diff --git a/.changeset/fix-empty-run-result-fallback.md b/.changeset/fix-empty-run-result-fallback.md
new file mode 100644
index 00000000..3c87be83
--- /dev/null
+++ b/.changeset/fix-empty-run-result-fallback.md
@@ -0,0 +1,5 @@
+---
+"@perstack/runtime": patch
+---
+
+Use fallback text instead of retry when LLM generates empty text in GeneratingRunResult
diff --git a/packages/runtime/src/state-machine/states/generating-run-result.test.ts b/packages/runtime/src/state-machine/states/generating-run-result.test.ts
index 508a0212..79d43ef0 100644
--- a/packages/runtime/src/state-machine/states/generating-run-result.test.ts
+++ b/packages/runtime/src/state-machine/states/generating-run-result.test.ts
@@ -302,7 +302,7 @@ describe("@perstack/runtime: StateMachineLogic['GeneratingRunResult']", () => {
     }
   })
 
-  it("retries when LLM generates empty text", async () => {
+  it("falls back to 'OK' when LLM generates empty text", async () => {
     const setting = createRunSetting()
     const checkpoint = createCheckpoint()
     const step = createStep({
@@ -332,46 +332,18 @@ describe("@perstack/runtime: StateMachineLogic['GeneratingRunResult']", () => {
       skillManager: createMockSkillManagerFromAdapters({}),
       llmExecutor: mockLLMExecutor as unknown as LLMExecutor,
     })
-    expect(event.type).toBe("retry")
-    if (event.type === "retry") {
-      expect(event.reason).toContain("No text generated")
-    }
-  })
-
-  it("returns stopRunByError when empty text retries exceed maxRetries", async () => {
-    const setting = createRunSetting({ maxRetries: 2 })
-    const checkpoint = createCheckpoint({ retryCount: 2 })
-    const step = createStep({
-      toolCalls: [
-        {
-          id: "tc_123",
-          skillName: "@perstack/base",
-          toolName: "attemptCompletion",
-          args: {},
-        },
-      ],
-      toolResults: [
-        {
-          id: "tc_123",
-          skillName: "@perstack/base",
-          toolName: "attemptCompletion",
-          result: [{ type: "textPart", text: JSON.stringify({}), id: createId() }],
-        },
-      ],
-    })
-    mockLLMExecutor.setMockResult(createMockResult(undefined))
-    const event = await StateMachineLogics.GeneratingRunResult({
-      setting,
-      checkpoint,
-      step,
-      eventListener: async () => {},
-      skillManager: createMockSkillManagerFromAdapters({}),
-      llmExecutor: mockLLMExecutor as unknown as LLMExecutor,
-    })
-    expect(event.type).toBe("stopRunByError")
-    if (event.type === "stopRunByError") {
-      expect(event.error.name).toBe("EmptyRunResult")
-      expect(event.error.message).toContain("No text generated for run result")
+    expect(event.type).toBe("completeRun")
+    if (event.type === "completeRun") {
+      expect(event.text).toBe("OK")
+      const lastMessage = event.checkpoint.messages[event.checkpoint.messages.length - 1]
+      expect(lastMessage.type).toBe("expertMessage")
+      if (lastMessage.type === "expertMessage") {
+        const textPart = lastMessage.contents.find((c) => c.type === "textPart")
+        expect(textPart).toBeDefined()
+        if (textPart?.type === "textPart") {
+          expect(textPart.text).toBe("OK")
+        }
+      }
     }
   })
 })
diff --git a/packages/runtime/src/state-machine/states/generating-run-result.ts b/packages/runtime/src/state-machine/states/generating-run-result.ts
index a1a37a7c..e390f9dc 100644
--- a/packages/runtime/src/state-machine/states/generating-run-result.ts
+++ b/packages/runtime/src/state-machine/states/generating-run-result.ts
@@ -144,46 +144,19 @@ export async function generatingRunResultLogic({
   const usage = usageFromGenerateTextResult(generationResult)
   const { text, reasoning } = generationResult
 
-  // Empty text = retry (LLM generated no visible text, possibly only reasoning/thinking)
-  if (!text) {
-    const currentRetryCount = checkpoint.retryCount ?? 0
-    if (currentRetryCount >= setting.maxRetries) {
-      return stopRunByError(setting, checkpoint, {
-        checkpoint: {
-          ...checkpoint,
-          status: "stoppedByError",
-        },
-        step: {
-          ...step,
-          finishedAt: Date.now(),
-        },
-        error: {
-          name: "EmptyRunResult",
-          message: `Max retries (${setting.maxRetries}) exceeded: No text generated for run result`,
-          isRetryable: false,
-        },
-      })
-    }
-    const reason = JSON.stringify({
-      error: "Error: No text generated",
-      message: "You must provide a text response summarizing the result. Try again.",
-    })
-    return retry(setting, checkpoint, {
-      reason,
-      newMessages: [toolMessage, createUserMessage([{ type: "textPart", text: reason }])],
-      usage,
-    })
-  }
-
   // Extract thinking from reasoning (Anthropic, Google)
   const thinkingParts = extractThinkingParts(reasoning as ReasoningPart[] | undefined)
   const thinkingText = extractThinkingText(reasoning as ReasoningPart[] | undefined)
 
+  // Fallback when LLM generates no visible text (e.g., only reasoning/thinking with extended thinking).
+  // Since GeneratingRunResult is only reached after a successful attemptCompletion (no remaining todos),
+  // the task is already complete — use "OK" as a minimal non-empty completion text.
+  const resultText = text || "OK"
+
   // Build ExpertMessage with ThinkingPart + TextPart
-  // Always include textPart even if empty - required for delegation result handling
   const expertContents: Array<Omit<ThinkingPart, "id"> | Omit<TextPart, "id">> = [
     ...thinkingParts,
-    { type: "textPart" as const, text: text ?? "" },
+    { type: "textPart" as const, text: resultText },
   ]
   const newMessages = [toolMessage, createExpertMessage(expertContents)]
 
@@ -217,7 +190,7 @@ export async function generatingRunResultLogic({
       finishedAt: Date.now(),
       usage: sumUsage(step.usage, usage),
     },
-    text,
+    text: resultText,
     usage,
   })
 }