fix(agents): gate auto reasoning by effective thinking level (openclaw#24335) thanks @Kay-051

src/auto-reply/reply.directive.directive-behavior.defaults-think-low-reasoning-capable-models-no.test.ts

@@ -244,7 +244,7 @@ describe("directive behavior", () => {
       expect(call?.model).toBe("claude-opus-4-5");
     });
   });
-  it("defaults thinking to low for reasoning-capable models during normal replies", async () => {
+  it("defaults thinking to low for reasoning-capable models without auto-enabling reasoning", async () => {
     await withTempHome(async (home) => {
       mockEmbeddedTextResult("done");
       vi.mocked(loadModelCatalog).mockResolvedValueOnce([
@@ -269,6 +269,38 @@ describe("directive behavior", () => {
       expect(runEmbeddedPiAgent).toHaveBeenCalledOnce();
       const call = vi.mocked(runEmbeddedPiAgent).mock.calls[0]?.[0];
       expect(call?.thinkLevel).toBe("low");
+      expect(call?.reasoningLevel).toBe("off");
+    });
+  });
+  it("keeps auto-reasoning enabled when thinking is explicitly off", async () => {
+    await withTempHome(async (home) => {
+      mockEmbeddedTextResult("done");
+      vi.mocked(loadModelCatalog).mockResolvedValueOnce([
+        {
+          id: "claude-opus-4-5",
+          name: "Opus 4.5",
+          provider: "anthropic",
+          reasoning: true,
+        },
+      ]);
+
+      await getReplyFromConfig(
+        {
+          Body: "hello",
+          From: "+1004",
+          To: "+2000",
+        },
+        {},
+        makeWhatsAppDirectiveConfig(home, {
+          model: { primary: "anthropic/claude-opus-4-5" },
+          thinkingDefault: "off",
+        }),
+      );
+
+      expect(runEmbeddedPiAgent).toHaveBeenCalledOnce();
+      const call = vi.mocked(runEmbeddedPiAgent).mock.calls[0]?.[0];
+      expect(call?.thinkLevel).toBe("off");
+      expect(call?.reasoningLevel).toBe("on");
     });
   });
   it("passes elevated defaults when sender is approved", async () => {

src/auto-reply/reply/get-reply-directives.ts

@@ -389,14 +389,16 @@ export async function resolveReplyDirectives(params: {
   provider = modelState.provider;
   model = modelState.model;
 
-  // When neither directive nor session set reasoning, default to model capability (e.g. OpenRouter with reasoning: true).
-  // Skip auto-enabling when thinking is already active — the model's internal
-  // thinking blocks would otherwise be formatted and delivered as visible
-  // "Reasoning:" messages, leaking internal content to the user.
+  // When neither directive nor session set reasoning, default to model capability
+  // (e.g. OpenRouter with reasoning: true). Skip auto-enabling when thinking is
+  // active, including model-inferred defaults, or internal thinking blocks can
+  // be emitted as visible "Reasoning:" messages.
   const reasoningExplicitlySet =
     directives.reasoningLevel !== undefined ||
     (sessionEntry?.reasoningLevel !== undefined && sessionEntry?.reasoningLevel !== null);
-  const thinkingActive = resolvedThinkLevel !== undefined && resolvedThinkLevel !== "off";
+  const effectiveThinkingForReasoning =
+    resolvedThinkLevel ?? (await modelState.resolveDefaultThinkingLevel());
+  const thinkingActive = effectiveThinkingForReasoning !== "off";
   if (!reasoningExplicitlySet && resolvedReasoningLevel === "off" && !thinkingActive) {
     resolvedReasoningLevel = await modelState.resolveDefaultReasoningLevel();
   }