fix(auto-reply): honor per-model thinking params

Auto-reply now uses the existing per-model model params thinking value before falling back to the global thinkingDefault, matching gateway/shared model selection behavior.\n\nVerified with targeted auto-reply and agents Vitest coverage plus formatting and diff checks.\n\nThanks @tynamite for the fix.

src/auto-reply/reply/model-selection.test.ts

@@ -108,6 +108,43 @@ describe("createModelSelectionState catalog loading", () => {
     expect(loadModelCatalog).not.toHaveBeenCalled();
   });
 
+  it("prefers per-model params.thinking over global thinkingDefault", async () => {
+    vi.mocked(loadModelCatalog).mockClear();
+    const cfg = {
+      agents: {
+        defaults: {
+          thinkingDefault: "low",
+          models: {
+            "openai-codex/gpt-5.4": {
+              params: { thinking: "high" },
+            },
+          },
+        },
+      },
+      models: {
+        providers: {
+          "openai-codex": {
+            baseUrl: "https://api.openai.com/v1",
+            models: [makeConfiguredModel()],
+          },
+        },
+      },
+    } as OpenClawConfig;
+
+    const state = await createModelSelectionState({
+      cfg,
+      agentCfg: cfg.agents?.defaults,
+      defaultProvider: "openai-codex",
+      defaultModel: "gpt-5.4",
+      provider: "openai-codex",
+      model: "gpt-5.4",
+      hasModelDirective: false,
+    });
+
+    await expect(state.resolveDefaultThinkingLevel()).resolves.toBe("high");
+    expect(loadModelCatalog).not.toHaveBeenCalled();
+  });
+
   it("uses the implicit model default when no global thinking default is configured", async () => {
     vi.mocked(loadModelCatalog).mockClear();
     const cfg = {

src/auto-reply/reply/model-selection.ts

@@ -7,6 +7,7 @@ import type { ModelCatalogEntry } from "../../agents/model-catalog.js";
 import { parseConfiguredModelVisibilityEntries } from "../../agents/model-selection-shared.js";
 import {
   buildConfiguredModelCatalog,
+  legacyModelKey,
   modelKey,
   normalizeModelRef,
   normalizeProviderId,
@@ -458,10 +459,32 @@ export async function createModelSelectionState(params: {
       return defaultThinkingLevel;
     }
     const agentThinkingDefault = agentEntry?.thinkingDefault as ThinkLevel | undefined;
+    if (agentThinkingDefault) {
+      defaultThinkingLevel = agentThinkingDefault;
+      return defaultThinkingLevel;
+    }
+    const configuredModels = cfg.agents?.defaults?.models;
+    const canonicalKey = modelKey(provider, model);
+    const legacyKey = legacyModelKey(provider, model);
+    const configuredModelThinkingDefault =
+      configuredModels?.[canonicalKey]?.params?.thinking ??
+      (legacyKey ? configuredModels?.[legacyKey]?.params?.thinking : undefined);
+    if (
+      configuredModelThinkingDefault === "off" ||
+      configuredModelThinkingDefault === "minimal" ||
+      configuredModelThinkingDefault === "low" ||
+      configuredModelThinkingDefault === "medium" ||
+      configuredModelThinkingDefault === "high" ||
+      configuredModelThinkingDefault === "xhigh" ||
+      configuredModelThinkingDefault === "adaptive" ||
+      configuredModelThinkingDefault === "max"
+    ) {
+      defaultThinkingLevel = configuredModelThinkingDefault;
+      return defaultThinkingLevel;
+    }
     const configuredThinkingDefault = agentCfg?.thinkingDefault as ThinkLevel | undefined;
-    const explicitThinkingDefault = agentThinkingDefault ?? configuredThinkingDefault;
-    if (explicitThinkingDefault) {
-      defaultThinkingLevel = explicitThinkingDefault;
+    if (configuredThinkingDefault) {
+      defaultThinkingLevel = configuredThinkingDefault;
       return defaultThinkingLevel;
     }
     const catalogForThinking = await resolveThinkingCatalog();