT_MAX_CONTINUATIONS,\n    buildContinuationPrompt,\n    continuationSystemPrompt = DEFAULT_CONTINUATION_SYSTEM_PROMPT,\n  } = options;\n\n  let rawResponse = '';\n  let continuationCount = 0;\n  const warnings: string[] = [];\n  let wasSalvaged = false;\n\n  console.log('[Continuation] Starting LLM call with continuation support');\n  console.log(\n    `[Continuation] Max tokens: ${maxTokens}, Max continuations: ${maxContinuations}`,\n  );\n\n  try {\n    const response = await client.callRawWithMetadata({\n      systemPrompt,\n      userPrompt,\n      maxTokens,\n    });\n\n    rawResponse = extractJsonFromText(response.raw) || response.raw;\n\n    console.log(\n      `[Continuation] Initial response: ${rawResponse.length} chars, finish_reason: ${response.finishReason}`,\n    );\n\n    let truncation = detectTruncation(rawResponse, response.finishReason);\n\n    while (truncation.isTruncated && continuationCount < maxContinuations) {\n      continuationCount++;\n      const warningMsg = `Response truncated (${truncation.reason}), continuing (attempt ${continuationCount}/${maxContinuations})`;\n      console.log(`[Continuation] ${warningMsg}`);\n      warnings.push(warningMsg);\n\n      const contPrompt = buildContinuationPrompt(\n        rawResponse,\n        continuationCount,\n      );\n\n      const contResponse = await client.callRawWithMetadata({\n        systemPrompt: continuationSystemPrompt,\n        userPrompt: contPrompt,\n        maxTokens,\n      });\n\n      console.log(\n        `[Continuation] Continuation response: ${contResponse.raw.length} chars, finish_reason: ${contResponse.finishReason}`,\n      );\n\n      const cleanedContResponse =\n        extractJsonFromText(contResponse.raw) || contResponse.raw;\n      rawResponse = mergeResponses(rawResponse, cleanedContResponse);\n\n      truncation = detectTruncation(rawResponse, contResponse.finishReason);\n    }\n\n    if (\n      continuationCount >= maxContinuations &&\n      truncation.isTruncated\n    ) {\n      console.warn(\n        `[Continuation] Reached max continuations (${maxContinuations}), attempting to salvage...`,\n      );\n      warnings.push(\n        `Reached max continuations - some content may be incomplete`,\n      );\n      wasSalvaged = true;\n    }\n\n    const cleanedResponse =\n      extractJsonFromText(rawResponse) || rawResponse;\n    let data: T;\n\n    try {\n      if (isValidJson(cleanedResponse)) {\n        data = JSON.parse(cleanedResponse) as T;\n      } else {\n        const closed = autoCloseJson(cleanedResponse);\n        data = JSON.parse(closed) as T;\n        if (!wasSalvaged) {\n          warnings.push('Response required auto-closing of JSON brackets');\n        }\n      }\n    } catch (parseError) {\n      const salvaged = salvagePartialResponse<T>(cleanedResponse);\n      if (salvaged) {\n        data = salvaged;\n        wasSalvaged = true;\n        warnings.push('Response was salvaged from partial data');\n      } else {\n        throw new Error(\n          `Failed to parse response after ${continuationCount} continuations: ${parseError}`,\n        );\n      }\n    }\n\n    if (schema) {\n      try {\n        data = schema.parse(data);\n      } catch (validationError) {\n        console.warn(\n          '[Continuation] Schema validation failed:',\n          validationError,\n        );\n        warnings.push(`Schema validation issue: ${validationError}`);\n      }\n    }\n\n    console.log(\n      `[Continuation] Complete. Continuations: ${continuationCount}, Warnings: ${warnings.length}`,\n    );\n\n    return {\n      data,\n      raw: rawResponse,\n      continuationCount,\n      warnings,\n      wasSalvaged,\n    };\n  } catch (error) {\n    console.error('[Continuation] Error during LLM call:', error);\n    throw error;\n  }\n}\n\n/**\n * Build a generic continuation prompt.\n *\n * Creates a prompt that instructs the LLM to continue from where\n * a previous truncated response left off.\n *\n * @param {string} context - Original context/prompt\n * @param {string} partialResponse - The partial response generated so far\n * @param {number} attempt - Current continuation attempt number\n * @param {number} [maxAttempts] - Maximum allowed continuation attempts\n * @returns {string} The formatted continuation prompt\n */\nexport function buildGenericContinuationPrompt(\n  context: string,\n  partialResponse: string,\n  attempt: number,\n  maxAttempts: number = DEFAULT_MAX_CONTINUATIONS,\n): string {\n  return `## CONTINUATION REQUEST (Attempt ${attempt}/${maxAttempts})\n\nYour previous response was truncated. Continue generating from where you left off.\n\n### ORIGINAL CONTEXT\n${context}\n\n### WHAT YOU GENERATED SO FAR\n\\`\\`\\`json\n${partialResponse}\n\\`\\`\\`\n\n### INSTRUCTIONS\n1. Continue from EXACTLY where the response was cut off\n2. Do NOT repeat any content already generated\n3. Complete the JSON structure properly\n4. Do NOT wrap your response in markdown code blocks\n\nContinue generating now:`;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAwDA,IAAM,qBAAqB;AAC3B,IAAM,yBAAyB,CAAC,KAAK,GAAI;AAGzC,IAAM,oBACJ;AAEF,IAAM,iBAAoD;AAAA,EACxD,QAAQ;AAAA,EACR,YAAY;AACd;AAEA,IAAM,oBAAuD;AAAA,EAC3D,QAAQ;AAAA,EACR,YAAY;AACd;AAEA,IAAM,mBAAsD;AAAA,EAC1D,QAAQ;AAAA,EACR,YAAY;AACd;AAEO,IAAM,kBAAN,MAAsB;AAAA,EAM3B,YAAY,MAA8B;AACxC,SAAK,WAAW,KAAK;AACrB,SAAK,QAAQ,KAAK,SAAS,eAAe,KAAK,QAAQ;AACvD,SAAK,WAAW,KAAK,WAAW,kBAAkB,KAAK,QAAQ,GAAG,QAAQ,OAAO,EAAE;AAEnF,UAAM,SAAS,KAAK,UAAU,KAAK,cAAc;AACjD,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR,0CAA0C,KAAK,QAAQ,oBAC9C,iBAAiB,KAAK,QAAQ,CAAC;AAAA,MAC1C;AAAA,IACF;AACA,SAAK,SAAS;AAAA,EAChB;AAAA;AAAA,EAGA,MAAM,MAAM,MAAwC;AAClD,UAAM,QAAQ,MAAM,KAAK,WAAW,CAAC,IAAI,CAAC;AAC1C,UAAM,YAAY,MAAM,WAAW,CAAC;AACpC,QAAI,CAAC,WAAW;AACd,YAAM,IAAI,MAAM,uDAAuD;AAAA,IACzE;AACA,WAAO,EAAE,WAAW,OAAO,MAAM,MAAM;AAAA,EACzC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,WAAW,OAA6D;AAC5E,QAAI,MAAM,WAAW,GAAG;AACtB,aAAO,EAAE,YAAY,CAAC,GAAG,OAAO,EAAE,cAAc,GAAG,aAAa,EAAE,EAAE;AAAA,IACtE;AAOA,QAAI,UAAU;AACd,QAAI,YAAmB,IAAI,MAAM,8BAA8B;AAC/D,aAAS,UAAU,GAAG,WAAW,oBAAoB,WAAW;AAC9D,YAAM,UAAU,MAAM,IAAI,CAAC,MAAO,EAAE,SAAS,UAAU,EAAE,MAAM,GAAG,OAAO,IAAI,CAAE;AAC/E,UAAI;AACF,eAAO,MAAM,KAAK,eAAe,MAAM,QAAQ,OAAO;AAAA,MACxD,SAAS,GAAG;AACV,oBAAY,aAAa,QAAQ,IAAI,IAAI,MAAM,OAAO,CAAC,CAAC;AACxD,YAAI,kBAAkB,KAAK,UAAU,OAAO,KAAK,UAAU,KAAK;AAC9D,oBAAU,KAAK,MAAM,UAAU,CAAC;AAChC;AAAA,QACF;AACA,cAAM,UAAU,uBAAuB,UAAU,CAAC;AAClD,YAAI,YAAY,OAAW;AAC3B,cAAM,IAAI,QAAQ,CAAC,MAAM,WAAW,GAAG,OAAO,CAAC;AAAA,MACjD;AAAA,IACF;AACA,UAAM;AAAA,EACR;AAAA,EAEA,MAAc,eACZ,YACA,SAC+B;AAC/B,UAAM,WAAW,MAAM,MAAM,GAAG,KAAK,OAAO,eAAe;AAAA,MACzD,QAAQ;AAAA,MACR,SAAS;AAAA,QACP,gBAAgB;AAAA,QAChB,eAAe,UAAU,KAAK,MAAM;AAAA,MACtC;AAAA,MACA,MAAM,KAAK,UAAU,EAAE,OAAO,KAAK,OAAO,OAAO,QAAQ,CAAC;AAAA,IAC5D,CAAC;AACD,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,YAAY,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,WAAW;AAC/D,YAAM,IAAI;AAAA,QACR,oBAAoB,KAAK,QAAQ,aAAa,SAAS,MAAM,IAAI,SAAS,UAAU,KAAK,SAAS;AAAA,MACpG;AAAA,IACF;AACA,UAAM,UAAU,MAAM,SAAS,KAAK;AACpC,QAAI;AACJ,QAAI;AACF,aAAO,KAAK,MAAM,OAAO;AAAA,IAC3B,QAAQ;AACN,YAAM,IAAI;AAAA,QACR,oBAAoB,KAAK,QAAQ,+BAA+B,OAAO;AAAA,MACzE;AAAA,IACF;AACA,QAAI,CAAC,MAAM,QAAQ,KAAK,IAAI,GAAG;AAI7B,YAAM,SAAS,KAAK,OAAO,WAAW;AACtC,YAAM,IAAI;AAAA,QACR,oBAAoB,KAAK,QAAQ,oCAAoC,MAAM;AAAA,MAC7E;AAAA,IACF;AACA,UAAM,UAAU,CAAC,GAAG,KAAK,IAAI,EAAE,KAAK,CAAC,GAAG,MAAM,EAAE,QAAQ,EAAE,KAAK;AAC/D,QAAI,QAAQ,WAAW,YAAY;AACjC,YAAM,IAAI;AAAA,QACR,sCAAsC,QAAQ,MAAM,mBAAmB,UAAU;AAAA,MACnF;AAAA,IACF;AACA,WAAO;AAAA,MACL,YAAY,QAAQ,IAAI,CAAC,MAAM,EAAE,SAAS;AAAA,MAC1C,OAAO;AAAA,QACL,cAAc,KAAK,OAAO,iBAAiB;AAAA,QAC3C,aAAa,KAAK,OAAO,gBAAgB;AAAA,MAC3C;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,gBAAoC;AAC1C,WAAO,QAAQ,IAAI,iBAAiB,KAAK,QAAQ,CAAC;AAAA,EACpD;AACF;AAWO,SAAS,iBAAiB,GAA0B,GAAkC;AAC3F,MAAI,EAAE,WAAW,EAAE,QAAQ;AACzB,UAAM,IAAI,MAAM,6CAA6C,EAAE,MAAM,OAAO,EAAE,MAAM,GAAG;AAAA,EACzF;AACA,MAAI,MAAM;AACV,MAAI,QAAQ;AACZ,MAAI,QAAQ;AACZ,WAAS,IAAI,GAAG,IAAI,EAAE,QAAQ,KAAK;AACjC,UAAM,KAAK,EAAE,CAAC;AACd,UAAM,KAAK,EAAE,CAAC;AACd,WAAO,KAAK;AACZ,aAAS,KAAK;AACd,aAAS,KAAK;AAAA,EAChB;AACA,MAAI,UAAU,KAAK,UAAU,EAAG,QAAO;AACvC,SAAO,OAAO,KAAK,KAAK,KAAK,IAAI,KAAK,KAAK,KAAK;AAClD;;;ACpLO,SAAS,iBACd,UACA,cACkB;AAClB,MAAI,iBAAiB,UAAU;AAC7B,UAAMA,eAAc,cAAc,QAAQ;AAC1C,WAAO;AAAA,MACL,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,gBAAgB;AAAA,MAChB,qBAAqB,wBAAwB,QAAQ;AAAA,MACrD,sBAAsBA,aAAY;AAAA,MAClC,oBAAoBA,aAAY;AAAA,IAClC;AAAA,EACF;AAEA,MAAI;AACF,SAAK,MAAM,QAAQ;AACnB,WAAO,EAAE,aAAa,OAAO,QAAQ,OAAO;AAAA,EAC9C,QAAQ;AAAA,EAER;AAEA,MAAI,iBAAiB,UAAU,iBAAiB,MAAM;AACpD,UAAM,UAAU,SAAS,KAAK;AAE9B,UAAM,eACJ,QAAQ,SAAS,GAAG,KACpB,QAAQ,SAAS,GAAG,KACpB,QAAQ,SAAS,KAAK,KACtB,QAAQ,KAAK,OAAO,KACpB,QAAQ,KAAK,OAAO;AAEtB,QAAI,cAAc;AAChB,YAAMA,eAAc,cAAc,QAAQ;AAC1C,aAAO;AAAA,QACL,aAAa;AAAA,QACb,QAAQ;AAAA,QACR,gBAAgB;AAAA,QAChB,qBAAqB,wBAAwB,QAAQ;AAAA,QACrD,sBAAsBA,aAAY;AAAA,QAClC,oBAAoBA,aAAY;AAAA,MAClC;AAAA,IACF;AAEA,QAAI;AACF,YAAM,SAAS,cAAc,OAAO;AACpC,WAAK,MAAM,MAAM;AACjB,aAAO,EAAE,aAAa,OAAO,QAAQ,OAAO;AAAA,IAC9C,QAAQ;AACN,aAAO,EAAE,aAAa,OAAO,QAAQ,OAAO;AAAA,IAC9C;AAAA,EACF;AAEA,QAAM,cAAc,cAAc,QAAQ;AAC1C,MACE,YAAY,uBAAuB,KACnC,YAAY,qBAAqB,GACjC;AACA,WAAO;AAAA,MACL,aAAa;AAAA,MACb,QAAQ;AAAA,MACR,gBAAgB;AAAA,MAChB,qBAAqB,wBAAwB,QAAQ;AAAA,MACrD,sBAAsB,YAAY;AAAA,MAClC,oBAAoB,YAAY;AAAA,IAClC;AAAA,EACF;AAEA,SAAO,EAAE,aAAa,OAAO,QAAQ,OAAO;AAC9C;AAMA,SAAS,cAAc,MAOrB;AACA,MAAI,WAAW;AACf,MAAI,UAAU;AACd,MAAI,eAAe;AACnB,MAAI,gBAAgB;AACpB,MAAI,aAAa;AACjB,MAAI,cAAc;AAElB,aAAW,QAAQ,MAAM;AACvB,QAAI,SAAS;AACX,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,QAAQ,UAAU;AAC7B,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,KAAK;AAChB,iBAAW,CAAC;AACZ;AAAA,IACF;AACA,QAAI,SAAU;AAEd,YAAQ,MAAM;AAAA,MACZ,KAAK;AACH;AACA;AAAA,MACF,KAAK;AACH;AACA;AAAA,MACF,KAAK;AACH;AACA;AAAA,MACF,KAAK;AACH;AACA;AAAA,IACJ;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,sBAAsB,KAAK,IAAI,GAAG,eAAe,aAAa;AAAA,IAC9D,oBAAoB,KAAK,IAAI,GAAG,aAAa,WAAW;AAAA,EAC1D;AACF;AAEO,SAAS,wBAAwB,MAA8B;AACpE,QAAM,aAAa,cAAc,IAAI;AACrC,MAAI;AACF,WAAO,KAAK,MAAM,UAAU;AAAA,EAC9B,QAAQ;AAAA,EAER;AAEA,QAAM,UAAU,KAAK,KAAK;AAE1B,MAAI,QAAQ,WAAW,GAAG,GAAG;AAC3B,UAAM,oBAAoB,6BAA6B,OAAO;AAC9D,QAAI,oBAAoB,GAAG;AACzB,YAAM,SAAS,QAAQ,UAAU,GAAG,iBAAiB,IAAI;AACzD,UAAI;AACF,eAAO,KAAK,MAAM,MAAM;AAAA,MAC1B,QAAQ;AAAA,MAER;AAAA,IACF;AAAA,EACF;AAEA,MAAI,QAAQ,WAAW,GAAG,GAAG;AAC3B,UAAM,SAAS,cAAc,OAAO;AACpC,QAAI;AACF,aAAO,KAAK,MAAM,MAAM;AAAA,IAC1B,QAAQ;AACN,YAAM,oBAAoB,+BAA+B,OAAO;AAChE,UAAI,oBAAoB,GAAG;AACzB,cAAM,SAAS,QAAQ,UAAU,GAAG,iBAAiB,IAAI;AACzD,YAAI;AACF,iBAAO,KAAK,MAAM,MAAM;AAAA,QAC1B,QAAQ;AAAA,QAER;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;AAEA,SAAS,6BAA6B,MAAsB;AAC1D,MAAI,QAAQ;AACZ,MAAI,WAAW;AACf,MAAI,UAAU;AACd,MAAI,yBAAyB;AAE7B,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,OAAO,KAAK,CAAC;AAEnB,QAAI,SAAS;AACX,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,QAAQ,UAAU;AAC7B,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,KAAK;AAChB,iBAAW,CAAC;AACZ;AAAA,IACF;AACA,QAAI,SAAU;AAEd,QAAI,SAAS,OAAO,SAAS,KAAK;AAChC;AAAA,IACF,WAAW,SAAS,OAAO,SAAS,KAAK;AACvC;AACA,UAAI,UAAU,GAAG;AACf,iCAAyB,IAAI;AAAA,MAC/B;AAAA,IACF,WAAW,SAAS,OAAO,UAAU,GAAG;AACtC,+BAAyB;AAAA,IAC3B;AAAA,EACF;AAEA,SAAO,yBAAyB,IAAI,yBAAyB;AAC/D;AAEA,SAAS,+BAA+B,MAAsB;AAC5D,MAAI,QAAQ;AACZ,MAAI,WAAW;AACf,MAAI,UAAU;AACd,MAAI,iBAAiB;AAErB,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,OAAO,KAAK,CAAC;AAEnB,QAAI,SAAS;AACX,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,QAAQ,UAAU;AAC7B,gBAAU;AACV;AAAA,IACF;AACA,QAAI,SAAS,KAAK;AAChB,iBAAW,CAAC;AACZ;AAAA,IACF;AACA,QAAI,SAAU;AAEd,QAAI,SAAS,OAAO,SAAS,KAAK;AAChC;AAAA,IACF,WAAW,SAAS,OAAO,SAAS,KAAK;AACvC;AAAA,IACF,WAAW,SAAS,OAAO,UAAU,GAAG;AACtC,uBAAiB;AAAA,IACnB;AAAA,EACF;AAEA,SAAO,iBAAiB,IAAI,iBAAiB;AAC/C;AAEO,SAAS,kBAAkB,SAA0B;AAC1D,QAAM,UAAU,QAAQ,KAAK;AAC7B,MAAI,CAAC,QAAS,QAAO;AAErB,QAAM,WAAW,cAAc,OAAO;AACtC,MACE,SAAS,uBAAuB,KAChC,SAAS,qBAAqB,GAC9B;AACA,WAAO;AAAA,EACT;AAEA,QAAM,gBAAgB;AAAA,IACpB;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AAEA,aAAW,WAAW,eAAe;AACnC,QAAI,QAAQ,KAAK,OAAO,EAAG,QAAO;AAAA,EACpC;AAEA,SAAO;AACT;;;AC7QA,IAAM,qBAAqB;AAC3B,IAAM,4BAA4B;AAMlC,IAAM,qCAAqC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAuBpC,SAAS,eACd,UACA,cACQ;AACR,QAAM,cAAc,SAAS,QAAQ;AACrC,QAAM,cAAc,aAAa,UAAU;AAE3C,MAAI,cAAc,YACf,QAAQ,iBAAiB,EAAE,EAC3B,QAAQ,YAAY,EAAE,EACtB,KAAK;AAER,MAAI,YAAY,WAAW,GAAG,GAAG;AAC/B,QAAI;AACF,YAAM,aAAa,KAAK,MAAM,cAAc,WAAW,CAAC;AACxD,YAAM,OAAO,OAAO,KAAK,UAAU;AACnC,UAAI,KAAK,WAAW,KAAK,MAAM,QAAQ,WAAW,KAAK,CAAC,CAAC,CAAC,GAAG;AAC3D,sBAAc,WAAW,KAAK,CAAC,CAAC,EAC7B,IAAI,CAAC,SAAkB,KAAK,UAAU,IAAI,CAAC,EAC3C,KAAK,KAAK;AAAA,MACf;AAAA,IACF,QAAQ;AAAA,IAER;AAAA,EACF;AAEA,MAAI,YAAY,WAAW,GAAG,KAAK,YAAY,WAAW,GAAG,GAAG;AAC9D,WAAO,cAAc;AAAA,EACvB;AAEA,QAAM,oBAAoB,cAAc,KAAK,WAAW;AACxD,QAAM,sBAAsB,YAAY,KAAK,WAAW;AAExD,MAAI,qBAAqB,qBAAqB;AAC5C,WAAO,cAAc,QAAQ;AAAA,EAC/B;AAEA,SAAO,cAAc;AACvB;AAEO,SAAS,uBAA0B,aAA+B;AACvE,UAAQ,KAAK,uDAAuD;AAEpE,MAAI;AACF,UAAM,kBAAkB,oBAAoB,WAAW,KAAK;AAC5D,UAAM,SAAS,cAAc,eAAe;AAC5C,UAAM,SAAS,KAAK,MAAM,MAAM;AAChC,YAAQ,IAAI,uDAAuD;AACnE,WAAO;AAAA,EACT,SAAS,OAAO;AACd,YAAQ,MAAM,8CAA8C,KAAK;AAAA,EACnE;AAEA,SAAO;AACT;AAgBA,eAAsB,qBACpB,SACgC;AAChC,QAAM;AAAA,IACJ;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,YAAY;AAAA,IACZ,mBAAmB;AAAA,IACnB;AAAA,IACA,2BAA2B;AAAA,EAC7B,IAAI;AAEJ,MAAI,cAAc;AAClB,MAAI,oBAAoB;AACxB,QAAM,WAAqB,CAAC;AAC5B,MAAI,cAAc;AAElB,UAAQ,IAAI,4DAA4D;AACxE,UAAQ;AAAA,IACN,8BAA8B,SAAS,wBAAwB,gBAAgB;AAAA,EACjF;AAEA,MAAI;AACF,UAAM,WAAW,MAAM,OAAO,oBAAoB;AAAA,MAChD;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAED,kBAAc,oBAAoB,SAAS,GAAG,KAAK,SAAS;AAE5D,YAAQ;AAAA,MACN,oCAAoC,YAAY,MAAM,0BAA0B,SAAS,YAAY;AAAA,IACvG;AAEA,QAAI,aAAa,iBAAiB,aAAa,SAAS,YAAY;AAEpE,WAAO,WAAW,eAAe,oBAAoB,kBAAkB;AACrE;AACA,YAAM,aAAa,uBAAuB,WAAW,MAAM,0BAA0B,iBAAiB,IAAI,gBAAgB;AAC1H,cAAQ,IAAI,kBAAkB,UAAU,EAAE;AAC1C,eAAS,KAAK,UAAU;AAExB,YAAM,aAAa;AAAA,QACjB;AAAA,QACA;AAAA,MACF;AAEA,YAAM,eAAe,MAAM,OAAO,oBAAoB;AAAA,QACpD,cAAc;AAAA,QACd,YAAY;AAAA,QACZ;AAAA,MACF,CAAC;AAED,cAAQ;AAAA,QACN,yCAAyC,aAAa,IAAI,MAAM,0BAA0B,aAAa,YAAY;AAAA,MACrH;AAEA,YAAM,sBACJ,oBAAoB,aAAa,GAAG,KAAK,aAAa;AACxD,oBAAc,eAAe,aAAa,mBAAmB;AAE7D,mBAAa,iBAAiB,aAAa,aAAa,YAAY;AAAA,IACtE;AAEA,QACE,qBAAqB,oBACrB,WAAW,aACX;AACA,cAAQ;AAAA,QACN,6CAA6C,gBAAgB;AAAA,MAC/D;AACA,eAAS;AAAA,QACP;AAAA,MACF;AACA,oBAAc;AAAA,IAChB;AAEA,UAAM,kBACJ,oBAAoB,WAAW,KAAK;AACtC,QAAI;AAEJ,QAAI;AACF,UAAI,YAAY,eAAe,GAAG;AAChC,eAAO,KAAK,MAAM,eAAe;AAAA,MACnC,OAAO;AACL,cAAM,SAAS,cAAc,eAAe;AAC5C,eAAO,KAAK,MAAM,MAAM;AACxB,YAAI,CAAC,aAAa;AAChB,mBAAS,KAAK,iDAAiD;AAAA,QACjE;AAAA,MACF;AAAA,IACF,SAAS,YAAY;AACnB,YAAM,WAAW,uBAA0B,eAAe;AAC1D,UAAI,UAAU;AACZ,eAAO;AACP,sBAAc;AACd,iBAAS,KAAK,yCAAyC;AAAA,MACzD,OAAO;AACL,cAAM,IAAI;AAAA,UACR,kCAAkC,iBAAiB,mBAAmB,UAAU;AAAA,QAClF;AAAA,MACF;AAAA,IACF;AAEA,QAAI,QAAQ;AACV,UAAI;AACF,eAAO,OAAO,MAAM,IAAI;AAAA,MAC1B,SAAS,iBAAiB;AACxB,gBAAQ;AAAA,UACN;AAAA,UACA;AAAA,QACF;AACA,iBAAS,KAAK,4BAA4B,eAAe,EAAE;AAAA,MAC7D;AAAA,IACF;AAEA,YAAQ;AAAA,MACN,2CAA2C,iBAAiB,eAAe,SAAS,MAAM;AAAA,IAC5F;AAEA,WAAO;AAAA,MACL;AAAA,MACA,KAAK;AAAA,MACL;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF,SAAS,OAAO;AACd,YAAQ,MAAM,yCAAyC,KAAK;AAC5D,UAAM;AAAA,EACR;AACF;AAcO,SAAS,+BACd,SACA,iBACA,SACA,cAAsB,2BACd;AACR,SAAO,oCAAoC,OAAO,IAAI,WAAW;AAAA;AAAA;AAAA;AAAA;AAAA,EAKjE,OAAO;AAAA;AAAA;AAAA;AAAA,EAIP,eAAe;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAUjB;","names":["bracketInfo"]}

package/dist/{rate-limiter-7-oOHrtX.d.ts → rate-limiter-C12cctao.d.ts}

@@ -24,8 +24,15 @@ declare class TokenTracker {
     private usage;
     /** Sum of provider-reported authoritative costs (e.g. OpenRouter `usage.cost`). */
     private authoritativeCostUSD;
-    /** Token buckets for calls WITHOUT an authoritative cost — priced cache-aware. */
-    private computed;
+    /** Token buckets for calls WITHOUT an authoritative cost — priced cache-aware.
+     *  Keyed by the model that MADE the calls: `setModel` flips the tracker
+     *  between models mid-process (coordinator on pro, subagents on flash), and
+     *  pricing the aggregate at whichever model is current when someone reads
+     *  the estimate mispriced the whole history by up to the models' rate ratio
+     *  (snapshot/diff consumers saw phantom multi-dollar deltas — or $0 after a
+     *  negative-clamp — per battery spec). Per-model buckets keep the
+     *  retroactive late-pricing property without cross-model contamination. */
+    private computedByModel;
     constructor(model?: string);
     /** Cache-aware cost for one (or an aggregate of) call(s), in USD. */
     private costFor;

package/dist/structured-output.d.ts

@@ -1,4 +1,4 @@
-import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-7-oOHrtX.js';
+import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-C12cctao.js';
 import { z } from 'zod';
 
 /**

package/dist/structured-output.js

@@ -4,8 +4,8 @@ import {
   getStructuredOutputClient,
   isStructuredOutputAvailable,
   resetStructuredOutputClient
-} from "./chunk-IDXSWM57.js";
-import "./chunk-SXSP6M24.js";
+} from "./chunk-S6KQIMQQ.js";
+import "./chunk-CZX533SU.js";
 export {
   STRUCTURED_OUTPUT_MODELS,
   StructuredOutputClient,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@almadar/llm",
-  "version": "2.31.0",
+  "version": "2.33.0",
   "description": "Multi-provider LLM client with rate limiting, token tracking, structured outputs, and continuation handling",
   "type": "module",
   "main": "./dist/index.js",
@@ -38,14 +38,14 @@
     "@langchain/openai": "^1.2.13",
     "openai": "^6.18.0",
     "zod": "^3.22.0",
-    "@almadar/logger": "^1.3.0"
+    "@almadar/logger": "^1.9.0"
   },
   "peerDependencies": {
-    "@almadar/core": "^10.11.0"
+    "@almadar/core": "^10.29.0"
   },
   "devDependencies": {
-    "@almadar/core": "^10.11.0",
-    "@almadar/eslint-plugin": ">=2.3.0",
+    "@almadar/core": "^10.29.0",
+    "@almadar/eslint-plugin": "^2.14.0",
     "@types/node": "^22.0.0",
     "@typescript-eslint/parser": "8.56.0",
     "eslint": "10.0.0",

package/src/embedding-client.ts

@@ -49,8 +49,18 @@ export interface EmbeddingBatchResult {
 interface OpenAIEmbeddingApiResponse {
   data: Array<{ embedding: number[]; index: number }>;
   usage?: { prompt_tokens?: number; total_tokens?: number };
+  /** OpenRouter reports provider hiccups as HTTP 200 with an error envelope and no `data`. */
+  error?: { message?: string; code?: number | string };
 }
 
+/** Transient-failure retry policy: attempts and backoff (ms) between them. */
+const EMBED_MAX_ATTEMPTS = 3;
+const EMBED_RETRY_BACKOFF_MS = [500, 2000];
+
+/** Provider rejects an input that exceeds the model's context window (bge: 512 tokens). */
+const TOKEN_OVERFLOW_RE =
+  /context length|maximum input length|reduce the length of the input|input_tokens/i;
+
 const DEFAULT_MODELS: Record<EmbeddingProvider, string> = {
   openai: 'text-embedding-3-small',
   openrouter: 'baai/bge-base-en-v1.5',
@@ -106,11 +116,36 @@ export class EmbeddingClient {
     if (texts.length === 0) {
       return { embeddings: [], usage: { promptTokens: 0, totalTokens: 0 } };
     }
-    // bge-base-en-v1.5 has a 512-token context limit; trim each input so
-    // providers (notably OpenRouter) don't return a malformed response body.
-    // 2000 chars ≈ 500 tokens for typical English text — same cap used by
-    // the runtime's tryEmbedRequest helper.
-    const trimmed = texts.map((t) => (t.length > 2000 ? t.slice(0, 2000) : t));
+    // bge-base-en-v1.5 has a 512-token context limit. Chars-per-token varies
+    // with density (~4 for typical English, ~2.5-3 for dense technical/code
+    // text), so no fixed char cap can guarantee a fit. Start at 2000 chars and,
+    // when the provider rejects an input for exceeding the context window,
+    // halve the cap and retry — the provider's own token count is the source
+    // of truth, not a chars-per-token guess.
+    let charCap = 2000;
+    let lastError: Error = new Error('EmbeddingClient: unreachable');
+    for (let attempt = 1; attempt <= EMBED_MAX_ATTEMPTS; attempt++) {
+      const trimmed = texts.map((t) => (t.length > charCap ? t.slice(0, charCap) : t));
+      try {
+        return await this.embedBatchOnce(texts.length, trimmed);
+      } catch (e) {
+        lastError = e instanceof Error ? e : new Error(String(e));
+        if (TOKEN_OVERFLOW_RE.test(lastError.message) && charCap > 500) {
+          charCap = Math.floor(charCap / 2);
+          continue;
+        }
+        const backoff = EMBED_RETRY_BACKOFF_MS[attempt - 1];
+        if (backoff === undefined) break;
+        await new Promise((r) => setTimeout(r, backoff));
+      }
+    }
+    throw lastError;
+  }
+
+  private async embedBatchOnce(
+    inputCount: number,
+    trimmed: ReadonlyArray<string>,
+  ): Promise<EmbeddingBatchResult> {
     const response = await fetch(`${this.baseUrl}/embeddings`, {
       method: 'POST',
       headers: {
@@ -125,14 +160,28 @@ export class EmbeddingClient {
         `EmbeddingClient: ${this.provider} returned ${response.status} ${response.statusText}: ${errorText}`,
       );
     }
-    const body = (await response.json()) as OpenAIEmbeddingApiResponse;
+    const rawBody = await response.text();
+    let body: OpenAIEmbeddingApiResponse;
+    try {
+      body = JSON.parse(rawBody) as OpenAIEmbeddingApiResponse;
+    } catch {
+      throw new Error(
+        `EmbeddingClient: ${this.provider} returned unparseable body: ${rawBody}`,
+      );
+    }
     if (!Array.isArray(body.data)) {
-      throw new Error('EmbeddingClient: provider response missing `data` array');
+      // OpenRouter reports provider hiccups as 200 + `{error}` with no data —
+      // surface the real body instead of a blind "missing data" (G9: this
+      // error was undiagnosable for weeks because the body was discarded).
+      const detail = body.error?.message ?? rawBody;
+      throw new Error(
+        `EmbeddingClient: ${this.provider} response has no \`data\` array: ${detail}`,
+      );
     }
     const ordered = [...body.data].sort((a, b) => a.index - b.index);
-    if (ordered.length !== texts.length) {
+    if (ordered.length !== inputCount) {
       throw new Error(
-        `EmbeddingClient: provider returned ${ordered.length} embeddings for ${texts.length} inputs`,
+        `EmbeddingClient: provider returned ${ordered.length} embeddings for ${inputCount} inputs`,
       );
     }
     return {

package/src/token-tracker.ts

@@ -201,16 +201,26 @@ export class TokenTracker {
 
   /** Sum of provider-reported authoritative costs (e.g. OpenRouter `usage.cost`). */
   private authoritativeCostUSD = 0;
-  /** Token buckets for calls WITHOUT an authoritative cost — priced cache-aware. */
-  private computed = { promptTokens: 0, completionTokens: 0, cachedPromptTokens: 0, cacheWriteTokens: 0 };
+  /** Token buckets for calls WITHOUT an authoritative cost — priced cache-aware.
+   *  Keyed by the model that MADE the calls: `setModel` flips the tracker
+   *  between models mid-process (coordinator on pro, subagents on flash), and
+   *  pricing the aggregate at whichever model is current when someone reads
+   *  the estimate mispriced the whole history by up to the models' rate ratio
+   *  (snapshot/diff consumers saw phantom multi-dollar deltas — or $0 after a
+   *  negative-clamp — per battery spec). Per-model buckets keep the
+   *  retroactive late-pricing property without cross-model contamination. */
+  private computedByModel = new Map<
+    string,
+    { promptTokens: number; completionTokens: number; cachedPromptTokens: number; cacheWriteTokens: number }
+  >();
 
   constructor(model: string = 'claude-sonnet-4-5-20250929') {
     this.model = model;
   }
 
   /** Cache-aware cost for one (or an aggregate of) call(s), in USD. */
-  private costFor(promptTokens: number, completionTokens: number, cached: number, written: number): number {
-    const costs = getCostForModel(this.model);
+  private costFor(model: string, promptTokens: number, completionTokens: number, cached: number, written: number): number {
+    const costs = getCostForModel(model);
     const cacheReadRate = costs.cacheReadCostPer1K ?? costs.promptCostPer1K;
     const cacheWriteRate = costs.cacheWriteCostPer1K ?? costs.promptCostPer1K;
     const uncached = Math.max(0, promptTokens - cached - written);
@@ -260,11 +270,18 @@ export class TokenTracker {
       this.authoritativeCostUSD += authoritative;
       estimatedCost = authoritative;
     } else {
-      this.computed.promptTokens += promptTokens;
-      this.computed.completionTokens += completionTokens;
-      this.computed.cachedPromptTokens += cached;
-      this.computed.cacheWriteTokens += written;
-      estimatedCost = this.costFor(promptTokens, completionTokens, cached, written);
+      const bucket = this.computedByModel.get(this.model) ?? {
+        promptTokens: 0,
+        completionTokens: 0,
+        cachedPromptTokens: 0,
+        cacheWriteTokens: 0,
+      };
+      bucket.promptTokens += promptTokens;
+      bucket.completionTokens += completionTokens;
+      bucket.cachedPromptTokens += cached;
+      bucket.cacheWriteTokens += written;
+      this.computedByModel.set(this.model, bucket);
+      estimatedCost = this.costFor(this.model, promptTokens, completionTokens, cached, written);
     }
 
     const entry: CallLogEntry = {
@@ -294,15 +311,17 @@ export class TokenTracker {
   }
 
   getEstimatedCost(): number {
-    return (
-      this.authoritativeCostUSD +
-      this.costFor(
-        this.computed.promptTokens,
-        this.computed.completionTokens,
-        this.computed.cachedPromptTokens,
-        this.computed.cacheWriteTokens,
-      )
-    );
+    let computed = 0;
+    for (const [model, b] of this.computedByModel) {
+      computed += this.costFor(
+        model,
+        b.promptTokens,
+        b.completionTokens,
+        b.cachedPromptTokens,
+        b.cacheWriteTokens,
+      );
+    }
+    return this.authoritativeCostUSD + computed;
   }
 
   getFormattedCost(): string {
@@ -334,7 +353,7 @@ export class TokenTracker {
       cacheWriteTokens: 0,
     };
     this.authoritativeCostUSD = 0;
-    this.computed = { promptTokens: 0, completionTokens: 0, cachedPromptTokens: 0, cacheWriteTokens: 0 };
+    this.computedByModel.clear();
   }
 
   setModel(model: string): void {

package/dist/chunk-SXSP6M24.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/rate-limiter.ts","../src/token-tracker.ts"],"sourcesContent":["/**\n * Rate Limiter for LLM API Calls\n *\n * Implements token bucket algorithm with:\n * - Configurable requests per minute/second\n * - Automatic backoff on 429 errors\n * - Queue for pending requests\n *\n * @packageDocumentation\n */\n\nexport interface RateLimiterOptions {\n  /** Maximum requests per minute (default: 60) */\n  requestsPerMinute?: number;\n  /** Maximum requests per second (default: 3) */\n  requestsPerSecond?: number;\n  /** Maximum concurrent requests (default: 5) */\n  maxConcurrent?: number;\n  /** Base delay for exponential backoff in ms (default: 1000) */\n  baseBackoffMs?: number;\n  /** Maximum backoff delay in ms (default: 60000) */\n  maxBackoffMs?: number;\n}\n\ninterface QueuedRequest<T> {\n  execute: () => Promise<T>;\n  resolve: (value: T) => void;\n  reject: (error: Error) => void;\n  retryCount: number;\n}\n\n/**\n * Rate limiter for LLM API calls using token bucket algorithm.\n *\n * @example\n * ```typescript\n * const limiter = new RateLimiter({ requestsPerMinute: 30 });\n * const result = await limiter.execute(() => llm.invoke(messages));\n * ```\n */\nexport class RateLimiter {\n  private requestsPerMinute: number;\n  private requestsPerSecond: number;\n  private maxConcurrent: number;\n  private baseBackoffMs: number;\n  private maxBackoffMs: number;\n\n  private minuteTokens: number;\n  private secondTokens: number;\n  private activeRequests: number = 0;\n  private queue: QueuedRequest<unknown>[] = [];\n  private lastMinuteReset: number = Date.now();\n  private lastSecondReset: number = Date.now();\n  private processing: boolean = false;\n  private currentBackoffMs: number = 0;\n\n  constructor(options: RateLimiterOptions = {}) {\n    this.requestsPerMinute = options.requestsPerMinute ?? 60;\n    this.requestsPerSecond = options.requestsPerSecond ?? 3;\n    this.maxConcurrent = options.maxConcurrent ?? 5;\n    this.baseBackoffMs = options.baseBackoffMs ?? 1000;\n    this.maxBackoffMs = options.maxBackoffMs ?? 60000;\n\n    this.minuteTokens = this.requestsPerMinute;\n    this.secondTokens = this.requestsPerSecond;\n  }\n\n  async execute<T>(fn: () => Promise<T>, _maxRetries: number = 3): Promise<T> {\n    return new Promise<T>((resolve, reject) => {\n      this.queue.push({\n        execute: fn as () => Promise<unknown>,\n        resolve: resolve as (value: unknown) => void,\n        reject,\n        retryCount: 0,\n      });\n      this.processQueue();\n    });\n  }\n\n  getStatus(): {\n    queueLength: number;\n    activeRequests: number;\n    minuteTokens: number;\n    secondTokens: number;\n    backoffMs: number;\n  } {\n    return {\n      queueLength: this.queue.length,\n      activeRequests: this.activeRequests,\n      minuteTokens: this.minuteTokens,\n      secondTokens: this.secondTokens,\n      backoffMs: this.currentBackoffMs,\n    };\n  }\n\n  reset(): void {\n    this.minuteTokens = this.requestsPerMinute;\n    this.secondTokens = this.requestsPerSecond;\n    this.activeRequests = 0;\n    this.queue = [];\n    this.currentBackoffMs = 0;\n    this.lastMinuteReset = Date.now();\n    this.lastSecondReset = Date.now();\n  }\n\n  private async processQueue(): Promise<void> {\n    if (this.processing) return;\n    this.processing = true;\n\n    while (this.queue.length > 0) {\n      this.refillTokens();\n\n      if (!this.canMakeRequest()) {\n        const waitTime = this.getWaitTime();\n        await this.sleep(waitTime);\n        continue;\n      }\n\n      if (this.currentBackoffMs > 0) {\n        await this.sleep(this.currentBackoffMs);\n        this.currentBackoffMs = 0;\n      }\n\n      const request = this.queue.shift();\n      if (!request) continue;\n\n      this.consumeTokens();\n      this.activeRequests++;\n\n      try {\n        const result = await request.execute();\n        request.resolve(result);\n        this.currentBackoffMs = 0;\n      } catch (error) {\n        const err = error instanceof Error ? error : new Error(String(error));\n\n        if (this.isRateLimitError(err)) {\n          this.currentBackoffMs = Math.min(\n            this.baseBackoffMs * Math.pow(2, request.retryCount),\n            this.maxBackoffMs,\n          );\n\n          console.warn(\n            `[RateLimiter] Rate limited. Backing off for ${this.currentBackoffMs}ms ` +\n              `(retry ${request.retryCount + 1})`,\n          );\n\n          if (request.retryCount < 3) {\n            this.queue.unshift({\n              ...request,\n              retryCount: request.retryCount + 1,\n            });\n          } else {\n            request.reject(\n              new Error(\n                `Rate limit exceeded after ${request.retryCount + 1} retries: ${err.message}`,\n              ),\n            );\n          }\n        } else {\n          request.reject(err);\n        }\n      } finally {\n        this.activeRequests--;\n      }\n    }\n\n    this.processing = false;\n  }\n\n  private refillTokens(): void {\n    const now = Date.now();\n    if (now - this.lastMinuteReset >= 60000) {\n      this.minuteTokens = this.requestsPerMinute;\n      this.lastMinuteReset = now;\n    }\n    if (now - this.lastSecondReset >= 1000) {\n      this.secondTokens = this.requestsPerSecond;\n      this.lastSecondReset = now;\n    }\n  }\n\n  private canMakeRequest(): boolean {\n    return (\n      this.minuteTokens > 0 &&\n      this.secondTokens > 0 &&\n      this.activeRequests < this.maxConcurrent\n    );\n  }\n\n  private consumeTokens(): void {\n    this.minuteTokens--;\n    this.secondTokens--;\n  }\n\n  private getWaitTime(): number {\n    const now = Date.now();\n    if (this.secondTokens <= 0) {\n      return Math.max(0, 1000 - (now - this.lastSecondReset));\n    }\n    if (this.minuteTokens <= 0) {\n      return Math.max(0, 60000 - (now - this.lastMinuteReset));\n    }\n    return 100;\n  }\n\n  private isRateLimitError(error: Error): boolean {\n    const message = error.message.toLowerCase();\n    return (\n      message.includes('429') ||\n      message.includes('rate limit') ||\n      message.includes('too many requests') ||\n      message.includes('quota exceeded')\n    );\n  }\n\n  private sleep(ms: number): Promise<void> {\n    return new Promise((resolve) => setTimeout(resolve, ms));\n  }\n}\n\n// Singleton instance\nlet globalRateLimiter: RateLimiter | null = null;\n\nexport function getGlobalRateLimiter(\n  options?: RateLimiterOptions,\n): RateLimiter {\n  if (!globalRateLimiter) {\n    globalRateLimiter = new RateLimiter(options);\n  }\n  return globalRateLimiter;\n}\n\nexport function resetGlobalRateLimiter(): void {\n  globalRateLimiter?.reset();\n  globalRateLimiter = null;\n}\n","/**\n * Token Tracker for LLM Usage\n *\n * Tracks token usage across multiple LLM calls for:\n * - Cost estimation (pricing fetched from OpenRouter models API)\n * - Usage monitoring\n * - Quota management\n * - Per-call JSONL logging\n *\n * @packageDocumentation\n */\n\nimport { appendFileSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';\nimport { dirname, join } from 'node:path';\n\nexport interface TokenUsage {\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  callCount: number;\n  /** Subset of `promptTokens` that were cache reads (billed at the cache-read rate). */\n  cachedPromptTokens: number;\n  /** Subset of `promptTokens` written to cache (Anthropic; billed at the cache-write rate). */\n  cacheWriteTokens: number;\n}\n\nexport interface TokenCost {\n  promptCostPer1K: number;\n  completionCostPer1K: number;\n  /** Per-1K rate for cache-read (cache-hit) prompt tokens. Falls back to prompt rate when absent. */\n  cacheReadCostPer1K?: number;\n  /** Per-1K rate for cache-write (cache-creation) prompt tokens. Falls back to prompt rate when absent. */\n  cacheWriteCostPer1K?: number;\n}\n\nexport interface CallLogEntry {\n  timestamp: string;\n  provider: string;\n  model: string;\n  promptTokens: number;\n  completionTokens: number;\n  totalTokens: number;\n  /** Cache-read (cache-hit) subset of promptTokens, billed at the discounted rate. */\n  cachedPromptTokens?: number;\n  /** Cache-write subset of promptTokens (Anthropic). */\n  cacheWriteTokens?: number;\n  estimatedCost: number;\n  durationMs?: number;\n  source: 'local-log';\n}\n\n// ---------------------------------------------------------------------------\n// Pricing: fetched from OpenRouter /api/v1/models, cached to disk for 24h\n// ---------------------------------------------------------------------------\n\nconst ALMADAR_ROOT = process.env['ALMADAR_ROOT'] ?? process.cwd();\nconst PRICING_CACHE_PATH = join(ALMADAR_ROOT, '.llm-pricing-cache.json');\nconst CALL_LOG_PATH = join(ALMADAR_ROOT, '.llm-call-log.jsonl');\nconst CACHE_TTL_MS = 24 * 60 * 60 * 1000; // 24 hours\n// Bump when the cached TokenCost shape changes so stale on-disk caches are\n// invalidated on upgrade. v2 added cacheReadCostPer1K / cacheWriteCostPer1K.\nconst PRICING_CACHE_VERSION = 2;\n\n/** Map from our local model name to OpenRouter model ID */\nconst MODEL_ID_MAP: Record<string, string> = {\n  // Anthropic\n  'claude-opus-4-5-20250929': 'anthropic/claude-opus-4.5',\n  'claude-sonnet-4-5-20250929': 'anthropic/claude-sonnet-4.5',\n  'claude-sonnet-4-20250514': 'anthropic/claude-sonnet-4',\n  'claude-3-5-haiku-20241022': 'anthropic/claude-3.5-haiku',\n  // DeepSeek — map to current versions on OpenRouter\n  'deepseek-chat': 'deepseek/deepseek-v4-flash',\n  'deepseek-coder': 'deepseek/deepseek-v4-flash',\n  'deepseek-reasoner': 'deepseek/deepseek-v4-flash',\n  'deepseek-v4-pro': 'deepseek/deepseek-v4-pro',\n  'deepseek-v4-flash': 'deepseek/deepseek-v4-flash',\n  // Kimi\n  'kimi-k2.5': 'moonshotai/kimi-k2.5',\n};\n\n// Fallback: zero cost — forces OpenRouter fetch for real pricing\nconst FALLBACK_COSTS: Record<string, TokenCost> = {};\n\ninterface PricingCache {\n  version?: number;\n  fetchedAt: number;\n  models: Record<string, TokenCost>;\n}\n\nlet pricingCache: PricingCache | null = null;\n\nfunction loadCachedPricing(): PricingCache | null {\n  try {\n    const raw = readFileSync(PRICING_CACHE_PATH, 'utf-8');\n    const parsed = JSON.parse(raw) as PricingCache;\n    if (parsed.version === PRICING_CACHE_VERSION && Date.now() - parsed.fetchedAt < CACHE_TTL_MS) {\n      return parsed;\n    }\n  } catch {\n    // No cache or expired\n  }\n  return null;\n}\n\nasync function fetchPricingFromOpenRouter(): Promise<Record<string, TokenCost>> {\n  const res = await fetch('https://openrouter.ai/api/v1/models');\n  if (!res.ok) throw new Error(`OpenRouter models API: HTTP ${res.status}`);\n  const json = await res.json() as {\n    data?: Array<{\n      id: string;\n      pricing?: {\n        prompt?: string;\n        completion?: string;\n        input_cache_read?: string;\n        input_cache_write?: string;\n      };\n    }>;\n  };\n  const models: Record<string, TokenCost> = {};\n  for (const m of json.data ?? []) {\n    const promptPerToken = parseFloat(m.pricing?.prompt ?? '0');\n    const completionPerToken = parseFloat(m.pricing?.completion ?? '0');\n    const cacheReadPerToken = parseFloat(m.pricing?.input_cache_read ?? '0');\n    const cacheWritePerToken = parseFloat(m.pricing?.input_cache_write ?? '0');\n    if (promptPerToken > 0 || completionPerToken > 0) {\n      models[m.id] = {\n        promptCostPer1K: promptPerToken * 1000,\n        completionCostPer1K: completionPerToken * 1000,\n        // 0 (field absent) → leave undefined so cost math falls back to the prompt rate.\n        ...(cacheReadPerToken > 0 ? { cacheReadCostPer1K: cacheReadPerToken * 1000 } : {}),\n        ...(cacheWritePerToken > 0 ? { cacheWriteCostPer1K: cacheWritePerToken * 1000 } : {}),\n      };\n    }\n  }\n  return models;\n}\n\n/**\n * Get pricing for all models. Uses 24h disk cache, fetches from OpenRouter on miss.\n * Non-blocking: returns cached/fallback immediately, refreshes in background if stale.\n */\nfunction getPricing(): Record<string, TokenCost> {\n  if (pricingCache) return pricingCache.models;\n\n  const diskCache = loadCachedPricing();\n  if (diskCache) {\n    pricingCache = diskCache;\n    return diskCache.models;\n  }\n\n  // Trigger background fetch, return fallback for now\n  refreshPricingCache();\n  return FALLBACK_COSTS;\n}\n\nfunction refreshPricingCache(): void {\n  fetchPricingFromOpenRouter()\n    .then((models) => {\n      pricingCache = { version: PRICING_CACHE_VERSION, fetchedAt: Date.now(), models };\n      try {\n        mkdirSync(dirname(PRICING_CACHE_PATH), { recursive: true });\n        writeFileSync(PRICING_CACHE_PATH, JSON.stringify(pricingCache));\n      } catch {\n        // Non-critical\n      }\n    })\n    .catch(() => {\n      // Silently fail, use fallback\n    });\n}\n\nfunction getCostForModel(model: string): TokenCost {\n  const pricing = getPricing();\n  // Try direct match on OpenRouter ID\n  const orId = MODEL_ID_MAP[model];\n  if (orId && pricing[orId]) return pricing[orId];\n  // Try direct key match (e.g., user passed \"openai/gpt-4o\")\n  if (pricing[model]) return pricing[model];\n  // Fuzzy: find first key containing the model name\n  for (const [key, cost] of Object.entries(pricing)) {\n    if (key.includes(model) || model.includes(key.split('/')[1] ?? '')) return cost;\n  }\n  // No pricing available — return zero (OpenRouter fetch pending or model not listed)\n  return { promptCostPer1K: 0, completionCostPer1K: 0 };\n}\n\n// ---------------------------------------------------------------------------\n// TokenTracker\n// ---------------------------------------------------------------------------\n\nexport class TokenTracker {\n  private model: string;\n  private usage: TokenUsage = {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n    callCount: 0,\n    cachedPromptTokens: 0,\n    cacheWriteTokens: 0,\n  };\n\n  /** Sum of provider-reported authoritative costs (e.g. OpenRouter `usage.cost`). */\n  private authoritativeCostUSD = 0;\n  /** Token buckets for calls WITHOUT an authoritative cost — priced cache-aware. */\n  private computed = { promptTokens: 0, completionTokens: 0, cachedPromptTokens: 0, cacheWriteTokens: 0 };\n\n  constructor(model: string = 'claude-sonnet-4-5-20250929') {\n    this.model = model;\n  }\n\n  /** Cache-aware cost for one (or an aggregate of) call(s), in USD. */\n  private costFor(promptTokens: number, completionTokens: number, cached: number, written: number): number {\n    const costs = getCostForModel(this.model);\n    const cacheReadRate = costs.cacheReadCostPer1K ?? costs.promptCostPer1K;\n    const cacheWriteRate = costs.cacheWriteCostPer1K ?? costs.promptCostPer1K;\n    const uncached = Math.max(0, promptTokens - cached - written);\n    return (\n      (uncached / 1000) * costs.promptCostPer1K +\n      (cached / 1000) * cacheReadRate +\n      (written / 1000) * cacheWriteRate +\n      (completionTokens / 1000) * costs.completionCostPer1K\n    );\n  }\n\n  /**\n   * Record one LLM call's usage. `promptTokens` is the TOTAL input count\n   * (cache reads + cache writes + uncached); `cachedPromptTokens` and\n   * `cacheWriteTokens` are subsets of it, priced at their own (cheaper /\n   * pricier) rates. Providers that don't report cache detail pass 0, which\n   * reduces to the previous flat-rate behaviour.\n   */\n  addUsage(\n    promptTokens: number,\n    completionTokens: number,\n    options?: {\n      provider?: string;\n      durationMs?: number;\n      cachedPromptTokens?: number;\n      cacheWriteTokens?: number;\n      /** Provider-reported authoritative cost (e.g. OpenRouter `usage.cost`). When set, used verbatim. */\n      costUSD?: number;\n    },\n  ): void {\n    const cached = Math.min(promptTokens, Math.max(0, options?.cachedPromptTokens ?? 0));\n    const written = Math.min(promptTokens - cached, Math.max(0, options?.cacheWriteTokens ?? 0));\n\n    this.usage.promptTokens += promptTokens;\n    this.usage.completionTokens += completionTokens;\n    this.usage.totalTokens += promptTokens + completionTokens;\n    this.usage.cachedPromptTokens += cached;\n    this.usage.cacheWriteTokens += written;\n    this.usage.callCount++;\n\n    // Prefer the provider's authoritative cost (already cache- and routing-\n    // adjusted). Otherwise bucket the tokens and price them cache-aware so\n    // late-arriving pricing still applies retroactively to the estimate.\n    const authoritative = options?.costUSD;\n    let estimatedCost: number;\n    if (authoritative != null && Number.isFinite(authoritative)) {\n      this.authoritativeCostUSD += authoritative;\n      estimatedCost = authoritative;\n    } else {\n      this.computed.promptTokens += promptTokens;\n      this.computed.completionTokens += completionTokens;\n      this.computed.cachedPromptTokens += cached;\n      this.computed.cacheWriteTokens += written;\n      estimatedCost = this.costFor(promptTokens, completionTokens, cached, written);\n    }\n\n    const entry: CallLogEntry = {\n      timestamp: new Date().toISOString(),\n      provider: options?.provider ?? 'unknown',\n      model: this.model,\n      promptTokens,\n      completionTokens,\n      totalTokens: promptTokens + completionTokens,\n      cachedPromptTokens: cached,\n      cacheWriteTokens: written,\n      estimatedCost,\n      durationMs: options?.durationMs,\n      source: 'local-log',\n    };\n\n    try {\n      mkdirSync(dirname(CALL_LOG_PATH), { recursive: true });\n      appendFileSync(CALL_LOG_PATH, JSON.stringify(entry) + '\\n');\n    } catch {\n      // Non-critical: don't break LLM calls if logging fails\n    }\n  }\n\n  getSummary(): TokenUsage {\n    return { ...this.usage };\n  }\n\n  getEstimatedCost(): number {\n    return (\n      this.authoritativeCostUSD +\n      this.costFor(\n        this.computed.promptTokens,\n        this.computed.completionTokens,\n        this.computed.cachedPromptTokens,\n        this.computed.cacheWriteTokens,\n      )\n    );\n  }\n\n  getFormattedCost(): string {\n    const cost = this.getEstimatedCost();\n    return `$${cost.toFixed(4)}`;\n  }\n\n  getReport(): string {\n    const summary = this.getSummary();\n    const cost = this.getEstimatedCost();\n    return [\n      `Token Usage Report (${this.model})`,\n      `─────────────────────────────`,\n      `Calls:              ${summary.callCount}`,\n      `Prompt Tokens:      ${summary.promptTokens.toLocaleString()}`,\n      `Completion Tokens:  ${summary.completionTokens.toLocaleString()}`,\n      `Total Tokens:       ${summary.totalTokens.toLocaleString()}`,\n      `Estimated Cost:     $${cost.toFixed(4)}`,\n    ].join('\\n');\n  }\n\n  reset(): void {\n    this.usage = {\n      promptTokens: 0,\n      completionTokens: 0,\n      totalTokens: 0,\n      callCount: 0,\n      cachedPromptTokens: 0,\n      cacheWriteTokens: 0,\n    };\n    this.authoritativeCostUSD = 0;\n    this.computed = { promptTokens: 0, completionTokens: 0, cachedPromptTokens: 0, cacheWriteTokens: 0 };\n  }\n\n  setModel(model: string): void {\n    this.model = model;\n  }\n}\n\n// Global tracker instance\nlet globalTracker: TokenTracker | null = null;\n\nexport function getGlobalTokenTracker(model?: string): TokenTracker {\n  if (!globalTracker) {\n    globalTracker = new TokenTracker(model);\n  } else if (model) {\n    globalTracker.setModel(model);\n  }\n  return globalTracker;\n}\n\nexport function resetGlobalTokenTracker(): void {\n  globalTracker?.reset();\n}\n\nexport function getCallLogPath(): string {\n  return CALL_LOG_PATH;\n}\n\n/** Force-refresh the pricing cache from OpenRouter. */\nexport async function refreshPricing(): Promise<void> {\n  const models = await fetchPricingFromOpenRouter();\n  pricingCache = { version: PRICING_CACHE_VERSION, fetchedAt: Date.now(), models };\n  mkdirSync(dirname(PRICING_CACHE_PATH), { recursive: true });\n  writeFileSync(PRICING_CACHE_PATH, JSON.stringify(pricingCache));\n}\n"],"mappings":";AAwCO,IAAM,cAAN,MAAkB;AAAA,EAgBvB,YAAY,UAA8B,CAAC,GAAG;AAP9C,SAAQ,iBAAyB;AACjC,SAAQ,QAAkC,CAAC;AAC3C,SAAQ,kBAA0B,KAAK,IAAI;AAC3C,SAAQ,kBAA0B,KAAK,IAAI;AAC3C,SAAQ,aAAsB;AAC9B,SAAQ,mBAA2B;AAGjC,SAAK,oBAAoB,QAAQ,qBAAqB;AACtD,SAAK,oBAAoB,QAAQ,qBAAqB;AACtD,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,eAAe,QAAQ,gBAAgB;AAE5C,SAAK,eAAe,KAAK;AACzB,SAAK,eAAe,KAAK;AAAA,EAC3B;AAAA,EAEA,MAAM,QAAW,IAAsB,cAAsB,GAAe;AAC1E,WAAO,IAAI,QAAW,CAAC,SAAS,WAAW;AACzC,WAAK,MAAM,KAAK;AAAA,QACd,SAAS;AAAA,QACT;AAAA,QACA;AAAA,QACA,YAAY;AAAA,MACd,CAAC;AACD,WAAK,aAAa;AAAA,IACpB,CAAC;AAAA,EACH;AAAA,EAEA,YAME;AACA,WAAO;AAAA,MACL,aAAa,KAAK,MAAM;AAAA,MACxB,gBAAgB,KAAK;AAAA,MACrB,cAAc,KAAK;AAAA,MACnB,cAAc,KAAK;AAAA,MACnB,WAAW,KAAK;AAAA,IAClB;AAAA,EACF;AAAA,EAEA,QAAc;AACZ,SAAK,eAAe,KAAK;AACzB,SAAK,eAAe,KAAK;AACzB,SAAK,iBAAiB;AACtB,SAAK,QAAQ,CAAC;AACd,SAAK,mBAAmB;AACxB,SAAK,kBAAkB,KAAK,IAAI;AAChC,SAAK,kBAAkB,KAAK,IAAI;AAAA,EAClC;AAAA,EAEA,MAAc,eAA8B;AAC1C,QAAI,KAAK,WAAY;AACrB,SAAK,aAAa;AAElB,WAAO,KAAK,MAAM,SAAS,GAAG;AAC5B,WAAK,aAAa;AAElB,UAAI,CAAC,KAAK,eAAe,GAAG;AAC1B,cAAM,WAAW,KAAK,YAAY;AAClC,cAAM,KAAK,MAAM,QAAQ;AACzB;AAAA,MACF;AAEA,UAAI,KAAK,mBAAmB,GAAG;AAC7B,cAAM,KAAK,MAAM,KAAK,gBAAgB;AACtC,aAAK,mBAAmB;AAAA,MAC1B;AAEA,YAAM,UAAU,KAAK,MAAM,MAAM;AACjC,UAAI,CAAC,QAAS;AAEd,WAAK,cAAc;AACnB,WAAK;AAEL,UAAI;AACF,cAAM,SAAS,MAAM,QAAQ,QAAQ;AACrC,gBAAQ,QAAQ,MAAM;AACtB,aAAK,mBAAmB;AAAA,MAC1B,SAAS,OAAO;AACd,cAAM,MAAM,iBAAiB,QAAQ,QAAQ,IAAI,MAAM,OAAO,KAAK,CAAC;AAEpE,YAAI,KAAK,iBAAiB,GAAG,GAAG;AAC9B,eAAK,mBAAmB,KAAK;AAAA,YAC3B,KAAK,gBAAgB,KAAK,IAAI,GAAG,QAAQ,UAAU;AAAA,YACnD,KAAK;AAAA,UACP;AAEA,kBAAQ;AAAA,YACN,+CAA+C,KAAK,gBAAgB,aACxD,QAAQ,aAAa,CAAC;AAAA,UACpC;AAEA,cAAI,QAAQ,aAAa,GAAG;AAC1B,iBAAK,MAAM,QAAQ;AAAA,cACjB,GAAG;AAAA,cACH,YAAY,QAAQ,aAAa;AAAA,YACnC,CAAC;AAAA,UACH,OAAO;AACL,oBAAQ;AAAA,cACN,IAAI;AAAA,gBACF,6BAA6B,QAAQ,aAAa,CAAC,aAAa,IAAI,OAAO;AAAA,cAC7E;AAAA,YACF;AAAA,UACF;AAAA,QACF,OAAO;AACL,kBAAQ,OAAO,GAAG;AAAA,QACpB;AAAA,MACF,UAAE;AACA,aAAK;AAAA,MACP;AAAA,IACF;AAEA,SAAK,aAAa;AAAA,EACpB;AAAA,EAEQ,eAAqB;AAC3B,UAAM,MAAM,KAAK,IAAI;AACrB,QAAI,MAAM,KAAK,mBAAmB,KAAO;AACvC,WAAK,eAAe,KAAK;AACzB,WAAK,kBAAkB;AAAA,IACzB;AACA,QAAI,MAAM,KAAK,mBAAmB,KAAM;AACtC,WAAK,eAAe,KAAK;AACzB,WAAK,kBAAkB;AAAA,IACzB;AAAA,EACF;AAAA,EAEQ,iBAA0B;AAChC,WACE,KAAK,eAAe,KACpB,KAAK,eAAe,KACpB,KAAK,iBAAiB,KAAK;AAAA,EAE/B;AAAA,EAEQ,gBAAsB;AAC5B,SAAK;AACL,SAAK;AAAA,EACP;AAAA,EAEQ,cAAsB;AAC5B,UAAM,MAAM,KAAK,IAAI;AACrB,QAAI,KAAK,gBAAgB,GAAG;AAC1B,aAAO,KAAK,IAAI,GAAG,OAAQ,MAAM,KAAK,gBAAgB;AAAA,IACxD;AACA,QAAI,KAAK,gBAAgB,GAAG;AAC1B,aAAO,KAAK,IAAI,GAAG,OAAS,MAAM,KAAK,gBAAgB;AAAA,IACzD;AACA,WAAO;AAAA,EACT;AAAA,EAEQ,iBAAiB,OAAuB;AAC9C,UAAM,UAAU,MAAM,QAAQ,YAAY;AAC1C,WACE,QAAQ,SAAS,KAAK,KACtB,QAAQ,SAAS,YAAY,KAC7B,QAAQ,SAAS,mBAAmB,KACpC,QAAQ,SAAS,gBAAgB;AAAA,EAErC;AAAA,EAEQ,MAAM,IAA2B;AACvC,WAAO,IAAI,QAAQ,CAAC,YAAY,WAAW,SAAS,EAAE,CAAC;AAAA,EACzD;AACF;AAGA,IAAI,oBAAwC;AAErC,SAAS,qBACd,SACa;AACb,MAAI,CAAC,mBAAmB;AACtB,wBAAoB,IAAI,YAAY,OAAO;AAAA,EAC7C;AACA,SAAO;AACT;AAEO,SAAS,yBAA+B;AAC7C,qBAAmB,MAAM;AACzB,sBAAoB;AACtB;;;AChOA,SAAS,gBAAgB,WAAW,cAAc,qBAAqB;AACvE,SAAS,SAAS,YAAY;AA0C9B,IAAM,eAAe,QAAQ,IAAI,cAAc,KAAK,QAAQ,IAAI;AAChE,IAAM,qBAAqB,KAAK,cAAc,yBAAyB;AACvE,IAAM,gBAAgB,KAAK,cAAc,qBAAqB;AAC9D,IAAM,eAAe,KAAK,KAAK,KAAK;AAGpC,IAAM,wBAAwB;AAG9B,IAAM,eAAuC;AAAA;AAAA,EAE3C,4BAA4B;AAAA,EAC5B,8BAA8B;AAAA,EAC9B,4BAA4B;AAAA,EAC5B,6BAA6B;AAAA;AAAA,EAE7B,iBAAiB;AAAA,EACjB,kBAAkB;AAAA,EAClB,qBAAqB;AAAA,EACrB,mBAAmB;AAAA,EACnB,qBAAqB;AAAA;AAAA,EAErB,aAAa;AACf;AAGA,IAAM,iBAA4C,CAAC;AAQnD,IAAI,eAAoC;AAExC,SAAS,oBAAyC;AAChD,MAAI;AACF,UAAM,MAAM,aAAa,oBAAoB,OAAO;AACpD,UAAM,SAAS,KAAK,MAAM,GAAG;AAC7B,QAAI,OAAO,YAAY,yBAAyB,KAAK,IAAI,IAAI,OAAO,YAAY,cAAc;AAC5F,aAAO;AAAA,IACT;AAAA,EACF,QAAQ;AAAA,EAER;AACA,SAAO;AACT;AAEA,eAAe,6BAAiE;AAC9E,QAAM,MAAM,MAAM,MAAM,qCAAqC;AAC7D,MAAI,CAAC,IAAI,GAAI,OAAM,IAAI,MAAM,+BAA+B,IAAI,MAAM,EAAE;AACxE,QAAM,OAAO,MAAM,IAAI,KAAK;AAW5B,QAAM,SAAoC,CAAC;AAC3C,aAAW,KAAK,KAAK,QAAQ,CAAC,GAAG;AAC/B,UAAM,iBAAiB,WAAW,EAAE,SAAS,UAAU,GAAG;AAC1D,UAAM,qBAAqB,WAAW,EAAE,SAAS,cAAc,GAAG;AAClE,UAAM,oBAAoB,WAAW,EAAE,SAAS,oBAAoB,GAAG;AACvE,UAAM,qBAAqB,WAAW,EAAE,SAAS,qBAAqB,GAAG;AACzE,QAAI,iBAAiB,KAAK,qBAAqB,GAAG;AAChD,aAAO,EAAE,EAAE,IAAI;AAAA,QACb,iBAAiB,iBAAiB;AAAA,QAClC,qBAAqB,qBAAqB;AAAA;AAAA,QAE1C,GAAI,oBAAoB,IAAI,EAAE,oBAAoB,oBAAoB,IAAK,IAAI,CAAC;AAAA,QAChF,GAAI,qBAAqB,IAAI,EAAE,qBAAqB,qBAAqB,IAAK,IAAI,CAAC;AAAA,MACrF;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAMA,SAAS,aAAwC;AAC/C,MAAI,aAAc,QAAO,aAAa;AAEtC,QAAM,YAAY,kBAAkB;AACpC,MAAI,WAAW;AACb,mBAAe;AACf,WAAO,UAAU;AAAA,EACnB;AAGA,sBAAoB;AACpB,SAAO;AACT;AAEA,SAAS,sBAA4B;AACnC,6BAA2B,EACxB,KAAK,CAAC,WAAW;AAChB,mBAAe,EAAE,SAAS,uBAAuB,WAAW,KAAK,IAAI,GAAG,OAAO;AAC/E,QAAI;AACF,gBAAU,QAAQ,kBAAkB,GAAG,EAAE,WAAW,KAAK,CAAC;AAC1D,oBAAc,oBAAoB,KAAK,UAAU,YAAY,CAAC;AAAA,IAChE,QAAQ;AAAA,IAER;AAAA,EACF,CAAC,EACA,MAAM,MAAM;AAAA,EAEb,CAAC;AACL;AAEA,SAAS,gBAAgB,OAA0B;AACjD,QAAM,UAAU,WAAW;AAE3B,QAAM,OAAO,aAAa,KAAK;AAC/B,MAAI,QAAQ,QAAQ,IAAI,EAAG,QAAO,QAAQ,IAAI;AAE9C,MAAI,QAAQ,KAAK,EAAG,QAAO,QAAQ,KAAK;AAExC,aAAW,CAAC,KAAK,IAAI,KAAK,OAAO,QAAQ,OAAO,GAAG;AACjD,QAAI,IAAI,SAAS,KAAK,KAAK,MAAM,SAAS,IAAI,MAAM,GAAG,EAAE,CAAC,KAAK,EAAE,EAAG,QAAO;AAAA,EAC7E;AAEA,SAAO,EAAE,iBAAiB,GAAG,qBAAqB,EAAE;AACtD;AAMO,IAAM,eAAN,MAAmB;AAAA,EAgBxB,YAAY,QAAgB,8BAA8B;AAd1D,SAAQ,QAAoB;AAAA,MAC1B,cAAc;AAAA,MACd,kBAAkB;AAAA,MAClB,aAAa;AAAA,MACb,WAAW;AAAA,MACX,oBAAoB;AAAA,MACpB,kBAAkB;AAAA,IACpB;AAGA;AAAA,SAAQ,uBAAuB;AAE/B;AAAA,SAAQ,WAAW,EAAE,cAAc,GAAG,kBAAkB,GAAG,oBAAoB,GAAG,kBAAkB,EAAE;AAGpG,SAAK,QAAQ;AAAA,EACf;AAAA;AAAA,EAGQ,QAAQ,cAAsB,kBAA0B,QAAgB,SAAyB;AACvG,UAAM,QAAQ,gBAAgB,KAAK,KAAK;AACxC,UAAM,gBAAgB,MAAM,sBAAsB,MAAM;AACxD,UAAM,iBAAiB,MAAM,uBAAuB,MAAM;AAC1D,UAAM,WAAW,KAAK,IAAI,GAAG,eAAe,SAAS,OAAO;AAC5D,WACG,WAAW,MAAQ,MAAM,kBACzB,SAAS,MAAQ,gBACjB,UAAU,MAAQ,iBAClB,mBAAmB,MAAQ,MAAM;AAAA,EAEtC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,SACE,cACA,kBACA,SAQM;AACN,UAAM,SAAS,KAAK,IAAI,cAAc,KAAK,IAAI,GAAG,SAAS,sBAAsB,CAAC,CAAC;AACnF,UAAM,UAAU,KAAK,IAAI,eAAe,QAAQ,KAAK,IAAI,GAAG,SAAS,oBAAoB,CAAC,CAAC;AAE3F,SAAK,MAAM,gBAAgB;AAC3B,SAAK,MAAM,oBAAoB;AAC/B,SAAK,MAAM,eAAe,eAAe;AACzC,SAAK,MAAM,sBAAsB;AACjC,SAAK,MAAM,oBAAoB;AAC/B,SAAK,MAAM;AAKX,UAAM,gBAAgB,SAAS;AAC/B,QAAI;AACJ,QAAI,iBAAiB,QAAQ,OAAO,SAAS,aAAa,GAAG;AAC3D,WAAK,wBAAwB;AAC7B,sBAAgB;AAAA,IAClB,OAAO;AACL,WAAK,SAAS,gBAAgB;AAC9B,WAAK,SAAS,oBAAoB;AAClC,WAAK,SAAS,sBAAsB;AACpC,WAAK,SAAS,oBAAoB;AAClC,sBAAgB,KAAK,QAAQ,cAAc,kBAAkB,QAAQ,OAAO;AAAA,IAC9E;AAEA,UAAM,QAAsB;AAAA,MAC1B,YAAW,oBAAI,KAAK,GAAE,YAAY;AAAA,MAClC,UAAU,SAAS,YAAY;AAAA,MAC/B,OAAO,KAAK;AAAA,MACZ;AAAA,MACA;AAAA,MACA,aAAa,eAAe;AAAA,MAC5B,oBAAoB;AAAA,MACpB,kBAAkB;AAAA,MAClB;AAAA,MACA,YAAY,SAAS;AAAA,MACrB,QAAQ;AAAA,IACV;AAEA,QAAI;AACF,gBAAU,QAAQ,aAAa,GAAG,EAAE,WAAW,KAAK,CAAC;AACrD,qBAAe,eAAe,KAAK,UAAU,KAAK,IAAI,IAAI;AAAA,IAC5D,QAAQ;AAAA,IAER;AAAA,EACF;AAAA,EAEA,aAAyB;AACvB,WAAO,EAAE,GAAG,KAAK,MAAM;AAAA,EACzB;AAAA,EAEA,mBAA2B;AACzB,WACE,KAAK,uBACL,KAAK;AAAA,MACH,KAAK,SAAS;AAAA,MACd,KAAK,SAAS;AAAA,MACd,KAAK,SAAS;AAAA,MACd,KAAK,SAAS;AAAA,IAChB;AAAA,EAEJ;AAAA,EAEA,mBAA2B;AACzB,UAAM,OAAO,KAAK,iBAAiB;AACnC,WAAO,IAAI,KAAK,QAAQ,CAAC,CAAC;AAAA,EAC5B;AAAA,EAEA,YAAoB;AAClB,UAAM,UAAU,KAAK,WAAW;AAChC,UAAM,OAAO,KAAK,iBAAiB;AACnC,WAAO;AAAA,MACL,uBAAuB,KAAK,KAAK;AAAA,MACjC;AAAA,MACA,uBAAuB,QAAQ,SAAS;AAAA,MACxC,uBAAuB,QAAQ,aAAa,eAAe,CAAC;AAAA,MAC5D,uBAAuB,QAAQ,iBAAiB,eAAe,CAAC;AAAA,MAChE,uBAAuB,QAAQ,YAAY,eAAe,CAAC;AAAA,MAC3D,wBAAwB,KAAK,QAAQ,CAAC,CAAC;AAAA,IACzC,EAAE,KAAK,IAAI;AAAA,EACb;AAAA,EAEA,QAAc;AACZ,SAAK,QAAQ;AAAA,MACX,cAAc;AAAA,MACd,kBAAkB;AAAA,MAClB,aAAa;AAAA,MACb,WAAW;AAAA,MACX,oBAAoB;AAAA,MACpB,kBAAkB;AAAA,IACpB;AACA,SAAK,uBAAuB;AAC5B,SAAK,WAAW,EAAE,cAAc,GAAG,kBAAkB,GAAG,oBAAoB,GAAG,kBAAkB,EAAE;AAAA,EACrG;AAAA,EAEA,SAAS,OAAqB;AAC5B,SAAK,QAAQ;AAAA,EACf;AACF;AAGA,IAAI,gBAAqC;AAElC,SAAS,sBAAsB,OAA8B;AAClE,MAAI,CAAC,eAAe;AAClB,oBAAgB,IAAI,aAAa,KAAK;AAAA,EACxC,WAAW,OAAO;AAChB,kBAAc,SAAS,KAAK;AAAA,EAC9B;AACA,SAAO;AACT;AAEO,SAAS,0BAAgC;AAC9C,iBAAe,MAAM;AACvB;","names":[]}