fix(agents): prioritize per-model thinking defaults (#30439)

* fix(agents): honor per-model thinking defaults

* fix(agents): preserve thinking fallback with model defaults

---------

Co-authored-by: Mark L <73659136+markliuyuxiang@users.noreply.github.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>

src/auto-reply/reply/directive-handling.levels.test.ts

@@ -0,0 +1,36 @@
+import { describe, expect, it, vi } from "vitest";
+import { resolveCurrentDirectiveLevels } from "./directive-handling.levels.js";
+
+describe("resolveCurrentDirectiveLevels", () => {
+  it("prefers resolved model default over agent thinkingDefault", async () => {
+    const resolveDefaultThinkingLevel = vi.fn().mockResolvedValue("high");
+
+    const result = await resolveCurrentDirectiveLevels({
+      sessionEntry: {},
+      agentCfg: {
+        thinkingDefault: "low",
+      },
+      resolveDefaultThinkingLevel,
+    });
+
+    expect(result.currentThinkLevel).toBe("high");
+    expect(resolveDefaultThinkingLevel).toHaveBeenCalledTimes(1);
+  });
+
+  it("keeps session thinking override without consulting defaults", async () => {
+    const resolveDefaultThinkingLevel = vi.fn().mockResolvedValue("high");
+
+    const result = await resolveCurrentDirectiveLevels({
+      sessionEntry: {
+        thinkingLevel: "minimal",
+      },
+      agentCfg: {
+        thinkingDefault: "low",
+      },
+      resolveDefaultThinkingLevel,
+    });
+
+    expect(result.currentThinkLevel).toBe("minimal");
+    expect(resolveDefaultThinkingLevel).not.toHaveBeenCalled();
+  });
+});

src/auto-reply/reply/directive-handling.levels.ts

@@ -21,8 +21,8 @@ export async function resolveCurrentDirectiveLevels(params: {
 }> {
   const resolvedDefaultThinkLevel =
     (params.sessionEntry?.thinkingLevel as ThinkLevel | undefined) ??
-    (params.agentCfg?.thinkingDefault as ThinkLevel | undefined) ??
-    (await params.resolveDefaultThinkingLevel());
+    (await params.resolveDefaultThinkingLevel()) ??
+    (params.agentCfg?.thinkingDefault as ThinkLevel | undefined);
   const currentThinkLevel = resolvedDefaultThinkLevel;
   const currentVerboseLevel =
     (params.sessionEntry?.verboseLevel as VerboseLevel | undefined) ??

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

@@ -339,9 +339,7 @@ export async function resolveReplyDirectives(params: {
   });
   const defaultActivation = defaultGroupActivation(requireMention);
   const resolvedThinkLevel =
-    directives.thinkLevel ??
-    (sessionEntry?.thinkingLevel as ThinkLevel | undefined) ??
-    (agentCfg?.thinkingDefault as ThinkLevel | undefined);
+    directives.thinkLevel ?? (sessionEntry?.thinkingLevel as ThinkLevel | undefined);
 
   const resolvedVerboseLevel =
     directives.verboseLevel ??
@@ -390,6 +388,10 @@ export async function resolveReplyDirectives(params: {
   });
   provider = modelState.provider;
   model = modelState.model;
+  const resolvedThinkLevelWithDefault =
+    resolvedThinkLevel ??
+    (await modelState.resolveDefaultThinkingLevel()) ??
+    (agentCfg?.thinkingDefault as ThinkLevel | undefined);
 
   // When neither directive nor session set reasoning, default to model capability
   // (e.g. OpenRouter with reasoning: true). Skip auto-enabling when thinking is
@@ -398,9 +400,7 @@ export async function resolveReplyDirectives(params: {
   const reasoningExplicitlySet =
     directives.reasoningLevel !== undefined ||
     (sessionEntry?.reasoningLevel !== undefined && sessionEntry?.reasoningLevel !== null);
-  const effectiveThinkingForReasoning =
-    resolvedThinkLevel ?? (await modelState.resolveDefaultThinkingLevel());
-  const thinkingActive = effectiveThinkingForReasoning !== "off";
+  const thinkingActive = resolvedThinkLevelWithDefault !== "off";
   if (!reasoningExplicitlySet && resolvedReasoningLevel === "off" && !thinkingActive) {
     resolvedReasoningLevel = await modelState.resolveDefaultReasoningLevel();
   }
@@ -477,7 +477,7 @@ export async function resolveReplyDirectives(params: {
       elevatedAllowed,
       elevatedFailures,
       defaultActivation,
-      resolvedThinkLevel,
+      resolvedThinkLevel: resolvedThinkLevelWithDefault,
       resolvedVerboseLevel,
       resolvedReasoningLevel,
       resolvedElevatedLevel,