@jambonz/llm 0.5.0 → 0.5.1

package/dist/{types-BWsmVtXv.d.cts → types-BEgQaQIb.d.cts}

@@ -105,6 +105,29 @@ type LlmEvent = {
      * `vendor_metadata`.
      */
     vendorMetadata?: Record<string, string | number>;
+    /**
+     * Optional client-side timing breakdown for diagnosing latency.
+     * Captured by the adapter using `process.hrtime`. Numbers are
+     * milliseconds. When present:
+     *   - `requestToHeadersMs`: from when the SDK call was initiated
+     *     to when response headers were received. Includes TLS/TCP
+     *     handshake (if connection was cold), network RTT in both
+     *     directions, and any vendor-side queueing/processing time
+     *     before the first byte.
+     *   - `headersToFirstTokenMs`: from response headers received to
+     *     first content chunk emitted. Usually small (<50ms) if the
+     *     vendor is streaming normally; larger values suggest SSE
+     *     buffering or chunk-transfer issues. Undefined if no token
+     *     was emitted (error or empty response).
+     *   - `totalMs`: full wall-clock from SDK call to stream end.
+     * Adapters that cannot measure these (different SDK shape) omit
+     * the field entirely.
+     */
+    clientTiming?: {
+        requestToHeadersMs: number;
+        headersToFirstTokenMs?: number;
+        totalMs: number;
+    };
 };
 /** Convenience alias for the fully-accumulated tool-call event shape. */
 type ToolCallEvent = Extract<LlmEvent, {

package/dist/{types-BWsmVtXv.d.ts → types-BEgQaQIb.d.ts}

@@ -105,6 +105,29 @@ type LlmEvent = {
      * `vendor_metadata`.
      */
     vendorMetadata?: Record<string, string | number>;
+    /**
+     * Optional client-side timing breakdown for diagnosing latency.
+     * Captured by the adapter using `process.hrtime`. Numbers are
+     * milliseconds. When present:
+     *   - `requestToHeadersMs`: from when the SDK call was initiated
+     *     to when response headers were received. Includes TLS/TCP
+     *     handshake (if connection was cold), network RTT in both
+     *     directions, and any vendor-side queueing/processing time
+     *     before the first byte.
+     *   - `headersToFirstTokenMs`: from response headers received to
+     *     first content chunk emitted. Usually small (<50ms) if the
+     *     vendor is streaming normally; larger values suggest SSE
+     *     buffering or chunk-transfer issues. Undefined if no token
+     *     was emitted (error or empty response).
+     *   - `totalMs`: full wall-clock from SDK call to stream end.
+     * Adapters that cannot measure these (different SDK shape) omit
+     * the field entirely.
+     */
+    clientTiming?: {
+        requestToHeadersMs: number;
+        headersToFirstTokenMs?: number;
+        totalMs: number;
+    };
 };
 /** Convenience alias for the fully-accumulated tool-call event shape. */
 type ToolCallEvent = Extract<LlmEvent, {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@jambonz/llm",
-  "version": "0.5.0",
+  "version": "0.5.1",
   "description": "Voice-oriented LLM adapter library — a uniform interface for streaming chat completions from multiple LLM vendors, with first-class support for tool calls, AbortSignal interruption, and lossless history round-tripping.",
   "type": "module",
   "main": "./dist/index.cjs",