feat(core): support glm-5.3-flash under glm-v family

apps/site/docs/en/model-common-config.mdx

@@ -400,6 +400,7 @@ MIDSCENE_INSIGHT_MODEL_FAMILY="xiaomi-mimo"
 
 | Model version | Commonly used model names | `MIDSCENE_MODEL_FAMILY` | Notes |
 | --- | --- | --- | --- |
+| GLM-5.3 series | `glm-5.3-flash` | `glm-v` | Always-thinking: `thinking.type` cannot be disabled, so disabling reasoning sends `reasoning_effort: low` instead. |
 | GLM-5V series | `glm-5v-turbo` | `glm-v` | — |
 | GLM-4.6 series | `glm-4.6v` | `glm-v` | `glm-4.6v` is open-source. |
 

apps/site/docs/zh/model-common-config.mdx

@@ -401,6 +401,7 @@ MIDSCENE_INSIGHT_MODEL_FAMILY="xiaomi-mimo"
 
 | 模型版本 | 常用模型名称 | `MIDSCENE_MODEL_FAMILY` | 备注 |
 | --- | --- | --- | --- |
+| GLM-5.3 系列 | `glm-5.3-flash` | `glm-v` | 始终思考：不支持关闭 `thinking.type`，关闭推理时会改发 `reasoning_effort: low`。 |
 | GLM-5V 系列 | `glm-5v-turbo` | `glm-v` | — |
 | GLM-4.6 系列 | `glm-4.6v` | `glm-v` | `glm-4.6v` 是开源模型。 |
 

packages/core/src/ai-model/model-adapter/types.ts

@@ -63,6 +63,12 @@ export type ChatCompletionUnsupportedUserConfig =
 
 export interface ChatCompletionCallInput {
   intent?: TIntent;
+  /**
+   * Model name resolved for this call. Adapters may use it to branch within a
+   * family when a specific model requires different request parameters (for
+   * example, always-thinking variants).
+   */
+  modelName?: string;
   userConfig?: ChatCompletionCallUserConfig;
   /**
    * Number of preceding semantic parsing failures for this request.
@@ -83,6 +89,10 @@ export interface ChatCompletionCallInput {
 
 export interface ChatCompletionCallContext {
   intent?: TIntent;
+  /**
+   * Model name resolved for this call; see ChatCompletionCallInput.
+   */
+  modelName?: string;
   userConfig: ChatCompletionCallUserConfig;
   semanticRetryAttempt?: number;
   requiresOriginalImageDetail?: boolean;

packages/core/src/ai-model/models/glm.ts

@@ -5,14 +5,31 @@ import type {
   ModelAdapterDefinition,
 } from '../model-adapter/types';
 
+const ALWAYS_THINKING_GLM_MODEL_PATTERN = /^glm-5\.3-flash\b/;
+
 const buildGlmChatCompletionParams = (
   input: ChatCompletionCallContext,
 ): ChatCompletionParamsResult => {
-  const { midsceneDefaults, userConfig } = input;
+  const { midsceneDefaults, userConfig, modelName } = input;
   const { reasoningEnabled } = userConfig;
   const commonOverrideConfig: Record<string, unknown> = {};
 
-  if (userConfig.temperature !== undefined) {
+  // GLM models that cannot turn thinking off. GLM-5.3-Flash rejects
+  // `thinking.type: 'disabled'` with error 1210 ("该模型始终思考，不支持关闭
+  // 思考"), so for these models the adapter keeps thinking enabled and steers
+  // its depth with `reasoning_effort` instead. Field-tested against the
+  // Zhipu Coding Plan endpoint in KSL-49 (2026-08-26).
+  const alwaysThinking = ALWAYS_THINKING_GLM_MODEL_PATTERN.test(
+    modelName ?? '',
+  );
+
+  if (alwaysThinking) {
+    // Officially recommended sampling parameters for GLM-5.3-Flash
+    // (https://docs.bigmodel.cn/cn/guide/models/vlm/glm-5.3-flash), and the
+    // values its vision integration was field-tested with.
+    commonOverrideConfig.temperature = userConfig.temperature ?? 1;
+    commonOverrideConfig.top_p = 0.95;
+  } else if (userConfig.temperature !== undefined) {
     commonOverrideConfig.temperature = userConfig.temperature;
   }
 
@@ -27,7 +44,21 @@ const buildGlmChatCompletionParams = (
 
   const modelSpecificConfig: Record<string, unknown> = {};
 
-  if (reasoningEnabled !== 'default') {
+  if (alwaysThinking) {
+    if (reasoningEnabled !== 'default') {
+      modelSpecificConfig.thinking = {
+        type: 'enabled',
+        clear_thinking: false,
+      };
+      // Thinking cannot be disabled, so a "no reasoning" intent maps to the
+      // cheapest effort level. `low` was sufficient for the KSL-49 vision
+      // probes; users can raise it via the reasoning-effort config.
+      modelSpecificConfig.reasoning_effort =
+        userConfig.reasoningEffort ?? 'low';
+    } else if (userConfig.reasoningEffort) {
+      modelSpecificConfig.reasoning_effort = userConfig.reasoningEffort;
+    }
+  } else if (reasoningEnabled !== 'default') {
     modelSpecificConfig.thinking = {
       type: (reasoningEnabled ?? false) ? 'enabled' : 'disabled',
     };
@@ -45,6 +76,8 @@ const buildGlmChatCompletionParams = (
 export const glmAdapters = {
   'glm-v': {
     chatCompletion: {
+      // reasoningEffort is honored only by the always-thinking GLM models
+      // (see above); reasoningBudget is not supported by the family at all.
       unsupportedUserConfig: ['reasoningEffort', 'reasoningBudget'],
       buildChatCompletionParams: buildGlmChatCompletionParams,
       useReasoningAsContentFallback: true,

packages/core/src/ai-model/service-caller/index.ts

@@ -403,6 +403,7 @@ export async function callAI(
     : undefined;
   const chatCompletionInput = {
     intent: modelConfig.intent,
+    modelName: modelConfig.modelName,
     userConfig: {
       temperature: modelConfig.temperature,
       reasoningEnabled: modelConfig.reasoningEnabled,

packages/core/tests/unit-test/model-adapter/glm.test.ts

@@ -110,3 +110,94 @@ describe('glm model adapter', () => {
     expect(result.config.response_format).toBeUndefined();
   });
 });
+
+describe('glm-5.3-flash always-thinking contract', () => {
+  it('never sends thinking.type=disabled for glm-5.3-flash', () => {
+    const result = glmAdapter.chatCompletion.buildChatCompletionParams({
+      modelName: 'glm-5.3-flash',
+      userConfig: {},
+    });
+    expect(result.config).toEqual({
+      temperature: 1,
+      top_p: 0.95,
+      thinking: { type: 'enabled', clear_thinking: false },
+      reasoning_effort: 'low',
+    });
+  });
+
+  it('maps a reasoning-disable intent to effort=low for glm-5.3-flash', () => {
+    const result = glmAdapter.chatCompletion.buildChatCompletionParams({
+      modelName: 'glm-5.3-flash',
+      userConfig: { reasoningEnabled: false },
+    });
+    expect(result.config).toEqual({
+      temperature: 1,
+      top_p: 0.95,
+      thinking: { type: 'enabled', clear_thinking: false },
+      reasoning_effort: 'low',
+    });
+  });
+
+  it('forwards an explicit reasoning effort for glm-5.3-flash', () => {
+    const result = glmAdapter.chatCompletion.buildChatCompletionParams({
+      modelName: 'glm-5.3-flash',
+      userConfig: { reasoningEnabled: true, reasoningEffort: 'high' },
+    });
+    expect(result.config).toEqual({
+      temperature: 1,
+      top_p: 0.95,
+      thinking: { type: 'enabled', clear_thinking: false },
+      reasoning_effort: 'high',
+    });
+  });
+
+  it('follows provider default for glm-5.3-flash when reasoningEnabled=default', () => {
+    const result = glmAdapter.chatCompletion.buildChatCompletionParams({
+      modelName: 'glm-5.3-flash',
+      userConfig: { reasoningEnabled: 'default' },
+    });
+    expect(result.config).toEqual({
+      temperature: 1,
+      top_p: 0.95,
+    });
+  });
+
+  it('keeps an explicit user temperature for glm-5.3-flash', () => {
+    const result = glmAdapter.chatCompletion.buildChatCompletionParams({
+      modelName: 'glm-5.3-flash',
+      userConfig: { temperature: 0 },
+    });
+    expect(result.config).toEqual({
+      temperature: 0,
+      top_p: 0.95,
+      thinking: { type: 'enabled', clear_thinking: false },
+      reasoning_effort: 'low',
+    });
+  });
+
+  it('combines json_object response format with always-thinking params', () => {
+    const result = glmAdapter.chatCompletion.buildChatCompletionParams({
+      modelName: 'glm-5.3-flash',
+      expectedJsonObjectResponse: true,
+      userConfig: {},
+    });
+    expect(result.config).toEqual({
+      temperature: 1,
+      top_p: 0.95,
+      response_format: { type: 'json_object' },
+      thinking: { type: 'enabled', clear_thinking: false },
+      reasoning_effort: 'low',
+    });
+  });
+
+  it('keeps toggleable thinking for non-flash glm-v models', () => {
+    const result = glmAdapter.chatCompletion.buildChatCompletionParams({
+      modelName: 'glm-5v-turbo',
+      userConfig: {},
+    });
+    expect(result.config).toEqual({
+      temperature: 0,
+      thinking: { type: 'disabled' },
+    });
+  });
+});