@hebo-ai/gateway 0.11.2 → 0.11.3

package/dist/endpoints/chat-completions/otel.js

@@ -160,6 +160,7 @@ export const getChatResponseAttributes = (completions, signalLevel) => {
             "gen_ai.usage.total_tokens": completions.usage?.total_tokens,
             "gen_ai.usage.input_tokens": completions.usage?.prompt_tokens,
             "gen_ai.usage.cache_read.input_tokens": completions.usage?.prompt_tokens_details?.cached_tokens,
+            "gen_ai.usage.cache_creation.input_tokens": completions.usage?.prompt_tokens_details?.cache_write_tokens,
             "gen_ai.usage.output_tokens": completions.usage?.completion_tokens,
             "gen_ai.usage.reasoning.output_tokens": completions.usage?.completion_tokens_details?.reasoning_tokens,
         });

package/dist/endpoints/messages/converters.js

@@ -625,14 +625,12 @@ export class MessagesTransformStream extends TransformStream {
                     }
                     case "finish": {
                         const stopReason = mapStopReason(part.finishReason);
-                        const totalOutputTokens = part.totalUsage?.outputTokens ?? 0;
-                        const totalInputTokens = part.totalUsage?.inputTokens ?? 0;
                         controller.enqueue({
                             event: "message_delta",
                             data: {
                                 type: "message_delta",
                                 delta: { stop_reason: stopReason, stop_sequence: null },
-                                usage: { output_tokens: totalOutputTokens, input_tokens: totalInputTokens },
+                                usage: mapUsage(part.totalUsage),
                             },
                         });
                         controller.enqueue({

package/dist/endpoints/messages/schema.d.ts

@@ -647,10 +647,7 @@ export type MessageDeltaEvent = SseFrame<{
         stop_reason: MessagesStopReason;
         stop_sequence: string | null;
     };
-    usage: {
-        output_tokens: number;
-        input_tokens?: number;
-    };
+    usage: MessagesUsage;
 }, "message_delta">;
 export type MessageStopEvent = SseFrame<{
     type: "message_stop";

package/dist/errors/ai-sdk.js

@@ -6,10 +6,19 @@ const normalizeApiCallError = (error) => {
     const statusText = `UPSTREAM_${STATUS_TEXT(status)}`;
     return new GatewayError(error, status, statusText, undefined, error.responseHeaders ?? undefined);
 };
+// `AbortError` / `TimeoutError` (raised by the AI SDK's internal `timeout` controller,
+// `AbortSignal.timeout`, or an aborted upstream `fetch`) reach us as plain DOMExceptions
+// that none of the AI SDK error classes match. Treat them as upstream gateway timeouts
+// so they surface as 504 with retry headers rather than defaulting to 500/502.
+// Inbound client disconnects are caught earlier in `lifecycle.ts` and overridden to 499.
+const isUpstreamAbortError = (error) => error instanceof Error && (error.name === "AbortError" || error.name === "TimeoutError");
 export const normalizeAiSdkError = (error) => {
     if (APICallError.isInstance(error)) {
         return normalizeApiCallError(error);
     }
+    if (isUpstreamAbortError(error)) {
+        return new GatewayError(error, 504, `UPSTREAM_${STATUS_TEXT(504)}`);
+    }
     if (RetryError.isInstance(error)) {
         if (APICallError.isInstance(error.lastError)) {
             return normalizeApiCallError(error.lastError);

package/dist/lifecycle.js

@@ -50,11 +50,12 @@ export const winterCgHandler = (run, config) => {
             else if (status === 200 && ctx.response?.status)
                 realStatus = ctx.response.status;
             if (realStatus !== 200) {
+                const err = reason ?? ctx.request.signal.reason;
                 logger[realStatus >= 500 ? "error" : "warn"]({
                     requestId: ctx.requestId,
-                    err: reason ?? ctx.request.signal.reason,
+                    err,
                 });
-                span.recordError(reason, true);
+                span.recordError(err, true);
             }
             span.setAttributes({ "http.response.status_code_effective": realStatus });
             if (ctx.operation === "chat" ||

package/dist/telemetry/gen-ai.js

@@ -1,5 +1,6 @@
 import { metrics } from "@opentelemetry/api";
 import { STATUS_TEXT } from "../errors/utils";
+import { logger } from "../logger";
 const getMeter = () => metrics.getMeter("@hebo/gateway");
 let requestDurationHistogram;
 let timePerOutputTokenHistogram;
@@ -93,20 +94,59 @@ export const recordTimePerOutputToken = (start, ttft, tokenAttrs, metricAttrs, s
         return;
     getTimePerOutputTokenHistogram().record((performance.now() - start - ttft) / 1000 / (outputTokens - 1), metricAttrs);
 };
+// Partitioning follows OTel semconv PR #3624:
+// https://github.com/open-telemetry/semantic-conventions/pull/3624
+// When a cache or reasoning breakdown is reported, partitioned data points sum
+// to the total and a bare {type} point MUST NOT be emitted alongside them.
 // FUTURE: record unsuccessful calls
 export const recordTokenUsage = (tokenAttrs, metricAttrs, signalLevel) => {
-    if (!signalLevel || (signalLevel !== "recommended" && signalLevel !== "full"))
+    if (signalLevel !== "recommended" && signalLevel !== "full")
         return;
-    const record = (value, tokenType) => {
-        if (typeof value !== "number")
-            return;
-        getTokenUsageHistogram().record(value, Object.assign({}, metricAttrs, { "gen_ai.token.type": tokenType }));
+    const histogram = getTokenUsageHistogram();
+    const emit = (value, extra) => {
+        if (value > 0)
+            histogram.record(value, { ...metricAttrs, ...extra });
     };
-    record(tokenAttrs["gen_ai.usage.input_tokens"], "input");
-    record(tokenAttrs["gen_ai.usage.output_tokens"], "output");
-    // FUTURE: "cached" and "reasoning" token types are not yet in the OTel standard — monitor:
-    // https://github.com/open-telemetry/semantic-conventions/issues/1959
-    // https://github.com/open-telemetry/semantic-conventions/issues/3341
-    record(tokenAttrs["gen_ai.usage.cache_read.input_tokens"], "cached");
-    record(tokenAttrs["gen_ai.usage.reasoning.output_tokens"], "reasoning");
+    emitInputTokens(emit, tokenAttrs);
+    emitOutputTokens(emit, tokenAttrs);
+};
+const emitInputTokens = (emit, tokenAttrs) => {
+    const total = tokenAttrs["gen_ai.usage.input_tokens"];
+    if (total === undefined)
+        return;
+    const cacheRead = tokenAttrs["gen_ai.usage.cache_read.input_tokens"];
+    const cacheCreation = tokenAttrs["gen_ai.usage.cache_creation.input_tokens"];
+    if (cacheRead === undefined && cacheCreation === undefined) {
+        emit(total, { "gen_ai.token.type": "input" });
+        return;
+    }
+    const read = cacheRead ?? 0;
+    const creation = cacheCreation ?? 0;
+    let uncached = total - read - creation;
+    if (uncached < 0) {
+        logger.warn({ inputTokens: total, cacheRead: read, cacheCreation: creation }, "[telemetry] input token cache partitions exceed total; clamping uncached to 0");
+        uncached = 0;
+    }
+    emit(read, { "gen_ai.token.type": "input", "gen_ai.token.cache": "read" });
+    emit(creation, { "gen_ai.token.type": "input", "gen_ai.token.cache": "creation" });
+    emit(uncached, { "gen_ai.token.type": "input", "gen_ai.token.cache": "uncached" });
+};
+const emitOutputTokens = (emit, tokenAttrs) => {
+    const total = tokenAttrs["gen_ai.usage.output_tokens"];
+    if (total === undefined)
+        return;
+    const reasoning = tokenAttrs["gen_ai.usage.reasoning.output_tokens"];
+    if (reasoning === undefined) {
+        emit(total, { "gen_ai.token.type": "output" });
+        return;
+    }
+    let reasoned = reasoning;
+    let nonReasoning = total - reasoning;
+    if (nonReasoning < 0) {
+        logger.warn({ outputTokens: total, reasoningTokens: reasoning }, "[telemetry] reasoning tokens exceed output total; clamping non-reasoning to 0");
+        reasoned = total;
+        nonReasoning = 0;
+    }
+    emit(reasoned, { "gen_ai.token.type": "output", "gen_ai.token.reasoning": true });
+    emit(nonReasoning, { "gen_ai.token.type": "output", "gen_ai.token.reasoning": false });
 };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hebo-ai/gateway",
-  "version": "0.11.2",
+  "version": "0.11.3",
   "description": "AI gateway as a framework. For full control over models, routing & lifecycle. OpenAI /chat/completions, OpenResponses /responses & Anthropic /messages.",
   "keywords": [
     "ai",
@@ -252,36 +252,36 @@
     "fix": "bun lint:staged && bun format:staged"
   },
   "dependencies": {
-    "@ai-sdk/provider": "^3.0.8",
-    "ai": "^6.0.168",
+    "@ai-sdk/provider": "^3.0.9",
+    "ai": "^6.0.169",
     "lru-cache": "^11.3.5",
     "uuid": "^14.0.0",
     "zod": "^4.3.6"
   },
   "devDependencies": {
-    "@ai-sdk/alibaba": "^1.0.17",
-    "@ai-sdk/amazon-bedrock": "^4.0.96",
-    "@ai-sdk/anthropic": "^3.0.71",
-    "@ai-sdk/cohere": "^3.0.30",
-    "@ai-sdk/deepinfra": "^2.0.45",
-    "@ai-sdk/deepseek": "^2.0.29",
-    "@ai-sdk/fireworks": "^2.0.46",
-    "@ai-sdk/google-vertex": "^4.0.112",
-    "@ai-sdk/groq": "^3.0.35",
-    "@ai-sdk/moonshotai": "^2.0.16",
-    "@ai-sdk/openai": "^3.0.53",
-    "@ai-sdk/togetherai": "^2.0.45",
-    "@ai-sdk/xai": "^3.0.83",
+    "@ai-sdk/alibaba": "^1.0.18",
+    "@ai-sdk/amazon-bedrock": "^4.0.97",
+    "@ai-sdk/anthropic": "^3.0.72",
+    "@ai-sdk/cohere": "^3.0.31",
+    "@ai-sdk/deepinfra": "^2.0.46",
+    "@ai-sdk/deepseek": "^2.0.30",
+    "@ai-sdk/fireworks": "^2.0.47",
+    "@ai-sdk/google-vertex": "^4.0.113",
+    "@ai-sdk/groq": "^3.0.36",
+    "@ai-sdk/moonshotai": "^2.0.17",
+    "@ai-sdk/openai": "^3.0.54",
+    "@ai-sdk/togetherai": "^2.0.46",
+    "@ai-sdk/xai": "^3.0.84",
     "@anthropic-ai/sdk": "^0.91.1",
-    "@aws-sdk/credential-providers": "^3.1037.0",
+    "@aws-sdk/credential-providers": "^3.1038.0",
     "@langfuse/otel": "^5.2.0",
     "@libsql/client": "^0.17.3",
     "@mjackson/node-fetch-server": "^0.7.0",
     "@opentelemetry/api": "^1.9.1",
     "@opentelemetry/context-async-hooks": "^2.7.0",
     "@opentelemetry/sdk-trace-base": "^2.7.0",
-    "@tanstack/react-router": "^1.168.24",
-    "@tanstack/react-start": "^1.167.49",
+    "@tanstack/react-router": "^1.168.25",
+    "@tanstack/react-start": "^1.167.50",
     "@types/better-sqlite3": "^7.6.13",
     "@types/bun": "1.3.13",
     "@types/pg": "^8.20.0",
@@ -294,10 +294,10 @@
     "lefthook": "^2.1.6",
     "mysql2": "^3.22.3",
     "next": "^16.2.4",
-    "openai": "^6.34.0",
+    "openai": "^6.35.0",
     "oxfmt": "^0.46.0",
-    "oxlint": "^1.61.0",
-    "oxlint-tsgolint": "^0.22.0",
+    "oxlint": "^1.62.0",
+    "oxlint-tsgolint": "^0.22.1",
     "pg": "^8.20.0",
     "pino": "^10.3.1",
     "postgres": "^3.4.9",
@@ -315,12 +315,12 @@
     "@ai-sdk/deepinfra": "^2.0.45",
     "@ai-sdk/deepseek": "^2.0.29",
     "@ai-sdk/fireworks": "^2.0.46",
-    "@ai-sdk/google": "^3.0.64",
+    "@ai-sdk/google": "^3.0.65",
     "@ai-sdk/google-vertex": "^4.0.80",
     "@ai-sdk/groq": "^3.0.29",
     "@ai-sdk/moonshotai": "^2.0.16",
     "@ai-sdk/openai": "^3.0.41",
-    "@ai-sdk/openai-compatible": "^2.0.41",
+    "@ai-sdk/openai-compatible": "^2.0.42",
     "@ai-sdk/togetherai": "^2.0.45",
     "@ai-sdk/xai": "^3.0.83",
     "@libsql/client": "^0.14.0",