fix: Capture parallel tool calls in eve llm span output

.changeset/tiny-ants-grin.md

@@ -0,0 +1,5 @@
+---
+"braintrust": patch
+---
+
+fix: Capture parallel tool calls in eve llm span output

e2e/scenarios/eve-instrumentation/agent/instructions.md

@@ -1,11 +1,16 @@
 You are a deterministic fixture agent for Braintrust Eve instrumentation tests.
 
-For every user task:
+For every user task, follow these steps exactly:
 
-1. Call the researcher subagent exactly once with the full user message as the
-   message.
-2. After the researcher result is available, call the read tool exactly once with
-   the URL https://eve.dev/docs/guides/instrumentation.
-3. After the read result is available, answer with a single sentence that starts
-   with "Final answer from read:" and includes the researcher result, read title,
-   URL, and read excerpt.
+1. Your first response MUST contain exactly these two tool calls and no text:
+   - researcher with only this argument: {"message":"<the full user message>"}.
+     Do not pass outputSchema.
+   - read with only this argument:
+     {"url":"https://eve.dev/docs/guides/instrumentation"}.
+     Emit both calls simultaneously in the same response, with researcher first
+     and read second, so Eve executes them as one parallel batch. The calls are
+     independent. NEVER call one by itself, wait for its result, or emit either
+     call in a separate response.
+2. Only after both results are available, answer with a single sentence that
+   starts with "Final answer from read:" and includes the researcher result, read
+   title, URL, and read excerpt.

e2e/scenarios/eve-instrumentation/scenario.test.ts

@@ -130,8 +130,19 @@ describe("eve instrumentation", () => {
     expect(root?.metrics?.tokens).toEqual(expect.any(Number));
     expect(root?.output).toContain("Final answer from read");
 
-    expect(steps).toHaveLength(3);
-    expect(steps.map((step) => step.span.type)).toEqual(["llm", "llm", "llm"]);
+    expect(steps).toHaveLength(2);
+    expect(steps.map((step) => step.span.type)).toEqual(["llm", "llm"]);
+    expect(steps[0]?.output).toMatchObject([
+      {
+        finish_reason: "tool_calls",
+        message: {
+          tool_calls: [
+            { function: { name: "researcher" }, type: "function" },
+            { function: { name: "read" }, type: "function" },
+          ],
+        },
+      },
+    ]);
     for (const step of steps) {
       expect(step.span.parentIds).toEqual([root?.span.id]);
       expect(Array.isArray(step.input)).toBe(true);
@@ -215,11 +226,18 @@ describe("eve instrumentation", () => {
     expect(secondRoot).toBeDefined();
     expect(secondRoot?.span.type).toBe("task");
     expect(secondRoot?.output).toContain("Final answer from read");
-    expect(secondSteps).toHaveLength(3);
-    expect(secondSteps.map((step) => step.span.type)).toEqual([
-      "llm",
-      "llm",
-      "llm",
+    expect(secondSteps).toHaveLength(2);
+    expect(secondSteps.map((step) => step.span.type)).toEqual(["llm", "llm"]);
+    expect(secondSteps[0]?.output).toMatchObject([
+      {
+        finish_reason: "tool_calls",
+        message: {
+          tool_calls: [
+            { function: { name: "researcher" }, type: "function" },
+            { function: { name: "read" }, type: "function" },
+          ],
+        },
+      },
     ]);
     expect(secondResearcher?.span.type).toBe("tool");
     expect(secondResearcher?.span.ended).toBe(true);

js/src/instrumentation/plugins/eve-plugin.test.ts

@@ -618,7 +618,7 @@ describe("braintrustEveHook", () => {
     ]);
   });
 
-  it("does not reconstruct later LLM inputs from earlier hook events", async () => {
+  it("merges incremental tool-call batches without reconstructing later LLM inputs", async () => {
     const wildcard = braintrustEveHook({ defineState }).events?.["*"];
     expect(wildcard).toBeDefined();
 
@@ -661,6 +661,22 @@ describe("braintrustEveHook", () => {
       },
       type: "actions.requested",
     });
+    await emit({
+      data: {
+        actions: [
+          {
+            callId: "call-search",
+            input: { query: "Updated Eve instrumentation" },
+            kind: "tool-call",
+            toolName: "search",
+          },
+        ],
+        sequence: 0,
+        stepIndex: 0,
+        turnId: "turn-incremental-tools",
+      },
+      type: "actions.requested",
+    });
     await emit({
       data: {
         actions: [
@@ -755,7 +771,28 @@ describe("braintrustEveHook", () => {
       {
         finish_reason: "tool_calls",
         message: {
-          tool_calls: [{ id: "call-read", function: { name: "read" } }],
+          tool_calls: [
+            {
+              function: {
+                arguments: JSON.stringify({
+                  query: "Updated Eve instrumentation",
+                }),
+                name: "search",
+              },
+              id: "call-search",
+              type: "function",
+            },
+            {
+              function: {
+                arguments: JSON.stringify({
+                  url: "https://eve.dev/docs/guides/instrumentation",
+                }),
+                name: "read",
+              },
+              id: "call-read",
+              type: "function",
+            },
+          ],
         },
       },
     ]);

js/src/instrumentation/plugins/eve-plugin.ts

@@ -694,27 +694,40 @@ class EveBridge {
       return;
     }
 
+    const toolCallsById = new Map<string, unknown>();
+    if (Array.isArray(step.output) && isObject(step.output[0])) {
+      const message = step.output[0]["message"];
+      if (isObject(message) && Array.isArray(message["tool_calls"])) {
+        for (const toolCall of message["tool_calls"]) {
+          if (isObject(toolCall) && typeof toolCall["id"] === "string") {
+            toolCallsById.set(toolCall["id"], toolCall);
+          }
+        }
+      }
+    }
+    for (const action of traceActions) {
+      const name =
+        action.kind === "tool-call"
+          ? action.toolName
+          : (action.subagentName ?? action.name ?? "agent");
+      toolCallsById.set(action.callId, {
+        function: {
+          arguments: JSON.stringify(action.input),
+          name,
+        },
+        id: action.callId,
+        type: "function",
+      });
+    }
+
     step.output = [
       {
         finish_reason: "tool_calls",
         index: 0,
         message: {
           content: null,
           role: "assistant",
-          tool_calls: traceActions.map((action) => {
-            const name =
-              action.kind === "tool-call"
-                ? action.toolName
-                : (action.subagentName ?? action.name ?? "agent");
-            return {
-              function: {
-                arguments: JSON.stringify(action.input),
-                name,
-              },
-              id: action.callId,
-              type: "function",
-            };
-          }),
+          tool_calls: [...toolCallsById.values()],
         },
       },
     ];