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

package/README.md

@@ -1,163 +1,221 @@
-# OpenAI Plugin
+# @elizaos/plugin-openai
 
-This plugin provides integration with OpenAI's models through the elizaOS platform.
+OpenAI model-provider plugin for [elizaOS](https://github.com/elizaos/eliza). Adds text generation, embeddings, image generation, image description, audio transcription, text-to-speech, and deep research to any Eliza agent by registering handlers for the standard `ModelType.*` slots on the `AgentRuntime`.
 
-## Usage
+## What this plugin does
 
-Add the plugin to your character configuration:
+- **Text generation** — multiple tiers (nano, small, medium, large, mega), plus dedicated response-handler and action-planner slots. Supports streaming and structured JSON output.
+- **Text embeddings** — `text-embedding-3-small` by default; dimension configurable.
+- **Image generation** — DALL-E 3 by default (`dall-e-3`).
+- **Image description** — vision model analyzes an image URL and returns `{ title, description }`.
+- **Audio transcription** — speech-to-text (`gpt-5-mini-transcribe` by default); accepts `Buffer`, `Blob`, `File`, or a URL.
+- **Text-to-speech** — returns an `ArrayBuffer` of audio. Six voices; mp3/wav/flac/opus/aac/pcm output.
+- **Deep research** — `ModelType.RESEARCH` via the OpenAI Responses API (`o3-deep-research` by default); returns annotated, multi-source research reports.
+- **Tokenizer** — encode/decode using js-tiktoken (browser-safe, no network calls).
 
-```json
-"plugins": ["@elizaos/plugin-openai"]
-```
+Works with any OpenAI-compatible endpoint: OpenAI, Cerebras, EvoLink, OpenRouter, local servers, etc.
 
-## Configuration
+## Enabling the plugin
 
-The plugin requires these environment variables (can be set in .env file or character settings):
+Add `@elizaos/plugin-openai` to your character's plugin list:
 
 ```json
-"settings": {
-  "OPENAI_API_KEY": "your_openai_api_key",
-  "OPENAI_BASE_URL": "optional_custom_endpoint",
-  "OPENAI_SMALL_MODEL": "gpt-5-mini",
-  "OPENAI_LARGE_MODEL": "gpt-5",
-  "OPENAI_EMBEDDING_MODEL": "text-embedding-3-small",
-  "OPENAI_EMBEDDING_API_KEY": "your_openai_api_key_for_embedding",
-  "OPENAI_EMBEDDING_URL": "optional_custom_endpoint",
-  "OPENAI_EMBEDDING_DIMENSIONS": "1536",
-  "OPENAI_IMAGE_DESCRIPTION_MODEL": "gpt-5-mini",
-  "OPENAI_IMAGE_DESCRIPTION_MAX_TOKENS": "8192",
-  "OPENAI_EXPERIMENTAL_TELEMETRY": "false",
-  "OPENAI_BROWSER_BASE_URL": "https://your-proxy.example.com/openai",
-  "OPENAI_BROWSER_EMBEDDING_URL": "https://your-proxy.example.com/openai"
+{
+  "plugins": ["@elizaos/plugin-openai"]
 }
 ```
 
-Or in `.env` file:
+The plugin auto-enables when `OPENAI_API_KEY`, `CEREBRAS_API_KEY`, or `EVOLINK_API_KEY` is present in the environment.
+
+## Required configuration
 
 ```
-OPENAI_API_KEY=your_openai_api_key
-# Optional overrides:
-OPENAI_BASE_URL=optional_custom_endpoint
-OPENAI_SMALL_MODEL=gpt-5-mini
-OPENAI_LARGE_MODEL=gpt-5
-OPENAI_EMBEDDING_MODEL=text-embedding-3-small
-OPENAI_EMBEDDING_API_KEY=your_openai_api_key_for_embedding
-OPENAI_EMBEDDING_URL=optional_custom_endpoint
-OPENAI_EMBEDDING_DIMENSIONS=1536
-OPENAI_IMAGE_DESCRIPTION_MODEL=gpt-5-mini
-OPENAI_IMAGE_DESCRIPTION_MAX_TOKENS=8192
-OPENAI_EXPERIMENTAL_TELEMETRY=false
-# Browser proxy (frontend builds only)
-OPENAI_BROWSER_BASE_URL=https://your-proxy.example.com/openai
-OPENAI_BROWSER_EMBEDDING_URL=https://your-proxy.example.com/openai
+OPENAI_API_KEY=sk-...
 ```
 
-### Configuration Options
+That is the only required setting. All other settings are optional overrides.
 
-- `OPENAI_API_KEY` (required): Your OpenAI API credentials
-- `OPENAI_BASE_URL`: Custom API endpoint (default: https://api.openai.com/v1)
-- `OPENAI_SMALL_MODEL`: Defaults to GPT-4o Mini ("gpt-5-mini")
-- `OPENAI_LARGE_MODEL`: Defaults to GPT-4o ("gpt-5")
-- `OPENAI_EMBEDDING_MODEL`: Defaults to text-embedding-3-small ("text-embedding-3-small")
-- `OPENAI_EMBEDDING_API_KEY`: Custom embedding api key (defaults to `OPENAI_API_KEY`)
-- `OPENAI_EMBEDDING_URL`: Custom embedding endpoint (defaults to `OPENAI_BASE_URL`)
-- `OPENAI_EMBEDDING_DIMENSIONS`: Defaults to 1536 (1536)
-- `OPENAI_IMAGE_DESCRIPTION_MODEL`: Model used for image description (default: "gpt-5-mini")
-- `OPENAI_IMAGE_DESCRIPTION_MAX_TOKENS`: Maximum tokens for image descriptions (default: 8192)
-- `OPENAI_EXPERIMENTAL_TELEMETRY`: Enable experimental telemetry features for enhanced debugging and usage analytics (default: false)
-- `OPENAI_BROWSER_BASE_URL`: Browser-only base URL to a proxy endpoint that forwards requests to OpenAI without exposing keys
-- `OPENAI_BROWSER_EMBEDDING_URL`: Browser-only embeddings endpoint base URL
+## Full configuration reference
 
-### Browser mode and proxying
+Set these as environment variables or in your character's `settings` object.
 
-When bundled for the browser, this plugin avoids sending Authorization headers. Set `OPENAI_BROWSER_BASE_URL` (and optionally `OPENAI_BROWSER_EMBEDDING_URL`) to a server-side proxy you control that injects the OpenAI API key. This prevents exposing secrets in frontend builds.
+### API endpoint
 
-Example minimal proxy (Express):
+| Variable | Default | Description |
+|---|---|---|
+| `OPENAI_BASE_URL` | `https://api.openai.com/v1` | Override for compatible endpoints |
+| `ELIZA_PROVIDER` | — | Set to `cerebras` or `evolink` to force a compatible-provider mode |
+| `EVOLINK_API_KEY` | — | Auth for EvoLink's OpenAI-compatible endpoint |
+| `EVOLINK_BASE_URL` | `https://direct.evolink.ai/v1` | EvoLink endpoint used when `EVOLINK_API_KEY` enables the plugin |
+| `EVOLINK_MODEL` | `gpt-5.2` | Text model used by EvoLink mode |
 
-```ts
-import express from "express";
-import fetch from "node-fetch";
+### Text generation models
 
-const app = express();
-app.use(express.json());
+| Variable | Default | Description |
+|---|---|---|
+| `OPENAI_SMALL_MODEL` / `SMALL_MODEL` | `gpt-5.4-mini` | Fast, cost-effective responses |
+| `OPENAI_NANO_MODEL` / `NANO_MODEL` | falls back to small | Fastest tier |
+| `OPENAI_MEDIUM_MODEL` / `MEDIUM_MODEL` | falls back to small | Mid-tier |
+| `OPENAI_LARGE_MODEL` / `LARGE_MODEL` | `gpt-5` | Complex reasoning |
+| `OPENAI_MEGA_MODEL` / `MEGA_MODEL` | falls back to large | Highest tier |
+| `OPENAI_RESPONSE_HANDLER_MODEL` | falls back to small | Response-handler slot |
+| `OPENAI_ACTION_PLANNER_MODEL` | falls back to medium | Action-planner slot |
+| `OPENAI_REASONING_EFFORT` | — | `minimal`/`low`/`medium`/`high` for o-series models |
 
-app.post("/openai/*", async (req, res) => {
-  const url = `https://api.openai.com/v1/${req.params[0]}`;
-  const r = await fetch(url, {
-    method: "POST",
-    headers: {
-      Authorization: `Bearer ${process.env.OPENAI_API_KEY}`,
-      "Content-Type": "application/json",
-    },
-    body: JSON.stringify(req.body),
-  });
-  res
-    .status(r.status)
-    .set(Object.fromEntries(r.headers))
-    .send(await r.text());
-});
+### Embeddings
 
-app.listen(3000);
-```
+| Variable | Default | Description |
+|---|---|---|
+| `OPENAI_EMBEDDING_MODEL` | `text-embedding-3-small` | Embedding model |
+| `OPENAI_EMBEDDING_URL` | `OPENAI_BASE_URL` | Separate endpoint for embeddings |
+| `OPENAI_EMBEDDING_API_KEY` | `OPENAI_API_KEY` | Separate key for embeddings |
+| `OPENAI_EMBEDDING_DIMENSIONS` | `1536` | Vector dimensions (must match model) |
 
-### Experimental Telemetry
+### Image generation and description
 
-When `OPENAI_EXPERIMENTAL_TELEMETRY` is set to `true`, the plugin enables advanced telemetry features that provide:
+| Variable | Default | Description |
+|---|---|---|
+| `OPENAI_IMAGE_MODEL` | `dall-e-3` | Image generation model |
+| `OPENAI_IMAGE_DESCRIPTION_MODEL` | `gpt-5-mini` | Vision/description model |
+| `OPENAI_IMAGE_DESCRIPTION_BASE_URL` | `OPENAI_BASE_URL` | Separate endpoint for vision |
+| `OPENAI_IMAGE_DESCRIPTION_API_KEY` | `OPENAI_API_KEY` | Separate key for vision |
+| `OPENAI_IMAGE_DESCRIPTION_MAX_TOKENS` | `8192` | Max tokens in description response |
 
-- Enhanced debugging capabilities for model performance issues
-- Detailed usage analytics for optimization
-- Better observability into OpenAI API interactions
-- Foundation for future monitoring and analytics features through Sentry or other frameworks
+### Audio
 
-**Note**: This feature is opt-in due to privacy considerations, as telemetry data may contain information about model usage patterns. Enable only when you need enhanced debugging or analytics capabilities.
+| Variable | Default | Description |
+|---|---|---|
+| `OPENAI_TRANSCRIPTION_MODEL` | `gpt-5-mini-transcribe` | Speech-to-text model |
+| `OPENAI_TTS_MODEL` | `tts-1` | Text-to-speech model |
+| `OPENAI_TTS_VOICE` | `nova` | Voice: alloy, echo, fable, onyx, nova, shimmer |
+| `OPENAI_TTS_INSTRUCTIONS` | — | Style instructions for TTS |
 
-The plugin provides these model classes:
+### Deep research
 
-- `TEXT_SMALL`: Optimized for fast, cost-effective responses
-- `TEXT_LARGE`: For complex tasks requiring deeper reasoning
-- `TEXT_EMBEDDING`: Text embedding model (text-embedding-3-small by default)
-- `IMAGE`: DALL-E image generation
-- `IMAGE_DESCRIPTION`: GPT-4o image analysis
-- `TRANSCRIPTION`: Whisper audio transcription
-- `TEXT_TOKENIZER_ENCODE`: Text tokenization
-- `TEXT_TOKENIZER_DECODE`: Token decoding
+| Variable | Default | Description |
+|---|---|---|
+| `OPENAI_RESEARCH_MODEL` | `o3-deep-research` | Research model (o3 or o4-mini variants) |
+| `OPENAI_RESEARCH_TIMEOUT` | `3600000` (1 hr) | Request timeout in milliseconds |
 
-## Additional Features
+### Browser and proxy
 
-### Image Generation
+| Variable | Default | Description |
+|---|---|---|
+| `OPENAI_BROWSER_BASE_URL` | — | Proxy URL for browser builds (keeps key server-side) |
+| `OPENAI_BROWSER_EMBEDDING_URL` | — | Proxy URL for browser embedding requests |
+| `OPENAI_ALLOW_BROWSER_API_KEY` | `false` | Send auth header in browser builds (opt-in) |
 
-```js
-await runtime.useModel(ModelType.IMAGE, {
-  prompt: "A sunset over mountains",
-  n: 1, // number of images
-  size: "1024x1024", // image resolution
+### Other
+
+| Variable | Default | Description |
+|---|---|---|
+| `OPENAI_EXPERIMENTAL_TELEMETRY` | `false` | Enable AI SDK experimental telemetry |
+
+## Usage examples
+
+```ts
+import { ModelType } from "@elizaos/core";
+
+// Text generation
+const reply = await runtime.useModel(ModelType.TEXT_LARGE, {
+  prompt: "Explain quantum entanglement in plain English.",
 });
-```
 
-### Audio Transcription
+// Streaming text
+const result = await runtime.useModel(ModelType.TEXT_LARGE, {
+  prompt: "Count from 1 to 10.",
+  stream: true,
+  onStreamChunk: (chunk) => process.stdout.write(chunk),
+});
 
-```js
-const transcription = await runtime.useModel(
-  ModelType.TRANSCRIPTION,
-  audioBuffer,
-);
-```
+// Structured JSON output
+const data = await runtime.useModel(ModelType.TEXT_LARGE, {
+  prompt: "Return a JSON object with name and age fields.",
+  responseSchema: {
+    type: "object",
+    properties: { name: { type: "string" }, age: { type: "number" } },
+    required: ["name", "age"],
+  },
+});
 
-### Image Analysis
+// Embedding
+const vector = await runtime.useModel(ModelType.TEXT_EMBEDDING, {
+  text: "text to embed",
+});
+
+// Image generation
+const images = await runtime.useModel(ModelType.IMAGE, {
+  prompt: "A sunset over mountains",
+  count: 1,
+  size: "1024x1024",
+});
 
-```js
+// Image description
 const { title, description } = await runtime.useModel(
   ModelType.IMAGE_DESCRIPTION,
-  "https://example.com/image.jpg",
+  "https://example.com/photo.jpg"
 );
+
+// Audio transcription (Buffer, Blob, File, or URL string all accepted)
+const transcript = await runtime.useModel(ModelType.TRANSCRIPTION, audioBuffer);
+
+// Text-to-speech
+const audio = await runtime.useModel(ModelType.TEXT_TO_SPEECH, {
+  text: "Hello, world.",
+  voice: "nova",
+  format: "mp3",
+});
+
+// Deep research (may take minutes)
+const report = await runtime.useModel(ModelType.RESEARCH, {
+  input: "What are the latest advances in fusion energy?",
+  tools: [{ type: "web_search_preview" }],
+});
+console.log(report.text, report.annotations);
 ```
 
-### Text Embeddings
+## Browser proxy setup
+
+In browser builds this plugin does not send `Authorization` headers by default, to avoid exposing API keys in frontend bundles. Point `OPENAI_BROWSER_BASE_URL` at a server-side proxy that injects the key:
 
-```js
-await runtime.useModel(ModelType.TEXT_EMBEDDING, "text to embed");
+```ts
+// Minimal Express proxy
+import express from "express";
+const app = express();
+app.use(express.json());
+
+app.all("/openai/*", async (req, res) => {
+  const url = `https://api.openai.com/v1/${req.params[0]}`;
+  const r = await fetch(url, {
+    method: req.method,
+    headers: { Authorization: `Bearer ${process.env.OPENAI_API_KEY}`, "Content-Type": "application/json" },
+    body: req.method !== "GET" ? JSON.stringify(req.body) : undefined,
+  });
+  res.status(r.status).send(await r.text());
+});
+
+app.listen(3000);
 ```
 
-### Tokenizer in browser
+Then set `OPENAI_BROWSER_BASE_URL=http://localhost:3000/openai`.
+
+## Cerebras compatibility
+
+Point `OPENAI_BASE_URL` at a Cerebras endpoint or set `ELIZA_PROVIDER=cerebras` and the plugin automatically adapts: structured output uses `json_object` mode, `reasoning_effort` defaults to `"low"` for reasoning-capable models (to prevent empty responses), and `CEREBRAS_API_KEY` is accepted as an alias for `OPENAI_API_KEY`. Embeddings fall back to a deterministic local hash when no explicit embedding URL is set, since Cerebras does not provide an embeddings endpoint.
+
+## EvoLink compatibility
 
-js-tiktoken is WASM and browser-safe; this plugin uses `encodingForModel` directly in both Node and browser builds.
+Set `EVOLINK_API_KEY` to use EvoLink through its OpenAI-compatible endpoint. The plugin defaults to `https://direct.evolink.ai/v1` and `gpt-5.2`; set `EVOLINK_BASE_URL` or `EVOLINK_MODEL` to override either value.
+
+## Prompt caching
+
+Pass `providerOptions.openai.promptCacheKey` and `promptCacheRetention` on any `GenerateTextParams` call to enable OpenAI prompt caching:
+
+```ts
+await runtime.useModel(ModelType.TEXT_LARGE, {
+  prompt: "...",
+  providerOptions: {
+    openai: { promptCacheKey: "my-key", promptCacheRetention: "24h" },
+  },
+});
+```

package/auto-enable.ts

@@ -6,9 +6,9 @@
 // auto-enable engine loads dozens of these per boot.
 import type { PluginAutoEnableContext } from "@elizaos/core";
 
-const ENV_KEYS = ["OPENAI_API_KEY"] as const;
+const ENV_KEYS = ["OPENAI_API_KEY", "CEREBRAS_API_KEY", "EVOLINK_API_KEY"] as const;
 
-/** Enable when an OpenAI API key is present in the environment. */
+/** Enable when an OpenAI-compatible API key is present in the environment. */
 export function shouldEnable(ctx: PluginAutoEnableContext): boolean {
   return ENV_KEYS.some((k) => {
     const v = ctx.env[k];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@elizaos/plugin-openai",
-  "version": "2.0.0-beta.1",
+  "version": "2.0.3-beta.2",
   "type": "module",
   "main": "dist/node/index.node.js",
   "module": "dist/node/index.node.js",
@@ -26,6 +26,12 @@
         "default": "./dist/node/index.node.js"
       },
       "default": "./dist/node/index.node.js"
+    },
+    "./*.css": "./dist/*.css",
+    "./*": {
+      "types": "./dist/*.d.ts",
+      "import": "./dist/*.js",
+      "default": "./dist/*.js"
     }
   },
   "files": [
@@ -53,7 +59,7 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.14",
-    "@elizaos/core": "2.0.0-beta.1",
+    "@elizaos/core": "2.0.3-beta.2",
     "@types/bun": "^1.3.8",
     "@types/json-schema": "^7.0.15",
     "@types/node": "^25.0.3",
@@ -62,7 +68,7 @@
     "vitest": "^4.0.0"
   },
   "peerDependencies": {
-    "@elizaos/core": "2.0.0-beta.1",
+    "@elizaos/core": "2.0.3-beta.2",
     "zod": "^4.4.3"
   },
   "scripts": {
@@ -73,9 +79,9 @@
     "clean": "rm -rf dist .turbo",
     "format": "bunx @biomejs/biome format --write .",
     "format:check": "bunx @biomejs/biome format .",
-    "typecheck": "tsc --noEmit",
-    "test": "vitest run --config ./vitest.config.ts --passWithNoTests",
-    "test:ts": "vitest run --config ./vitest.config.ts --passWithNoTests",
+    "typecheck": "tsgo --noEmit",
+    "test": "vitest run --config ./vitest.config.ts",
+    "test:ts": "vitest run --config ./vitest.config.ts",
     "build": "bun run build.ts"
   },
   "publishConfig": {
@@ -97,6 +103,26 @@
         "default": "https://api.openai.com/v1",
         "sensitive": false
       },
+      "EVOLINK_API_KEY": {
+        "type": "string",
+        "description": "API key used to authenticate requests to EvoLink's OpenAI-compatible API.",
+        "required": false,
+        "sensitive": true
+      },
+      "EVOLINK_BASE_URL": {
+        "type": "string",
+        "description": "Base URL for EvoLink's OpenAI-compatible API.",
+        "required": false,
+        "default": "https://direct.evolink.ai/v1",
+        "sensitive": false
+      },
+      "EVOLINK_MODEL": {
+        "type": "string",
+        "description": "Identifier of the EvoLink chat model to use when EVOLINK_API_KEY enables this plugin.",
+        "required": false,
+        "default": "gpt-5.2",
+        "sensitive": false
+      },
       "OPENAI_SMALL_MODEL": {
         "type": "string",
         "description": "Identifier of the small language model to be used (overrides SMALL_MODEL).",
@@ -149,6 +175,18 @@
         "default": 1536,
         "sensitive": false
       },
+      "OPENAI_IMAGE_DESCRIPTION_API_KEY": {
+        "type": "string",
+        "description": "Optional API key for image-description requests. Falls back to the general OpenAI API key when unset.",
+        "required": false,
+        "sensitive": true
+      },
+      "OPENAI_IMAGE_DESCRIPTION_BASE_URL": {
+        "type": "string",
+        "description": "Optional OpenAI-compatible base URL for image-description requests. Useful when text models use a non-vision-compatible endpoint.",
+        "required": false,
+        "sensitive": false
+      },
       "OPENAI_IMAGE_DESCRIPTION_MODEL": {
         "type": "string",
         "description": "Identifier of the model used for describing images.",
@@ -204,7 +242,7 @@
       }
     }
   },
-  "gitHead": "646c632924826e2b75c2304a75ee56959fe4a460",
+  "gitHead": "82fe0f44215954c2417328203f5bd6510985c1fc",
   "eliza": {
     "platforms": [
       "browser",

package/dist/browser/index.browser.js

@@ -1,3 +0,0 @@
-import{logger as W,ModelType as V}from"@elizaos/core";import{logger as u}from"@elizaos/core";import{logger as v}from"@elizaos/core";function J0(k){if(typeof process>"u"||!process.env)return;let I=process.env[k];return I===void 0?void 0:String(I)}function H(k,I,$){let K=k.getSetting(I);if(K!==void 0&&K!==null)return String(K);return J0(I)??$}function m(k,I,$){let K=H(k,I);if(K===void 0)return $;let X=Number.parseInt(K,10);if(!Number.isFinite(X))throw Error(`Setting '${I}' must be a valid integer, got: ${K}`);return X}function _k(k,I,$){let K=H(k,I);if(K===void 0)return $;let X=K.toLowerCase();return X==="true"||X==="1"||X==="yes"}function w(){return typeof globalThis<"u"&&typeof globalThis.document<"u"}function zk(k){return w()&&!!H(k,"OPENAI_BROWSER_BASE_URL")}function P(k){let I=H(k,"MILADY_PROVIDER");if(I&&I.toLowerCase()==="cerebras")return!0;let $=H(k,"OPENAI_BASE_URL");if($&&/(^|\.)cerebras\.ai(\/|$)/i.test($))return!0;return!1}function L(k){if(P(k)){let I=H(k,"CEREBRAS_API_KEY");if(I)return I}return H(k,"OPENAI_API_KEY")}function j0(k){let I=H(k,"OPENAI_EMBEDDING_API_KEY");if(I)return v.debug("[OpenAI] Using specific embedding API key"),I;return v.debug("[OpenAI] Falling back to general API key for embeddings"),L(k)}function F(k,I=!1){if(w()&&!_k(k,"OPENAI_ALLOW_BROWSER_API_KEY",!1))return{};let $=I?j0(k):L(k);return $?{Authorization:`Bearer ${$}`}:{}}function R(k){let I=H(k,"OPENAI_BROWSER_BASE_URL"),$=w()&&I?I:H(k,"OPENAI_BASE_URL")??"https://api.openai.com/v1";return v.debug(`[OpenAI] Base URL: ${$}`),$}function Rk(k){let I=w()?H(k,"OPENAI_BROWSER_EMBEDDING_URL")??H(k,"OPENAI_BROWSER_BASE_URL"):H(k,"OPENAI_EMBEDDING_URL");if(I)return v.debug(`[OpenAI] Using embedding base URL: ${I}`),I;return v.debug("[OpenAI] Falling back to general base URL for embeddings"),R(k)}function T(k){return H(k,"OPENAI_SMALL_MODEL")??H(k,"SMALL_MODEL")??"gpt-5.4-mini"}function Fk(k){return H(k,"OPENAI_NANO_MODEL")??H(k,"NANO_MODEL")??T(k)}function s(k){return H(k,"OPENAI_MEDIUM_MODEL")??H(k,"MEDIUM_MODEL")??T(k)}function g(k){return H(k,"OPENAI_LARGE_MODEL")??H(k,"LARGE_MODEL")??"gpt-5"}function Wk(k){return H(k,"OPENAI_MEGA_MODEL")??H(k,"MEGA_MODEL")??g(k)}function Dk(k){return H(k,"OPENAI_RESPONSE_HANDLER_MODEL")??H(k,"OPENAI_SHOULD_RESPOND_MODEL")??H(k,"RESPONSE_HANDLER_MODEL")??H(k,"SHOULD_RESPOND_MODEL")??T(k)}function Ek(k){return H(k,"OPENAI_ACTION_PLANNER_MODEL")??H(k,"OPENAI_PLANNER_MODEL")??H(k,"ACTION_PLANNER_MODEL")??H(k,"PLANNER_MODEL")??s(k)}function wk(k){return H(k,"OPENAI_EMBEDDING_MODEL")??"text-embedding-3-small"}function Lk(k){return H(k,"OPENAI_IMAGE_DESCRIPTION_MODEL")??"gpt-5-mini"}function Nk(k){return H(k,"OPENAI_TRANSCRIPTION_MODEL")??"gpt-5-mini-transcribe"}function Ok(k){return H(k,"OPENAI_TTS_MODEL")??"tts-1"}function Pk(k){return H(k,"OPENAI_TTS_VOICE")??"nova"}function Tk(k){return H(k,"OPENAI_TTS_INSTRUCTIONS")??""}function Bk(k){return H(k,"OPENAI_IMAGE_MODEL")??"dall-e-3"}function fk(k){return _k(k,"OPENAI_EXPERIMENTAL_TELEMETRY",!1)}function bk(k){return m(k,"OPENAI_EMBEDDING_DIMENSIONS",1536)}function xk(k){return m(k,"OPENAI_IMAGE_DESCRIPTION_MAX_TOKENS",8192)}function Mk(k){return H(k,"OPENAI_RESEARCH_MODEL")??"o3-deep-research"}function Ck(k){return m(k,"OPENAI_RESEARCH_TIMEOUT",3600000)}globalThis.AI_SDK_LOG_WARNINGS??=!1;function yk(k,I){G0(I)}async function G0(k){if(w()){u.debug("[OpenAI] Skipping API validation in browser environment");return}if(!L(k)){u.warn("[OpenAI] OPENAI_API_KEY is not configured. OpenAI functionality will fail until a valid API key is provided.");return}try{let $=R(k),K=await fetch(`${$}/models`,{headers:F(k)});if(!K.ok){u.warn(`[OpenAI] API key validation failed: ${K.status} ${K.statusText}. Please verify your OPENAI_API_KEY is correct.`);return}}catch($){let K=$ instanceof Error?$.message:String($);u.warn(`[OpenAI] API validation error: ${K}. OpenAI functionality may be limited.`)}}import{logger as y,recordLlmCall as ck}from"@elizaos/core";import{logger as H0}from"@elizaos/core";var B={WAV:{HEADER:[82,73,70,70],IDENTIFIER:[87,65,86,69]},MP3_ID3:[73,68,51],OGG:[79,103,103,83],FLAC:[102,76,97,67],FTYP:[102,116,121,112],WEBM_EBML:[26,69,223,163]},Y0=12;function f(k,I,$){for(let K=0;K<$.length;K++){let X=$[K];if(X===void 0||k[I+K]!==X)return!1}return!0}function r(k){if(k.length<Y0)return"application/octet-stream";if(f(k,0,B.WAV.HEADER)&&f(k,8,B.WAV.IDENTIFIER))return"audio/wav";let I=k[0],$=k[1];if(f(k,0,B.MP3_ID3)||I===255&&$!==void 0&&($&224)===224)return"audio/mpeg";if(f(k,0,B.OGG))return"audio/ogg";if(f(k,0,B.FLAC))return"audio/flac";if(f(k,4,B.FTYP))return"audio/mp4";if(f(k,0,B.WEBM_EBML))return"audio/webm";return H0.warn("Could not detect audio format from buffer, using generic binary type"),"application/octet-stream"}function U0(k){switch(k){case"audio/wav":return"wav";case"audio/mpeg":return"mp3";case"audio/ogg":return"ogg";case"audio/flac":return"flac";case"audio/mp4":return"m4a";case"audio/webm":return"webm";case"application/octet-stream":return"bin"}}function Sk(k){return`recording.${U0(k)}`}function vk(k){return k instanceof Blob||k instanceof File}function t(k){return Buffer.isBuffer(k)}function q0(k){return typeof k==="object"&&k!==null&&"audio"in k&&(vk(k.audio)||t(k.audio))}function V0(k){return typeof k==="object"&&k!==null&&"audioUrl"in k&&typeof k.audioUrl==="string"}async function hk(k){let I=await fetch(k);if(!I.ok)throw Error(`Failed to fetch audio from URL: ${I.status}`);return I.blob()}async function a(k,I){let $=Nk(k),K,X={};if(typeof I==="string")y.debug(`[OpenAI] Fetching audio from URL: ${I}`),K=await hk(I);else if(vk(I))K=I;else if(t(I)){let G=r(I);y.debug(`[OpenAI] Auto-detected audio MIME type: ${G}`),K=new Blob([new Uint8Array(I)],{type:G})}else if(q0(I)){if(X=I,I.model)$=I.model;if(t(I.audio)){let G=I.mimeType??r(I.audio);y.debug(`[OpenAI] Using MIME type: ${G}`),K=new Blob([new Uint8Array(I.audio)],{type:G})}else K=I.audio}else if(V0(I))y.debug(`[OpenAI] Fetching audio from URL: ${I.audioUrl}`),K=await hk(I.audioUrl),X={prompt:I.prompt};else throw Error("TRANSCRIPTION expects Blob, File, Buffer, URL string, or TranscriptionParams object");y.debug(`[OpenAI] Using TRANSCRIPTION model: ${$}`);let A=K.type||"audio/webm",J=K.name||Sk(A.startsWith("audio/")?A:"audio/webm"),Q=new FormData;if(Q.append("file",K,J),Q.append("model",$),X.language)Q.append("language",X.language);if(X.responseFormat)Q.append("response_format",X.responseFormat);if(X.prompt)Q.append("prompt",X.prompt);if(X.temperature!==void 0)Q.append("temperature",String(X.temperature));if(X.timestampGranularities)for(let G of X.timestampGranularities)Q.append("timestamp_granularities[]",G);let Z=R(k),j={model:$,systemPrompt:X.prompt??"",userPrompt:[`audio transcription request: filename=${J}`,`mimeType=${A}`,X.language?`language=${X.language}`:"",X.responseFormat?`responseFormat=${X.responseFormat}`:""].filter(Boolean).join(" "),temperature:X.temperature??0,maxTokens:0,purpose:"external_llm",actionType:"openai.audio.transcriptions.create"};return(await ck(k,j,async()=>{let G=await fetch(`${Z}/audio/transcriptions`,{method:"POST",headers:F(k),body:Q});if(!G.ok){let _=await G.text().catch(()=>"Unknown error");throw Error(`OpenAI transcription failed: ${G.status} ${G.statusText} - ${_}`)}let U=await G.json();return j.response=U.text,U})).text}async function e(k,I){let $,K,X="mp3",A,J;if(typeof I==="string")$=I,K=void 0;else{if($=I.text,K=I.voice,"format"in I&&I.format)X=I.format;if("model"in I&&I.model)A=I.model;if("instructions"in I&&I.instructions)J=I.instructions}if(A=A??Ok(k),K=K??Pk(k),J=J??Tk(k),y.debug(`[OpenAI] Using TEXT_TO_SPEECH model: ${A}`),!$||$.trim().length===0)throw Error("TEXT_TO_SPEECH requires non-empty text");if($.length>4096)throw Error("TEXT_TO_SPEECH text exceeds 4096 character limit");let Q=["alloy","echo","fable","onyx","nova","shimmer"];if(K&&!Q.includes(K))throw Error(`Invalid voice: ${K}. Must be one of: ${Q.join(", ")}`);let Z=R(k),j={model:A,voice:K,input:$,response_format:X};if(J&&J.length>0)j.instructions=J;let Y={model:A,systemPrompt:J??"",userPrompt:$,temperature:0,maxTokens:0,purpose:"external_llm",actionType:"openai.audio.speech.create"};return ck(k,Y,async()=>{let G=await fetch(`${Z}/audio/speech`,{method:"POST",headers:{...F(k),"Content-Type":"application/json",...X==="mp3"?{Accept:"audio/mpeg"}:{}},body:JSON.stringify(j)});if(!G.ok){let _=await G.text().catch(()=>"Unknown error");throw Error(`OpenAI TTS failed: ${G.status} ${G.statusText} - ${_}`)}let U=await G.arrayBuffer();return Y.response=`[audio bytes=${U.byteLength} format=${X}]`,U})}import{logger as d,ModelType as F0,VECTOR_DIMS as W0}from"@elizaos/core";import{EventType as _0}from"@elizaos/core";var gk=200;function z0(k){if(k.length<=gk)return k;return`${k.slice(0,gk)}…`}function R0(k){if("promptTokens"in k){let I="promptTokensDetails"in k?k.promptTokensDetails:void 0,$=k.cachedPromptTokens??I?.cachedTokens;return{promptTokens:k.promptTokens??0,completionTokens:k.completionTokens??0,totalTokens:k.totalTokens??(k.promptTokens??0)+(k.completionTokens??0),cachedPromptTokens:$}}if("inputTokens"in k||"outputTokens"in k){let I=k.inputTokens??0,$=k.outputTokens??0,K=k.totalTokens??I+$;return{promptTokens:I,completionTokens:$,totalTokens:K,cachedPromptTokens:k.cachedInputTokens}}return{promptTokens:0,completionTokens:0,totalTokens:0}}function S(k,I,$,K){let X=R0(K),A={runtime:k,source:"openai",provider:"openai",type:I,prompt:z0($),tokens:{prompt:X.promptTokens,completion:X.completionTokens,total:X.totalTokens,...X.cachedPromptTokens!==void 0?{cached:X.cachedPromptTokens}:{}}};k.emitEvent(_0.MODEL_USED,A)}function D0(k){let I=Object.values(W0);if(!I.includes(k))throw Error(`Invalid embedding dimension: ${k}. Must be one of: ${I.join(", ")}`);return k}function E0(k){if(k===null)return null;if(typeof k==="string")return k;if(typeof k==="object"&&typeof k.text==="string")return k.text;throw Error("Invalid embedding params: expected string, { text: string }, or null")}function w0(k){let I=w()?"OPENAI_BROWSER_EMBEDDING_URL":"OPENAI_EMBEDDING_URL",$=H(k,I);return typeof $==="string"&&$.trim().length>0}function L0(k){return P(k)&&!w0(k)}function dk(k){let I=2166136261;for(let $=0;$<k.length;$+=1)I^=k.charCodeAt($),I=Math.imul(I,16777619);return I>>>0}function N0(k,I){let $=Array(I).fill(0),K=k.toLowerCase(),X=K.match(/[a-z0-9]+(?:[_-][a-z0-9]+)*/g)??[K],A=(Q,Z)=>{let j=dk(Q),Y=j%I,G=(j&1)===0?1:-1;$[Y]+=G*Z;let U=dk(`b:${Q}`),_=U%I,z=(U&1)===0?1:-1;$[_]+=z*Z*0.5};X.forEach((Q,Z)=>{if(A(Q,1),Z>0)A(`${X[Z-1]} ${Q}`,0.35)}),A(K.slice(0,512),0.15);let J=Math.sqrt($.reduce((Q,Z)=>Q+Z*Z,0));if(J===0)return $[0]=1,$;return $.map((Q)=>Q/J)}async function kk(k,I){let $=wk(k),K=D0(bk(k)),X=E0(I);if(X===null){d.debug("[OpenAI] Creating test embedding for initialization");let _=Array(K).fill(0);return _[0]=0.1,_}let A=X.trim();if(A.length===0)throw Error("Cannot generate embedding for empty text");let J=32000;if(A.length>J)d.warn(`[OpenAI] Embedding input too long (~${Math.ceil(A.length/4)} tokens), truncating to ~8000 tokens`),A=A.slice(0,J);if(L0(k))return d.debug("[OpenAI] Using deterministic local embedding fallback for Cerebras mode"),N0(A,K);let Z=`${Rk(k)}/embeddings`;d.debug(`[OpenAI] Generating embedding with model: ${$}`);let j=await fetch(Z,{method:"POST",headers:{...F(k,!0),"Content-Type":"application/json"},body:JSON.stringify({model:$,input:A})});if(!j.ok){let _=await j.text().catch(()=>"Unknown error");throw Error(`OpenAI embedding API error: ${j.status} ${j.statusText} - ${_}`)}let Y=await j.json(),G=Y?.data?.[0];if(!G?.embedding)throw Error("OpenAI API returned invalid embedding response structure");let U=G.embedding;if(U.length!==K)throw Error(`Embedding dimension mismatch: got ${U.length}, expected ${K}. Check OPENAI_EMBEDDING_DIMENSIONS setting.`);if(Y.usage)S(k,F0.TEXT_EMBEDDING,A,{promptTokens:Y.usage.prompt_tokens,completionTokens:0,totalTokens:Y.usage.total_tokens});return d.debug(`[OpenAI] Generated embedding with ${U.length} dimensions`),U}import{logger as pk,ModelType as O0,recordLlmCall as ok}from"@elizaos/core";var lk="Please analyze this image and provide a title and detailed description.";async function Ik(k,I){let $=Bk(k),K=I.count??1,X=I.size??"1024x1024",A=I;if(pk.debug(`[OpenAI] Using IMAGE model: ${$}`),I.prompt.trim().length===0)throw Error("IMAGE generation requires a non-empty prompt");if(K<1||K>10)throw Error("IMAGE count must be between 1 and 10");let J=R(k),Q={model:$,prompt:I.prompt,n:K,size:X};if(A.quality)Q.quality=A.quality;if(A.style)Q.style=A.style;let Z={model:$,systemPrompt:"",userPrompt:I.prompt,temperature:0,maxTokens:0,purpose:"external_llm",actionType:"openai.images.generate"},j=await ok(k,Z,async()=>{let Y=await fetch(`${J}/images/generations`,{method:"POST",headers:{...F(k),"Content-Type":"application/json"},body:JSON.stringify(Q)});if(!Y.ok){let U=await Y.text().catch(()=>"Unknown error");throw Error(`OpenAI image generation failed: ${Y.status} ${Y.statusText} - ${U}`)}let G=await Y.json();return Z.response=JSON.stringify(G.data),G});if(j.data.length===0)throw Error("OpenAI API returned no images");return j.data.map((Y)=>({url:Y.url,revisedPrompt:Y.revised_prompt}))}function P0(k){return k.match(/title[:\s]+(.+?)(?:\n|$)/i)?.[1]?.trim()??"Image Analysis"}function T0(k){return k.replace(/title[:\s]+(.+?)(?:\n|$)/i,"").trim()}async function $k(k,I){let $=Lk(k),K=xk(k);pk.debug(`[OpenAI] Using IMAGE_DESCRIPTION model: ${$}`);let X,A;if(typeof I==="string")X=I,A=lk;else X=I.imageUrl,A=I.prompt??lk;if(!X||X.trim().length===0)throw Error("IMAGE_DESCRIPTION requires a valid image URL");let J=R(k),Q={model:$,messages:[{role:"user",content:[{type:"text",text:A},{type:"image_url",image_url:{url:X}}]}],max_tokens:K},Z={model:$,systemPrompt:"",userPrompt:A,temperature:0,maxTokens:K,purpose:"external_llm",actionType:"openai.chat.completions.create"},j=await ok(k,Z,async()=>{let U=await fetch(`${J}/chat/completions`,{method:"POST",headers:{...F(k),"Content-Type":"application/json"},body:JSON.stringify(Q)});if(!U.ok){let z=await U.text().catch(()=>"Unknown error");throw Error(`OpenAI image description failed: ${U.status} ${U.statusText} - ${z}`)}let _=await U.json();if(Z.response=_.choices?.[0]?.message?.content??"",_.usage)Z.promptTokens=_.usage.prompt_tokens,Z.completionTokens=_.usage.completion_tokens;return _});if(j.usage)S(k,O0.IMAGE_DESCRIPTION,typeof I==="string"?I:I.prompt??"",{promptTokens:j.usage.prompt_tokens,completionTokens:j.usage.completion_tokens,totalTokens:j.usage.total_tokens});let G=j.choices?.[0]?.message?.content;if(!G)throw Error("OpenAI API returned empty image description");return{title:P0(G),description:T0(G)}}import{logger as N,recordLlmCall as B0}from"@elizaos/core";function f0(k){switch(k.type){case"web_search_preview":return{type:"web_search_preview"};case"file_search":return{type:"file_search",vector_store_ids:k.vectorStoreIds};case"code_interpreter":return{type:"code_interpreter",container:k.container??{type:"auto"}};case"mcp":return{type:"mcp",server_label:k.serverLabel,server_url:k.serverUrl,require_approval:k.requireApproval??"never"};default:throw Error(`Unknown research tool type: ${k.type}`)}}function b0(k){switch(k.type){case"web_search_call":return{id:k.id??"",type:"web_search_call",status:k.status??"completed",action:{type:k.action?.type??"search",query:k.action?.query,url:k.action?.url}};case"file_search_call":return{id:k.id??"",type:"file_search_call",status:k.status??"completed",query:k.query??"",results:k.results?.map((I)=>({fileId:I.file_id,fileName:I.file_name,score:I.score}))};case"code_interpreter_call":return{id:k.id??"",type:"code_interpreter_call",status:k.status??"completed",code:k.code??"",output:k.output};case"mcp_tool_call":return{id:k.id??"",type:"mcp_tool_call",status:k.status??"completed",serverLabel:k.server_label??"",toolName:k.tool_name??"",arguments:k.arguments??{},result:k.result};case"message":return{type:"message",content:k.content?.map((I)=>({type:"output_text",text:I.text,annotations:I.annotations?.map(($)=>({url:$.url,title:$.title,startIndex:$.start_index,endIndex:$.end_index}))??[]}))??[]};default:return null}}function x0(k){if(k.output_text){let K=[];if(k.output){for(let X of k.output)if(X.type==="message"&&X.content){for(let A of X.content)if(A.annotations)for(let J of A.annotations)K.push({url:J.url,title:J.title,startIndex:J.start_index,endIndex:J.end_index})}}return{text:k.output_text,annotations:K}}let I="",$=[];if(k.output){for(let K of k.output)if(K.type==="message"&&K.content){for(let X of K.content)if(I+=X.text,X.annotations)for(let A of X.annotations)$.push({url:A.url,title:A.title,startIndex:A.start_index,endIndex:A.end_index})}}return{text:I,annotations:$}}async function Kk(k,I){let $=L(k);if(!$)throw Error("OPENAI_API_KEY is required for deep research. Set it in your environment variables or runtime settings.");let K=R(k),X=I.model??Mk(k),A=Ck(k);N.debug(`[OpenAI] Starting deep research with model: ${X}`),N.debug(`[OpenAI] Research input: ${I.input.substring(0,100)}...`);let J=I.tools?.filter((z)=>z.type==="web_search_preview"||z.type==="file_search"||z.type==="mcp");if(!J||J.length===0)N.debug("[OpenAI] No data source tools specified, defaulting to web_search_preview"),I.tools=[{type:"web_search_preview"},...I.tools??[]];let Q={model:X,input:I.input};if(I.instructions)Q.instructions=I.instructions;if(I.background!==void 0)Q.background=I.background;if(I.tools&&I.tools.length>0)Q.tools=I.tools.map(f0);if(I.maxToolCalls!==void 0)Q.max_tool_calls=I.maxToolCalls;if(I.reasoningSummary)Q.reasoning={summary:I.reasoningSummary};N.debug(`[OpenAI] Research request body: ${JSON.stringify(Q,null,2)}`);let Z={model:X,systemPrompt:I.instructions??"",userPrompt:I.input,temperature:0,maxTokens:0,purpose:"external_llm",actionType:"openai.responses.create"},j=await B0(k,Z,async()=>{let z=await fetch(`${K}/responses`,{method:"POST",headers:{Authorization:`Bearer ${$}`,"Content-Type":"application/json"},body:JSON.stringify(Q),signal:AbortSignal.timeout(A)});if(!z.ok){let l=await z.text();throw N.error(`[OpenAI] Research request failed: ${z.status} ${l}`),Error(`Deep research request failed: ${z.status} ${z.statusText}`)}let O=await z.json();return Z.response=O.output_text??"",O});if(j.error)throw N.error(`[OpenAI] Research API error: ${j.error.message}`),Error(`Deep research error: ${j.error.message}`);N.debug(`[OpenAI] Research response received. Status: ${j.status??"completed"}`);let{text:Y,annotations:G}=x0(j),U=[];if(j.output)for(let z of j.output){let O=b0(z);if(O)U.push(O)}let _={id:j.id,text:Y,annotations:G,outputItems:U,status:j.status};return N.info(`[OpenAI] Research completed. Text length: ${Y.length}, Annotations: ${G.length}, Output items: ${U.length}`),_}import{buildCanonicalSystemPrompt as C0,dropDuplicateLeadingSystemMessage as y0,logger as S0,ModelType as x,normalizeSchemaForCerebras as h0,recordLlmCall as nk,resolveEffectiveSystemPrompt as c0,sanitizeFunctionNameForCerebras as v0}from"@elizaos/core";import{generateText as g0,jsonSchema as sk,Output as d0,streamText as l0}from"ai";import{createOpenAI as uk}from"@ai-sdk/openai";var M0="sk-proxy";function Xk(k){let I=R(k),$=L(k);if(!$&&zk(k))return uk({apiKey:M0,baseURL:I});if(!$)throw Error("OPENAI_API_KEY is required. Set it in your environment variables or runtime settings.");return uk({apiKey:$,baseURL:I})}var p0=x.TEXT_NANO??"TEXT_NANO",o0=x.TEXT_MEDIUM??"TEXT_MEDIUM",u0=x.TEXT_MEGA??"TEXT_MEGA",n0=x.RESPONSE_HANDLER??"RESPONSE_HANDLER",i0=x.ACTION_PLANNER??"ACTION_PLANNER";function m0(k){let I=[{type:"text",text:k.prompt}];for(let $ of k.attachments??[])I.push({type:"file",data:$.data,mediaType:$.mediaType,...$.filename?{filename:$.filename}:{}});return I}function rk(k){if(!k)return;let I=k.inputTokens??0,$=k.outputTokens??0,K=k,X=ik(K.cacheReadInputTokens,K.cachedInputTokens,K.inputTokenDetails?.cacheReadTokens,K.inputTokenDetails?.cachedInputTokens,K.input_tokens_details?.cache_read_input_tokens,K.input_tokens_details?.cached_tokens,K.prompt_tokens_details?.cached_tokens)??void 0,A=ik(K.cacheCreationInputTokens,K.cacheWriteInputTokens,K.inputTokenDetails?.cacheCreationInputTokens,K.inputTokenDetails?.cacheCreationTokens,K.inputTokenDetails?.cacheWriteTokens,K.input_tokens_details?.cache_creation_input_tokens);return{promptTokens:I,completionTokens:$,totalTokens:I+$,cachedPromptTokens:X,cacheReadInputTokens:X,cacheCreationInputTokens:A}}function ik(...k){for(let I of k){if(typeof I==="number"&&Number.isFinite(I))return I;if(typeof I==="string"&&I.trim().length>0){let $=Number(I);if(Number.isFinite($))return $}}return}function s0(k){let I=k;return{promptCacheKey:I.providerOptions?.openai?.promptCacheKey,promptCacheRetention:I.providerOptions?.openai?.promptCacheRetention}}function r0(k,I){let K=k.providerOptions,X=s0(k);if(!K&&!X.promptCacheKey&&!X.promptCacheRetention)return;let A=P(I),{agentName:J,openai:Q,...Z}=K??{},Y={...(()=>{if(!Q||typeof Q!=="object")return Q;if(!A)return Q;let{promptCacheRetention:U,..._}=Q;return _})()??{},...X.promptCacheKey?{promptCacheKey:X.promptCacheKey}:{},...!A&&X.promptCacheRetention?{promptCacheRetention:X.promptCacheRetention}:{}},G={...Z,...Object.keys(Y).length>0?{openai:Y}:{}};return Object.keys(G).length>0?G:void 0}function t0(k){if(k&&typeof k==="object"&&"responseFormat"in k&&"parseCompleteOutput"in k)return k;let I=k&&typeof k==="object"&&"schema"in k?k:{schema:k};return d0.object({schema:sk(h(I.schema,!0)),...I.name?{name:I.name}:{},...I.description?{description:I.description}:{}})}function a0(k,I={}){if(!k)return;if(!Array.isArray(k))return k;let $={};for(let K of k){let X=b(K),A=b(X.function),J=E(X.name,A.name);if(!J)throw Error("[OpenAI] Native tool definition is missing a name.");let Q=E(X.description,A.description),Z=X.parameters??A.parameters??{type:"object"},j=h(Z,!0);if(I.cerebrasMode)j=h0(j);let Y=I.cerebrasMode?v0(J):J;$[Y]={...Q?{description:Q}:{},inputSchema:sk(j)}}return Object.keys($).length>0?$:void 0}function e0(k){if(!Array.isArray(k))return;return k.map((I)=>kI(I))}function kI(k){let I=b(k),$=ZI(I.providerOptions);if(I.role==="system")return{role:"system",content:tk(I.content),...$?{providerOptions:$}:{}};if(I.role==="assistant")return{role:"assistant",content:II(I),...$?{providerOptions:$}:{}};if(I.role==="tool")return{role:"tool",content:$I(I),...$?{providerOptions:$}:{}};return{role:"user",content:KI(I.content),...$?{providerOptions:$}:{}}}function II(k){let I=Array.isArray(k.toolCalls)?k.toolCalls:[];if(I.length===0){if(Array.isArray(k.content)||typeof k.content==="string")return k.content;return""}let $=[];if(typeof k.content==="string"&&k.content.length>0)$.push({type:"text",text:k.content});else if(Array.isArray(k.content))$.push(...k.content);for(let K of I){let X=b(K),A=b(X.function),J=E(X.toolCallId,X.id),Q=E(X.toolName,X.name,A.name);if(!J||!Q)continue;$.push({type:"tool-call",toolCallId:J,toolName:Q,input:XI(X,A)})}return $}function $I(k){if(Array.isArray(k.content))return k.content;let I=E(k.toolCallId,k.id)??"tool-call",$=E(k.toolName,k.name)??"tool",K=ak(k.content);return[{type:"tool-result",toolCallId:I,toolName:$,output:typeof K==="string"?{type:"text",value:K}:{type:"json",value:K}}]}function KI(k){if(Array.isArray(k))return k;return tk(k)}function tk(k){if(typeof k==="string")return k;if(k==null)return"";return typeof k==="object"?JSON.stringify(k):String(k)}function XI(k,I){if("input"in k)return k.input;return ak(k.arguments??I.arguments??{})}function ak(k){if(typeof k!=="string")return k??"";try{return JSON.parse(k)}catch{return k}}function AI(k){if(!k)return;if(typeof k==="string"&&(k==="auto"||k==="none"||k==="required"))return k;let I=b(k);if(I.type==="tool"){if(typeof I.toolName==="string"&&I.toolName.length>0)return k;let K=E(I.toolName,I.name);if(K)return{type:"tool",toolName:K}}if(I.type==="function"){let K=b(I.function),X=E(K.name);if(X)return{type:"tool",toolName:X}}let $=E(I.name);if($)return{type:"tool",toolName:$};return k}function h(k,I=!1){if(!k||typeof k!=="object"||Array.isArray(k))return{type:"object"};let K={...k};if(typeof K.type!=="string"){let X=QI(K,I);if(X)K.type=X}if(K.properties&&typeof K.properties==="object"&&!Array.isArray(K.properties)){let X={};for(let[Q,Z]of Object.entries(K.properties))X[Q]=h(Z);K.properties=X;let A=Object.keys(X),J=Array.isArray(K.required)?K.required.filter((Q)=>typeof Q==="string"):[];K.required=[...new Set([...J,...A])]}if(K.type==="object"&&K.additionalProperties!==!1)K.additionalProperties=!1;if(K.items)K.items=Array.isArray(K.items)?K.items.map((X)=>h(X)):h(K.items);for(let X of["anyOf","oneOf","allOf"]){let A=K[X];if(Array.isArray(A))K[X]=A.map((J)=>h(J))}return K}function QI(k,I){if("properties"in k||"required"in k||"additionalProperties"in k||I)return"object";if("items"in k)return"array";if(Array.isArray(k.enum)&&k.enum.length>0){let $=new Set(k.enum.map((K)=>typeof K));if($.size===1){let[K]=[...$];if(K==="string"||K==="number"||K==="boolean")return K}}return}function b(k){return k&&typeof k==="object"&&!Array.isArray(k)?k:{}}function ZI(k){return k&&typeof k==="object"&&!Array.isArray(k)?k:void 0}function E(...k){for(let I of k)if(typeof I==="string"&&I.length>0)return I;return}function JI(k){return Boolean(k.messages||k.tools||k.toolChoice||k.responseSchema)}function jI(k,I){return{text:k.text,toolCalls:k.toolCalls??[],finishReason:k.finishReason,usage:rk(k.usage),providerMetadata:GI(k.providerMetadata,I)}}function GI(k,I){if(!I)return k;if(k&&typeof k==="object"&&!Array.isArray(k))return{...k,modelName:I};return{modelName:I}}function mk(k,I,$,K,X,A,J){let Q=I,Z=J,j=Z&&"prompt"in Z?Z.prompt:void 0,Y=Z&&"messages"in Z&&Array.isArray(Z.messages)?Z.messages:void 0,G=typeof Z?.system==="string"?Z.system:$;return{model:k,modelType:X,provider:"vercel-ai-sdk",systemPrompt:G??"",userPrompt:typeof j==="string"?j:typeof I.prompt==="string"?I.prompt:"",prompt:typeof j==="string"?j:void 0,messages:Y,tools:Z?.tools??Q.tools,toolChoice:Z?.toolChoice??Q.toolChoice,output:Z?.output!==void 0?HI(Q.responseSchema,Z.output):void 0,responseSchema:Q.responseSchema,providerOptions:A??Z?.providerOptions??Q.providerOptions,temperature:I.temperature??0,maxTokens:typeof Z?.maxOutputTokens==="number"?Z.maxOutputTokens:I.maxTokens??8192,purpose:"external_llm",actionType:K}}function HI(k,I){if(k!==void 0)return{type:"object",schema:k};return YI(I)}function YI(k){try{return JSON.parse(JSON.stringify(k,(I,$)=>{if(typeof $==="function")return;if(typeof $==="bigint")return $.toString();return $}))}catch{return String(k)}}function UI(k,I){if(!I)return;k.promptTokens=I.inputTokens??0,k.completionTokens=I.outputTokens??0}async function M(k,I,$,K){let X=I,A=Xk(k),J=K(k);S0.debug(`[OpenAI] Using ${$} model: ${J}`);let Q=r0(I,k),j=(X.attachments?.length??0)>0?m0(X):void 0,Y=JI(X),G=c0({params:X,fallback:C0({character:k.character})}),U=X.providerOptions?.agentName,_={isEnabled:fk(k),functionId:U?`agent:${U}`:void 0,metadata:U?{agentName:U}:void 0},z=A.chat(J),O=P(k),l=a0(X.tools,{cerebrasMode:O}),qk=AI(X.toolChoice),Vk=e0(X.messages),n=y0(Vk,G),i=n&&n.length>0?n:Vk,A0=typeof I.prompt==="string"&&I.prompt.length>0?I.prompt:"",Q0=i&&i.length>0?{messages:i}:j?{messages:[{role:"user",content:j}]}:{prompt:A0},p={model:z,...Q0,system:G,allowSystemInMessages:!0,maxOutputTokens:I.maxTokens??8192,experimental_telemetry:_,...l?{tools:l}:{},...qk?{toolChoice:qk}:{},...X.responseSchema&&!P(k)?{output:t0(X.responseSchema)}:{},...Q?{providerOptions:Q}:{}};if(I.stream){let D=mk(J,I,G,"ai.streamText",$,Q,p);D.response="";let c=await nk(k,D,()=>l0(p));return{textStream:c.textStream,text:Promise.resolve(c.text),...Y?{toolCalls:Promise.resolve(c.toolCalls)}:{},usage:Promise.resolve(c.usage).then(rk),finishReason:Promise.resolve(c.finishReason).then((Z0)=>Z0)}}let C=mk(J,I,G,"ai.generateText",$,Q,p),o=await nk(k,C,async()=>{let D=await g0(p);return C.response=D.text,C.toolCalls=D.toolCalls??[],C.finishReason=D.finishReason,C.providerMetadata=D.providerMetadata,UI(C,D.usage),D});if(o.usage)S(k,$,I.prompt??"",o.usage);if(Y)return jI(o,J);return o.text}async function Ak(k,I){return M(k,I,x.TEXT_SMALL,T)}async function Qk(k,I){return M(k,I,p0,Fk)}async function Zk(k,I){return M(k,I,o0,s)}async function Jk(k,I){return M(k,I,x.TEXT_LARGE,g)}async function jk(k,I){return M(k,I,u0,Wk)}async function Gk(k,I){return M(k,I,n0,Dk)}async function Hk(k,I){return M(k,I,i0,Ek)}import{ModelType as K0}from"@elizaos/core";import{ModelType as qI}from"@elizaos/core";import{encodingForModel as VI,getEncoding as _I}from"js-tiktoken";function ek(k){let $=k.toLowerCase().includes("4o")?"o200k_base":"cl100k_base";try{return VI(k)}catch{return _I($)}}function k0(k,I){if(I===qI.TEXT_SMALL)return T(k);return g(k)}function I0(k,I,$){let K=k0(k,I);return ek(K).encode($)}function $0(k,I,$){let K=k0(k,I);return ek(K).decode($)}async function Yk(k,I){if(!I.prompt)throw Error("Tokenization requires a non-empty prompt");let $=I.modelType??K0.TEXT_LARGE;return I0(k,$,I.prompt)}async function Uk(k,I){if(!I.tokens||!Array.isArray(I.tokens))throw Error("Detokenization requires a valid tokens array");if(I.tokens.length===0)return"";for(let K=0;K<I.tokens.length;K++){let X=I.tokens[K];if(typeof X!=="number"||!Number.isFinite(X))throw Error(`Invalid token at index ${K}: expected number`)}let $=I.modelType??K0.TEXT_LARGE;return $0(k,$,I.tokens)}function zI(){if(typeof process>"u")return{};return process.env}var q=zI(),RI=V.TEXT_NANO??"TEXT_NANO",FI=V.TEXT_MEDIUM??"TEXT_MEDIUM",WI=V.TEXT_MEGA??"TEXT_MEGA",DI=V.RESPONSE_HANDLER??"RESPONSE_HANDLER",EI=V.ACTION_PLANNER??"ACTION_PLANNER",wI={name:"openai",description:"OpenAI API integration for text, image, audio, and embedding models",autoEnable:{envKeys:["OPENAI_API_KEY"]},config:{OPENAI_API_KEY:q.OPENAI_API_KEY??null,OPENAI_BASE_URL:q.OPENAI_BASE_URL??null,OPENAI_NANO_MODEL:q.OPENAI_NANO_MODEL??null,OPENAI_MEDIUM_MODEL:q.OPENAI_MEDIUM_MODEL??null,OPENAI_SMALL_MODEL:q.OPENAI_SMALL_MODEL??null,OPENAI_LARGE_MODEL:q.OPENAI_LARGE_MODEL??null,OPENAI_MEGA_MODEL:q.OPENAI_MEGA_MODEL??null,OPENAI_RESPONSE_HANDLER_MODEL:q.OPENAI_RESPONSE_HANDLER_MODEL??null,OPENAI_SHOULD_RESPOND_MODEL:q.OPENAI_SHOULD_RESPOND_MODEL??null,OPENAI_ACTION_PLANNER_MODEL:q.OPENAI_ACTION_PLANNER_MODEL??null,OPENAI_PLANNER_MODEL:q.OPENAI_PLANNER_MODEL??null,NANO_MODEL:q.NANO_MODEL??null,MEDIUM_MODEL:q.MEDIUM_MODEL??null,SMALL_MODEL:q.SMALL_MODEL??null,LARGE_MODEL:q.LARGE_MODEL??null,MEGA_MODEL:q.MEGA_MODEL??null,RESPONSE_HANDLER_MODEL:q.RESPONSE_HANDLER_MODEL??null,SHOULD_RESPOND_MODEL:q.SHOULD_RESPOND_MODEL??null,ACTION_PLANNER_MODEL:q.ACTION_PLANNER_MODEL??null,PLANNER_MODEL:q.PLANNER_MODEL??null,OPENAI_EMBEDDING_MODEL:q.OPENAI_EMBEDDING_MODEL??null,OPENAI_EMBEDDING_API_KEY:q.OPENAI_EMBEDDING_API_KEY??null,OPENAI_EMBEDDING_URL:q.OPENAI_EMBEDDING_URL??null,OPENAI_EMBEDDING_DIMENSIONS:q.OPENAI_EMBEDDING_DIMENSIONS??null,OPENAI_IMAGE_DESCRIPTION_MODEL:q.OPENAI_IMAGE_DESCRIPTION_MODEL??null,OPENAI_IMAGE_DESCRIPTION_MAX_TOKENS:q.OPENAI_IMAGE_DESCRIPTION_MAX_TOKENS??null,OPENAI_EXPERIMENTAL_TELEMETRY:q.OPENAI_EXPERIMENTAL_TELEMETRY??null,OPENAI_RESEARCH_MODEL:q.OPENAI_RESEARCH_MODEL??null,OPENAI_RESEARCH_TIMEOUT:q.OPENAI_RESEARCH_TIMEOUT??null},async init(k,I){yk(k,I)},models:{[V.TEXT_EMBEDDING]:async(k,I)=>{return kk(k,I)},[V.TEXT_TOKENIZER_ENCODE]:async(k,I)=>{return Yk(k,I)},[V.TEXT_TOKENIZER_DECODE]:async(k,I)=>{return Uk(k,I)},[V.TEXT_SMALL]:async(k,I)=>{return Ak(k,I)},[RI]:async(k,I)=>{return Qk(k,I)},[FI]:async(k,I)=>{return Zk(k,I)},[V.TEXT_LARGE]:async(k,I)=>{return Jk(k,I)},[WI]:async(k,I)=>{return jk(k,I)},[DI]:async(k,I)=>{return Gk(k,I)},[EI]:async(k,I)=>{return Hk(k,I)},[V.IMAGE]:async(k,I)=>{return Ik(k,I)},[V.IMAGE_DESCRIPTION]:async(k,I)=>{return $k(k,I)},[V.TRANSCRIPTION]:async(k,I)=>{return a(k,I)},[V.TEXT_TO_SPEECH]:async(k,I)=>{return e(k,I)},[V.RESEARCH]:async(k,I)=>{return Kk(k,I)}},tests:[{name:"openai_plugin_tests",tests:[{name:"openai_test_api_connectivity",fn:async(k)=>{let I=R(k),$=await fetch(`${I}/models`,{headers:F(k)});if(!$.ok)throw Error(`API connectivity test failed: ${$.status} ${$.statusText}`);let K=await $.json();W.info(`[OpenAI Test] API connected. ${K.data?.length??0} models available.`)}},{name:"openai_test_text_embedding",fn:async(k)=>{let I=await k.useModel(V.TEXT_EMBEDDING,{text:"Hello, world!"});if(!Array.isArray(I)||I.length===0)throw Error("Embedding should return a non-empty array");W.info(`[OpenAI Test] Generated embedding with ${I.length} dimensions`)}},{name:"openai_test_text_small",fn:async(k)=>{let I=await k.useModel(V.TEXT_SMALL,{prompt:"Say hello in exactly 5 words."});if(typeof I!=="string"||I.length===0)throw Error("TEXT_SMALL should return non-empty string");W.info(`[OpenAI Test] TEXT_SMALL generated: "${I.substring(0,50)}..."`)}},{name:"openai_test_text_large",fn:async(k)=>{let I=await k.useModel(V.TEXT_LARGE,{prompt:"Explain quantum computing in 2 sentences."});if(typeof I!=="string"||I.length===0)throw Error("TEXT_LARGE should return non-empty string");W.info(`[OpenAI Test] TEXT_LARGE generated: "${I.substring(0,50)}..."`)}},{name:"openai_test_tokenizer_roundtrip",fn:async(k)=>{let $=await k.useModel(V.TEXT_TOKENIZER_ENCODE,{prompt:"Hello, tokenizer test!",modelType:V.TEXT_SMALL});if(!Array.isArray($)||$.length===0)throw Error("Tokenization should return non-empty token array");let K=await k.useModel(V.TEXT_TOKENIZER_DECODE,{tokens:$,modelType:V.TEXT_SMALL});if(K!=="Hello, tokenizer test!")throw Error(`Tokenizer roundtrip failed: expected "Hello, tokenizer test!", got "${K}"`);W.info(`[OpenAI Test] Tokenizer roundtrip successful (${$.length} tokens)`)}},{name:"openai_test_streaming",fn:async(k)=>{let I=[],$=await k.useModel(V.TEXT_LARGE,{prompt:"Count from 1 to 5, one number per line.",stream:!0,onStreamChunk:(K)=>{I.push(K)}});if(typeof $!=="string"||$.length===0)throw Error("Streaming should return non-empty result");if(I.length===0)throw Error("No streaming chunks received");W.info(`[OpenAI Test] Streaming test: ${I.length} chunks received`)}},{name:"openai_test_image_description",fn:async(k)=>{let $=await k.useModel(V.IMAGE_DESCRIPTION,"https://upload.wikimedia.org/wikipedia/commons/thumb/a/a7/Camponotus_flavomarginatus_ant.jpg/440px-Camponotus_flavomarginatus_ant.jpg");if(!$||typeof $!=="object"||!("title"in $)||!("description"in $))throw Error("Image description should return { title, description }");W.info(`[OpenAI Test] Image described: "${$.title}"`)}},{name:"openai_test_transcription",fn:async(k)=>{let K=await(await fetch("https://upload.wikimedia.org/wikipedia/commons/2/25/En-Open_Source.ogg")).arrayBuffer(),X=Buffer.from(new Uint8Array(K)),A=await k.useModel(V.TRANSCRIPTION,X);if(typeof A!=="string")throw Error("Transcription should return a string");W.info(`[OpenAI Test] Transcription: "${A.substring(0,50)}..."`)}},{name:"openai_test_text_to_speech",fn:async(k)=>{let I=await k.useModel(V.TEXT_TO_SPEECH,{text:"Hello, this is a text-to-speech test."});if(!(I instanceof ArrayBuffer)||I.byteLength===0)throw Error("TTS should return non-empty ArrayBuffer");W.info(`[OpenAI Test] TTS generated ${I.byteLength} bytes of audio`)}},{name:"openai_test_structured_output_via_text_large",fn:async(k)=>{let I=await k.useModel(V.TEXT_LARGE,{prompt:"Return a JSON object with exactly these fields: name (string), age (number), active (boolean)",responseSchema:{type:"object",properties:{name:{type:"string"},age:{type:"number"},active:{type:"boolean"}},required:["name","age","active"]}});if(!I||typeof I!=="object"&&typeof I!=="string")throw Error("Structured output should return an object or text");W.info(`[OpenAI Test] Structured output: ${JSON.stringify(I).substring(0,100)}`)}},{name:"openai_test_research",fn:async(k)=>{let I=await k.useModel(V.RESEARCH,{input:"What is the current date and time?",tools:[{type:"web_search_preview"}],maxToolCalls:3});if(!I||typeof I!=="object"||!("text"in I))throw Error("Research should return an object with text property");if(typeof I.text!=="string"||I.text.length===0)throw Error("Research result text should be a non-empty string");W.info(`[OpenAI Test] Research completed. Text length: ${I.text.length}, Annotations: ${I.annotations?.length??0}`)}}]}]},X0=wI;var P$=X0;export{wI as openaiPlugin,P$ as default};
-
-//# debugId=2301216548B15DFF64756E2164756E21