fix: handle explicit silent assistant replies (#92073)

Signed-off-by: sallyom <somalley@redhat.com>

src/agents/embedded-agent-runner/result-fallback-classifier.test.ts

@@ -75,6 +75,23 @@ describe("classifyEmbeddedAgentRunResultForModelFallback", () => {
     expect(result).toBeNull();
   });
 
+  it("does not fallback on deliberate silent terminal replies after payload filtering", () => {
+    const result = classifyEmbeddedAgentRunResultForModelFallback({
+      provider: "openai",
+      model: "gpt-5.5",
+      result: {
+        payloads: [],
+        meta: {
+          durationMs: 42,
+          finalAssistantRawText: "NO_REPLY",
+          finalAssistantVisibleText: "NO_REPLY",
+        },
+      },
+    });
+
+    expect(result).toBeNull();
+  });
+
   it("uses provider-scoped failover matching for business-denial payloads", () => {
     const result = classifyEmbeddedAgentRunResultForModelFallback({
       provider: "openrouter",

src/agents/embedded-agent-runner/run.incomplete-turn.test.ts

@@ -2706,6 +2706,62 @@ describe("runEmbeddedAgent incomplete-turn safety", () => {
     ).toBe(false);
   });
 
+  it("treats exact NO_REPLY assistant turns as silent only when the caller allows it", () => {
+    const attempt = makeAttemptResult({
+      assistantTexts: ["NO_REPLY"],
+      lastAssistant: {
+        role: "assistant",
+        stopReason: "stop",
+        provider: "openai",
+        model: "gpt-5.5",
+        content: [{ type: "text", text: "NO_REPLY" }],
+      } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+    });
+
+    expect(
+      shouldTreatEmptyAssistantReplyAsSilent({
+        allowEmptyAssistantReplyAsSilent: true,
+        payloadCount: 0,
+        aborted: false,
+        timedOut: false,
+        attempt,
+      }),
+    ).toBe(true);
+    expect(
+      shouldTreatEmptyAssistantReplyAsSilent({
+        allowEmptyAssistantReplyAsSilent: false,
+        payloadCount: 0,
+        aborted: false,
+        timedOut: false,
+        attempt,
+      }),
+    ).toBe(false);
+  });
+
+  it("treats post-tool exact NO_REPLY assistant turns as intentional silence", () => {
+    const attempt = makeAttemptResult({
+      assistantTexts: ["NO_REPLY"],
+      toolMetas: [{ toolName: "process.poll", meta: "pid=123" }],
+      lastAssistant: {
+        role: "assistant",
+        stopReason: "stop",
+        provider: "openai",
+        model: "gpt-5.5",
+        content: [{ type: "text", text: "NO_REPLY" }],
+      } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+    });
+
+    expect(
+      shouldTreatEmptyAssistantReplyAsSilent({
+        allowEmptyAssistantReplyAsSilent: true,
+        payloadCount: 0,
+        aborted: false,
+        timedOut: false,
+        attempt,
+      }),
+    ).toBe(true);
+  });
+
   it("does not treat error or side-effect empty turns as silent", () => {
     const errorAttempt = makeAttemptResult({
       assistantTexts: [],
@@ -2717,6 +2773,16 @@ describe("runEmbeddedAgent incomplete-turn safety", () => {
         content: [],
       } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
     });
+    const silentErrorAttempt = makeAttemptResult({
+      assistantTexts: ["NO_REPLY"],
+      lastAssistant: {
+        role: "assistant",
+        stopReason: "error",
+        provider: "openai",
+        model: "gpt-5.5",
+        content: [{ type: "text", text: "NO_REPLY" }],
+      } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+    });
     const sideEffectAttempt = makeAttemptResult({
       assistantTexts: [],
       didSendViaMessagingTool: true,
@@ -2751,6 +2817,15 @@ describe("runEmbeddedAgent incomplete-turn safety", () => {
         attempt: errorAttempt,
       }),
     ).toBe(false);
+    expect(
+      shouldTreatEmptyAssistantReplyAsSilent({
+        allowEmptyAssistantReplyAsSilent: true,
+        payloadCount: 0,
+        aborted: false,
+        timedOut: false,
+        attempt: silentErrorAttempt,
+      }),
+    ).toBe(false);
     expect(
       shouldTreatEmptyAssistantReplyAsSilent({
         allowEmptyAssistantReplyAsSilent: true,
@@ -2803,6 +2878,47 @@ describe("runEmbeddedAgent incomplete-turn safety", () => {
     expect(result.meta.livenessState).toBe("working");
   });
 
+  it("returns NO_REPLY without retrying exact silent assistant replies when silence is allowed", async () => {
+    mockedClassifyFailoverReason.mockReturnValue(null);
+    mockedRunEmbeddedAttempt.mockResolvedValue(
+      makeAttemptResult({
+        assistantTexts: ["NO_REPLY"],
+        lastAssistant: {
+          role: "assistant",
+          stopReason: "stop",
+          provider: "openai",
+          model: "gpt-5.5",
+          content: [
+            {
+              type: "thinking",
+              thinking: "internal reasoning",
+              thinkingSignature: JSON.stringify({ id: "rs_exact_silent", type: "reasoning" }),
+            },
+            { type: "text", text: "NO_REPLY" },
+          ],
+        } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+      }),
+    );
+
+    const result = await runEmbeddedAgent({
+      ...overflowBaseRunParams,
+      allowEmptyAssistantReplyAsSilent: true,
+      provider: "openai",
+      model: "gpt-5.5",
+      runId: "run-exact-silent-assistant-reply",
+    });
+
+    expect(mockedRunEmbeddedAttempt).toHaveBeenCalledTimes(1);
+    const onlyCall = runAttemptCall(0);
+    expect(onlyCall.prompt).not.toContain(REASONING_ONLY_RETRY_INSTRUCTION);
+    expect(onlyCall.prompt).not.toContain(EMPTY_RESPONSE_RETRY_INSTRUCTION);
+    expectNoWarnMessageWith("empty response detected");
+    expectNoWarnMessageWith("incomplete turn detected");
+    expect(result.payloads).toEqual([{ text: "NO_REPLY" }]);
+    expect(result.meta.terminalReplyKind).toBe("silent-empty");
+    expect(result.meta.livenessState).toBe("working");
+  });
+
   it("retries post-tool openai-compatible empty stop turns even when empty silence is allowed", async () => {
     mockedClassifyFailoverReason.mockReturnValue(null);
     mockedResolveModelAsync.mockResolvedValue({
@@ -2862,6 +2978,54 @@ describe("runEmbeddedAgent incomplete-turn safety", () => {
     expectWarnMessageWith("empty response detected");
   });
 
+  it("returns NO_REPLY without retrying post-tool exact silent assistant replies", async () => {
+    mockedClassifyFailoverReason.mockReturnValue(null);
+    mockedResolveModelAsync.mockResolvedValue({
+      model: {
+        id: "step-router-v1",
+        provider: "stepfun",
+        contextWindow: 200000,
+        api: "openai-completions",
+      },
+      error: null,
+      authStorage: {
+        setRuntimeApiKey: vi.fn(),
+      },
+      modelRegistry: {},
+    });
+    mockedRunEmbeddedAttempt.mockResolvedValueOnce(
+      makeAttemptResult({
+        assistantTexts: ["NO_REPLY"],
+        toolMetas: [{ toolName: "process.poll", meta: "pid=123" }],
+        lastAssistant: {
+          role: "assistant",
+          api: "openai-completions",
+          stopReason: "stop",
+          provider: "stepfun",
+          model: "step-router-v1",
+          content: [{ type: "text", text: "NO_REPLY" }],
+        } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+      }),
+    );
+
+    const result = await runEmbeddedAgent({
+      ...overflowBaseRunParams,
+      allowEmptyAssistantReplyAsSilent: true,
+      provider: "stepfun",
+      model: "step-router-v1",
+      runId: "run-post-tool-exact-silent-retry",
+    });
+
+    expect(mockedRunEmbeddedAttempt).toHaveBeenCalledTimes(1);
+    const onlyCall = runAttemptCall(0);
+    expect(onlyCall.prompt).not.toContain(EMPTY_RESPONSE_RETRY_INSTRUCTION);
+    expectNoWarnMessageWith("empty response detected");
+    expectNoWarnMessageWith("incomplete turn detected");
+    expect(result.payloads).toEqual([{ text: "NO_REPLY" }]);
+    expect(result.meta.terminalReplyKind).toBe("silent-empty");
+    expect(result.meta.livenessState).toBe("working");
+  });
+
   it("keeps retrying and surfacing clean empty assistant turns without the silence flag", async () => {
     mockedClassifyFailoverReason.mockReturnValue(null);
     mockedRunEmbeddedAttempt.mockResolvedValue(

src/agents/embedded-agent-runner/run/incomplete-turn.ts

@@ -635,7 +635,7 @@ function shouldSkipPlanningOnlyRetry(params: {
   );
 }
 
-/** Allows configured silent handling for replay-safe empty or reasoning-only assistant turns. */
+/** Allows configured silent handling for replay-safe empty, reasoning-only, or explicit silent turns. */
 export function shouldTreatEmptyAssistantReplyAsSilent(params: {
   allowEmptyAssistantReplyAsSilent?: boolean;
   payloadCount: number;
@@ -649,6 +649,14 @@ export function shouldTreatEmptyAssistantReplyAsSilent(params: {
   if (hasCommittedMessagingToolDeliveryEvidence(params.attempt)) {
     return false;
   }
+  const assistant = params.attempt.currentAttemptAssistant ?? params.attempt.lastAssistant;
+  if (
+    params.payloadCount === 0 &&
+    assistant?.stopReason !== "error" &&
+    hasOnlySilentAssistantReply(params.attempt.assistantTexts)
+  ) {
+    return true;
+  }
   // Post-tool empty stops are ambiguous provider failures, not intentional silence.
   // Let the retry/incomplete-turn paths decide whether replay is safe.
   if (