@almadar/llm 2.50.0 → 2.52.0

package/dist/{chunk-RPG3SUIB.js → chunk-T6AKOBX3.js}

@@ -1,169 +1,3 @@
-// src/rate-limiter.ts
-var RateLimiter = class {
-  constructor(options = {}) {
-    this.activeRequests = 0;
-    this.queue = [];
-    this.lastMinuteReset = Date.now();
-    this.lastSecondReset = Date.now();
-    this.pumping = false;
-    this.currentBackoffMs = 0;
-    this.wakeTimer = null;
-    this.requestsPerMinute = options.requestsPerMinute ?? 60;
-    this.requestsPerSecond = options.requestsPerSecond ?? 3;
-    this.maxConcurrent = options.maxConcurrent ?? 5;
-    this.baseBackoffMs = options.baseBackoffMs ?? 1e3;
-    this.maxBackoffMs = options.maxBackoffMs ?? 6e4;
-    this.minuteTokens = this.requestsPerMinute;
-    this.secondTokens = this.requestsPerSecond;
-  }
-  async execute(fn, _maxRetries = 3) {
-    return new Promise((resolve, reject) => {
-      this.queue.push({
-        execute: fn,
-        resolve,
-        reject,
-        retryCount: 0
-      });
-      this.pump();
-    });
-  }
-  getStatus() {
-    return {
-      queueLength: this.queue.length,
-      activeRequests: this.activeRequests,
-      minuteTokens: this.minuteTokens,
-      secondTokens: this.secondTokens,
-      backoffMs: this.currentBackoffMs
-    };
-  }
-  reset() {
-    this.minuteTokens = this.requestsPerMinute;
-    this.secondTokens = this.requestsPerSecond;
-    this.activeRequests = 0;
-    this.queue = [];
-    this.currentBackoffMs = 0;
-    this.lastMinuteReset = Date.now();
-    this.lastSecondReset = Date.now();
-    if (this.wakeTimer !== null) {
-      clearTimeout(this.wakeTimer);
-      this.wakeTimer = null;
-    }
-  }
-  /**
-   * Start queued requests up to the rate/concurrency budget, then return —
-   * in-flight requests re-enter the pump as they settle, and a wake timer
-   * covers token-bucket refills and 429 backoff. Never awaits a request.
-   */
-  pump() {
-    if (this.pumping) return;
-    this.pumping = true;
-    try {
-      this.refillTokens();
-      while (this.queue.length > 0 && this.currentBackoffMs === 0 && this.canMakeRequest()) {
-        const request = this.queue.shift();
-        if (!request) break;
-        this.consumeTokens();
-        this.activeRequests++;
-        void this.settle(request);
-      }
-    } finally {
-      this.pumping = false;
-    }
-    if (this.queue.length === 0) return;
-    if (this.currentBackoffMs > 0) {
-      this.scheduleWake(this.currentBackoffMs, true);
-    } else if (this.activeRequests < this.maxConcurrent) {
-      this.scheduleWake(this.getWaitTime(), false);
-    }
-  }
-  scheduleWake(ms, clearsBackoff) {
-    if (this.wakeTimer !== null) return;
-    this.wakeTimer = setTimeout(() => {
-      this.wakeTimer = null;
-      if (clearsBackoff) this.currentBackoffMs = 0;
-      this.pump();
-    }, ms);
-  }
-  async settle(request) {
-    try {
-      const result = await request.execute();
-      request.resolve(result);
-      this.currentBackoffMs = 0;
-    } catch (error) {
-      const err = error instanceof Error ? error : new Error(String(error));
-      if (this.isRateLimitError(err)) {
-        this.currentBackoffMs = Math.min(
-          this.baseBackoffMs * Math.pow(2, request.retryCount),
-          this.maxBackoffMs
-        );
-        console.warn(
-          `[RateLimiter] Rate limited. Backing off for ${this.currentBackoffMs}ms (retry ${request.retryCount + 1})`
-        );
-        if (request.retryCount < 3) {
-          this.queue.unshift({
-            ...request,
-            retryCount: request.retryCount + 1
-          });
-        } else {
-          request.reject(
-            new Error(
-              `Rate limit exceeded after ${request.retryCount + 1} retries: ${err.message}`
-            )
-          );
-        }
-      } else {
-        request.reject(err);
-      }
-    } finally {
-      this.activeRequests--;
-      this.pump();
-    }
-  }
-  refillTokens() {
-    const now = Date.now();
-    if (now - this.lastMinuteReset >= 6e4) {
-      this.minuteTokens = this.requestsPerMinute;
-      this.lastMinuteReset = now;
-    }
-    if (now - this.lastSecondReset >= 1e3) {
-      this.secondTokens = this.requestsPerSecond;
-      this.lastSecondReset = now;
-    }
-  }
-  canMakeRequest() {
-    return this.minuteTokens > 0 && this.secondTokens > 0 && this.activeRequests < this.maxConcurrent;
-  }
-  consumeTokens() {
-    this.minuteTokens--;
-    this.secondTokens--;
-  }
-  getWaitTime() {
-    const now = Date.now();
-    if (this.secondTokens <= 0) {
-      return Math.max(0, 1e3 - (now - this.lastSecondReset));
-    }
-    if (this.minuteTokens <= 0) {
-      return Math.max(0, 6e4 - (now - this.lastMinuteReset));
-    }
-    return 100;
-  }
-  isRateLimitError(error) {
-    const message = error.message.toLowerCase();
-    return message.includes("429") || message.includes("rate limit") || message.includes("too many requests") || message.includes("quota exceeded");
-  }
-};
-var globalRateLimiter = null;
-function getGlobalRateLimiter(options) {
-  if (!globalRateLimiter) {
-    globalRateLimiter = new RateLimiter(options);
-  }
-  return globalRateLimiter;
-}
-function resetGlobalRateLimiter() {
-  globalRateLimiter?.reset();
-  globalRateLimiter = null;
-}
-
 // src/token-tracker.ts
 import { appendFileSync, mkdirSync, readFileSync, writeFileSync } from "fs";
 import { dirname, join } from "path";
@@ -243,7 +77,7 @@ function refreshPricingCache() {
   }).catch(() => {
   });
 }
-function getCostForModel(model) {
+function getCostForModelIfKnown(model) {
   const pricing = getPricing();
   const orId = MODEL_ID_MAP[model];
   if (orId && pricing[orId]) return pricing[orId];
@@ -251,7 +85,27 @@ function getCostForModel(model) {
   for (const [key, cost] of Object.entries(pricing)) {
     if (key.includes(model) || model.includes(key.split("/")[1] ?? "")) return cost;
   }
-  return { promptCostPer1K: 0, completionCostPer1K: 0 };
+  return void 0;
+}
+function getCostForModel(model) {
+  return getCostForModelIfKnown(model) ?? { promptCostPer1K: 0, completionCostPer1K: 0 };
+}
+function priceTokens(costs, promptTokens, completionTokens, cached, written) {
+  const cacheReadRate = costs.cacheReadCostPer1K ?? costs.promptCostPer1K;
+  const cacheWriteRate = costs.cacheWriteCostPer1K ?? costs.promptCostPer1K;
+  const uncached = Math.max(0, promptTokens - cached - written);
+  return uncached / 1e3 * costs.promptCostPer1K + cached / 1e3 * cacheReadRate + written / 1e3 * cacheWriteRate + completionTokens / 1e3 * costs.completionCostPer1K;
+}
+function estimateCostUSD(model, tokens) {
+  const costs = getCostForModelIfKnown(model);
+  if (costs === void 0) return void 0;
+  return priceTokens(
+    costs,
+    tokens.promptTokens,
+    tokens.completionTokens,
+    Math.max(0, tokens.cachedPromptTokens ?? 0),
+    Math.max(0, tokens.cacheWriteTokens ?? 0)
+  );
 }
 var TokenTracker = class {
   constructor(model = "claude-sonnet-4-5-20250929") {
@@ -278,11 +132,7 @@ var TokenTracker = class {
   }
   /** Cache-aware cost for one (or an aggregate of) call(s), in USD. */
   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);
-    return uncached / 1e3 * costs.promptCostPer1K + cached / 1e3 * cacheReadRate + written / 1e3 * cacheWriteRate + completionTokens / 1e3 * costs.completionCostPer1K;
+    return priceTokens(getCostForModel(model), promptTokens, completionTokens, cached, written);
   }
   /**
    * Record one LLM call's usage. `promptTokens` is the TOTAL input count
@@ -401,11 +251,9 @@ function resetGlobalTokenTracker() {
 }
 
 export {
-  RateLimiter,
-  getGlobalRateLimiter,
-  resetGlobalRateLimiter,
+  estimateCostUSD,
   TokenTracker,
   getGlobalTokenTracker,
   resetGlobalTokenTracker
 };
-//# sourceMappingURL=chunk-RPG3SUIB.js.map
+//# sourceMappingURL=chunk-T6AKOBX3.js.map

package/dist/chunk-T6AKOBX3.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/token-tracker.ts"],"sourcesContent":["/**\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\n/**\n * Look up a model's pricing row without a zero-cost fallback — `undefined`\n * means \"no pricing known yet\" (OpenRouter fetch pending, or the model\n * isn't listed), distinct from a model that is genuinely free. Single\n * lookup path: `getCostForModel` (zero-fallback, for the tracker's own\n * running totals) and `estimateCostUSD` (pure, callers outside the\n * tracker) both resolve through this.\n */\nfunction getCostForModelIfKnown(model: string): TokenCost | undefined {\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  return undefined;\n}\n\nfunction getCostForModel(model: string): TokenCost {\n  // No pricing available — return zero (OpenRouter fetch pending or model not listed)\n  return getCostForModelIfKnown(model) ?? { promptCostPer1K: 0, completionCostPer1K: 0 };\n}\n\n/** Cache-aware cost formula shared by `TokenTracker.costFor` (instance,\n *  zero-fallback pricing) and `estimateCostUSD` (pure, `undefined` pricing\n *  propagates to the caller instead of silently pricing at $0). */\nfunction priceTokens(\n  costs: TokenCost,\n  promptTokens: number,\n  completionTokens: number,\n  cached: number,\n  written: number,\n): number {\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\nexport interface EstimateCostTokens {\n  promptTokens: number;\n  completionTokens: number;\n  /** Subset of `promptTokens` served from the provider's prefix cache. */\n  cachedPromptTokens?: number;\n  /** Subset of `promptTokens` written to cache (Anthropic cache-write). */\n  cacheWriteTokens?: number;\n}\n\n/**\n * Pure per-call cost estimate from token counts alone, priced from the SAME\n * OpenRouter-fetched table `TokenTracker` uses for the whole-run total (the\n * 24h disk cache in `getPricing` — no second pricing source). Callers with\n * tokens + a model name but no `TokenTracker` instance (e.g. a trace-event\n * emitter) use this instead of reimplementing the cache-aware math.\n *\n * Returns `undefined` when the model has no known pricing row — the\n * caller's job to render that as \"unknown\", never as `$0`.\n */\nexport function estimateCostUSD(model: string, tokens: EstimateCostTokens): number | undefined {\n  const costs = getCostForModelIfKnown(model);\n  if (costs === undefined) return undefined;\n  return priceTokens(\n    costs,\n    tokens.promptTokens,\n    tokens.completionTokens,\n    Math.max(0, tokens.cachedPromptTokens ?? 0),\n    Math.max(0, tokens.cacheWriteTokens ?? 0),\n  );\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    return priceTokens(getCostForModel(model), promptTokens, completionTokens, cached, written);\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":";AAYA,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;AAUA,SAAS,uBAAuB,OAAsC;AACpE,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;AACA,SAAO;AACT;AAEA,SAAS,gBAAgB,OAA0B;AAEjD,SAAO,uBAAuB,KAAK,KAAK,EAAE,iBAAiB,GAAG,qBAAqB,EAAE;AACvF;AAKA,SAAS,YACP,OACA,cACA,kBACA,QACA,SACQ;AACR,QAAM,gBAAgB,MAAM,sBAAsB,MAAM;AACxD,QAAM,iBAAiB,MAAM,uBAAuB,MAAM;AAC1D,QAAM,WAAW,KAAK,IAAI,GAAG,eAAe,SAAS,OAAO;AAC5D,SACG,WAAW,MAAQ,MAAM,kBACzB,SAAS,MAAQ,gBACjB,UAAU,MAAQ,iBAClB,mBAAmB,MAAQ,MAAM;AAEtC;AAqBO,SAAS,gBAAgB,OAAe,QAAgD;AAC7F,QAAM,QAAQ,uBAAuB,KAAK;AAC1C,MAAI,UAAU,OAAW,QAAO;AAChC,SAAO;AAAA,IACL;AAAA,IACA,OAAO;AAAA,IACP,OAAO;AAAA,IACP,KAAK,IAAI,GAAG,OAAO,sBAAsB,CAAC;AAAA,IAC1C,KAAK,IAAI,GAAG,OAAO,oBAAoB,CAAC;AAAA,EAC1C;AACF;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,WAAO,YAAY,gBAAgB,KAAK,GAAG,cAAc,kBAAkB,QAAQ,OAAO;AAAA,EAC5F;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/{client-SYMpnw-w.d.ts → client-DfzMDgkm.d.ts}

@@ -1,4 +1,4 @@
-import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-CXaf8aAy.js';
+import { R as RateLimiterOptions, T as TokenUsage } from './rate-limiter-Bz0iSJgZ.js';
 import { ChatOpenAI } from '@langchain/openai';
 import { ChatAnthropic } from '@langchain/anthropic';
 import Anthropic from '@anthropic-ai/sdk';
@@ -133,6 +133,15 @@ interface LLMUsage {
      *  `prompt_tokens_details.cached_tokens`, deepseek-native
      *  `prompt_cache_hit_tokens`). Absent when the provider reports neither. */
     cachedPromptTokens?: number;
+    /**
+     * USD cost of this one call. OpenRouter's authoritative `usage.cost`
+     * (real, routing+cache-adjusted charge) when the provider is
+     * `openrouter`; otherwise `estimateCostUSD` priced from the same
+     * OpenRouter-fetched table `TokenTracker` uses for the whole-run total.
+     * Absent — never `0` — when neither is available (pricing fetch
+     * pending, or the model isn't in OpenRouter's catalog).
+     */
+    costUSD?: number;
 }
 type LLMFinishReason = 'stop' | 'length' | 'content_filter' | 'tool_calls' | null;
 interface LLMResponse<T> {

package/dist/client.d.ts

@@ -1,7 +1,7 @@
-import './rate-limiter-CXaf8aAy.js';
+import './rate-limiter-Bz0iSJgZ.js';
 import '@langchain/openai';
 import '@langchain/anthropic';
 import '@anthropic-ai/sdk';
 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 PROVIDER_CATALOG, r as ProviderCatalogEntry, s as ProviderConfig, V as VisionCallOptions, t as VisionImageMediaType, u as VisionImagePart, v as buildVisionMessageContent, w as createAnthropicClient, x as createCreativeClient, y as createDeepSeekClient, z as createFixClient, B as createKimiClient, E as createOpenAIClient, F as createOpenRouterClient, G as createRequirementsClient, H as createZhipuClient, I as getAvailableProvider, J as getSharedLLMClient, M as isProviderAvailable, N as listLmstudioModels, Q as listSelectableProviders, S as resetSharedLLMClient } from './client-SYMpnw-w.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 PROVIDER_CATALOG, r as ProviderCatalogEntry, s as ProviderConfig, V as VisionCallOptions, t as VisionImageMediaType, u as VisionImagePart, v as buildVisionMessageContent, w as createAnthropicClient, x as createCreativeClient, y as createDeepSeekClient, z as createFixClient, B as createKimiClient, E as createOpenAIClient, F as createOpenRouterClient, G as createRequirementsClient, H as createZhipuClient, I as getAvailableProvider, J as getSharedLLMClient, M as isProviderAvailable, N as listLmstudioModels, Q as listSelectableProviders, S as resetSharedLLMClient } from './client-DfzMDgkm.js';
 import '@almadar/core';

package/dist/client.js

@@ -22,9 +22,10 @@ import {
   listLmstudioModels,
   listSelectableProviders,
   resetSharedLLMClient
-} from "./chunk-RBT22YSX.js";
+} from "./chunk-DLEZ7FGQ.js";
 import "./chunk-P4VCT25B.js";
-import "./chunk-RPG3SUIB.js";
+import "./chunk-MOIECDMB.js";
+import "./chunk-T6AKOBX3.js";
 export {
   ANTHROPIC_MODELS,
   DEEPSEEK_MODELS,

package/dist/index.d.ts

@@ -1,11 +1,11 @@
-import { L as LLMFinishReason, a as LLMClient } from './client-SYMpnw-w.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 PROVIDER_CATALOG, r as ProviderCatalogEntry, s as ProviderConfig, V as VisionCallOptions, t as VisionImageMediaType, u as VisionImagePart, v as buildVisionMessageContent, w as createAnthropicClient, x as createCreativeClient, y as createDeepSeekClient, z as createFixClient, B as createKimiClient, E as createOpenAIClient, F as createOpenRouterClient, G as createRequirementsClient, H as createZhipuClient, I as getAvailableProvider, J as getSharedLLMClient, M as isProviderAvailable, N as listLmstudioModels, Q as listSelectableProviders, R as parseChatCompletionResponse, S as resetSharedLLMClient } from './client-SYMpnw-w.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-CXaf8aAy.js';
+import { L as LLMFinishReason, a as LLMClient } from './client-DfzMDgkm.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 PROVIDER_CATALOG, r as ProviderCatalogEntry, s as ProviderConfig, V as VisionCallOptions, t as VisionImageMediaType, u as VisionImagePart, v as buildVisionMessageContent, w as createAnthropicClient, x as createCreativeClient, y as createDeepSeekClient, z as createFixClient, B as createKimiClient, E as createOpenAIClient, F as createOpenRouterClient, G as createRequirementsClient, H as createZhipuClient, I as getAvailableProvider, J as getSharedLLMClient, M as isProviderAvailable, N as listLmstudioModels, Q as listSelectableProviders, R as parseChatCompletionResponse, S as resetSharedLLMClient } from './client-DfzMDgkm.js';
+export { E as EstimateCostTokens, a as RateLimiter, R as RateLimiterOptions, b as TokenTracker, T as TokenUsage, e as estimateCostUSD, g as getGlobalRateLimiter, c as getGlobalTokenTracker, r as resetGlobalRateLimiter, d as resetGlobalTokenTracker } from './rate-limiter-Bz0iSJgZ.js';
 import { JsonValue, ServiceParams, ServiceContract } from '@almadar/core';
 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';
-export { ErrorPrediction, GFlowNetResult, GoalSpec, MasarError, MasarGenerateOptions, MasarGenerateResult, MasarHealthResult, MasarProvider, MasarProviderOptions, PredictErrorsResult, RankEditsResult, RankedEdit, getMasarProvider, resetMasarProvider } from './providers/index.js';
+export { ErrorPrediction, GFlowNetResult, GoalSpec, JEV_DECISIONS_URL, JEV_MODELS, JevAnswer, JevChoiceAnswer, JevChoiceQuestion, JevDecideRequest, JevDecideResult, JevError, JevHealthResult, JevJsonValue, JevModelId, JevNoulAnswer, JevNoulQuestion, JevProvider, JevProviderOptions, JevQuestion, JevScoreAnswer, JevScoreQuestion, JevState, JevUsage, MasarError, MasarGenerateOptions, MasarGenerateResult, MasarHealthResult, MasarProvider, MasarProviderOptions, PredictErrorsResult, RankEditsResult, RankedEdit, getJevProvider, getMasarProvider, isJevAvailable, resetJevProvider, resetMasarProvider } from './providers/index.js';
 import '@langchain/openai';
 import '@langchain/anthropic';
 import '@anthropic-ai/sdk';
@@ -132,11 +132,13 @@ declare const LOCAL_IMAGE_MODELS: {
     readonly Z_IMAGE_TURBO: "z-image-turbo";
     readonly FLUX_SCHNELL: "flux-schnell";
     readonly FLUX_DEV: "flux-dev";
+    readonly FLUX_KONTEXT: "flux-kontext";
+    readonly FLUX2_KLEIN: "flux2-klein";
 };
 declare const LOCAL_IMAGE_DEFAULTS: {
     /** Z-Image-Turbo is the default: public on Hugging Face, no token required (FLUX is gated). */
-    readonly sheet: "z-image-turbo";
-    readonly preview: "z-image-turbo";
+    sheet: string;
+    preview: string;
 };
 interface ImageClientOptions {
     provider?: ImageProvider;
@@ -159,6 +161,10 @@ interface ImageGenerateOptions {
     quality?: 'auto' | 'low' | 'medium' | 'high';
     background?: 'auto' | 'transparent' | 'opaque';
     seed?: number;
+    /** Diffusion step count (local shim; 1–100). */
+    steps?: number;
+    /** img2img denoise strength 0–1 (local shim `image_strength`), used with `references`. */
+    strength?: number;
     references?: ReadonlyArray<ImageReference>;
     provider?: {
         order?: string[];

package/dist/index.js

@@ -23,7 +23,7 @@ import {
   listSelectableProviders,
   parseChatCompletionResponse,
   resetSharedLLMClient
-} from "./chunk-RBT22YSX.js";
+} from "./chunk-DLEZ7FGQ.js";
 import {
   autoCloseJson,
   extractJsonFromText,
@@ -37,21 +37,31 @@ import {
   getStructuredOutputClient,
   isStructuredOutputAvailable,
   resetStructuredOutputClient
-} from "./chunk-5AM54OAM.js";
+} from "./chunk-SJE3GTGZ.js";
 import {
   RateLimiter,
-  TokenTracker,
   getGlobalRateLimiter,
-  getGlobalTokenTracker,
-  resetGlobalRateLimiter,
-  resetGlobalTokenTracker
-} from "./chunk-RPG3SUIB.js";
+  resetGlobalRateLimiter
+} from "./chunk-MOIECDMB.js";
 import {
+  JEV_DECISIONS_URL,
+  JEV_MODELS,
+  JevError,
+  JevProvider,
   MasarError,
   MasarProvider,
+  getJevProvider,
   getMasarProvider,
+  isJevAvailable,
+  resetJevProvider,
   resetMasarProvider
-} from "./chunk-MUTXGY6D.js";
+} from "./chunk-OXFWONZP.js";
+import {
+  TokenTracker,
+  estimateCostUSD,
+  getGlobalTokenTracker,
+  resetGlobalTokenTracker
+} from "./chunk-T6AKOBX3.js";
 
 // src/embedding-client.ts
 var EMBEDDING_PROVIDERS = ["openai", "openrouter", "lmstudio"];
@@ -219,12 +229,14 @@ var OPENROUTER_IMAGE_DEFAULTS = {
 var LOCAL_IMAGE_MODELS = {
   Z_IMAGE_TURBO: "z-image-turbo",
   FLUX_SCHNELL: "flux-schnell",
-  FLUX_DEV: "flux-dev"
+  FLUX_DEV: "flux-dev",
+  FLUX_KONTEXT: "flux-kontext",
+  FLUX2_KLEIN: "flux2-klein"
 };
 var LOCAL_IMAGE_DEFAULTS = {
   /** Z-Image-Turbo is the default: public on Hugging Face, no token required (FLUX is gated). */
-  sheet: LOCAL_IMAGE_MODELS.Z_IMAGE_TURBO,
-  preview: LOCAL_IMAGE_MODELS.Z_IMAGE_TURBO
+  sheet: process.env.LOCAL_IMAGE_SHEET_MODEL ?? LOCAL_IMAGE_MODELS.Z_IMAGE_TURBO,
+  preview: process.env.LOCAL_IMAGE_PREVIEW_MODEL ?? LOCAL_IMAGE_MODELS.Z_IMAGE_TURBO
 };
 var IMAGE_MAX_ATTEMPTS = 3;
 var IMAGE_RETRY_BACKOFF_MS = [500, 2e3];
@@ -243,6 +255,8 @@ var API_KEY_ENV_VARS2 = {
 var IMAGE_CHECKED_PARAM_NAMES = {
   background: "background",
   seed: "seed",
+  steps: "steps",
+  strength: "image_strength",
   references: "input_references"
 };
 function isJsonObject(value) {
@@ -315,7 +329,11 @@ var _ImageClient = class _ImageClient {
     const cached = this.capabilitiesCache.get(model);
     if (cached) return cached;
     if (this.provider === "local") {
-      const local = { model, supportedParameters: /* @__PURE__ */ new Set(["seed", IMAGE_CHECKED_PARAM_NAMES.references]), raw: {} };
+      const local = {
+        model,
+        supportedParameters: /* @__PURE__ */ new Set(["seed", IMAGE_CHECKED_PARAM_NAMES.steps, IMAGE_CHECKED_PARAM_NAMES.strength, IMAGE_CHECKED_PARAM_NAMES.references]),
+        raw: {}
+      };
       this.capabilitiesCache.set(model, local);
       return local;
     }
@@ -353,6 +371,14 @@ var _ImageClient = class _ImageClient {
       delete filtered.seed;
       dropped.push(IMAGE_CHECKED_PARAM_NAMES.seed);
     }
+    if (filtered.steps !== void 0 && !caps.supportedParameters.has(IMAGE_CHECKED_PARAM_NAMES.steps)) {
+      delete filtered.steps;
+      dropped.push(IMAGE_CHECKED_PARAM_NAMES.steps);
+    }
+    if (filtered.strength !== void 0 && !caps.supportedParameters.has(IMAGE_CHECKED_PARAM_NAMES.strength)) {
+      delete filtered.strength;
+      dropped.push(IMAGE_CHECKED_PARAM_NAMES.strength);
+    }
     if (filtered.references !== void 0 && filtered.references.length > 0 && !caps.supportedParameters.has(IMAGE_CHECKED_PARAM_NAMES.references)) {
       delete filtered.references;
       dropped.push(IMAGE_CHECKED_PARAM_NAMES.references);
@@ -375,6 +401,7 @@ var _ImageClient = class _ImageClient {
     if (opts.quality !== void 0) requestBody.quality = opts.quality;
     if (opts.background !== void 0) requestBody.background = opts.background;
     if (opts.seed !== void 0) requestBody.seed = opts.seed;
+    if (opts.steps !== void 0) requestBody.steps = opts.steps;
     if (opts.references !== void 0 && opts.references.length > 0) {
       requestBody.input_references = opts.references.map((r) => ({ type: "image_url", image_url: { url: r.url } }));
     }
@@ -398,6 +425,8 @@ var _ImageClient = class _ImageClient {
     if (opts.n !== void 0) requestBody.n = opts.n;
     if (opts.size !== void 0) requestBody.size = opts.size;
     if (opts.seed !== void 0) requestBody.seed = opts.seed;
+    if (opts.steps !== void 0) requestBody.steps = opts.steps;
+    if (opts.strength !== void 0) requestBody.image_strength = opts.strength;
     if (opts.references !== void 0 && opts.references.length > 0) {
       requestBody.input_references = opts.references.map((r) => ({ type: "image_url", image_url: { url: r.url } }));
     }
@@ -910,6 +939,10 @@ export {
   EmbeddingClient,
   IMAGE_PROVIDER_CATALOG,
   ImageClient,
+  JEV_DECISIONS_URL,
+  JEV_MODELS,
+  JevError,
+  JevProvider,
   KIMI_MODELS,
   LLMClient,
   LOCAL_IMAGE_DEFAULTS,
@@ -940,14 +973,17 @@ export {
   createRequirementsClient,
   createZhipuClient,
   detectTruncation,
+  estimateCostUSD,
   extractJsonFromText,
   findLastCompleteElement,
   getAvailableProvider,
   getGlobalRateLimiter,
   getGlobalTokenTracker,
+  getJevProvider,
   getMasarProvider,
   getSharedLLMClient,
   getStructuredOutputClient,
+  isJevAvailable,
   isLikelyTruncated,
   isProviderAvailable,
   isStructuredOutputAvailable,
@@ -959,6 +995,7 @@ export {
   parseJsonResponse,
   resetGlobalRateLimiter,
   resetGlobalTokenTracker,
+  resetJevProvider,
   resetMasarProvider,
   resetSharedLLMClient,
   resetStructuredOutputClient,