fix(dashboard-agent): replace fixed 30ms eval wait with deterministic poll

Mono-RevId: dd197d9c6b561976f0a005411fa44c4bdc8e799c

Author: carderne

internal-packages/dashboard-agent/src/test-support.ts

@@ -281,6 +281,26 @@ export function fakeEvalTrigger(): { trigger: DashboardAgentEvalTrigger; calls:
   };
 }
 
+/**
+ * Waits until `calls` has at least `count` entries, with a generous deadline.
+ * Use this instead of a fixed setTimeout after sendMessage when asserting evals,
+ * because onTurnComplete fires after the turn-complete chunk and the gap is not
+ * bounded at 30 ms under load.
+ */
+export async function waitForEvals(
+  calls: unknown[],
+  count: number,
+  timeoutMs = 5_000
+): Promise<void> {
+  const deadline = Date.now() + timeoutMs;
+  while (calls.length < count) {
+    if (Date.now() > deadline) {
+      throw new Error(`timed out waiting for ${count} eval(s); got ${calls.length}`);
+    }
+    await new Promise((r) => setTimeout(r, 10));
+  }
+}
+
 /** Stands in for the org opt-out check, so no test depends on a network call. */
 export function fakeEvalPolicy(allowed = true): DashboardAgentEvalPolicyCheck {
   return async () => allowed;

internal-packages/dashboard-agent/src/watch-actions.test.ts

@@ -26,6 +26,7 @@ import {
   fakeEvalPolicy,
   fakeEvalTrigger,
   fakeStore,
+  waitForEvals,
   finish,
   mockModel,
   savedPuts,
@@ -429,8 +430,7 @@ describe("watch wake narration", () => {
       await harness.sendAction(FAILED_RUN_WAKE);
       settlementDown = false;
       await harness.sendMessage(userMessage("what happened?"));
-      // onTurnComplete enqueues after the turn-complete chunk, so give it a tick.
-      await new Promise((r) => setTimeout(r, 30));
+      await waitForEvals(evals, 1);
 
       // The wake ran as a bounded small-model call; the typed turn ran uncapped.
       expect(calls[0]).toEqual({ tools: 0, maxOutputTokens: 300 });