@elizaos/plugin-elizacloud 2.0.0-beta.1 → 2.0.3-beta.2

package/dist/models/index.js

@@ -19,7 +19,7 @@ var __require = /* @__PURE__ */ createRequire(import.meta.url);
 // src/utils/config.ts
 import { logger } from "@elizaos/core";
 import { DEFAULT_ELIZA_CLOUD_TEXT_MODEL } from "@elizaos/core";
-var DEFAULT_ELIZA_CLOUD_LARGE_MODEL = "deepseek/deepseek-v4-pro";
+var DEFAULT_ELIZA_CLOUD_LARGE_MODEL = "zai-glm-4.7";
 function getEnvValue(key) {
   if (typeof process === "undefined") {
     return;
@@ -109,9 +109,18 @@ function getExperimentalTelemetry(runtime) {
   const setting = getSetting(runtime, "ELIZAOS_CLOUD_EXPERIMENTAL_TELEMETRY", "false");
   return String(setting).toLowerCase() === "true";
 }
+function resolveCloudTimeoutMs(envKey, defaultMs) {
+  const raw = typeof process !== "undefined" ? process.env[envKey] : undefined;
+  if (raw === undefined || raw.trim() === "")
+    return defaultMs;
+  const parsed = Number.parseInt(raw, 10);
+  if (!Number.isFinite(parsed))
+    return defaultMs;
+  return parsed <= 0 ? undefined : parsed;
+}
 
 // src/utils/sdk-client.ts
-import { CloudApiClient, ElizaCloudClient } from "@elizaos/cloud-sdk";
+import { ElizaCloudClient } from "@elizaos/cloud-sdk";
 function trimTrailingSlash(value) {
   return value.replace(/\/+$/, "");
 }
@@ -126,7 +135,11 @@ function apiKeyForRuntime(runtime, embedding = false) {
 }
 function createCloudApiClient(runtime, embedding = false) {
   const baseUrl = embedding ? getEmbeddingBaseURL(runtime) : getBaseURL(runtime);
-  return new CloudApiClient(trimTrailingSlash(baseUrl), apiKeyForRuntime(runtime, embedding));
+  return new ElizaCloudClient({
+    apiBaseUrl: trimTrailingSlash(baseUrl),
+    baseUrl: apiBaseToSiteBaseUrl(baseUrl),
+    apiKey: apiKeyForRuntime(runtime, embedding)
+  }).v1;
 }
 function createElizaCloudClient(runtime) {
   const apiBaseUrl = trimTrailingSlash(getBaseURL(runtime));
@@ -141,7 +154,7 @@ function createElizaCloudClient(runtime) {
 import {
   EventType
 } from "@elizaos/core";
-function emitModelUsageEvent(runtime, type, _prompt, usage) {
+function emitModelUsageEvent(runtime, type, _prompt, usage, meta = {}) {
   const inputTokens = Number(usage.inputTokens || 0);
   const outputTokens = Number(usage.outputTokens || 0);
   const totalTokens = Number(usage.totalTokens != null ? usage.totalTokens : inputTokens + outputTokens);
@@ -153,14 +166,34 @@ function emitModelUsageEvent(runtime, type, _prompt, usage) {
       prompt: inputTokens,
       completion: outputTokens,
       total: totalTokens
-    }
+    },
+    ...meta.modelName ? { modelName: meta.modelName } : {},
+    ...typeof meta.costUsd === "number" && Number.isFinite(meta.costUsd) ? { costUsd: meta.costUsd } : {}
   };
   runtime.emitEvent(EventType.MODEL_USED, payload);
 }
 
 // src/models/embeddings.ts
-import { logger as logger2, ModelType, VECTOR_DIMS } from "@elizaos/core";
+import {
+  logger as logger2,
+  ModelType,
+  timeInferenceSpan,
+  VECTOR_DIMS
+} from "@elizaos/core";
 var MAX_BATCH_SIZE = 100;
+var EMBED_MAX_ATTEMPTS = 2;
+var EMBED_BACKOFF_BASE_MS = 1000;
+var EMBED_BACKOFF_CAP_MS = 8000;
+var EMBED_REQUEST_TIMEOUT_MS = 60000;
+function embeddingBackoffMs(attempt, retryAfterSec) {
+  const exp = EMBED_BACKOFF_BASE_MS * 2 ** attempt;
+  const serverHint = typeof retryAfterSec === "number" && retryAfterSec > 0 ? retryAfterSec * 1000 : 0;
+  const base = Math.min(EMBED_BACKOFF_CAP_MS, Math.max(exp, serverHint));
+  return Math.round(base * (1 + Math.random() * 0.25));
+}
+function sleep(ms) {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
 function extractRateLimitInfo(response) {
   return {
     remainingRequests: parseInt(response.headers.get("x-ratelimit-remaining-requests") || "", 10) || undefined,
@@ -182,16 +215,16 @@ function getEmbeddingConfig(runtime) {
   }
   return { embeddingModelName, embeddingDimension };
 }
-function createErrorVector(dimension, marker) {
+function createInitProbeVector(dimension) {
   const vector = Array(dimension).fill(0);
-  vector[0] = marker;
+  vector[0] = 0.1;
   return vector;
 }
 async function handleTextEmbedding(runtime, params) {
   const { embeddingDimension } = getEmbeddingConfig(runtime);
   if (params === null) {
     logger2.debug("Creating test embedding for initialization");
-    return createErrorVector(embeddingDimension, 0.1);
+    return createInitProbeVector(embeddingDimension);
   }
   let text;
   if (typeof params === "string") {
@@ -199,12 +232,10 @@ async function handleTextEmbedding(runtime, params) {
   } else if (typeof params === "object" && params.text) {
     text = params.text;
   } else {
-    logger2.warn("Invalid input format for embedding");
-    return createErrorVector(embeddingDimension, 0.2);
+    throw new Error("Invalid input format for embedding: expected string or { text: string }");
   }
   if (!text.trim()) {
-    logger2.warn("Empty text for embedding");
-    return createErrorVector(embeddingDimension, 0.3);
+    throw new Error("Cannot generate embedding for empty text");
   }
   const results = await handleBatchTextEmbedding(runtime, [text]);
   return results[0];
@@ -213,84 +244,75 @@ async function handleBatchTextEmbedding(runtime, texts) {
   const { embeddingModelName, embeddingDimension } = getEmbeddingConfig(runtime);
   const client = createCloudApiClient(runtime, true);
   if (!texts || texts.length === 0) {
-    logger2.warn("[BatchEmbeddings] Empty texts array");
     return [];
   }
   const validTexts = [];
-  const results = new Array(texts.length);
   for (let i = 0;i < texts.length; i++) {
     const text = texts[i]?.trim();
-    if (text) {
-      validTexts.push({ text, originalIndex: i });
-    } else {
-      results[i] = createErrorVector(embeddingDimension, 0.3);
+    if (!text) {
+      throw new Error(`Cannot generate embedding for empty text at index ${i}`);
     }
+    validTexts.push({ text, originalIndex: i });
   }
-  if (validTexts.length === 0) {
-    logger2.warn("[BatchEmbeddings] All texts were empty");
-    return results;
-  }
+  const results = new Array(texts.length);
   for (let batchStart = 0;batchStart < validTexts.length; batchStart += MAX_BATCH_SIZE) {
     const batchEnd = Math.min(batchStart + MAX_BATCH_SIZE, validTexts.length);
     const batch = validTexts.slice(batchStart, batchEnd);
     const batchTexts = batch.map((b) => b.text);
     logger2.info(`[BatchEmbeddings] Processing batch ${Math.floor(batchStart / MAX_BATCH_SIZE) + 1}/${Math.ceil(validTexts.length / MAX_BATCH_SIZE)}: ${batch.length} texts`);
     try {
-      const response = await client.requestRaw("POST", "/embeddings", {
-        json: {
-          model: embeddingModelName,
-          input: batchTexts
-        }
-      });
-      const rateLimitInfo = extractRateLimitInfo(response);
-      if (rateLimitInfo.remainingRequests !== undefined && rateLimitInfo.remainingRequests < 50) {
-        logger2.warn(`[BatchEmbeddings] Rate limit: ${rateLimitInfo.remainingRequests}/${rateLimitInfo.limitRequests} requests remaining`);
-      }
-      if (response.status === 429) {
-        const retryAfter = rateLimitInfo.retryAfter || 30;
-        logger2.warn(`[BatchEmbeddings] Rate limited, waiting ${retryAfter}s...`);
-        await new Promise((resolve) => setTimeout(resolve, retryAfter * 1000));
-        const retryResponse = await client.requestRaw("POST", "/embeddings", {
+      let response = null;
+      for (let attempt = 0;attempt < EMBED_MAX_ATTEMPTS; attempt++) {
+        const resp = await timeInferenceSpan("cloud.embedding", () => client.requestRaw("POST", "/embeddings", {
           json: {
             model: embeddingModelName,
-            input: batchTexts
-          }
-        });
-        if (!retryResponse.ok) {
-          logger2.error(`[BatchEmbeddings] Retry failed: ${retryResponse.status}`);
-          for (const item of batch) {
-            results[item.originalIndex] = createErrorVector(embeddingDimension, 0.4);
-          }
-          continue;
+            input: batchTexts,
+            dimensions: embeddingDimension
+          },
+          timeoutMs: EMBED_REQUEST_TIMEOUT_MS
+        }), { batch: batchTexts.length, attempt });
+        const rateLimitInfo = extractRateLimitInfo(resp);
+        if (rateLimitInfo.remainingRequests !== undefined && rateLimitInfo.remainingRequests < 50) {
+          logger2.warn(`[BatchEmbeddings] Rate limit: ${rateLimitInfo.remainingRequests}/${rateLimitInfo.limitRequests} requests remaining`);
         }
-        const retryData = await retryResponse.json();
-        if (retryData?.data) {
-          for (const item of retryData.data) {
-            const originalIndex = batch[item.index].originalIndex;
-            results[originalIndex] = item.embedding;
-          }
-          logger2.info(`[BatchEmbeddings] Retry successful for ${batch.length} embeddings`);
+        const transient = resp.status === 429 || resp.status === 502 || resp.status === 503 || resp.status === 504;
+        if (transient && attempt < EMBED_MAX_ATTEMPTS - 1) {
+          const delay = embeddingBackoffMs(attempt, rateLimitInfo.retryAfter);
+          logger2.warn(`[BatchEmbeddings] ${resp.status} (attempt ${attempt + 1}/${EMBED_MAX_ATTEMPTS}) — backing off ${delay}ms`);
+          await resp.text().catch(() => {
+            return;
+          });
+          await sleep(delay);
+          continue;
         }
-        continue;
+        response = resp;
+        break;
+      }
+      if (!response) {
+        throw new Error("[BatchEmbeddings] No response after retry loop");
       }
       if (!response.ok) {
-        logger2.error(`[BatchEmbeddings] API error: ${response.status} - ${response.statusText}`);
-        for (const item of batch) {
-          results[item.originalIndex] = createErrorVector(embeddingDimension, 0.4);
+        if (response.status === 401 || response.status === 403) {
+          throw new Error(`[BatchEmbeddings] Authentication failed (${response.status}). ` + `Check ELIZAOS_CLOUD_API_KEY or ELIZAOS_CLOUD_EMBEDDING_API_KEY — ` + `the current key is not authorized for the embedding endpoint.`);
         }
-        continue;
+        throw new Error(`[BatchEmbeddings] API error: ${response.status} ${response.statusText}`);
       }
       const data = await response.json();
       if (!data?.data || !Array.isArray(data.data)) {
-        logger2.error("[BatchEmbeddings] API returned invalid structure");
-        for (const item of batch) {
-          results[item.originalIndex] = createErrorVector(embeddingDimension, 0.5);
-        }
-        continue;
+        throw new Error("[BatchEmbeddings] API returned invalid response structure");
+      }
+      if (data.data.length !== batch.length) {
+        throw new Error(`[BatchEmbeddings] expected ${batch.length} embeddings, got ${data.data.length}`);
       }
       for (const item of data.data) {
-        const originalIndex = batch[item.index].originalIndex;
-        results[originalIndex] = item.embedding;
+        const slot = typeof item.index === "number" ? batch[item.index] : undefined;
+        if (!slot) {
+          throw new Error(`[BatchEmbeddings] response index out of range: ${String(item.index)} (batch size ${batch.length})`);
+        }
+        if (!Array.isArray(item.embedding) || item.embedding.length !== embeddingDimension) {
+          throw new Error(`[BatchEmbeddings] dimension mismatch: model returned ${Array.isArray(item.embedding) ? item.embedding.length : "non-array"}d but agent is configured for ${embeddingDimension}d`);
+        }
+        results[slot.originalIndex] = item.embedding;
       }
       if (data.usage) {
         const usage = {
@@ -300,13 +322,11 @@ async function handleBatchTextEmbedding(runtime, texts) {
         };
         emitModelUsageEvent(runtime, ModelType.TEXT_EMBEDDING, `batch:${batch.length}`, usage);
       }
-      logger2.debug(`[BatchEmbeddings] Got ${batch.length} embeddings (${embeddingDimension}d), remaining: ${rateLimitInfo.remainingRequests ?? "unknown"}`);
+      logger2.debug(`[BatchEmbeddings] Got ${batch.length} embeddings (${embeddingDimension}d)`);
     } catch (error) {
       const message = error instanceof Error ? error.message : String(error);
-      logger2.error(`[BatchEmbeddings] Error: ${message}`);
-      for (const item of batch) {
-        results[item.originalIndex] = createErrorVector(embeddingDimension, 0.6);
-      }
+      logger2.error(`[BatchEmbeddings] Batch failed: ${message}`);
+      throw error instanceof Error ? error : new Error(message);
     }
   }
   return results;
@@ -424,7 +444,10 @@ async function handleImageGeneration(runtime, params) {
 }
 async function handleImageDescription(runtime, params) {
   const disableSetting = getSetting(runtime, "DISABLE_IMAGE_DESCRIPTION", "");
-  const disabled = disableSetting === "true" || disableSetting === "1" || process.env.DISABLE_IMAGE_DESCRIPTION === "true" || process.env.DISABLE_IMAGE_DESCRIPTION === "1";
+  const disabled = [disableSetting, process.env.DISABLE_IMAGE_DESCRIPTION].some((value) => {
+    const normalized = value?.trim().toLowerCase();
+    return normalized === "1" || normalized === "true" || normalized === "yes" || normalized === "on";
+  });
   if (disabled) {
     logger4.debug("[ELIZAOS_CLOUD] IMAGE_DESCRIPTION skipped — DISABLE_IMAGE_DESCRIPTION is set");
     return {
@@ -463,16 +486,21 @@ async function handleImageDescription(runtime, params) {
     let response = null;
     let attemptedRetry = false;
     for (let attempt = 0;attempt < 2; attempt++) {
-      response = await client.routes.postApiV1ChatCompletionsRaw({
-        json: requestBody
+      const attemptResponse = await client.routes.postApiV1ChatCompletionsRaw({
+        json: requestBody,
+        timeoutMs: resolveCloudTimeoutMs("ELIZAOS_CLOUD_IMAGE_TIMEOUT_MS", 120000)
       });
-      if (response.status !== 429 || attemptedRetry)
+      if (!attemptResponse) {
+        continue;
+      }
+      response = attemptResponse;
+      if (attemptResponse.status !== 429 || attemptedRetry)
         break;
-      const headerValue = response.headers.get("retry-after");
+      const headerValue = attemptResponse.headers.get("retry-after");
       const headerRetryAfter = headerValue !== null && Number.isFinite(Number(headerValue)) ? Number(headerValue) : undefined;
       let bodyRetryAfter;
       try {
-        const peek = await response.clone().json();
+        const peek = await attemptResponse.clone().json();
         bodyRetryAfter = typeof peek?.retryAfter === "number" && Number.isFinite(peek.retryAfter) ? peek.retryAfter : undefined;
       } catch {}
       const retryAfter = headerRetryAfter ?? bodyRetryAfter ?? 0;
@@ -485,8 +513,12 @@ async function handleImageDescription(runtime, params) {
       logger4.warn(`[ELIZAOS_CLOUD] Image analysis rate-limited (429); upstream retryAfter=${retryAfter || "unknown"}s — failing fast`);
       break;
     }
-    if (!response?.ok) {
-      const status = response?.status ?? 0;
+    if (!response) {
+      throw new Error("ElizaOS Cloud API did not return a response");
+    }
+    const finalResponse = response;
+    if (!finalResponse.ok) {
+      const status = finalResponse.status;
       if (status === 402) {
         throw new Error("Eliza Cloud credits exhausted — top up at https://www.elizacloud.ai/dashboard/settings?tab=billing");
       }
@@ -495,7 +527,7 @@ async function handleImageDescription(runtime, params) {
       }
       throw new Error(`ElizaOS Cloud API error: ${status}`);
     }
-    const typedResult = await response.json();
+    const typedResult = await finalResponse.json();
     const content = typedResult.choices?.[0]?.message?.content;
     if (typedResult.usage) {
       emitModelUsageEvent(runtime, ModelType2.IMAGE_DESCRIPTION, typeof params === "string" ? params : params.prompt || "", {
@@ -523,6 +555,7 @@ async function handleImageDescription(runtime, params) {
 
 // src/models/research.ts
 import { logger as logger5, ModelType as ModelType3 } from "@elizaos/core";
+var DEFAULT_RESEARCH_TIMEOUT_MS = 600000;
 function normalizeInput(input) {
   if (typeof input !== "string") {
     return input;
@@ -645,7 +678,8 @@ async function handleResearch(runtime, params) {
     requestBody.reasoning = { summary: params.reasoningSummary };
   }
   const response = await createCloudApiClient(runtime).requestRaw("POST", "/responses", {
-    json: requestBody
+    json: requestBody,
+    timeoutMs: resolveCloudTimeoutMs("ELIZAOS_CLOUD_RESEARCH_TIMEOUT_MS", DEFAULT_RESEARCH_TIMEOUT_MS)
   });
   if (!response.ok) {
     const errorText = await response.text();
@@ -682,17 +716,59 @@ async function handleResearch(runtime, params) {
 }
 
 // src/models/speech.ts
-import { logger as logger6 } from "@elizaos/core";
+import { isCloudConnected, logger as logger6, toRuntimeSettings } from "@elizaos/core";
+var cloudTtsClientFactory = (runtime) => createElizaCloudClient(runtime);
+function setCloudTtsClientFactoryForTesting(factory) {
+  if (factory === null) {
+    cloudTtsClientFactory = (runtime) => createElizaCloudClient(runtime);
+  } else {
+    cloudTtsClientFactory = factory;
+  }
+}
+
+class CloudTtsUnavailableError extends Error {
+  constructor(message = "Eliza Cloud is not connected") {
+    super(message);
+    this.name = "CloudTtsUnavailableError";
+  }
+}
+function normalizeTextInput(input) {
+  if (typeof input === "string")
+    return { text: input };
+  return input;
+}
+function resolveModelId(options) {
+  if (options.modelId && options.modelId.trim()) {
+    return options.modelId.trim();
+  }
+  const model = options.model?.trim();
+  if (!model)
+    return;
+  if (model.startsWith("elevenlabs/")) {
+    return model.split("/").slice(1).join("/");
+  }
+  if (model.startsWith("eleven_")) {
+    return model;
+  }
+  return;
+}
+function resolveVoiceId(options) {
+  if (options.voiceId && options.voiceId.trim()) {
+    return options.voiceId.trim();
+  }
+  const voice = options.voice?.trim();
+  if (!voice)
+    return;
+  if (voice === "nova")
+    return;
+  return voice;
+}
 async function fetchTextToSpeech(runtime, options) {
-  const defaultModel = getSetting(runtime, "ELIZAOS_CLOUD_TTS_MODEL", "gpt-5-mini-tts");
-  const defaultVoice = getSetting(runtime, "ELIZAOS_CLOUD_TTS_VOICE", "nova");
-  const model = options.model || defaultModel;
-  const voice = options.voice || defaultVoice;
   const format = options.format || "mp3";
-  const modelId = model.startsWith("elevenlabs/") ? model.split("/").slice(1).join("/") : model.startsWith("eleven_") ? model : undefined;
-  const voiceId = voice && voice !== "nova" ? voice : undefined;
+  const modelId = resolveModelId(options);
+  const voiceId = resolveVoiceId(options);
   try {
-    const res = await createElizaCloudClient(runtime).routes.postApiV1VoiceTts({
+    const res = await cloudTtsClientFactory(runtime).routes.postApiV1VoiceTts({
       headers: {
         ...format === "mp3" ? { Accept: "audio/mpeg" } : {}
       },
@@ -700,7 +776,8 @@ async function fetchTextToSpeech(runtime, options) {
         text: options.text,
         ...voiceId ? { voiceId } : {},
         ...modelId ? { modelId } : {}
-      }
+      },
+      timeoutMs: resolveCloudTimeoutMs("ELIZAOS_CLOUD_TTS_TIMEOUT_MS", 60000)
     });
     if (!res.ok) {
       const err = await res.text();
@@ -718,30 +795,117 @@ async function fetchTextToSpeech(runtime, options) {
     throw new Error(`Failed to fetch speech from ElizaOS Cloud TTS: ${message}`);
   }
 }
+function toUint8Array(chunk) {
+  if (chunk instanceof Uint8Array)
+    return chunk;
+  if (chunk instanceof ArrayBuffer)
+    return new Uint8Array(chunk);
+  if (typeof chunk === "string")
+    return new TextEncoder().encode(chunk);
+  throw new TypeError(`Unexpected TTS chunk type: ${typeof chunk}`);
+}
+function concatChunks(chunks) {
+  const total = chunks.reduce((sum, chunk) => sum + chunk.byteLength, 0);
+  const out = new Uint8Array(total);
+  let offset = 0;
+  for (const chunk of chunks) {
+    out.set(chunk, offset);
+    offset += chunk.byteLength;
+  }
+  return out;
+}
+async function webStreamToUint8Array(stream) {
+  const reader = stream.getReader();
+  const chunks = [];
+  try {
+    while (true) {
+      const result = await reader.read();
+      if (result.done)
+        break;
+      chunks.push(toUint8Array(result.value));
+    }
+  } finally {
+    reader.releaseLock();
+  }
+  return concatChunks(chunks);
+}
+async function nodeStreamToUint8Array(stream) {
+  const chunks = [];
+  for await (const chunk of stream) {
+    chunks.push(toUint8Array(chunk));
+  }
+  return concatChunks(chunks);
+}
+function isReadableStream(stream) {
+  return typeof stream.getReader === "function";
+}
+async function ttsStreamToBytes(stream) {
+  if (isReadableStream(stream)) {
+    return webStreamToUint8Array(stream);
+  }
+  return nodeStreamToUint8Array(stream);
+}
+function buildAudioStreamResult(stream, mimeType) {
+  const collected = [];
+  let resolveBytes;
+  let rejectBytes;
+  const bytes = new Promise((resolve, reject) => {
+    resolveBytes = resolve;
+    rejectBytes = reject;
+  });
+  async function* generate() {
+    try {
+      if (isReadableStream(stream)) {
+        const reader = stream.getReader();
+        try {
+          for (;; ) {
+            const { value, done } = await reader.read();
+            if (done)
+              break;
+            const chunk = toUint8Array(value);
+            collected.push(chunk);
+            yield chunk;
+          }
+        } finally {
+          reader.releaseLock();
+        }
+      } else {
+        for await (const value of stream) {
+          const chunk = toUint8Array(value);
+          collected.push(chunk);
+          yield chunk;
+        }
+      }
+      resolveBytes(concatChunks(collected));
+    } catch (err) {
+      rejectBytes(err);
+      throw err;
+    }
+  }
+  return { audioStream: generate(), bytes, mimeType };
+}
 async function handleTextToSpeech(runtime, input) {
-  const options = typeof input === "string" ? { text: input } : input;
-  const resolvedModel = options.model || getSetting(runtime, "ELIZAOS_CLOUD_TTS_MODEL", "gpt-5-mini-tts");
+  if (!isCloudConnected(toRuntimeSettings(runtime))) {
+    throw new CloudTtsUnavailableError("Eliza Cloud is not connected — falling through to next TTS handler");
+  }
+  const options = normalizeTextInput(input);
+  const wantsStream = typeof input === "object" && input !== null && input.audioStream === true;
+  const resolvedModel = options.modelId || options.model || getSetting(runtime, "ELIZAOS_CLOUD_TTS_MODEL", "eleven_flash_v2_5");
   logger6.log(`[ELIZAOS_CLOUD] Using TEXT_TO_SPEECH model: ${resolvedModel}`);
   try {
     const speechStream = await fetchTextToSpeech(runtime, options);
-    return speechStream;
+    if (wantsStream) {
+      const format = options.format || "mp3";
+      const mimeType = format === "mp3" ? "audio/mpeg" : `audio/${format}`;
+      return buildAudioStreamResult(speechStream, mimeType);
+    }
+    return ttsStreamToBytes(speechStream);
   } catch (error) {
     const message = error instanceof Error ? error.message : String(error);
     logger6.error(`Error in TEXT_TO_SPEECH: ${message}`);
     throw error;
   }
 }
-// src/providers/openai.ts
-import { createOpenAI } from "@ai-sdk/openai";
-function createOpenAIClient(runtime) {
-  const baseURL = getBaseURL(runtime);
-  const apiKey = getApiKey(runtime) ?? (isProxyMode(runtime) ? "eliza-proxy" : undefined);
-  return createOpenAI({
-    apiKey: apiKey ?? "",
-    baseURL
-  });
-}
-
 // src/utils/responses-output.ts
 function asRecord(value) {
   if (!value || typeof value !== "object" || Array.isArray(value)) {
@@ -827,8 +991,11 @@ import {
   buildCanonicalSystemPrompt,
   logger as logger7,
   ModelType as ModelType4,
+  recordInferenceSpan,
   renderChatMessagesForPrompt,
-  resolveEffectiveSystemPrompt
+  resolveEffectiveSystemPrompt,
+  Semaphore,
+  timeInferenceSpan as timeInferenceSpan2
 } from "@elizaos/core";
 var TEXT_NANO_MODEL_TYPE = ModelType4.TEXT_NANO ?? "TEXT_NANO";
 var TEXT_MEDIUM_MODEL_TYPE = ModelType4.TEXT_MEDIUM ?? "TEXT_MEDIUM";
@@ -837,6 +1004,61 @@ var TEXT_LARGE_MODEL_TYPE = ModelType4.TEXT_LARGE;
 var TEXT_MEGA_MODEL_TYPE = ModelType4.TEXT_MEGA ?? "TEXT_MEGA";
 var RESPONSE_HANDLER_MODEL_TYPE = ModelType4.RESPONSE_HANDLER ?? "RESPONSE_HANDLER";
 var ACTION_PLANNER_MODEL_TYPE = ModelType4.ACTION_PLANNER ?? "ACTION_PLANNER";
+var NATIVE_CONCURRENCY_ENV = "ELIZAOS_CLOUD_NATIVE_CONCURRENCY";
+var DEFAULT_NATIVE_CONCURRENCY = 8;
+var TEXT_TIMEOUT_ENV = "ELIZAOS_CLOUD_TEXT_TIMEOUT_MS";
+var DEFAULT_TEXT_TIMEOUT_MS = 120000;
+function resolveTextTimeoutMs() {
+  const raw = typeof process !== "undefined" ? process.env[TEXT_TIMEOUT_ENV] : undefined;
+  if (raw === undefined || raw.trim() === "")
+    return DEFAULT_TEXT_TIMEOUT_MS;
+  const parsed = Number.parseInt(raw, 10);
+  if (!Number.isFinite(parsed))
+    return DEFAULT_TEXT_TIMEOUT_MS;
+  return parsed <= 0 ? undefined : parsed;
+}
+var STREAMING_ENV = "ELIZAOS_CLOUD_STREAMING";
+function resolveStreamingEnabled() {
+  const raw = typeof process !== "undefined" ? process.env[STREAMING_ENV] : undefined;
+  if (raw === undefined)
+    return true;
+  const v = raw.trim().toLowerCase();
+  return v !== "0" && v !== "false" && v !== "off";
+}
+function buildStreamAbortSignal(abortSignal, timeoutMs) {
+  const timeoutSig = typeof timeoutMs === "number" && timeoutMs > 0 ? AbortSignal.timeout(timeoutMs) : undefined;
+  if (abortSignal && timeoutSig)
+    return AbortSignal.any([abortSignal, timeoutSig]);
+  return abortSignal ?? timeoutSig;
+}
+var nativeChatLimiter = null;
+function resolveNativeConcurrency() {
+  const raw = typeof process !== "undefined" ? process.env[NATIVE_CONCURRENCY_ENV] : undefined;
+  const parsed = raw ? Number.parseInt(raw, 10) : Number.NaN;
+  return Number.isFinite(parsed) && parsed > 0 ? parsed : DEFAULT_NATIVE_CONCURRENCY;
+}
+function getNativeChatLimiter() {
+  if (!nativeChatLimiter) {
+    nativeChatLimiter = new Semaphore(resolveNativeConcurrency());
+  }
+  return nativeChatLimiter;
+}
+async function withNativeChatLimit(fn, label = "native") {
+  const limiter = getNativeChatLimiter();
+  const waitStartedAt = Date.now();
+  await limiter.acquire();
+  recordInferenceSpan("cloud.semaphore-wait", Date.now() - waitStartedAt, {
+    route: label
+  });
+  try {
+    return await timeInferenceSpan2(`cloud.http:${label}`, fn, { route: label });
+  } finally {
+    limiter.release();
+  }
+}
+function __resetNativeChatLimiterForTests() {
+  nativeChatLimiter = null;
+}
 var REASONING_MODEL_PATTERNS = [
   "o1",
   "o3",
@@ -847,27 +1069,51 @@ var REASONING_MODEL_PATTERNS = [
   "claude-opus-4-7",
   "gpt-5"
 ];
-var RESPONSES_ROUTED_PREFIXES = ["openai/", "anthropic/"];
-function buildUserContent(params) {
-  const content = [{ type: "text", text: params.prompt ?? "" }];
-  for (const attachment of params.attachments ?? []) {
-    content.push({
-      type: "file",
-      data: attachment.data,
-      mediaType: attachment.mediaType,
-      ...attachment.filename ? { filename: attachment.filename } : {}
-    });
+var SPAN_SAMPLER_HONORING_MODEL_PREFIXES = [
+  "vast/eliza-1-",
+  "elizaos/eliza-1-",
+  "eliza-1-"
+];
+function isSpanSamplerHonoringModel(modelName) {
+  const lower = modelName.toLowerCase();
+  return SPAN_SAMPLER_HONORING_MODEL_PREFIXES.some((prefix) => lower.startsWith(prefix));
+}
+function buildSpanSamplerHeader(plan) {
+  if (!plan || plan.overrides.length === 0)
+    return;
+  const overrides = plan.overrides.map((o) => {
+    const wire = {
+      span_index: o.spanIndex,
+      temperature: o.temperature
+    };
+    if (typeof o.topK === "number")
+      wire.top_k = o.topK;
+    if (typeof o.topP === "number")
+      wire.top_p = o.topP;
+    return wire;
+  });
+  const body = { overrides };
+  if (plan.strict === true)
+    body.strict = true;
+  return JSON.stringify(body);
+}
+function extractCostUsd(usage, response) {
+  const fromBody = firstNumber(asRecord2(usage).cost_usd, asRecord2(usage).costUsd, asRecord2(usage).cost);
+  if (typeof fromBody === "number" && Number.isFinite(fromBody)) {
+    return fromBody;
+  }
+  const header = response?.headers?.get?.("X-Eliza-Cost-Usd");
+  if (header) {
+    const parsed = Number(header);
+    if (Number.isFinite(parsed))
+      return parsed;
   }
-  return content;
+  return;
 }
 function isReasoningModel(modelName) {
   const lower = modelName.toLowerCase();
   return REASONING_MODEL_PATTERNS.some((pattern) => lower.includes(pattern));
 }
-function supportsStopSequences(modelName) {
-  const lower = modelName.toLowerCase();
-  return !RESPONSES_ROUTED_PREFIXES.some((prefix) => lower.startsWith(prefix));
-}
 function isRecord(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
 }
@@ -944,26 +1190,39 @@ function unwrapJsonSchema(value) {
   const record = asRecord2(value);
   return record.schema ?? record.jsonSchema ?? value;
 }
+function normalizeNativeToolEntry(rawTool, fallbackName) {
+  const tool = asRecord2(rawTool);
+  const nested = asRecord2(tool.function);
+  const name = firstString(nested.name, tool.name, fallbackName);
+  if (!name) {
+    return;
+  }
+  const description = firstString(nested.description, tool.description);
+  const inputSchema = unwrapJsonSchema(nested.parameters ?? tool.inputSchema ?? tool.parameters ?? tool.schema ?? { type: "object" });
+  return {
+    type: "function",
+    function: {
+      name,
+      ...description ? { description } : {},
+      parameters: inputSchema
+    }
+  };
+}
 function normalizeNativeTools(tools) {
   if (!tools) {
     return;
   }
   if (Array.isArray(tools)) {
-    return tools;
+    const normalized2 = tools.map((tool) => normalizeNativeToolEntry(tool)).filter((tool) => tool !== undefined);
+    return normalized2.length > 0 ? normalized2 : undefined;
   }
   const toolSet = asRecord2(tools);
   const normalized = [];
   for (const [name, rawTool] of Object.entries(toolSet)) {
-    const tool = asRecord2(rawTool);
-    const inputSchema = unwrapJsonSchema(tool.inputSchema ?? tool.parameters ?? tool.schema ?? { type: "object" });
-    normalized.push({
-      type: "function",
-      function: {
-        name,
-        ...typeof tool.description === "string" ? { description: tool.description } : {},
-        parameters: inputSchema
-      }
-    });
+    const entry = normalizeNativeToolEntry(rawTool, name);
+    if (entry) {
+      normalized.push(entry);
+    }
   }
   return normalized.length > 0 ? normalized : undefined;
 }
@@ -1189,16 +1448,8 @@ function getModelNameForType(runtime, modelType) {
   }
 }
 function buildGenerateParams(runtime, modelType, params) {
-  const paramsWithAttachments = params;
   const prompt = params.prompt ?? "";
-  const maxTokens = params.maxTokens ?? 8192;
-  const openai = createOpenAIClient(runtime);
   const modelName = getModelNameForType(runtime, modelType);
-  const experimentalTelemetry = getExperimentalTelemetry(runtime);
-  const userContent = (paramsWithAttachments.attachments?.length ?? 0) > 0 ? buildUserContent(paramsWithAttachments) : undefined;
-  const model = openai.chat(modelName);
-  const reasoning = isReasoningModel(modelName);
-  const stopSequences = !reasoning && supportsStopSequences(modelName) && Array.isArray(params.stopSequences) && params.stopSequences.length > 0 ? params.stopSequences : undefined;
   const systemPrompt = resolveEffectiveSystemPrompt({
     params,
     fallback: buildCanonicalSystemPrompt({ character: runtime.character })
@@ -1206,28 +1457,24 @@ function buildGenerateParams(runtime, modelType, params) {
   const promptText = renderChatMessagesForPrompt(params.messages, {
     omitDuplicateSystem: systemPrompt
   }) ?? prompt;
-  const generateParams = {
-    model,
-    ...userContent ? { messages: [{ role: "user", content: userContent }] } : { prompt: promptText },
-    system: systemPrompt,
-    ...stopSequences ? { stopSequences } : {},
-    maxOutputTokens: maxTokens,
-    experimental_telemetry: {
-      isEnabled: experimentalTelemetry
-    }
-  };
-  return { generateParams, modelName, modelType, prompt: promptText, systemPrompt };
+  return { modelName, modelType, prompt: promptText, systemPrompt };
 }
 async function generateTextWithModel(runtime, modelType, params) {
   const { modelName, prompt, systemPrompt } = buildGenerateParams(runtime, modelType, params);
   const paramsWithNative = params;
   logger7.debug(`[ELIZAOS_CLOUD] Generating text with ${modelType} model: ${modelName}`);
-  if (params.stream) {
-    logger7.debug("[ELIZAOS_CLOUD] Streaming text disabled for responses compatibility; falling back to buffered response.");
-  }
+  const paramsStreaming = params;
+  const wantsStream = Boolean(paramsStreaming.stream) && paramsStreaming.streamStructured === true && resolveStreamingEnabled();
   logger7.log(`[ELIZAOS_CLOUD] Using ${modelType} model: ${modelName}`);
   logger7.log(prompt);
   if (hasNativeTransportOptions(paramsWithNative)) {
+    if (wantsStream) {
+      return streamNativeChatCompletion(runtime, modelType, paramsWithNative, {
+        modelName,
+        prompt,
+        systemPrompt
+      });
+    }
     const nativeResult = await generateNativeChatCompletion(runtime, modelType, paramsWithNative, {
       modelName,
       prompt,
@@ -1255,13 +1502,21 @@ async function generateTextWithModel(runtime, modelType, params) {
   if (!reasoning && typeof params.temperature === "number") {
     requestBody.temperature = params.temperature;
   }
-  const response = await createCloudApiClient(runtime).requestRaw("POST", "/responses", {
-    headers: {
-      "X-Eliza-Llm-Purpose": getPurposeForModelType(modelType),
-      "X-Eliza-Model-Type": modelType
-    },
-    json: requestBody
-  });
+  const responsesHeaders = {
+    "X-Eliza-Llm-Purpose": getPurposeForModelType(modelType),
+    "X-Eliza-Model-Type": modelType
+  };
+  if (isSpanSamplerHonoringModel(modelName)) {
+    const samplerHeader = buildSpanSamplerHeader(params.spanSamplerPlan);
+    if (samplerHeader) {
+      responsesHeaders["x-eliza-span-samplers"] = samplerHeader;
+    }
+  }
+  const response = await withNativeChatLimit(() => createCloudApiClient(runtime).requestRaw("POST", "/responses", {
+    headers: responsesHeaders,
+    json: requestBody,
+    timeoutMs: resolveTextTimeoutMs()
+  }), "responses");
   const responseText = await response.text();
   let data = {};
   if (responseText) {
@@ -1286,6 +1541,12 @@ async function generateTextWithModel(runtime, modelType, params) {
       inputTokens: data.usage.input_tokens ?? 0,
       outputTokens: data.usage.output_tokens ?? 0,
       totalTokens: data.usage.total_tokens ?? 0
+    }, {
+      modelName: getModelNameForType(runtime, modelType),
+      ...(() => {
+        const costUsd = extractCostUsd(data.usage, response);
+        return typeof costUsd === "number" ? { costUsd } : {};
+      })()
     });
   }
   const text = extractResponsesOutputText(data);
@@ -1296,13 +1557,21 @@ async function generateTextWithModel(runtime, modelType, params) {
 }
 async function generateNativeChatCompletion(runtime, modelType, params, context) {
   const requestBody = buildNativeRequestBody(params, context.modelName, context.prompt, context.systemPrompt);
-  const response = await createCloudApiClient(runtime).requestRaw("POST", "/chat/completions", {
-    headers: {
-      "X-Eliza-Llm-Purpose": getPurposeForModelType(modelType),
-      "X-Eliza-Model-Type": modelType
-    },
-    json: requestBody
-  });
+  const headers = {
+    "X-Eliza-Llm-Purpose": getPurposeForModelType(modelType),
+    "X-Eliza-Model-Type": modelType
+  };
+  if (isSpanSamplerHonoringModel(context.modelName)) {
+    const samplerHeader = buildSpanSamplerHeader(params.spanSamplerPlan);
+    if (samplerHeader) {
+      headers["x-eliza-span-samplers"] = samplerHeader;
+    }
+  }
+  const response = await withNativeChatLimit(() => createCloudApiClient(runtime).requestRaw("POST", "/chat/completions", {
+    headers,
+    json: requestBody,
+    timeoutMs: resolveTextTimeoutMs()
+  }), "chat/completions");
   const responseText = await response.text();
   let data = {};
   if (responseText) {
@@ -1324,7 +1593,13 @@ async function generateNativeChatCompletion(runtime, modelType, params, context)
   }
   const usage = convertNativeUsage(data.usage);
   if (usage) {
-    emitModelUsageEvent(runtime, modelType, context.prompt, usage);
+    emitModelUsageEvent(runtime, modelType, context.prompt, usage, {
+      modelName: context.modelName,
+      ...(() => {
+        const costUsd = extractCostUsd(data.usage, response);
+        return typeof costUsd === "number" ? { costUsd } : {};
+      })()
+    });
   }
   const text = extractChatCompletionText(data);
   const toolCalls = extractNativeToolCalls(data);
@@ -1342,6 +1617,251 @@ async function generateNativeChatCompletion(runtime, modelType, params, context)
     }
   };
 }
+function deferred() {
+  let resolve;
+  const promise = new Promise((r) => {
+    resolve = r;
+  });
+  return { promise, resolve };
+}
+async function* parseOpenAiSseStream(body) {
+  const reader = body.getReader();
+  const decoder = new TextDecoder;
+  let buffer = "";
+  const handle = (line) => {
+    const trimmed = line.trimStart();
+    if (!trimmed.startsWith("data:"))
+      return null;
+    const payload = trimmed.slice(5).trim();
+    if (payload === "")
+      return null;
+    if (payload === "[DONE]")
+      return "DONE";
+    try {
+      return JSON.parse(payload);
+    } catch {
+      return null;
+    }
+  };
+  try {
+    for (;; ) {
+      const { value, done } = await reader.read();
+      if (done)
+        break;
+      buffer += decoder.decode(value, { stream: true });
+      let nl;
+      while ((nl = buffer.indexOf(`
+`)) >= 0) {
+        const line = buffer.slice(0, nl);
+        buffer = buffer.slice(nl + 1);
+        const frame = handle(line);
+        if (frame === "DONE")
+          return;
+        if (frame)
+          yield frame;
+      }
+    }
+    const tail = handle(buffer);
+    if (tail && tail !== "DONE")
+      yield tail;
+  } finally {
+    try {
+      await reader.cancel();
+    } catch {}
+  }
+}
+function accumulateToolCallDeltas(acc, deltas) {
+  if (!Array.isArray(deltas))
+    return;
+  for (const raw of deltas) {
+    const d = asRecord2(raw);
+    const index = typeof d.index === "number" ? d.index : 0;
+    const cur = acc.get(index) ?? { args: "" };
+    const id = firstString(d.id);
+    if (id)
+      cur.id = id;
+    const fn = recordAt(d, "function");
+    const name = firstString(fn.name);
+    if (name)
+      cur.name = name;
+    if (typeof fn.arguments === "string")
+      cur.args += fn.arguments;
+    acc.set(index, cur);
+  }
+}
+function finalizeStreamedToolCalls(acc) {
+  const out = [];
+  for (const [index, c] of [...acc.entries()].sort((a, b) => a[0] - b[0])) {
+    if (!c.name)
+      continue;
+    out.push({
+      type: "tool-call",
+      toolCallId: c.id ?? `call_${c.name}_${index}`,
+      toolName: c.name,
+      input: parseJsonIfPossible(c.args.trim() === "" ? "{}" : c.args)
+    });
+  }
+  return out;
+}
+async function streamNativeChatCompletion(runtime, modelType, params, context) {
+  const requestBody = buildNativeRequestBody(params, context.modelName, context.prompt, context.systemPrompt);
+  requestBody.stream = true;
+  requestBody.stream_options = { include_usage: true };
+  const headers = {
+    "X-Eliza-Llm-Purpose": getPurposeForModelType(modelType),
+    "X-Eliza-Model-Type": modelType
+  };
+  if (isSpanSamplerHonoringModel(context.modelName)) {
+    const samplerHeader = buildSpanSamplerHeader(params.spanSamplerPlan);
+    if (samplerHeader) {
+      headers["x-eliza-span-samplers"] = samplerHeader;
+    }
+  }
+  const abortSignal = params.signal;
+  const signal = buildStreamAbortSignal(abortSignal, resolveTextTimeoutMs());
+  const limiter = getNativeChatLimiter();
+  const waitStartedAt = Date.now();
+  await limiter.acquire();
+  recordInferenceSpan("cloud.semaphore-wait", Date.now() - waitStartedAt, {
+    route: "chat/completions:stream"
+  });
+  let permitReleased = false;
+  const releasePermit = () => {
+    if (!permitReleased) {
+      permitReleased = true;
+      limiter.release();
+    }
+  };
+  let response;
+  try {
+    response = await createCloudApiClient(runtime).requestRaw("POST", "/chat/completions", {
+      headers,
+      json: requestBody,
+      ...signal ? { signal } : {}
+    });
+  } catch (err) {
+    releasePermit();
+    throw err;
+  }
+  if (!response.ok) {
+    let errorBody;
+    try {
+      const errText = await response.text();
+      if (errText) {
+        errorBody = JSON.parse(errText).error;
+      }
+    } catch {}
+    releasePermit();
+    const message = typeof errorBody?.message === "string" && errorBody.message.trim() ? errorBody.message.trim() : `elizaOS Cloud error ${response.status}`;
+    const requestError = new Error(message);
+    requestError.status = response.status;
+    if (errorBody)
+      requestError.error = errorBody;
+    throw requestError;
+  }
+  const contentType = response.headers.get("content-type") ?? "";
+  const isSse = contentType.includes("text/event-stream") && response.body !== null;
+  if (!isSse) {
+    const bufferedText = await response.text();
+    releasePermit();
+    let data = {};
+    if (bufferedText) {
+      try {
+        data = JSON.parse(bufferedText);
+      } catch (parseErr) {
+        logger7.error(`[ELIZAOS_CLOUD] Failed to parse buffered chat completions JSON: ${parseErr instanceof Error ? parseErr.message : String(parseErr)}`);
+      }
+    }
+    const text = extractChatCompletionText(data);
+    const toolCalls = extractNativeToolCalls(data);
+    const usage = convertNativeUsage(data.usage);
+    if (usage) {
+      emitModelUsageEvent(runtime, modelType, context.prompt, usage, {
+        modelName: context.modelName,
+        ...(() => {
+          const costUsd = extractCostUsd(data.usage, response);
+          return typeof costUsd === "number" ? { costUsd } : {};
+        })()
+      });
+    }
+    if (!text.trim() && toolCalls.length === 0) {
+      throw new Error("elizaOS Cloud returned no text or tool calls");
+    }
+    async function* single() {
+      if (text)
+        yield text;
+    }
+    return {
+      textStream: single(),
+      text: Promise.resolve(text),
+      usage: Promise.resolve(usage),
+      finishReason: Promise.resolve(data.choices?.[0]?.finish_reason),
+      toolCalls: Promise.resolve(toolCalls),
+      providerMetadata: { modelName: context.modelName, usage: data.usage }
+    };
+  }
+  const body = response.body;
+  const toolAcc = new Map;
+  let accumulated = "";
+  let nativeUsage;
+  let rawUsage;
+  let finishReason;
+  const textD = deferred();
+  const usageD = deferred();
+  const finishD = deferred();
+  const toolCallsD = deferred();
+  async function* generate() {
+    try {
+      for await (const frame of parseOpenAiSseStream(body)) {
+        if (frame.error) {
+          const message = asRecord2(frame.error).message;
+          throw new Error(typeof message === "string" && message.trim() ? message.trim() : "elizaOS Cloud stream error");
+        }
+        const choices = Array.isArray(frame.choices) ? frame.choices : [];
+        const choice = asRecord2(choices[0]);
+        const delta = recordAt(choice, "delta");
+        if (typeof delta.content === "string" && delta.content.length > 0) {
+          accumulated += delta.content;
+          yield delta.content;
+        }
+        if (delta.tool_calls) {
+          accumulateToolCallDeltas(toolAcc, delta.tool_calls);
+        }
+        const fr = firstString(choice.finish_reason);
+        if (fr)
+          finishReason = fr;
+        if (frame.usage) {
+          rawUsage = frame.usage;
+          nativeUsage = convertNativeUsage(frame.usage);
+        }
+      }
+    } finally {
+      releasePermit();
+      const toolCalls = finalizeStreamedToolCalls(toolAcc);
+      textD.resolve(accumulated);
+      usageD.resolve(nativeUsage);
+      finishD.resolve(finishReason);
+      toolCallsD.resolve(toolCalls);
+      if (nativeUsage) {
+        emitModelUsageEvent(runtime, modelType, context.prompt, nativeUsage, {
+          modelName: context.modelName,
+          ...(() => {
+            const costUsd = extractCostUsd(rawUsage, response);
+            return typeof costUsd === "number" ? { costUsd } : {};
+          })()
+        });
+      }
+    }
+  }
+  return {
+    textStream: generate(),
+    text: textD.promise,
+    usage: usageD.promise,
+    finishReason: finishD.promise,
+    toolCalls: toolCallsD.promise,
+    providerMetadata: { modelName: context.modelName }
+  };
+}
 async function handleTextSmall(runtime, params) {
   return generateTextWithModel(runtime, TEXT_SMALL_MODEL_TYPE, params);
 }
@@ -1383,4 +1903,4 @@ export {
   fetchTextToSpeech
 };
 
-//# debugId=06464BB61850EEE264756E2164756E21
+//# debugId=4FFC473B568B95F864756E2164756E21