Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。
TypeScript agentai-agentai-assistantandroidclaude-codecodexelectroniosllmmacosreact-nativetypescriptwindows
/**
* visionCapability 三态判定测试。
*
* 覆盖:deepseek 系列 → no-vision(含带命名空间 / 裸 id / [1m] 后缀各种形态);
* claude/gpt/gemini → vision;未知 → unknown;codex/ 前缀归一化。
*/
import { describe, expect, it } from 'vitest';
import {
classifyVisionCapability,
isKnownNoVisionModel,
normalizeVisionModelId,
} from '../visionCapability.js';
describe('normalizeVisionModelId', () => {
it('strips [1m] suffix', () => {
expect(normalizeVisionModelId('deepseek/deepseek-v4-flash[1m]')).toBe(
'deepseek/deepseek-v4-flash',
);
});
it('strips codex/ prefix', () => {
expect(normalizeVisionModelId('codex/gpt-5.5')).toBe('gpt-5.5');
expect(normalizeVisionModelId('openai-codex/gpt-6.1-sol')).toBe('gpt-6.1-sol');
});
});
describe('classifyVisionCapability', () => {
it('classifies deepseek v4 series as no-vision (namespace + bare + suffix)', () => {
expect(classifyVisionCapability('deepseek/deepseek-v4-pro')).toBe('no-vision');
expect(classifyVisionCapability('deepseek/deepseek-v4-flash')).toBe('no-vision');
expect(classifyVisionCapability('deepseek/deepseek-v4-flash[1m]')).toBe('no-vision');
// 裸 id(部分 runtime body.model 形态)
expect(classifyVisionCapability('deepseek-v4-flash')).toBe('no-vision');
});
it('classifies known vision models', () => {
expect(classifyVisionCapability('anthropic/claude-opus-4-8')).toBe('vision');
expect(classifyVisionCapability('openai/gpt-5.5')).toBe('vision');
expect(classifyVisionCapability('google/gemini-3.5-flash')).toBe('vision');
expect(classifyVisionCapability('xai/grok-4.5')).toBe('vision');
});
it('classifies glm-5.2 as no-vision (namespace + bare + suffix)', () => {
expect(classifyVisionCapability('z-ai/glm-5.2')).toBe('no-vision');
// 裸 id / 带 [1m] 后缀的 runtime body.model 形态(normalize 已剥 [1m])。
expect(classifyVisionCapability('glm-5.2')).toBe('no-vision');
expect(classifyVisionCapability('glm-5.2[1m]')).toBe('no-vision');
});
/**
* 回归(2026-09-04):家族级 no-vision 前缀曾把同家族的**视觉变体**一并判成不能看图 ——
* `deepseek-` / `glm-5.2` 是家族前缀,而 `deepseek-vl2`、`deepseek-v4-flash-vision-exp`、
* `glm-5.2-vision` 都以它们开头。当时四个型号全部误判,后果不只是设置页标错:
* isKnownNoVisionModel 是视觉桥「未自定义时默认勾选」的依据,误判会让本可直接吃图的
* 模型被迫走桥,图片降级成文字描述。
*/
it('lets an explicit vision marker override the family-level no-vision verdict', () => {
expect(classifyVisionCapability('deepseek/deepseek-vl2')).toBe('vision');
expect(classifyVisionCapability('deepseek/deepseek-v4-flash-vision-exp')).toBe('vision');
expect(classifyVisionCapability('deepseek-v4-flash-vision-exp')).toBe('vision');
expect(classifyVisionCapability('z-ai/glm-5.2-vision')).toBe('vision');
// 同家族的非视觉型号不受影响 —— 标记必须真的出现在 id 里。
expect(classifyVisionCapability('deepseek/deepseek-v4-flash')).toBe('no-vision');
expect(classifyVisionCapability('z-ai/glm-5.2')).toBe('no-vision');
// [1m] 后缀先被 normalize 剥掉,不影响标记匹配。
expect(classifyVisionCapability('deepseek/deepseek-vl2[1m]')).toBe('vision');
});
it('does not treat vision / vl inside a longer word as a marker', () => {
// 段边界约束:标记必须自成一段(前后是 ^ / - / / 或 . / 结尾),
// 否则「名字里恰好含这两个字母」的纯文本模型会被误判成多模态。
expect(classifyVisionCapability('deepseek/deepseek-vllm-runtime')).toBe('no-vision');
expect(classifyVisionCapability('foo/invision-model')).toBe('unknown');
expect(classifyVisionCapability('foo/bar-vlad')).toBe('unknown');
});
it('classifies unknown models as unknown', () => {
expect(classifyVisionCapability('foo/bar-model')).toBe('unknown');
expect(classifyVisionCapability('qwen/qwen3.7-max')).toBe('unknown');
});
it('isKnownNoVisionModel helper', () => {
expect(isKnownNoVisionModel('deepseek/deepseek-v4-flash')).toBe(true);
expect(isKnownNoVisionModel('anthropic/claude-opus-4-8')).toBe(false);
});
});