[codex] Surface tool call steps in flow logs

packages/core/lib/v3/llm/aisdk.ts

@@ -376,8 +376,13 @@ You must respond in JSON format. respond WITH JSON. Do not include any other tex
       model: this.model.modelId,
       output:
         textResponse.text ||
-        (transformedToolCalls.length > 0
-          ? `[${transformedToolCalls.length} tool calls]`
+        ((textResponse.toolCalls?.length ?? 0) > 0
+          ? summarizeToolCalls(
+              textResponse.toolCalls!.map((toolCall) => ({
+                name: toolCall.toolName,
+                input: toolCall.input,
+              })),
+            )
           : ""),
       inputTokens: textResponse.usage.inputTokens,
       outputTokens: textResponse.usage.outputTokens,
@@ -407,3 +412,33 @@ You must respond in JSON format. respond WITH JSON. Do not include any other tex
     return result;
   }
 }
+
+function summarizeToolCalls(
+  toolCalls: Array<{ name?: string; input?: unknown }>,
+): string {
+  if (toolCalls.length === 0) {
+    return "";
+  }
+
+  const summaries = toolCalls.map((toolCall) => {
+    const name = toolCall.name || "unknown";
+    return `${name}(${truncateToolArguments(toolCall.input)})`;
+  });
+
+  return toolCalls.length === 1
+    ? `tool call: ${summaries[0]}`
+    : `tool calls: ${summaries.join("; ")}`;
+}
+
+function truncateToolArguments(toolInput: unknown, maxLength = 160): string {
+  if (toolInput === undefined) {
+    return "";
+  }
+
+  const serializedArguments =
+    typeof toolInput === "string" ? toolInput : JSON.stringify(toolInput);
+
+  return serializedArguments.length > maxLength
+    ? `${serializedArguments.slice(0, maxLength - 3)}...`
+    : serializedArguments;
+}

packages/core/tests/integration/flowLogger.spec.ts

@@ -525,6 +525,15 @@ test.describe("flow logger integration", () => {
       expect(
         directChildrenOfType(events, actRoot, "LlmResponseEvent"),
       ).toHaveLength(1);
+      const agentRootLlmResponses = directChildrenOfType(
+        events,
+        agentRoot,
+        "LlmResponseEvent",
+      );
+      expect(agentRootLlmResponses[0].data.output).toContain("tool call: act(");
+      expect(agentRootLlmResponses[1].data.output).toContain(
+        "tool call: done(",
+      );
       assertNoFloatingLlmEvents(events);
       assertNoFloatingCdpEvents(events);
     } finally {
@@ -822,6 +831,9 @@ test.describe("flow logger integration", () => {
       expectDirectParent(pageScreenshotCompleted, pageScreenshot);
       expect(llmRequests).toHaveLength(3);
       expect(llmResponses).toHaveLength(3);
+      expect(llmResponses[0].data.output).toContain("tool call: goto(");
+      expect(llmResponses[1].data.output).toContain("tool call: screenshot(");
+      expect(llmResponses[2].data.output).toContain("tool call: done(");
 
       for (const event of [...llmRequests, ...llmResponses]) {
         expect(event.eventParentIds).toEqual([root.eventId]);