@axlsdk/eval 0.15.0 → 0.16.0

package/README.md

@@ -344,7 +344,7 @@ for (const item of results.items) {
   // Per-item trace events (when run with { captureTraces: true })
   if (item.traces) {
     console.log(`${item.traces.length} trace events captured`);
-    const agentCalls = item.traces.filter(e => e.type === 'agent_call');
+    const agentCalls = item.traces.filter(e => e.type === 'agent_call_end');
     console.log(`${agentCalls.length} LLM turns`);
     // Failure-path traces are also populated (recovered from axlCapturedTraces
     // on the thrown error) — captureTraces is especially useful for debugging
@@ -353,6 +353,8 @@ for (const item of results.items) {
 }
 ```
 
+`captureTraces` strips verbose-mode `agent_call_end.data.messages` snapshots and high-volume events (`token`, `partial_object`) from the captured array to keep memory bounded. The structural events you'd want for debugging — `agent_call_start`/`agent_call_end`, `tool_call_*`, gate events, `pipeline`, `verify`, `handoff_*` — are all retained.
+
 ## Comparing Results
 
 Compare two runs to detect regressions and improvements. Runs must use the same dataset and scorers.

package/dist/index.d.cts

@@ -1,5 +1,5 @@
 import { z } from 'zod';
-import { TraceEvent, AxlRuntime } from '@axlsdk/axl';
+import { AxlEvent, AxlRuntime } from '@axlsdk/axl';
 
 type DatasetItem<TInput, TAnnotations = undefined> = {
     input: TInput;
@@ -135,9 +135,9 @@ type EvalItem = {
     metadata?: Record<string, unknown>;
     /** Trace events captured during this item's execution. Only populated when
      *  `runEval` was called with `{ captureTraces: true }`. Verbose-mode
-     *  `agent_call.data.messages` snapshots are stripped to keep memory bounded;
+     *  `agent_call_start.data.messages` snapshots are stripped to keep memory bounded;
      *  subscribe to `runtime.on('trace', ...)` directly if you need those. */
-    traces?: TraceEvent[];
+    traces?: AxlEvent[];
 };
 type EvalSummary = {
     count: number;
@@ -238,14 +238,14 @@ type RunEvalOptions = {
     /** Abort signal — checked before starting each item. */
     signal?: AbortSignal;
     /**
-     * Capture per-item `TraceEvent[]` from the runtime and store them on
+     * Capture per-item `AxlEvent[]` from the runtime and store them on
      * `EvalItem.traces`. Off by default because traces multiply memory with
      * dataset size × turns × agents. When on, the runner wraps the user-provided
      * `executeWorkflow` with `runtime.trackExecution({ captureTraces: true })`,
      * so any `runtime.execute()` / `ctx.ask()` activity inside the callback is
      * captured and scoped to the current item.
      *
-     * Note: verbose-mode `agent_call.data.messages` snapshots are stripped from
+     * Note: verbose-mode `agent_call_start.data.messages` snapshots are stripped from
      * captured traces to keep memory bounded — if you need the full verbose
      * payload, subscribe to `runtime.on('trace', ...)` directly.
      */

package/dist/index.d.ts

@@ -1,5 +1,5 @@
 import { z } from 'zod';
-import { TraceEvent, AxlRuntime } from '@axlsdk/axl';
+import { AxlEvent, AxlRuntime } from '@axlsdk/axl';
 
 type DatasetItem<TInput, TAnnotations = undefined> = {
     input: TInput;
@@ -135,9 +135,9 @@ type EvalItem = {
     metadata?: Record<string, unknown>;
     /** Trace events captured during this item's execution. Only populated when
      *  `runEval` was called with `{ captureTraces: true }`. Verbose-mode
-     *  `agent_call.data.messages` snapshots are stripped to keep memory bounded;
+     *  `agent_call_start.data.messages` snapshots are stripped to keep memory bounded;
      *  subscribe to `runtime.on('trace', ...)` directly if you need those. */
-    traces?: TraceEvent[];
+    traces?: AxlEvent[];
 };
 type EvalSummary = {
     count: number;
@@ -238,14 +238,14 @@ type RunEvalOptions = {
     /** Abort signal — checked before starting each item. */
     signal?: AbortSignal;
     /**
-     * Capture per-item `TraceEvent[]` from the runtime and store them on
+     * Capture per-item `AxlEvent[]` from the runtime and store them on
      * `EvalItem.traces`. Off by default because traces multiply memory with
      * dataset size × turns × agents. When on, the runner wraps the user-provided
      * `executeWorkflow` with `runtime.trackExecution({ captureTraces: true })`,
      * so any `runtime.execute()` / `ctx.ask()` activity inside the callback is
      * captured and scoped to the current item.
      *
-     * Note: verbose-mode `agent_call.data.messages` snapshots are stripped from
+     * Note: verbose-mode `agent_call_start.data.messages` snapshots are stripped from
      * captured traces to keep memory bounded — if you need the full verbose
      * payload, subscribe to `runtime.on('trace', ...)` directly.
      */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@axlsdk/eval",
-  "version": "0.15.0",
+  "version": "0.16.0",
   "description": "Evaluation framework for Axl agentic workflows",
   "type": "module",
   "main": "./dist/index.cjs",
@@ -25,7 +25,7 @@
     "dist"
   ],
   "dependencies": {
-    "@axlsdk/axl": "0.15.0"
+    "@axlsdk/axl": "0.16.0"
   },
   "peerDependencies": {
     "zod": "^4.0.0"