@agentionai/agents 1.2.0 → 1.4.0

package/dist/agents/AgentConfig.d.ts

@@ -16,6 +16,27 @@ export interface CommonAgentConfig {
     description: string;
     /** API key for authenticating with the LLM provider */
     apiKey: string;
+    /**
+     * Extra HTTP headers sent with every request to the provider.
+     *
+     * Useful for gateway and proxy attribution, tracing, or corporate egress
+     * requirements. OpenRouter, for example, uses `HTTP-Referer` and `X-Title`
+     * to attribute traffic to your app:
+     *
+     * ```typescript
+     * defaultHeaders: {
+     *   "HTTP-Referer": "https://myapp.example",
+     *   "X-Title": "My App",
+     * }
+     * ```
+     *
+     * These override headers the agent would otherwise set, including
+     * `Authorization` / `x-api-key`. That follows the Anthropic and OpenAI SDKs'
+     * own `defaultHeaders` behaviour and is deliberate — it lets a gateway swap
+     * in its own auth scheme. The flip side is that setting an auth header here
+     * replaces `apiKey`, so do it only when that is what you mean.
+     */
+    defaultHeaders?: Record<string, string>;
     /** Enable debug logging for troubleshooting (default: false) */
     debug?: boolean;
     /** Maximum number of messages to retain in conversation history */

package/dist/agents/BaseAgent.d.ts

@@ -56,6 +56,14 @@ export type TokenUsage = {
      * Generation throughput: `output_tokens` over `generationMs`. Falls back to
      * `totalMs` when the first-token time is unknown (an unstreamed call), in
      * which case it is an end-to-end rate rather than a pure generation rate.
+     *
+     * Accurate wherever thinking is streamed, because the first thinking chunk
+     * starts the generation window — verified on Anthropic extended thinking and
+     * on DeepSeek via OpenRouter. It over-reports on OpenAI's Responses API,
+     * which does not stream raw reasoning: there the thinking finishes before the
+     * first visible token, so it lands inside `timeToFirstTokenMs` while its
+     * tokens still count toward `output_tokens`. For the visible-output rate on
+     * that path, divide `output_tokens - reasoning_tokens` by `generationMs`.
      */
     outputTokensPerSecond?: number;
 };

package/dist/agents/anthropic/ClaudeAgent.js

@@ -40,10 +40,10 @@ class ClaudeAgent extends BaseAgent_1.BaseAgent {
         const authType = config.authType ?? vendorConfig.authType ?? "apiKey";
         const thinkingBudgetTokens = config.thinkingBudgetTokens ?? vendorConfig.thinkingBudgetTokens;
         this.client = new sdk_1.Anthropic(authType === "oauth"
-            ? { authToken: config.apiKey }
-            : { apiKey: config.apiKey });
+            ? { authToken: config.apiKey, defaultHeaders: config.defaultHeaders }
+            : { apiKey: config.apiKey, defaultHeaders: config.defaultHeaders });
         this.config = {
-            model: config.model || "claude-3-5-haiku-latest",
+            model: config.model || "claude-haiku-4-5",
             maxTokens: config.maxTokens || 1024,
             disableParallelToolUse,
             metadata,

package/dist/agents/google/GeminiAgent.js

@@ -35,7 +35,7 @@ class GeminiAgent extends BaseAgent_1.BaseAgent {
         const responseMimeType = config.responseMimeType ?? vendorConfig.responseMimeType;
         const responseSchema = config.responseSchema ?? vendorConfig.responseSchema;
         this.config = {
-            model: config.model || "gemini-2.0-flash",
+            model: config.model || "gemini-flash-latest",
             maxTokens: config.maxTokens || 1024,
             apiKey: config.apiKey,
             temperature: config.temperature,
@@ -47,9 +47,7 @@ class GeminiAgent extends BaseAgent_1.BaseAgent {
             responseSchema,
         };
         // Initialize the model
-        this.generativeModel = this.client.getGenerativeModel({
-            model: this.config.model,
-        });
+        this.generativeModel = this.client.getGenerativeModel({ model: this.config.model }, config.defaultHeaders ? { customHeaders: config.defaultHeaders } : undefined);
         // Add system message to history (skips if already exists with same content)
         this.addSystemMessage(this.getSystemMessage());
     }

package/dist/agents/mistral/MistralAgent.d.ts

@@ -26,6 +26,16 @@ type AgentConfig = BaseAgentConfig & {
  * const response = await agent.execute("Hello!");
  * ```
  */
+/**
+ * Build a `beforeRequest` hook that adds custom headers to every request.
+ *
+ * The Mistral SDK has no `defaultHeaders` option like the Anthropic and OpenAI
+ * clients, so headers are injected at the HTTP layer instead. They overwrite
+ * headers the SDK already set, so that `defaultHeaders` means the same thing on
+ * every provider — see `CommonAgentConfig.defaultHeaders`. Verified against the
+ * OpenAI SDK on the wire: its `defaultHeaders` win over the client's own auth.
+ */
+export declare function defaultHeadersHook(headers: Record<string, string>): (request: Request) => void;
 export declare class MistralAgent extends BaseAgent {
     private client;
     protected config: Partial<AgentConfig>;

package/dist/agents/mistral/MistralAgent.js

@@ -1,7 +1,9 @@
 "use strict";
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.MistralAgent = void 0;
+exports.defaultHeadersHook = defaultHeadersHook;
 const mistralai_1 = require("@mistralai/mistralai");
+const http_1 = require("@mistralai/mistralai/lib/http");
 const BaseAgent_1 = require("../BaseAgent");
 const AgentEvent_1 = require("../AgentEvent");
 const AgentError_1 = require("../errors/AgentError");
@@ -25,13 +27,34 @@ const VizConfig_1 = require("../../viz/VizConfig");
  * const response = await agent.execute("Hello!");
  * ```
  */
+/**
+ * Build a `beforeRequest` hook that adds custom headers to every request.
+ *
+ * The Mistral SDK has no `defaultHeaders` option like the Anthropic and OpenAI
+ * clients, so headers are injected at the HTTP layer instead. They overwrite
+ * headers the SDK already set, so that `defaultHeaders` means the same thing on
+ * every provider — see `CommonAgentConfig.defaultHeaders`. Verified against the
+ * OpenAI SDK on the wire: its `defaultHeaders` win over the client's own auth.
+ */
+function defaultHeadersHook(headers) {
+    return (request) => {
+        for (const [name, value] of Object.entries(headers)) {
+            request.headers.set(name, value);
+        }
+    };
+}
 class MistralAgent extends BaseAgent_1.BaseAgent {
     constructor(config, history) {
         super({ ...config, vendor: "mistral" }, history);
         /** Count of tool calls in current execution */
         this.currentToolCallCount = 0;
+        const httpClient = new http_1.HTTPClient();
+        if (config.defaultHeaders) {
+            httpClient.addHook("beforeRequest", defaultHeadersHook(config.defaultHeaders));
+        }
         this.client = new mistralai_1.Mistral({
             apiKey: config.apiKey,
+            httpClient,
         });
         // Merge flat config (deprecated) with nested vendorConfig
         // Flat config takes precedence for backward compatibility

package/dist/agents/model-types.d.ts

@@ -8,19 +8,19 @@
  * You can also provide any custom string for newer models not yet listed.
  * @see https://docs.anthropic.com/en/docs/about-claude/models
  */
-export type ClaudeModel = "claude-opus-4-7" | "claude-opus-4-6" | "claude-sonnet-4-6" | "claude-opus-4-5" | "claude-sonnet-4-5" | "claude-haiku-4-5" | "claude-haiku-4-5-20251001" | (string & Record<never, never>);
+export type ClaudeModel = "claude-opus-5" | "claude-sonnet-5" | "claude-fable-5" | "claude-opus-4-8" | "claude-opus-4-7" | "claude-opus-4-6" | "claude-sonnet-4-6" | "claude-opus-4-5" | "claude-opus-4-5-20251101" | "claude-sonnet-4-5" | "claude-sonnet-4-5-20250929" | "claude-haiku-4-5" | "claude-haiku-4-5-20251001" | (string & Record<never, never>);
 /**
  * Supported Google Gemini models.
  * You can also provide any custom string for newer models not yet listed.
  * @see https://ai.google.dev/gemini-api/docs/models/gemini
  */
-export type GeminiModel = "gemini-flash-latest" | "gemini-flash-lite-latest" | "gemini-3.0-pro" | "gemini-3.0-flash" | "gemini-2.5-flash-lite" | "gemini-2.0-flash-exp" | "gemini-2.0-flash" | "gemini-2.0-flash-lite" | (string & {});
+export type GeminiModel = "gemini-pro-latest" | "gemini-flash-latest" | "gemini-flash-lite-latest" | "gemini-3.6-flash" | "gemini-3.5-flash" | "gemini-3.5-flash-lite" | "gemini-3.1-pro-preview" | "gemini-3.1-flash-lite" | "gemini-3.1-flash-lite-preview" | "gemini-3-pro-preview" | "gemini-3-flash-preview" | "gemini-2.5-pro" | "gemini-2.5-flash" | "gemini-2.5-flash-lite" | "gemini-2.0-flash" | "gemini-2.0-flash-001" | "gemini-2.0-flash-lite" | "gemini-2.0-flash-lite-001" | (string & {});
 /**
  * Supported Mistral models.
  * You can also provide any custom string for newer models not yet listed.
  * @see https://docs.mistral.ai/getting-started/models/
  */
-export type MistralModel = "mistral-large-latest" | "mistral-small-latest" | "ministral-8b-latest" | "ministral-8b-2410" | "ministral-3b-latest" | "ministral-3b-2410" | "codestral-latest" | "codestral-2405" | "mistral-embed" | "mistral-moderation-latest" | "mistral-moderation-2411" | (string & {});
+export type MistralModel = "mistral-large-latest" | "mistral-large-2512" | "mistral-medium-latest" | "mistral-medium-3.5" | "mistral-small-latest" | "mistral-small-2603" | "magistral-small-latest" | "ministral-14b-latest" | "ministral-8b-latest" | "ministral-8b-2512" | "ministral-3b-latest" | "ministral-3b-2512" | "codestral-latest" | "codestral-2508" | "devstral-latest" | "devstral-medium-latest" | "mistral-embed" | "codestral-embed" | "mistral-moderation-latest" | "mistral-moderation-2603" | "mistral-ocr-latest" | (string & {});
 /**
  * Popular Ollama models (locally hosted).
  * You can also provide any custom string for models you have pulled.

package/dist/agents/ollama/OllamaAgent.js

@@ -80,6 +80,7 @@ class OllamaAgent extends BaseAgent_1.BaseAgent {
         this.config = {
             model: config.model || "llama3.2",
             host,
+            defaultHeaders: config.defaultHeaders,
             maxTokens: config.maxTokens,
             temperature: config.temperature,
             topP: config.topP,
@@ -99,7 +100,10 @@ class OllamaAgent extends BaseAgent_1.BaseAgent {
                 if (!OllamaClass) {
                     throw new Error("Could not find Ollama class in ollama package");
                 }
-                this._client = new OllamaClass({ host: this.config.host });
+                this._client = new OllamaClass({
+                    host: this.config.host,
+                    headers: this.config.defaultHeaders,
+                });
             }
             catch (err) {
                 throw new AgentError_1.ExecutionError(`Failed to load 'ollama' package. Install it with: npm install ollama\n${err instanceof Error ? err.message : String(err)}`);

package/dist/agents/openai/OpenAiAgent.js

@@ -58,6 +58,7 @@ class OpenAiAgent extends BaseAgent_1.BaseAgent {
         this.currentToolCallCount = 0;
         this.client = new openai_1.default({
             apiKey: config.apiKey,
+            defaultHeaders: config.defaultHeaders,
         });
         // Merge flat config (deprecated) with nested vendorConfig
         // Flat config takes precedence for backward compatibility

package/dist/agents/openai-compatible/OpenAICompatibleAgent.js

@@ -28,6 +28,7 @@ class OpenAICompatibleAgent extends BaseAgent_1.BaseAgent {
         this.client = new openai_1.default({
             apiKey: config.apiKey || "not-needed",
             baseURL: config.baseURL,
+            defaultHeaders: config.defaultHeaders,
         });
         this.config = {
             model: config.model,
@@ -320,13 +321,17 @@ class OpenAICompatibleAgent extends BaseAgent_1.BaseAgent {
         let reasoningContent = "";
         const toolCallAcc = new Map();
         let finishReason = null;
+        let streamUsage;
         for await (const chunk of stream) {
-            // Final chunk carrying usage (choices is empty)
-            if (chunk.choices.length === 0) {
-                if (chunk.usage)
-                    this.accumulateStreamUsage(chunk.usage);
+            // Usage can ride on any chunk: OpenAI sends it on a final choice-less
+            // chunk, OpenRouter attaches it to the last content chunk (the one
+            // carrying finish_reason). Keep the most recent and fold it in once the
+            // stream ends — it is a running total for the turn, not a delta, so
+            // taking the last one covers both layouts without double-counting.
+            if (chunk.usage)
+                streamUsage = chunk.usage;
+            if (chunk.choices.length === 0)
                 continue;
-            }
             const choice = chunk.choices[0];
             finishReason = choice.finish_reason ?? finishReason;
             const delta = choice.delta;
@@ -366,6 +371,10 @@ class OpenAICompatibleAgent extends BaseAgent_1.BaseAgent {
                 }
             }
         }
+        // Before any early return below, so a turn that hits the token limit or
+        // continues into a tool call still reports what it spent.
+        if (streamUsage)
+            this.accumulateStreamUsage(streamUsage);
         if (finishReason === "length") {
             const error = new AgentError_1.MaxTokensExceededError("Response exceeded maximum token limit", this.config.maxTokens || 1024);
             this.emit(AgentEvent_1.AgentEvent.MAX_TOKENS_EXCEEDED, error);

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@agentionai/agents",
   "author": "Laurent Zuijdwijk",
-  "version": "1.2.0",
+  "version": "1.4.0",
   "description": "Agent Library",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",