fix(text): omit enable_thinking by default and retry when API requires false

Non-streaming chat no longer forces enable_thinking=false, which breaks
thinking-only models. Retry once with false only when the server demands it.

Author: hopelynd

packages/commands/src/commands/auth/login-api-key.ts

@@ -4,6 +4,9 @@ import {
   chatPath,
   requestJson,
   normalizeModelBaseUrl,
+  applyChatEnableThinking,
+  resolveChatEnableThinking,
+  withEnableThinkingRetry,
   type AuthPersistPatch,
   type AuthStore,
   type Identity,
@@ -58,32 +61,48 @@ export async function validateAndPersistApiKey(
     ? normalizeModelBaseUrl(profile.persistBaseUrl)
     : undefined;
   const validationModel = profile.defaultTextModel || "qwen3.7-max";
+  const body: {
+    model: string;
+    messages: Array<{ role: string; content: string }>;
+    max_tokens: number;
+    stream: boolean;
+    enable_thinking?: boolean;
+  } = {
+    model: validationModel,
+    messages: [{ role: "user", content: "hi" }],
+    max_tokens: 1,
+    stream: false,
+  };
+
   const requestOpts = {
     url: baseUrl + chatPath(),
     method: "POST",
     headers: { Authorization: `Bearer ${key}` },
     timeout: Math.min(deps.settings.timeout, 30),
-    body: {
-      model: validationModel,
-      messages: [{ role: "user", content: "hi" }],
-      max_tokens: 1,
-      stream: false,
-      enable_thinking: validationModel === "qwen3.8-max-preview",
-    },
+    body,
   };
 
-  for (let attempt = 1; attempt <= 3; attempt++) {
-    try {
-      await requestJson<unknown>(httpDeps, requestOpts);
-      break;
-    } catch (error) {
-      if (attempt >= 3 || !canRetry(error)) {
-        process.stderr.write("Failed\n");
-        throw error;
-      }
-      const delayMs = RETRY_DELAY_BASE_MS * 2 ** (attempt - 1);
-      await new Promise((resolve) => setTimeout(resolve, delayMs));
-    }
+  try {
+    await withEnableThinkingRetry({
+      // Validation requests are always non-streaming.
+      initial: resolveChatEnableThinking({ stream: false }),
+      apply: (value) => applyChatEnableThinking(body, value),
+      run: async () => {
+        for (let attempt = 1; attempt <= 3; attempt++) {
+          try {
+            await requestJson<unknown>(httpDeps, requestOpts);
+            return;
+          } catch (error) {
+            if (attempt >= 3 || !canRetry(error)) throw error;
+            const delayMs = RETRY_DELAY_BASE_MS * 2 ** (attempt - 1);
+            await new Promise((resolve) => setTimeout(resolve, delayMs));
+          }
+        }
+      },
+    });
+  } catch (error) {
+    process.stderr.write("Failed\n");
+    throw error;
   }
 
   process.stderr.write("Valid\n");

packages/commands/src/commands/text/chat.ts

@@ -4,6 +4,9 @@ import {
   parseSSE,
   detectOutputFormat,
   readTextFromPathOrStdin,
+  applyChatEnableThinking,
+  resolveChatEnableThinking,
+  withEnableThinkingRetry,
   type ChatMessage,
   type ChatRequest,
   type ChatResponse,
@@ -124,7 +127,8 @@ export default defineCommand({
     const { system, messages } = parseMessages(flags);
 
     const model = flags.model || settings.defaultTextModel || "qwen3.7-max";
-    const shouldStream = flags.stream || process.stdout.isTTY;
+    // Coerce isTTY (may be undefined) so stream:false is serialized.
+    const shouldStream = Boolean(flags.stream || process.stdout.isTTY);
     const format = detectOutputFormat(settings.output);
 
     // Build messages array with system prompt
@@ -144,16 +148,13 @@ export default defineCommand({
     if (flags.temperature !== undefined) body.temperature = flags.temperature;
     if (flags.topP !== undefined) body.top_p = flags.topP;
 
-    if (flags.enableThinking) {
-      body.enable_thinking = true;
-      if (flags.thinkingBudget !== undefined) {
-        body.thinking_budget = flags.thinkingBudget;
-      }
-    } else if (!shouldStream) {
-      // DashScope qwen3 models default to enable_thinking=true server-side, but
-      // non-streaming calls require it to be explicitly false. Stream calls
-      // support thinking, so leave the field unset there (server handles it).
-      body.enable_thinking = false;
+    const enableThinking = resolveChatEnableThinking({
+      enableThinking: flags.enableThinking,
+      stream: shouldStream,
+    });
+    applyChatEnableThinking(body, enableThinking);
+    if (enableThinking === true && flags.thinkingBudget !== undefined) {
+      body.thinking_budget = flags.thinkingBudget;
     }
 
     if (flags.tool) {
@@ -229,10 +230,15 @@ export default defineCommand({
         resultOut.write("\n");
       }
     } else {
-      const response = await ctx.client.requestJson<ChatResponse>({
-        path: chatPath(),
-        method: "POST",
-        body,
+      const response = await withEnableThinkingRetry({
+        initial: enableThinking,
+        apply: (value) => applyChatEnableThinking(body, value),
+        run: () =>
+          ctx.client.requestJson<ChatResponse>({
+            path: chatPath(),
+            method: "POST",
+            body,
+          }),
       });
 
       const text = response.choices?.[0]?.message?.content ?? "";

packages/commands/tests/e2e/auth.e2e.test.ts

@@ -317,7 +317,7 @@ describe("e2e: auth", () => {
         body: {
           model: "qwen3.8-max-preview",
           stream: false,
-          enable_thinking: true,
+          enable_thinking: false,
         },
       });
 

packages/commands/tests/e2e/text-chat.e2e.test.ts

@@ -34,18 +34,25 @@ describe.skipIf(!isDashScopeE2EReady())("e2e: text chat（DashScope）", () => {
       "--model",
       "qwen3.7-max",
       "--message",
-      "干跑",
+      "dry-run",
       "--max-tokens",
       "8",
       "--output",
       "json",
     ]);
     expect(exitCode, stderr).toBe(0);
     const data = parseStdoutJson<{
-      request?: { model?: string; messages?: Array<{ content?: string }> };
+      request?: {
+        model?: string;
+        messages?: Array<{ content?: string }>;
+        enable_thinking?: boolean;
+        stream?: boolean;
+      };
     }>(stdout);
     expect(data.request?.model).toBe("qwen3.7-max");
-    expect(data.request?.messages?.some((m) => m.content === "干跑")).toBe(true);
+    expect(data.request?.messages?.some((message) => message.content === "dry-run")).toBe(true);
+    expect(data.request?.stream).toBe(false);
+    expect(data.request?.enable_thinking).toBe(false);
   });
 
   test("【qwen3.7-max】文本对话", async () => {

packages/core/src/index.ts

@@ -14,5 +14,6 @@ export * from "./finetune/index.ts";
 export * from "./deploy/index.ts";
 export * from "./types/index.ts";
 export * from "./utils/index.ts";
+export * from "./models/index.ts";
 export * from "./telemetry/index.ts";
 export * from "./advisor/index.ts";

packages/core/src/models/index.ts

@@ -0,0 +1,7 @@
+export {
+  adjustEnableThinkingAfterError,
+  applyChatEnableThinking,
+  resolveChatEnableThinking,
+  withEnableThinkingRetry,
+  type EnableThinkingAdjustResult,
+} from "./thinking.ts";

packages/core/src/models/thinking.ts

@@ -0,0 +1,76 @@
+/** resolve / adjust / retry helpers for chat `enable_thinking`. */
+
+/** Resolve the initial `enable_thinking` value (`undefined` omits the field). */
+export function resolveChatEnableThinking(options: {
+  enableThinking?: boolean;
+  /** Whether the request is streaming. */
+  stream?: boolean;
+}): boolean | undefined {
+  if (options.enableThinking) return true;
+  if (options.stream === false) return false;
+  return undefined;
+}
+
+export type EnableThinkingAdjustResult =
+  | { kind: "retry"; value: boolean | undefined }
+  | { kind: "none" };
+
+/** Map clear `enable_thinking` constraint errors to a one-shot retry adjustment. */
+export function adjustEnableThinkingAfterError(
+  current: boolean | undefined,
+  errorMessage: string,
+): EnableThinkingAdjustResult {
+  if (current !== true && /enable_thinking parameter is restricted to\s*true/i.test(errorMessage)) {
+    return { kind: "retry", value: true };
+  }
+
+  if (current === undefined && /enable_thinking must be set to false/i.test(errorMessage)) {
+    return { kind: "retry", value: false };
+  }
+
+  if (current !== undefined && /does not support enable_thinking/i.test(errorMessage)) {
+    return { kind: "retry", value: undefined };
+  }
+
+  return { kind: "none" };
+}
+
+/** Set or remove `enable_thinking`; clear `thinking_budget` when disabled or omitted. */
+export function applyChatEnableThinking(
+  body: { enable_thinking?: boolean; thinking_budget?: number },
+  value: boolean | undefined,
+): void {
+  if (value === undefined) {
+    delete body.enable_thinking;
+    delete body.thinking_budget;
+    return;
+  }
+  if (value === false) {
+    body.enable_thinking = false;
+    delete body.thinking_budget;
+    return;
+  }
+  body.enable_thinking = true;
+}
+
+function errorMessageOf(error: unknown): string {
+  if (error instanceof Error) return error.message;
+  return String(error);
+}
+
+/** Run once, then retry once if the error indicates an `enable_thinking` constraint. */
+export async function withEnableThinkingRetry<T>(options: {
+  initial: boolean | undefined;
+  apply: (value: boolean | undefined) => void;
+  run: () => Promise<T>;
+}): Promise<T> {
+  options.apply(options.initial);
+  try {
+    return await options.run();
+  } catch (error) {
+    const adjusted = adjustEnableThinkingAfterError(options.initial, errorMessageOf(error));
+    if (adjusted.kind === "none") throw error;
+    options.apply(adjusted.value);
+    return await options.run();
+  }
+}