ai-lcr 0.5.0 → 0.5.2

package/README.md

@@ -15,7 +15,7 @@
 </p>
 
 <p align="center">
-  <img src="assets/ai-lcr-hero.svg" alt="ai-lcr routes each model to its own cheapest provider — Gemini to Kunavo, DeepSeek to OpenRouter, Seedream to fal, Flux Schnell to Runware — and falls back on failure" width="820">
+  <img src="assets/ai-lcr-hero.svg" alt="ai-lcr keeps a cheapest-first list of providers per model — serves the cheapest (saving ~40%), fails over to the next on error, and snaps back to the cheapest after ~60s" width="720">
 </p>
 
 The same model costs different amounts on different providers — and no single provider is cheapest for everything. `ai-lcr` keeps a cheapest-first list per model, routes to the cheapest healthy one (⭐ below), and falls through on failure — the way phone carriers have done [Least Cost Routing](https://en.wikipedia.org/wiki/Least-cost_routing) for decades.
@@ -144,10 +144,6 @@ DeepInfra carries open weights only — no first-party Claude / GPT / Gemini. Fo
 2. **Fall through on failure.** On a retryable error — rate limit, 5xx, timeout, or a **billing cap** (402 / out-of-credit / quota) — it advances to the next provider, streaming-safe. A caller's own bad request (e.g. 400, 422) passes through immediately.
 3. **Recover.** After an idle window (`resetIntervalMs`, default 60s) it snaps back to the cheapest provider.
 
-<p align="center">
-  <img src="assets/ai-lcr-routing.svg" alt="routing diagram: cheapest first, fallback on failure, recover after idle" width="820">
-</p>
-
 ## See what happened (`onCall`)
 
 `onError`/`onCost` fire separately and uncorrelated, so a failover is hard to read after the fact. `onCall` gives you **one record per request** — the full chain, the winner, the reason for each failed hop, latency, and cost — and `formatCallRecord` turns it into a one-liner you can scan:

package/README.zh-CN.md

@@ -15,7 +15,7 @@
 </p>
 
 <p align="center">
-  <img src="assets/ai-lcr-hero.svg" alt="ai-lcr 把每个模型路由到各自最便宜的 provider——Gemini 走 Kunavo，DeepSeek 走 OpenRouter，Seedream 走 fal，Flux Schnell 走 Runware——失败时自动 fallback" width="820">
+  <img src="assets/ai-lcr-hero.svg" alt="ai-lcr 为每个模型维护一份「最便宜优先」的 provider 列表——默认走最便宜的（省约 40%），出错时切到下一个，约 60 秒后自动切回最便宜" width="720">
 </p>
 
 同一个模型在不同 provider 上的价格不同——而且没有任何单一 provider 在所有模型上都最便宜。`ai-lcr` 为每个模型维护一份「最便宜优先」的列表，路由到其中最便宜且健康的 provider（下表中的 ⭐），失败时向下穿透——这正是电话运营商几十年来一直在做的 [最低成本路由（Least Cost Routing）](https://en.wikipedia.org/wiki/Least-cost_routing)。
@@ -144,10 +144,6 @@ DeepInfra 只承载开源权重——没有第一方 Claude / GPT / Gemini。那
 2. **失败时向下穿透。** 遇到可重试的错误（限流、5xx、超时）时，前进到下一个 provider，且对流式安全。硬错误（400、401、403、422）会直接透传，不做重试。
 3. **恢复。** 在一段空闲窗口（`resetIntervalMs`，默认 60s）之后，自动回到最便宜的 provider。
 
-<p align="center">
-  <img src="assets/ai-lcr-routing.svg" alt="路由示意图：最便宜优先、失败时 fallback、空闲后恢复" width="820">
-</p>
-
 ## 支持的 provider
 
 任何 OpenAI 兼容的 endpoint 都可用——任何 AI SDK 的 provider 包也都可用，包括模型厂商自己的官方 API。

package/dist/index.cjs

@@ -41,6 +41,12 @@ __export(index_exports, {
 module.exports = __toCommonJS(index_exports);
 
 // src/fallback.ts
+var EmptyCompletionError = class extends Error {
+  constructor(provider) {
+    super(`ai-lcr: provider "${provider}" returned an empty completion (0 output tokens, no content)`);
+    this.name = "EmptyCompletionError";
+  }
+};
 var RETRYABLE_STATUS = /* @__PURE__ */ new Set([401, 402, 403, 408, 409, 413, 429, 498, 500]);
 var RETRYABLE_PATTERNS = [
   "overloaded",
@@ -153,6 +159,7 @@ function isRetryableError(error) {
   return RETRYABLE_PATTERNS.some((p) => text.includes(p));
 }
 function classifyError(error) {
+  if (error instanceof EmptyCompletionError) return "empty_completion";
   const e = error;
   const status = e?.statusCode ?? e?.status;
   if (typeof status === "number") return String(status);
@@ -175,6 +182,7 @@ var BILLING_PATTERNS = [
   "\u6263\u6B3E"
 ];
 function classifyErrorKind(error) {
+  if (error instanceof EmptyCompletionError) return "empty";
   const e = error;
   const status = e?.statusCode ?? e?.status;
   const { text } = errorSignals(error);
@@ -194,10 +202,27 @@ function costForUsage(cost, inputTokens, outputTokens, cacheReadTokens) {
   const cachedRate = cost.cacheRead ?? cost.input;
   return fullInput / 1e6 * cost.input + cached / 1e6 * cachedRate + outputTokens / 1e6 * cost.output;
 }
+function cacheSavingForUsage(cost, inputTokens, cacheReadTokens) {
+  if (cost.cacheRead === void 0) return 0;
+  const cached = Math.min(Math.max(cacheReadTokens, 0), inputTokens);
+  return cached / 1e6 * (cost.input - cost.cacheRead);
+}
 function requestIdFrom(options) {
   const raw = options.providerOptions?.lcr?.requestId;
   return typeof raw === "string" && raw.length > 0 ? raw : void 0;
 }
+var CONTENT_PART_TYPES = /* @__PURE__ */ new Set([
+  "text-delta",
+  "reasoning-delta",
+  "tool-call",
+  "tool-input-start",
+  "tool-input-delta",
+  "tool-input-end",
+  "file",
+  "source",
+  "tool-result",
+  "raw"
+]);
 var LcrFallbackModel = class {
   constructor(opts) {
     this.opts = opts;
@@ -298,19 +323,22 @@ var LcrFallbackModel = class {
     });
   }
   /**
-   * Baseline = what this same usage would have cost on the most expensive
-   * *priced* provider in the chain (typically the OpenRouter fallback leg). The
-   * winner's savings is `baselineUsd - costUsd`. Undefined when no provider in
-   * the chain carries a price (nothing to compare against).
+   * Baseline = what this same usage would have cost on the always-on fallback:
+   * the LAST priced leg of the chain (by convention the list-price provider you'd
+   * use without routing — e.g. OpenRouter, always last). The winner's saving is
+   * `baselineUsd - costUsd`. We take the last priced leg, NOT the most expensive
+   * one: prompt caching can make a sticker-cheaper provider (no `cacheRead`) cost
+   * MORE on a cache-heavy call, and a max-of-chain baseline would then fabricate a
+   * "saving" even on calls the fallback itself served. Undefined when no provider
+   * in the chain carries a price (nothing to compare against).
    */
   baselineUsd(inputTokens, outputTokens, cacheReadTokens) {
-    let max;
+    let baseline;
     for (const p of this.opts.providers) {
       if (!p.cost) continue;
-      const c = costForUsage(p.cost, inputTokens, outputTokens, cacheReadTokens);
-      if (max === void 0 || c > max) max = c;
+      baseline = costForUsage(p.cost, inputTokens, outputTokens, cacheReadTokens);
     }
-    return max;
+    return baseline;
   }
   /** Winner settled: record the attempt, fire `onCost` (compat) + `onCall`. */
   finalizeOk(ctx, provider, attemptStart, usage, ttftMs) {
@@ -319,7 +347,9 @@ var LcrFallbackModel = class {
     const outputTokens = usage?.outputTokens?.total ?? 0;
     const cacheReadTokens = usage?.inputTokens?.cacheRead ?? 0;
     const costUsd = provider.cost ? costForUsage(provider.cost, inputTokens, outputTokens, cacheReadTokens) : 0;
+    const cachedSavingUsd = provider.cost ? cacheSavingForUsage(provider.cost, inputTokens, cacheReadTokens) : 0;
     const usageMissing = inputTokens === 0 && outputTokens === 0;
+    const emptyCompletion = inputTokens > 0 && outputTokens === 0;
     this.emitCost({
       model: this.opts.modelName,
       provider: provider.label,
@@ -341,8 +371,10 @@ var LcrFallbackModel = class {
       ...cacheReadTokens > 0 ? { cachedInputTokens: cacheReadTokens } : {},
       costUsd,
       baselineUsd: this.baselineUsd(inputTokens, outputTokens, cacheReadTokens),
+      ...cachedSavingUsd > 0 ? { cachedSavingUsd } : {},
       ...ctx.requestId ? { requestId: ctx.requestId } : {},
-      ...usageMissing ? { usageMissing: true } : {}
+      ...usageMissing ? { usageMissing: true } : {},
+      ...emptyCompletion ? { emptyCompletion: true } : {}
     });
   }
   /** Every provider failed: fire `onCall` with no winner. */
@@ -373,6 +405,15 @@ var LcrFallbackModel = class {
       const attemptStart = Date.now();
       try {
         const result = await provider.model.doGenerate(options);
+        const out = result.usage?.outputTokens?.total ?? 0;
+        const inp = result.usage?.inputTokens?.total ?? 0;
+        if (inp > 0 && out === 0 && tried < n - 1) {
+          const emptyErr = new EmptyCompletionError(provider.label);
+          lastError = emptyErr;
+          this.emitError(emptyErr, provider.label);
+          this.recordFail(ctx, provider, attemptStart, emptyErr);
+          continue;
+        }
         this.settleSticky(idx);
         this.finalizeOk(ctx, provider, attemptStart, result.usage);
         return result;
@@ -434,7 +475,7 @@ var LcrFallbackModel = class {
     const servingIdx = idx;
     const triedBeforeServing = tried;
     let usage;
-    let streamedAny = false;
+    let contentStreamed = false;
     let ttftMs;
     const stream = new ReadableStream({
       async start(controller) {
@@ -443,17 +484,24 @@ var LcrFallbackModel = class {
           reader = result.stream.getReader();
           for (; ; ) {
             const { done, value } = await reader.read();
-            if (!streamedAny && value && typeof value === "object" && "error" in value) {
+            if (!contentStreamed && value && typeof value === "object" && "error" in value) {
               const err = value.error;
               if (self.shouldRetry(err)) throw err;
             }
             if (done) break;
-            if (value.type === "finish") usage = value.usage;
+            if (value.type === "finish") {
+              usage = value.usage;
+              const out = value.usage?.outputTokens?.total ?? 0;
+              const inp = value.usage?.inputTokens?.total ?? 0;
+              if (inp > 0 && out === 0 && !contentStreamed && triedBeforeServing + 1 < n) {
+                throw new EmptyCompletionError(servingProvider.label);
+              }
+            }
             if (ttftMs === void 0 && (value.type === "text-delta" || value.type === "reasoning-delta")) {
               ttftMs = Date.now() - servingAttemptStart;
             }
             controller.enqueue(value);
-            if (value.type !== "stream-start") streamedAny = true;
+            if (CONTENT_PART_TYPES.has(value.type)) contentStreamed = true;
           }
           self.settleSticky(servingIdx);
           self.finalizeOk(ctx, servingProvider, servingAttemptStart, usage, ttftMs);
@@ -461,7 +509,7 @@ var LcrFallbackModel = class {
         } catch (error) {
           self.emitError(error, servingProvider.label);
           self.recordFail(ctx, servingProvider, servingAttemptStart, error);
-          if (!streamedAny) {
+          if (!contentStreamed) {
             const nextTried = triedBeforeServing + 1;
             if (nextTried >= n) {
               self.finalizeFail(ctx);
@@ -524,6 +572,7 @@ function formatCallRecord(record, opts = {}) {
     line += `  (saved $${(record.baselineUsd - record.costUsd).toFixed(4)})`;
   }
   if (record.usageMissing) line += `  \u26A0no-usage`;
+  if (record.emptyCompletion) line += `  \u26A0empty`;
   const failed = record.attempts.filter((a) => !a.ok);
   if (failed.length > 0) {
     const reasons = failed.map((a) => `${a.provider} ${a.errorClass ?? "error"}`).join(", ");

package/dist/index.d.cts

@@ -49,8 +49,15 @@ interface CostEvent {
  *   - "auth":      401 / 403 — a misconfigured or revoked key.
  *   - "billing":   402 / out-of-credit / quota — account needs topping up.
  *   - "client":    a non-retryable caller error (e.g. 400 bad request).
+ *   - "empty":     provider returned a clean 200 but generated nothing
+ *                  (zero output tokens, no content) — a *content*-integrity
+ *                  failure, not a transport one. The provider looks healthy to
+ *                  every status/network check yet hands the user a blank. We
+ *                  fail over on it like a transient error, but tag it separately
+ *                  so a run of `"empty"` attempts (a quietly degraded model)
+ *                  doesn't hide inside the transient noise.
  */
-type ErrorKind = "transient" | "auth" | "billing" | "client";
+type ErrorKind = "transient" | "auth" | "billing" | "client" | "empty";
 /** One provider attempt within a single request. */
 interface RouteAttempt {
     /** Provider label that was tried (e.g. "tokenmart"). */
@@ -109,12 +116,25 @@ interface CallRecord {
     /** Computed from the winner's `cost`; 0 if no price was given or the call failed. */
     costUsd: number;
     /**
-     * What the same request would have cost on the most expensive *priced*
-     * provider in the chain, on identical token usage — the savings baseline
-     * (`baselineUsd - costUsd`). Set by both routers whenever at least one
-     * provider carries a `cost`; undefined only when no provider was priced.
+     * What this same usage would have cost on the savings baseline, so
+     * `baselineUsd - costUsd` is what routing actually saved. Text router: the
+     * always-on fallback leg — the LAST priced provider in the chain, i.e. the
+     * list-price provider you'd fall back to without routing (e.g. OpenRouter).
+     * Media router: the model-maker's official direct price. NOT the most
+     * expensive leg of the chain: prompt caching can make a sticker-cheaper
+     * provider cost more on a cache-heavy call, and a max-of-chain baseline would
+     * fabricate a "saving" on calls the fallback itself served. Undefined only
+     * when no provider was priced.
      */
     baselineUsd?: number;
+    /**
+     * The slice of `costUsd` that prompt-cache reads saved versus paying the full
+     * input rate for those same tokens (`cachedTokens × (input − cacheRead)`).
+     * Present only when > 0. This is the serving provider's own caching benefit —
+     * it happens with or without routing — so it is NOT a routing saving and must
+     * be surfaced separately, never folded into `baselineUsd - costUsd`.
+     */
+    cachedSavingUsd?: number;
     /**
      * Caller-supplied correlation id, read from `providerOptions.lcr.requestId`
      * on the call. Multi-step tool loops emit one record per `doStream`/
@@ -129,6 +149,21 @@ interface CallRecord {
      * other signal. Treat a flagged record as "cost unknown", not "free".
      */
     usageMissing?: boolean;
+    /**
+     * True when the winner served a clean, error-free response that nonetheless
+     * generated **nothing**: zero output tokens with a non-empty prompt (and, for
+     * streams, not one content part). The user asked and got a blank. Distinct
+     * from {@link usageMissing} (which is input *and* output both zero — usage not
+     * reported); here the prompt was billed but the model produced no output.
+     *
+     * Set only when this empty response is what the caller actually received —
+     * i.e. every provider in the chain came back empty, so failover couldn't
+     * rescue it. (When an earlier provider returns empty but a later one produces
+     * content, that earlier attempt is recorded as a failed `empty_completion` hop
+     * and this flag stays unset, because the winner did produce output.) Alert on
+     * it: a provider that quietly returns blanks passes every health check.
+     */
+    emptyCompletion?: boolean;
 }
 /**
  * Normalize an error into a short, log-friendly class for {@link CallRecord}.

package/dist/index.d.ts

@@ -49,8 +49,15 @@ interface CostEvent {
  *   - "auth":      401 / 403 — a misconfigured or revoked key.
  *   - "billing":   402 / out-of-credit / quota — account needs topping up.
  *   - "client":    a non-retryable caller error (e.g. 400 bad request).
+ *   - "empty":     provider returned a clean 200 but generated nothing
+ *                  (zero output tokens, no content) — a *content*-integrity
+ *                  failure, not a transport one. The provider looks healthy to
+ *                  every status/network check yet hands the user a blank. We
+ *                  fail over on it like a transient error, but tag it separately
+ *                  so a run of `"empty"` attempts (a quietly degraded model)
+ *                  doesn't hide inside the transient noise.
  */
-type ErrorKind = "transient" | "auth" | "billing" | "client";
+type ErrorKind = "transient" | "auth" | "billing" | "client" | "empty";
 /** One provider attempt within a single request. */
 interface RouteAttempt {
     /** Provider label that was tried (e.g. "tokenmart"). */
@@ -109,12 +116,25 @@ interface CallRecord {
     /** Computed from the winner's `cost`; 0 if no price was given or the call failed. */
     costUsd: number;
     /**
-     * What the same request would have cost on the most expensive *priced*
-     * provider in the chain, on identical token usage — the savings baseline
-     * (`baselineUsd - costUsd`). Set by both routers whenever at least one
-     * provider carries a `cost`; undefined only when no provider was priced.
+     * What this same usage would have cost on the savings baseline, so
+     * `baselineUsd - costUsd` is what routing actually saved. Text router: the
+     * always-on fallback leg — the LAST priced provider in the chain, i.e. the
+     * list-price provider you'd fall back to without routing (e.g. OpenRouter).
+     * Media router: the model-maker's official direct price. NOT the most
+     * expensive leg of the chain: prompt caching can make a sticker-cheaper
+     * provider cost more on a cache-heavy call, and a max-of-chain baseline would
+     * fabricate a "saving" on calls the fallback itself served. Undefined only
+     * when no provider was priced.
      */
     baselineUsd?: number;
+    /**
+     * The slice of `costUsd` that prompt-cache reads saved versus paying the full
+     * input rate for those same tokens (`cachedTokens × (input − cacheRead)`).
+     * Present only when > 0. This is the serving provider's own caching benefit —
+     * it happens with or without routing — so it is NOT a routing saving and must
+     * be surfaced separately, never folded into `baselineUsd - costUsd`.
+     */
+    cachedSavingUsd?: number;
     /**
      * Caller-supplied correlation id, read from `providerOptions.lcr.requestId`
      * on the call. Multi-step tool loops emit one record per `doStream`/
@@ -129,6 +149,21 @@ interface CallRecord {
      * other signal. Treat a flagged record as "cost unknown", not "free".
      */
     usageMissing?: boolean;
+    /**
+     * True when the winner served a clean, error-free response that nonetheless
+     * generated **nothing**: zero output tokens with a non-empty prompt (and, for
+     * streams, not one content part). The user asked and got a blank. Distinct
+     * from {@link usageMissing} (which is input *and* output both zero — usage not
+     * reported); here the prompt was billed but the model produced no output.
+     *
+     * Set only when this empty response is what the caller actually received —
+     * i.e. every provider in the chain came back empty, so failover couldn't
+     * rescue it. (When an earlier provider returns empty but a later one produces
+     * content, that earlier attempt is recorded as a failed `empty_completion` hop
+     * and this flag stays unset, because the winner did produce output.) Alert on
+     * it: a provider that quietly returns blanks passes every health check.
+     */
+    emptyCompletion?: boolean;
 }
 /**
  * Normalize an error into a short, log-friendly class for {@link CallRecord}.

package/dist/index.js

@@ -1,4 +1,10 @@
 // src/fallback.ts
+var EmptyCompletionError = class extends Error {
+  constructor(provider) {
+    super(`ai-lcr: provider "${provider}" returned an empty completion (0 output tokens, no content)`);
+    this.name = "EmptyCompletionError";
+  }
+};
 var RETRYABLE_STATUS = /* @__PURE__ */ new Set([401, 402, 403, 408, 409, 413, 429, 498, 500]);
 var RETRYABLE_PATTERNS = [
   "overloaded",
@@ -111,6 +117,7 @@ function isRetryableError(error) {
   return RETRYABLE_PATTERNS.some((p) => text.includes(p));
 }
 function classifyError(error) {
+  if (error instanceof EmptyCompletionError) return "empty_completion";
   const e = error;
   const status = e?.statusCode ?? e?.status;
   if (typeof status === "number") return String(status);
@@ -133,6 +140,7 @@ var BILLING_PATTERNS = [
   "\u6263\u6B3E"
 ];
 function classifyErrorKind(error) {
+  if (error instanceof EmptyCompletionError) return "empty";
   const e = error;
   const status = e?.statusCode ?? e?.status;
   const { text } = errorSignals(error);
@@ -152,10 +160,27 @@ function costForUsage(cost, inputTokens, outputTokens, cacheReadTokens) {
   const cachedRate = cost.cacheRead ?? cost.input;
   return fullInput / 1e6 * cost.input + cached / 1e6 * cachedRate + outputTokens / 1e6 * cost.output;
 }
+function cacheSavingForUsage(cost, inputTokens, cacheReadTokens) {
+  if (cost.cacheRead === void 0) return 0;
+  const cached = Math.min(Math.max(cacheReadTokens, 0), inputTokens);
+  return cached / 1e6 * (cost.input - cost.cacheRead);
+}
 function requestIdFrom(options) {
   const raw = options.providerOptions?.lcr?.requestId;
   return typeof raw === "string" && raw.length > 0 ? raw : void 0;
 }
+var CONTENT_PART_TYPES = /* @__PURE__ */ new Set([
+  "text-delta",
+  "reasoning-delta",
+  "tool-call",
+  "tool-input-start",
+  "tool-input-delta",
+  "tool-input-end",
+  "file",
+  "source",
+  "tool-result",
+  "raw"
+]);
 var LcrFallbackModel = class {
   constructor(opts) {
     this.opts = opts;
@@ -256,19 +281,22 @@ var LcrFallbackModel = class {
     });
   }
   /**
-   * Baseline = what this same usage would have cost on the most expensive
-   * *priced* provider in the chain (typically the OpenRouter fallback leg). The
-   * winner's savings is `baselineUsd - costUsd`. Undefined when no provider in
-   * the chain carries a price (nothing to compare against).
+   * Baseline = what this same usage would have cost on the always-on fallback:
+   * the LAST priced leg of the chain (by convention the list-price provider you'd
+   * use without routing — e.g. OpenRouter, always last). The winner's saving is
+   * `baselineUsd - costUsd`. We take the last priced leg, NOT the most expensive
+   * one: prompt caching can make a sticker-cheaper provider (no `cacheRead`) cost
+   * MORE on a cache-heavy call, and a max-of-chain baseline would then fabricate a
+   * "saving" even on calls the fallback itself served. Undefined when no provider
+   * in the chain carries a price (nothing to compare against).
    */
   baselineUsd(inputTokens, outputTokens, cacheReadTokens) {
-    let max;
+    let baseline;
     for (const p of this.opts.providers) {
       if (!p.cost) continue;
-      const c = costForUsage(p.cost, inputTokens, outputTokens, cacheReadTokens);
-      if (max === void 0 || c > max) max = c;
+      baseline = costForUsage(p.cost, inputTokens, outputTokens, cacheReadTokens);
     }
-    return max;
+    return baseline;
   }
   /** Winner settled: record the attempt, fire `onCost` (compat) + `onCall`. */
   finalizeOk(ctx, provider, attemptStart, usage, ttftMs) {
@@ -277,7 +305,9 @@ var LcrFallbackModel = class {
     const outputTokens = usage?.outputTokens?.total ?? 0;
     const cacheReadTokens = usage?.inputTokens?.cacheRead ?? 0;
     const costUsd = provider.cost ? costForUsage(provider.cost, inputTokens, outputTokens, cacheReadTokens) : 0;
+    const cachedSavingUsd = provider.cost ? cacheSavingForUsage(provider.cost, inputTokens, cacheReadTokens) : 0;
     const usageMissing = inputTokens === 0 && outputTokens === 0;
+    const emptyCompletion = inputTokens > 0 && outputTokens === 0;
     this.emitCost({
       model: this.opts.modelName,
       provider: provider.label,
@@ -299,8 +329,10 @@ var LcrFallbackModel = class {
       ...cacheReadTokens > 0 ? { cachedInputTokens: cacheReadTokens } : {},
       costUsd,
       baselineUsd: this.baselineUsd(inputTokens, outputTokens, cacheReadTokens),
+      ...cachedSavingUsd > 0 ? { cachedSavingUsd } : {},
       ...ctx.requestId ? { requestId: ctx.requestId } : {},
-      ...usageMissing ? { usageMissing: true } : {}
+      ...usageMissing ? { usageMissing: true } : {},
+      ...emptyCompletion ? { emptyCompletion: true } : {}
     });
   }
   /** Every provider failed: fire `onCall` with no winner. */
@@ -331,6 +363,15 @@ var LcrFallbackModel = class {
       const attemptStart = Date.now();
       try {
         const result = await provider.model.doGenerate(options);
+        const out = result.usage?.outputTokens?.total ?? 0;
+        const inp = result.usage?.inputTokens?.total ?? 0;
+        if (inp > 0 && out === 0 && tried < n - 1) {
+          const emptyErr = new EmptyCompletionError(provider.label);
+          lastError = emptyErr;
+          this.emitError(emptyErr, provider.label);
+          this.recordFail(ctx, provider, attemptStart, emptyErr);
+          continue;
+        }
         this.settleSticky(idx);
         this.finalizeOk(ctx, provider, attemptStart, result.usage);
         return result;
@@ -392,7 +433,7 @@ var LcrFallbackModel = class {
     const servingIdx = idx;
     const triedBeforeServing = tried;
     let usage;
-    let streamedAny = false;
+    let contentStreamed = false;
     let ttftMs;
     const stream = new ReadableStream({
       async start(controller) {
@@ -401,17 +442,24 @@ var LcrFallbackModel = class {
           reader = result.stream.getReader();
           for (; ; ) {
             const { done, value } = await reader.read();
-            if (!streamedAny && value && typeof value === "object" && "error" in value) {
+            if (!contentStreamed && value && typeof value === "object" && "error" in value) {
               const err = value.error;
               if (self.shouldRetry(err)) throw err;
             }
             if (done) break;
-            if (value.type === "finish") usage = value.usage;
+            if (value.type === "finish") {
+              usage = value.usage;
+              const out = value.usage?.outputTokens?.total ?? 0;
+              const inp = value.usage?.inputTokens?.total ?? 0;
+              if (inp > 0 && out === 0 && !contentStreamed && triedBeforeServing + 1 < n) {
+                throw new EmptyCompletionError(servingProvider.label);
+              }
+            }
             if (ttftMs === void 0 && (value.type === "text-delta" || value.type === "reasoning-delta")) {
               ttftMs = Date.now() - servingAttemptStart;
             }
             controller.enqueue(value);
-            if (value.type !== "stream-start") streamedAny = true;
+            if (CONTENT_PART_TYPES.has(value.type)) contentStreamed = true;
           }
           self.settleSticky(servingIdx);
           self.finalizeOk(ctx, servingProvider, servingAttemptStart, usage, ttftMs);
@@ -419,7 +467,7 @@ var LcrFallbackModel = class {
         } catch (error) {
           self.emitError(error, servingProvider.label);
           self.recordFail(ctx, servingProvider, servingAttemptStart, error);
-          if (!streamedAny) {
+          if (!contentStreamed) {
             const nextTried = triedBeforeServing + 1;
             if (nextTried >= n) {
               self.finalizeFail(ctx);
@@ -482,6 +530,7 @@ function formatCallRecord(record, opts = {}) {
     line += `  (saved $${(record.baselineUsd - record.costUsd).toFixed(4)})`;
   }
   if (record.usageMissing) line += `  \u26A0no-usage`;
+  if (record.emptyCompletion) line += `  \u26A0empty`;
   const failed = record.attempts.filter((a) => !a.ok);
   if (failed.length > 0) {
     const reasons = failed.map((a) => `${a.provider} ${a.errorClass ?? "error"}`).join(", ");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai-lcr",
-  "version": "0.5.0",
+  "version": "0.5.2",
   "description": "Least Cost Routing for LLMs — route every model call to the cheapest available provider, fall back automatically, and track real cost. Built for the Vercel AI SDK.",
   "keywords": [
     "ai",