fix(agents): detect unsigned thinking-only stall when reasoning payload inflates payloadCount

When a model (e.g. Qwen3 via llama.cpp) produces stopReason="stop" with
only an unsigned thinking block, the reasoning text is pushed into
replyItems as { isReasoning: true }, making payloadCount=1. The early-
return guard `payloadCount !== 0 && !toolUseTerminal` in
resolveIncompleteTurnPayloadText then returned null before any stall
detector ran, leaving the session with livenessState="working" and no
recovery path — a silent, permanent stall.

Add isUnsignedThinkingOnlyAssistantTurn helper and wire it into:
- resolveIncompleteTurnPayloadText: bypass the visible-text guard so
  stall detection fires below
- resolveReasoningOnlyRetryInstruction: retry unsigned thinking-only
  turns with REASONING_ONLY_RETRY_INSTRUCTION before surfacing error

Fixes #89787

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

@@ -1519,6 +1519,63 @@ describe("runEmbeddedAgent incomplete-turn safety", () => {
     expect(incompleteTurnText).toBeNull();
   });
 
+  it("surfaces stall on clean stop with only an unsigned thinking payload (payloadCount=1, no visible text)", () => {
+    // Regression: unsigned thinking payloads increment payloadCount but carry no
+    // user-visible content. The visible-text guard must not suppress incomplete-turn
+    // detection when the model produced only a thinking block and no answer. (#89787)
+    const incompleteTurnText = resolveIncompleteTurnPayloadText({
+      payloadCount: 1,
+      aborted: false,
+      timedOut: false,
+      attempt: makeAttemptResult({
+        assistantTexts: [],
+        lastAssistant: {
+          role: "assistant",
+          stopReason: "stop",
+          provider: "openai",
+          model: "qwen3.6-35b-a3b",
+          content: [
+            {
+              type: "thinking",
+              thinking: "let me plan the tool calls I need to make...",
+              // no signature — unsigned thinking block
+            },
+          ],
+        } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+      }),
+    });
+
+    expect(incompleteTurnText).toContain("couldn't generate a response");
+  });
+
+  it("does not surface a stall when unsigned thinking accompanies visible text (payloadCount=1)", () => {
+    // When the model emits both a thinking block and a visible text answer, the turn
+    // succeeded and no stall should be surfaced even though thinking is unsigned.
+    const incompleteTurnText = resolveIncompleteTurnPayloadText({
+      payloadCount: 1,
+      aborted: false,
+      timedOut: false,
+      attempt: makeAttemptResult({
+        assistantTexts: ["Here is the answer to your question."],
+        lastAssistant: {
+          role: "assistant",
+          stopReason: "stop",
+          provider: "openai",
+          model: "qwen3.6-35b-a3b",
+          content: [
+            {
+              type: "thinking",
+              thinking: "let me answer this...",
+            },
+            { type: "text", text: "Here is the answer to your question." },
+          ],
+        } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+      }),
+    });
+
+    expect(incompleteTurnText).toBeNull();
+  });
+
   it("surfaces an error for tool-use terminal turn with pre-tool text via runEmbeddedAgent (#76477)", async () => {
     mockedClassifyFailoverReason.mockReturnValue(null);
     mockedRunEmbeddedAttempt.mockResolvedValueOnce(
@@ -1687,6 +1744,59 @@ describe("runEmbeddedAgent incomplete-turn safety", () => {
     expect(retryInstruction).toBe(REASONING_ONLY_RETRY_INSTRUCTION);
   });
 
+  it("retries unsigned thinking-only turns via the reasoning-only path (openai-completions)", () => {
+    const retryInstruction = resolveReasoningOnlyRetryInstruction({
+      provider: "openai",
+      modelId: "qwen3.6-35b-a3b",
+      modelApi: "openai-completions",
+      aborted: false,
+      timedOut: false,
+      attempt: makeAttemptResult({
+        assistantTexts: [],
+        lastAssistant: {
+          role: "assistant",
+          stopReason: "stop",
+          provider: "openai",
+          model: "qwen3.6-35b-a3b",
+          content: [
+            {
+              type: "thinking",
+              thinking: "let me plan the tool calls I need to make...",
+            },
+          ],
+        } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+      }),
+    });
+
+    expect(retryInstruction).toBe(REASONING_ONLY_RETRY_INSTRUCTION);
+  });
+
+  it("retries unsigned thinking-only Ollama turns via the reasoning-only path", () => {
+    const retryInstruction = resolveReasoningOnlyRetryInstruction({
+      provider: "ollama",
+      modelId: "gemma4:31b",
+      aborted: false,
+      timedOut: false,
+      attempt: makeAttemptResult({
+        assistantTexts: [],
+        lastAssistant: {
+          role: "assistant",
+          stopReason: "end_turn",
+          provider: "ollama",
+          model: "gemma4:31b",
+          content: [
+            {
+              type: "thinking",
+              thinking: "internal reasoning",
+            },
+          ],
+        } as unknown as EmbeddedRunAttemptResult["lastAssistant"],
+      }),
+    });
+
+    expect(retryInstruction).toBe(REASONING_ONLY_RETRY_INSTRUCTION);
+  });
+
   it("retries unsigned-thinking Ollama turns via the empty-response path", () => {
     const retryInstruction = resolveEmptyResponseRetryInstruction({
       provider: "ollama",

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

@@ -266,9 +266,17 @@ export function resolveIncompleteTurnPayloadText(params: {
   // turn check in that case — the final post-tool response was never
   // produced. (#76477)
   const toolUseTerminal = params.attempt.lastAssistant?.stopReason === "toolUse";
+  const assistant = params.attempt.currentAttemptAssistant ?? params.attempt.lastAssistant;
+  // Unsigned thinking payloads count toward payloadCount but carry no user-visible
+  // content; bypass the visible-text guard when unsigned thinking was the only output
+  // so that incomplete-turn stall detection fires below. (#89787)
+  const unsignedThinkingOnlyTerminal =
+    params.payloadCount !== 0 &&
+    !joinAssistantTexts(params.attempt.assistantTexts).length &&
+    isUnsignedThinkingOnlyAssistantTurn(assistant);
 
   if (
-    (params.payloadCount !== 0 && !toolUseTerminal) ||
+    (params.payloadCount !== 0 && !toolUseTerminal && !unsignedThinkingOnlyTerminal) ||
     (params.aborted && params.externalAbort) ||
     params.timedOut ||
     params.attempt.clientToolCalls ||
@@ -300,9 +308,7 @@ export function resolveIncompleteTurnPayloadText(params: {
     hasAssistantVisibleText: params.payloadCount > 0,
     lastAssistant: params.attempt.lastAssistant,
   });
-  const reasoningOnlyAssistant = isReasoningOnlyAssistantTurn(
-    params.attempt.currentAttemptAssistant ?? params.attempt.lastAssistant,
-  );
+  const reasoningOnlyAssistant = isReasoningOnlyAssistantTurn(assistant);
   const emptyResponseAssistant = isEmptyResponseAssistantTurn({
     payloadCount: params.payloadCount,
     attempt: params.attempt,
@@ -310,6 +316,7 @@ export function resolveIncompleteTurnPayloadText(params: {
   if (
     !incompleteTerminalAssistant &&
     !reasoningOnlyAssistant &&
+    !unsignedThinkingOnlyTerminal &&
     !emptyResponseAssistant &&
     stopReason !== "error"
   ) {
@@ -508,6 +515,20 @@ function isReasoningOnlyAssistantTurn(message: unknown): boolean {
   return assessLastAssistantMessage(message as AgentMessage) === "incomplete-text";
 }
 
+// Unsigned thinking blocks have no cryptographic signature; assessLastAssistantMessage
+// returns "incomplete-thinking" for them. Empty content also returns "incomplete-thinking",
+// so the content.length > 0 guard is required to distinguish the two cases.
+function isUnsignedThinkingOnlyAssistantTurn(message: unknown): boolean {
+  if (message == null || typeof message !== "object") {
+    return false;
+  }
+  const content = (message as { content?: unknown }).content;
+  if (!Array.isArray(content) || content.length === 0) {
+    return false;
+  }
+  return assessLastAssistantMessage(message as AgentMessage) === "incomplete-thinking";
+}
+
 function isEmptyResponseAssistantTurn(params: {
   payloadCount: number;
   attempt: Pick<
@@ -638,7 +659,7 @@ export function resolveReasoningOnlyRetryInstruction(params: {
   if (assistant?.stopReason === "error") {
     return null;
   }
-  if (!isReasoningOnlyAssistantTurn(assistant)) {
+  if (!isReasoningOnlyAssistantTurn(assistant) && !isUnsignedThinkingOnlyAssistantTurn(assistant)) {
     return null;
   }