fix(agent-core): suppress the spurious error event for step-capped goal turns

.changeset/goal-step-cap-error-banner.md

@@ -0,0 +1,5 @@
+---
+"@moonshot-ai/kimi-code": patch
+---
+
+Stop showing a spurious error banner when a goal turn reaches the per-turn step limit (`loop_control.max_steps_per_turn`) during autonomous goal pursuit; the limit is expected control flow there and still surfaces for ordinary turns.

packages/agent-core/src/agent/turn/index.ts

@@ -633,7 +633,14 @@ export class TurnFlow {
           inputData: { errorType: summary.name, errorMessage: summary.message },
         });
         ended = { type: 'turn.ended', turnId, reason: 'failed', error: summary, durationMs: Date.now() - startedAt };
-        errorEvent = { type: 'error', ...summary };
+        // A goal-driven turn cut by the per-turn step limit is expected
+        // control flow — the goal driver immediately continues with a fresh
+        // continuation turn — so it must not raise the standalone error event
+        // that hosts render as an actionable failure. `turn.ended` and the
+        // `turn.interrupted` loop event still report the cap.
+        if (!(isMaxStepsExceededError(error) && this.agent.goal.getActiveGoal() !== null)) {
+          errorEvent = { type: 'error', ...summary };
+        }
         if (this.shouldTrackApiError(turnId)) {
           const classification = classifyApiError(error, summary);
           const properties: Record<string, TelemetryPropertyValue> = {

packages/agent-core/test/agent/turn.test.ts

@@ -1875,6 +1875,15 @@ describe('Agent turn flow', () => {
         }),
       }),
     );
+    // Without an active goal the cap is a user-facing failure: the standalone
+    // error event (suppressed for goal-driven turns) is emitted right after
+    // turn.ended.
+    expect(ctx.newEvents()).toContainEqual(
+      expect.objectContaining({
+        event: 'error',
+        args: expect.objectContaining({ code: 'loop.max_steps_exceeded' }),
+      }),
+    );
   });
 
   describe('loop control env overrides', () => {

packages/agent-core/test/harness/goal-session.test.ts

@@ -523,6 +523,9 @@ describe('goal session end-to-end', () => {
           ErrorCodes.LOOP_MAX_STEPS_EXCEEDED,
     );
     expect(maxStepEnds).toHaveLength(2);
+    // Capped goal turns are expected control flow: no standalone error event
+    // is emitted for hosts to render as an actionable failure.
+    expect(events.filter((event) => event['type'] === 'error')).toEqual([]);
     // The cap never pauses or blocks the goal: no stopped status is ever
     // emitted, and the goal completes (record cleared) in the third turn.
     expect(