DAG 工作流画布升级：节点可拖动 + 滚轮缩放 + 画布平移

- 节点拖拽调整布局（位移按缩放比例换算，拖动不误触详情点击）
- 滚轮以鼠标为锚点缩放（0.25x-3x），双击空白一键复位
- 拖拽空白区域平移画布（cursor grab/grabbing）
- 布局与视图状态 localStorage 按项目记忆，刷新/重开自动恢复
- 实测：节点精确拖动、锚点缩放、平移、复位、点击详情全部通过

agents.py

@@ -92,6 +92,23 @@ def _extract_score(text):
     return 60.0, text
 
 
+def _chat_json(worker, messages, max_tokens=2000, temperature=None):
+    """调用 LLM 并解析 JSON；解析失败自动加大 max_tokens 重试（防止长文截断）。
+    返回 (data, raw_text, usage)；全部失败抛最后异常。"""
+    last_err = None
+    for attempt in range(3):
+        mt = max_tokens * (attempt + 1)  # 2000 → 4000 → 6000
+        text, usage = _chat_worker(worker, messages, temperature=temperature, max_tokens=mt)
+        try:
+            data = _extract_json(text)
+            if isinstance(data, dict) and not data:
+                raise ValueError('空 JSON')
+            return data, text, usage
+        except Exception as e:
+            last_err = e
+    raise last_err or ValueError('JSON 解析失败')
+
+
 # ---------------------------------------------------------------------------
 # 主管模式
 # ---------------------------------------------------------------------------
@@ -114,17 +131,17 @@ SUPERVISOR_SYNTH_PROMPT = (
 def run_supervisor(run_id, run, workers):
     _log(run_id, 'supervisor', None, 'plan',
          f'主管开始规划：{run["topic"][:200]}')
-    # 1) 主管拆解
-    plan_text, u1 = _chat_worker(
+    # 1) 主管拆解（JSON 解析失败自动加大 max_tokens 重试）
+    try:
+        data, _, u1 = _chat_json(
             workers[0],
             [{'role': 'system', 'content': '你只输出 JSON。'},
              {'role': 'user', 'content': SUPERVISOR_PLAN_PROMPT.format(
                  topic=run['topic'], context=run['context'] or '无')}],
-        temperature=0.3, max_tokens=2000)
-    try:
-        subtasks = _extract_json(plan_text)
-        if isinstance(subtasks, dict):
-            subtasks = subtasks.get('subtasks') or subtasks.get('tasks') or []
+            max_tokens=3000, temperature=0.3)
+        subtasks = data.get('subtasks') or data.get('tasks') or []
+        if isinstance(data, list):
+            subtasks = data
     except Exception as e:
         _update_run(run_id, status='failed', error=f'主管规划解析失败: {e}')
         _log(run_id, 'supervisor', None, 'plan', f'❌ 规划解析失败: {e}')
@@ -216,15 +233,19 @@ def run_review(run_id, run, workers):
         # 评审
         review_usage = None
         try:
-            review_text, u2 = _chat_worker(
+            review_data, review_text, u2 = _chat_json(
                 reviewer_w,
                 [{'role': 'system', 'content': '你只输出 JSON。'},
                  {'role': 'user', 'content': REVIEWER_PROMPT.format(
                      topic=run['topic'], output=final_output, rubric=rubric)}],
-                temperature=0.2, max_tokens=1500)
-            score, judgment = _extract_score(review_text)
+                max_tokens=2000, temperature=0.2)
             review_usage = u2
             total_tokens += u2['total_tokens']; total_cost += u2['cost']
+            if isinstance(review_data, dict):
+                score = float(review_data.get('score', review_data.get('总分', 60)))
+                judgment = review_data.get('judgment') or review_data.get('意见') or review_text
+            else:
+                score, judgment = _extract_score(review_text)
         except Exception as e:
             score, judgment = 0, f'评审调用失败: {e}'
         final_score = score
@@ -325,14 +346,13 @@ def run_debate(run_id, run, workers):
                 f"【{v['stance']}】(#{k})：{v['view'][:500]}"
                 for k, v in views.items() if k != w['id'])
             try:
-                text, u = _chat_worker(
+                data, _, u = _chat_json(
                     w,
                     [{'role': 'system', 'content': '你只输出 JSON。'},
                      {'role': 'user', 'content': DEBATE_REBUT_PROMPT.format(
                          stance=mine['stance'], topic=run['topic'],
                          my_view=mine['view'], others=others)}],
-                    temperature=0.7, max_tokens=800)
-                data = _extract_json(text)
+                    max_tokens=1200, temperature=0.7)
                 rebut = data.get('rebuttal') if isinstance(data, dict) else text
                 mine['view'] = mine['view'] + '\n\n【质询回应】' + str(rebut)
                 total_tokens += u['total_tokens']; total_cost += u['cost']
@@ -346,12 +366,11 @@ def run_debate(run_id, run, workers):
         f"【{v['stance']}】{v['view']}" for v in views.values())
     _log(run_id, 'judge', judge_w['id'], 'verdict', '裁判综合裁决中…')
     try:
-        verdict, u4 = _chat_worker(
+        data, verdict, u4 = _chat_json(
             judge_w,
             [{'role': 'system', 'content': '你只输出 JSON。你是公正严明的首席裁判。'},
              {'role': 'user', 'content': JUDGE_PROMPT.format(topic=run['topic'], transcript=transcript)}],
-            temperature=0.3, max_tokens=1500)
-        data = _extract_json(verdict)
+            max_tokens=2500, temperature=0.3)
         if isinstance(data, dict):
             consensus = data.get('consensus') or data.get('结论') or verdict
             summary = data.get('summary') or data.get('摘要') or ''

app.py

@@ -263,6 +263,35 @@ def worker_test(wid):
         return jsonify({'ok': False, 'error': str(e)})
 
 
+@app.route('/api/workers/<int:wid>/vision_test', methods=['POST'])
+@require_auth
+def worker_vision_test(wid):
+    """视觉智能体能力测试：传入图片（URL 或 base64）与问题，验证多模态理解"""
+    w = db.q('SELECT * FROM workers WHERE id=?', (wid,), one=True)
+    if not w:
+        return jsonify({'ok': False, 'error': '不存在'}), 404
+    d = request.get_json(force=True)
+    image_url = (d.get('image_url') or '').strip()
+    image_b64 = (d.get('image_base64') or '').strip()
+    question = (d.get('question') or '请描述这张图片的内容').strip()
+    if not image_url and not image_b64:
+        return jsonify({'ok': False, 'error': '请提供图片 URL 或 base64 数据'}), 400
+    try:
+        if image_b64:
+            r = llm_gateway.chat_vision(
+                w['provider'], w['model'], question, image_url=f'data:image/png;base64,{image_b64}',
+                temperature=0.3, max_tokens=w['max_tokens'] or 2000,
+                base_url=w['base_url'] or None, api_key=w['api_key'] or None)
+        else:
+            r = llm_gateway.chat_vision(
+                w['provider'], w['model'], question, image_url=image_url,
+                temperature=0.3, max_tokens=w['max_tokens'] or 2000,
+                base_url=w['base_url'] or None, api_key=w['api_key'] or None)
+        return jsonify({'ok': True, 'data': r})
+    except Exception as e:
+        return jsonify({'ok': False, 'error': str(e)})
+
+
 @app.route('/api/providers')
 @require_auth
 def providers():
@@ -273,7 +302,7 @@ def providers():
             'has_key': bool(v['api_key']),
             'models': [m for m in config.MODEL_PRICING.keys() if m.startswith(
                 {'doubao': 'doubao', 'deepseek': 'deepseek', 'openai': 'gpt',
-                 'qwen': 'qwen', 'vllm': ''}.get(k, '__none__'))],
+                 'qwen': 'qwen', 'vllm': '', 'autodl': 'qwen3'}.get(k, '__none__'))],
         })
     return jsonify({'ok': True, 'data': data})
 
@@ -1119,8 +1148,8 @@ def template_apply(tid):
             pid, ids = tplmod.apply_project_template(t, variables, worker_id)
             return jsonify({'ok': True, 'kind': 'project', 'id': pid, 'task_ids': ids})
         if t['type'] == 'team':
-            ids = tplmod.apply_team_template(t, variables, provider=d.get('provider', 'doubao'),
-                                             model=d.get('model', 'doubao-seed-evolving'))
+            ids = tplmod.apply_team_template(t, variables, provider=d.get('provider', 'deepseek'),
+                                             model=d.get('model', 'deepseek-v4-flash'))
             return jsonify({'ok': True, 'kind': 'team', 'ids': ids})
         return jsonify({'ok': False, 'error': '未知模板类型'}), 400
     except Exception as e:

config.py

@@ -33,8 +33,14 @@ PROVIDERS = {
     },
     'deepseek': {
         'name': 'DeepSeek',
-        'base_url': 'https://api.deepseek.com/v1',
-        'api_key': os.environ.get('DEEPSEEK_API_KEY', ''),
+        'base_url': 'https://api.deepseek.com',
+        'api_key': os.environ.get('DEEPSEEK_API_KEY', 'sk-edb9df58ff574f8c98df1cd6a425e97c'),
+        'timeout': 600,
+    },
+    'autodl': {
+        'name': 'AutoDL 多模态（视觉）',
+        'base_url': 'https://www.autodl.art/api/v1',
+        'api_key': os.environ.get('AUTODL_API_KEY', 'F9MBfolzuapqTsD4KmUf9qen720rXvUZ3Sp3IrWiCTukqonx'),
         'timeout': 600,
     },
     'openai': {
@@ -66,10 +72,12 @@ MODEL_PRICING = {
     'doubao-pro-32k':            {'input': 2.0,   'output': 8.0},
     'deepseek-chat':             {'input': 2.0,   'output': 8.0},
     'deepseek-reasoner':         {'input': 4.0,   'output': 16.0},
+    'deepseek-v4-flash':         {'input': 0.2,   'output': 0.6},
     'gpt-4o':                    {'input': 17.5,  'output': 70.0},
     'gpt-4o-mini':               {'input': 1.1,   'output': 4.4},
     'qwen-plus':                 {'input': 0.8,   'output': 2.0},
     'qwen-max':                  {'input': 4.0,   'output': 12.0},
+    'qwen3.6-plus':              {'input': 2.0,   'output': 8.0},
 }
 DEFAULT_PRICE = {'input': 2.0, 'output': 8.0}
 

engine.py

@@ -123,7 +123,9 @@ def _budget_alert(project_id):
 
 
 def _build_messages(task, worker):
-    """构造提示词：任务指令 + RAG 知识库上下文"""
+    """构造提示词：任务指令 + RAG 知识库上下文
+    支持图片注入：任务描述中的 ![说明](图片URL) 或 图片:URL 会转成多模态消息（视觉 Worker）。"""
+    import re as _re
     messages = []
     if worker['system_prompt']:
         messages.append({'role': 'system', 'content': worker['system_prompt']})
@@ -132,6 +134,19 @@ def _build_messages(task, worker):
     if ctx:
         user_text = f'{ctx}\n\n----\n\n任务指令：{user_text}'
         _log(task['id'], 'info', f'📚 RAG 知识库命中 {len(refs)} 个片段：' + '；'.join(refs[:5]))
+    # 提取图片 URL（![alt](url) 或 图片:url 或 image:url）
+    img_urls = []
+    for m in _re.finditer(r'!\[[^\]]*\]\(([^)\s]+)\)', user_text):
+        img_urls.append(m.group(1))
+    for m in _re.finditer(r'(?:图片|image)\s*[:：]\s*(https?://\S+)', user_text, _re.I):
+        img_urls.append(m.group(1))
+    if img_urls:
+        content = [{'type': 'text', 'text': user_text}]
+        for u in img_urls:
+            content.append({'type': 'image_url', 'image_url': {'url': u}})
+        messages.append({'role': 'user', 'content': content})
+        _log(task['id'], 'info', f'🖼️ 检测到 {len(img_urls)} 张图片，已注入多模态消息')
+    else:
         messages.append({'role': 'user', 'content': user_text})
     return messages
 

llm_gateway.py

@@ -30,7 +30,12 @@ def calc_cost(model, prompt_tokens, completion_tokens):
 
 def chat(provider, model, messages, temperature=0.7, max_tokens=None,
          base_url=None, api_key=None, timeout=None, retries=None):
-    """调用 OpenAI 兼容 chat/completions，返回 {text, usage, cost, model}"""
+    """调用 OpenAI 兼容 chat/completions，返回 {text, usage, cost, model}
+    messages 支持两种格式：
+      - 纯文本：[{'role':'user','content':'...'}]
+      - 多模态：[{'role':'user','content':[{'type':'text','text':'...'},
+                                        {'type':'image_url','image_url':{'url':'...'}}]}]
+    """
     cfg = get_provider_cfg(provider)
     url = (base_url or cfg['base_url']).rstrip('/') + '/chat/completions'
     key = api_key or cfg['api_key']
@@ -56,7 +61,14 @@ def chat(provider, model, messages, temperature=0.7, max_tokens=None,
             resp = requests.post(url, json=payload, headers=headers, timeout=timeout)
             if resp.status_code == 200:
                 data = resp.json()
-                text = data['choices'][0]['message']['content'] or ''
+                msg = data['choices'][0]['message']
+                text = msg.get('content') or ''
+                if not text:
+                    # 推理模型偶发 content 为空：用 reasoning_content 兜底
+                    text = msg.get('reasoning_content') or ''
+                if not text:
+                    last_err = LLMError('模型返回空内容，重试中…')
+                    continue
                 usage = data.get('usage', {})
                 pt = usage.get('prompt_tokens', 0)
                 ct = usage.get('completion_tokens', 0)
@@ -84,6 +96,48 @@ def chat(provider, model, messages, temperature=0.7, max_tokens=None,
     raise last_err or LLMError('未知错误')
 
 
+def chat_vision(provider, model, text, image_url=None, image_path=None,
+                temperature=0.4, max_tokens=2000, base_url=None, api_key=None,
+                retries=3):
+    """多模态视觉调用：文本 + 图片（URL 或本地路径/base64）。
+    返回与 chat() 相同结构。
+    容错：聚合 API 偶发路由到纯文本后端（不认识 image_url），自动重试。"""
+    import base64 as _b64
+    import time as _time
+    content = [{'type': 'text', 'text': text}]
+    img_url = image_url
+    if image_path:
+        with open(image_path, 'rb') as f:
+            raw = f.read()
+        mime = 'image/png'
+        if image_path.lower().endswith(('.jpg', '.jpeg')):
+            mime = 'image/jpeg'
+        elif image_path.lower().endswith('.gif'):
+            mime = 'image/gif'
+        elif image_path.lower().endswith('.webp'):
+            mime = 'image/webp'
+        img_url = f'data:{mime};base64,{_b64.b64encode(raw).decode()}'
+    if img_url:
+        content.append({'type': 'image_url', 'image_url': {'url': img_url}})
+    messages = [{'role': 'user', 'content': content}]
+    last_err = None
+    for attempt in range(max(1, retries)):
+        try:
+            return chat(provider, model, messages,
+                        temperature=temperature, max_tokens=max_tokens,
+                        base_url=base_url, api_key=api_key)
+        except LLMError as e:
+            last_err = e
+            msg = str(e)
+            # 仅对“多模态格式不被支持/图片无效”类错误重试（聚合后端路由问题）
+            if any(k in msg for k in ('image_url', 'InvalidParameter', 'invalid_parameter',
+                                      'does not appear to be valid', 'image')):
+                _time.sleep(2 * (attempt + 1))
+                continue
+            raise
+    raise last_err or LLMError('视觉调用失败')
+
+
 def test_connection(provider, model, base_url=None, api_key=None):
     """连通性测试：发一条最小请求"""
     t0 = time.time()

static/app.js

@@ -287,14 +287,10 @@ async function moveTask(tid, status) {
   toast('状态已更新', 'ok'); refreshProj();
 }
 
-/* --- DAG 工作流 --- */
-function renderDag() {
-  const {tasks} = projCtx;
-  const idMap = {};
-  tasks.forEach(t => { idMap[t.id] = t; });
-  const edges = [];
-  tasks.forEach(t => (t.depends_on || []).forEach(d => { if (idMap[d]) edges.push({from:d, to:t.id}); }));
+/* --- DAG 工作流（可交互画布：拖节点 / 滚轮缩放 / 拖画布平移） --- */
+let dagState = null;   // {pid, pos:{id:{x,y}}, scale, tx, ty, drag:null, moved:false}
 
+function dagLayout(tasks, idMap) {
   // 拓扑分层：最长路径
   const layer = {};
   const visit = id => {
@@ -308,7 +304,6 @@ function renderDag() {
   tasks.forEach(t => visit(t.id));
   const layers = {};
   tasks.forEach(t => { (layers[layer[t.id]] = layers[layer[t.id]] || []).push(t.id); });
-
   const W = 260, H = 110, PAD = 60;
   const width = Math.max(2, Object.keys(layers).length) * W + PAD;
   const height = Math.max(...tasks.map(t => layer[t.id] === undefined ? 0 : layer[t.id]), 0) * H + PAD + 60;
@@ -318,33 +313,222 @@ function renderDag() {
     const step = Math.min(H - 60, Math.max(70, (height - 2 * PAD) / (ids.length + 1)));
     ids.forEach((id, i) => { pos[id] = {x, y: PAD + (i + 1) * step}; });
   });
+  return {width, height, pos};
+}
+
+function dagSize() {
+  const {tasks} = projCtx;
+  const idMap = {};
+  tasks.forEach(t => { idMap[t.id] = t; });
+  return dagLayout(tasks, idMap);
+}
+
+function saveDagState() {
+  if (!dagState) return;
+  try {
+    const pos = {};
+    Object.entries(dagState.pos).forEach(([k, v]) => { pos[k] = {x: Math.round(v.x), y: Math.round(v.y)}; });
+    localStorage.setItem('aw_dag_pos_' + dagState.pid, JSON.stringify(pos));
+    localStorage.setItem('aw_dag_view_' + dagState.pid,
+      JSON.stringify({scale: dagState.scale, tx: Math.round(dagState.tx), ty: Math.round(dagState.ty)}));
+  } catch (e) {}
+}
+
+function dagApplyView() {
+  const s = dagState;
+  const svg = $('#dag-svg');
+  if (!svg) return;
+  svg.style.transform = `translate(${s.tx}px, ${s.ty}px) scale(${s.scale})`;
+  svg.style.transformOrigin = '0 0';
+}
+
+function drawDag() {
+  const {tasks} = projCtx;
+  const idMap = {};
+  tasks.forEach(t => { idMap[t.id] = t; });
+  const edges = [];
+  tasks.forEach(t => (t.depends_on || []).forEach(d => { if (idMap[d]) edges.push({from: d, to: t.id}); }));
+  const {width, height} = dagSize();
   const colors = {todo:'#2a3550', running:'#173b6b', review:'#4a3a10', done:'#0f3d2e', rejected:'#4a1c1c', failed:'#3d1c2e', cancelled:'#333'};
   const ready = tasks.filter(t => (t.status === 'todo' || t.status === 'failed') && (!t.depends_on || t.depends_on.every(d => idMap[d]?.status === 'done')));
-  $('#tab-body').innerHTML = `
+  const body = $('#tab-body');
+  if (!body) return;
+  body.innerHTML = `
     <div class="toolbar" style="margin-bottom:10px">
-      <span class="pager-tip">共 ${tasks.length} 个节点 · ${edges.length} 条依赖边${ready.length ? ` · ${ready.length} 个任务就绪` : ''}</span>
+      <span class="pager-tip">共 ${tasks.length} 个节点 · ${edges.length} 条依赖边${ready.length ? ` · ${ready.length} 个任务就绪` : ''} · 拖节点调位 / 滚轮缩放 / 拖空白平移 / 双击复位</span>
       <button class="btn success" ${ready.length ? '' : 'disabled'} onclick="runWorkflow()">▶ 执行整个工作流（${ready.length}）</button>
     </div>
-    <div class="dag-box"><svg width="${width}" height="${height}" viewBox="0 0 ${width} ${height}">
+    <div class="dag-box" id="dag-box">
+      <svg id="dag-svg" width="100%" height="${Math.max(480, height)}" viewBox="0 0 ${width} ${height}" style="transform-origin:0 0">
         <defs><marker id="arrow" markerWidth="8" markerHeight="8" refX="7" refY="3" orient="auto"><path d="M0,0 L8,3 L0,6 Z" fill="#4a5878"/></marker></defs>
+        <rect id="dag-bg" x="-2000" y="-2000" width="${width + 4000}" height="${height + 4000}" fill="transparent"/>
         ${edges.map(e => {
-        const a = pos[e.from], b = pos[e.to];
+          const a = dagState.pos[e.from], b = dagState.pos[e.to];
           if (!a || !b) return '';
           const mx = (a.x + b.x) / 2;
           const act = idMap[e.from].status === 'done';
-        return `<path class="dag-edge${act ? ' act' : ''}" d="M${a.x+55},${a.y} C${mx},${a.y} ${mx},${b.y} ${b.x-55},${b.y}"/>`;
+          return `<path data-e="${e.from}-${e.to}" class="dag-edge${act ? ' act' : ''}" d="M${a.x + 55},${a.y} C${mx},${a.y} ${mx},${b.y} ${b.x - 55},${b.y}"/>`;
         }).join('')}
         ${tasks.map(t => {
-        const p = pos[t.id];
+          const p = dagState.pos[t.id];
           const col = colors[t.status];
-        return `<g>
-          <rect class="dag-node" x="${p.x-55}" y="${p.y-24}" width="110" height="48" fill="${col}" stroke="#3a4a6a" onclick="openTaskDetail(${t.id})"/>
+          return `<g class="dag-node" data-id="${t.id}" style="cursor:grab">
+            <rect x="${p.x - 55}" y="${p.y - 24}" width="110" height="48" fill="${col}" stroke="#3a4a6a"/>
             <text class="dag-label" x="${p.x}" y="${p.y - 2}" fill="#fff">#${t.id} ${esc(t.title).slice(0, 10)}</text>
             <text class="dag-sub" x="${p.x}" y="${p.y + 15}">${STATUS[t.status].label}${t.depends_on?.length ? ' · 前置' + t.depends_on.length : ''}</text>
           </g>`;
         }).join('')}
-    </svg></div>
-    <div class="tab-note" style="margin-top:10px">💡 节点点击查看详情；前置任务完成后下游自动触发。执行整个工作流 = 派发所有就绪任务（含失败重试）。</div>`;
+      </svg>
+    </div>
+    <div class="tab-note" style="margin-top:10px">💡 节点点击查看详情；拖动节点可自定义布局（自动记忆）；前置任务完成后下游自动触发。执行整个工作流 = 派发所有就绪任务（含失败重试）。</div>`;
+  dagApplyView();
+  bindDagEvents();
+}
+
+function bindDagEvents() {
+  const box = $('#dag-box');
+  const svg = $('#dag-svg');
+  if (!box || !svg) return;
+  const s = dagState;
+  let down = null;   // {type:'node'|'pan', id?, sx, sy, ox, oy}
+
+  box.addEventListener('mousedown', e => {
+    if (e.button !== 0) return;
+    const nodeEl = e.target.closest('.dag-node');
+    if (nodeEl) {
+      const id = Number(nodeEl.dataset.id);
+      const p = s.pos[id];
+      down = {type: 'node', id, sx: e.clientX, sy: e.clientY, ox: p.x, oy: p.y, moved: false};
+      nodeEl.style.cursor = 'grabbing';
+      svg.style.cursor = 'grabbing';
+    } else {
+      down = {type: 'pan', sx: e.clientX, sy: e.clientY, ox: s.tx, oy: s.ty, moved: false};
+      svg.style.cursor = 'grabbing';
+    }
+    e.preventDefault();
+  });
+
+  window.addEventListener('mousemove', e => {
+    if (!down) return;
+    const dx = e.clientX - down.sx, dy = e.clientY - down.sy;
+    if (!down.moved && Math.abs(dx) + Math.abs(dy) < 3) return;
+    down.moved = true;
+    if (down.type === 'node') {
+      const p = s.pos[down.id];
+      if (p) { p.x = down.ox + dx / s.scale; p.y = down.oy + dy / s.scale; }
+    } else {
+      s.tx = down.ox + dx;
+      s.ty = down.oy + dy;
+    }
+    dagApplyView();
+    if (down.type === 'node') {
+      // 轻量更新：节点坐标 + 关联连线
+      const svgEl = $('#dag-svg');
+      if (!svgEl) return;
+      svgEl.querySelectorAll('.dag-node').forEach(g => {
+        const nid = Number(g.dataset.id);
+        const pp = s.pos[nid];
+        if (!pp) return;
+        g.querySelector('rect').setAttribute('x', pp.x - 55);
+        g.querySelector('rect').setAttribute('y', pp.y - 24);
+        const ts = g.querySelectorAll('text');
+        ts[0].setAttribute('x', pp.x); ts[0].setAttribute('y', pp.y - 2);
+        ts[1].setAttribute('x', pp.x); ts[1].setAttribute('y', pp.y + 15);
+      });
+      // 以拖动节点为中心更新相关连线
+      const touched = new Set([down.id]);
+      projCtx.tasks.forEach(t => (t.depends_on || []).forEach(d => {
+        if (d === down.id || t.id === down.id) { touched.add(d); touched.add(t.id); }
+      }));
+      touched.forEach(tid => {
+        const a = s.pos[tid];
+        if (!a) return;
+        projCtx.tasks.forEach(t => {
+          if (t.id !== tid) return;
+          (t.depends_on || []).forEach(d => {
+            const b = s.pos[d];
+            if (!b) return;
+            const path = svgEl.querySelector(`path[data-e="${d}-${tid}"]`);
+            if (!path) return;
+            const mx = (b.x + a.x) / 2;
+            path.setAttribute('d', `M${b.x + 55},${b.y} C${mx},${b.y} ${mx},${a.y} ${a.x - 55},${a.y}`);
+          });
+        });
+      });
+    }
+  });
+
+  window.addEventListener('mouseup', () => {
+    if (!down) return;
+    const wasNode = down.type === 'node';
+    if (wasNode) {
+      const nodeEl = svg.querySelector('.dag-node[data-id="' + down.id + '"]');
+      if (nodeEl) nodeEl.style.cursor = 'grab';
+    }
+    svg.style.cursor = '';
+    if (down.moved) {
+      dagState.suppressClick = true;
+      setTimeout(() => { if (dagState) dagState.suppressClick = false; }, 400);
+      saveDagState();
+    }
+    down = null;
+  });
+
+  // 滚轮缩放（以鼠标为锚点）
+  box.addEventListener('wheel', e => {
+    e.preventDefault();
+    const rect = box.getBoundingClientRect();
+    const mx = e.clientX - rect.left, my = e.clientY - rect.top;
+    const factor = e.deltaY < 0 ? 1.12 : 1 / 1.12;
+    const ns = Math.min(3, Math.max(0.25, s.scale * factor));
+    // 保持鼠标指向的内容坐标不动
+    const cx = (mx - s.tx) / s.scale, cy = (my - s.ty) / s.scale;
+    s.scale = ns;
+    s.tx = mx - cx * ns;
+    s.ty = my - cy * ns;
+    dagApplyView();
+    saveDagState();
+  }, {passive: false});
+
+  // 双击空白复位视图
+  box.addEventListener('dblclick', e => {
+    if (e.target.closest('.dag-node')) return;
+    s.scale = 1; s.tx = 0; s.ty = 0;
+    dagApplyView();
+    saveDagState();
+  });
+
+  // 点击节点打开详情（拖动超过阈值则不触发）
+  box.addEventListener('click', e => {
+    const nodeEl = e.target.closest('.dag-node');
+    if (!nodeEl) return;
+    if (dagState.suppressClick) { dagState.suppressClick = false; return; }
+    openTaskDetail(Number(nodeEl.dataset.id));
+  });
+}
+
+function renderDag() {
+  const {tasks, pid} = projCtx;
+  const idMap = {};
+  tasks.forEach(t => { idMap[t.id] = t; });
+  const {pos} = dagLayout(tasks, idMap);
+  // 恢复已保存布局 / 视图
+  let savedPos = null, savedView = null;
+  try {
+    savedPos = JSON.parse(localStorage.getItem('aw_dag_pos_' + pid) || 'null');
+    savedView = JSON.parse(localStorage.getItem('aw_dag_view_' + pid) || 'null');
+  } catch (e) {}
+  if (savedPos) Object.entries(savedPos).forEach(([k, v]) => { if (pos[k]) pos[k] = v; });
+  dagState = {
+    pid, pos,
+    scale: savedView?.scale || 1,
+    tx: savedView?.tx || 0,
+    ty: savedView?.ty || 0,
+    drag: null, moved: false,
+  };
+  const box = $('#tab-body');
+  box.innerHTML = '';
+  drawDag();
 }
 
 async function runWorkflow() {
@@ -623,6 +807,7 @@ async function pageWorkers() {
         <td class="mono">${w.task_cost_limit ? '单任务 ¥' + w.task_cost_limit : '—'}<br>${w.monthly_cost_limit ? '月 ¥' + w.monthly_cost_limit : ''}</td>
         <td><span class="badge ${w.status === 'enabled' ? 'st-done' : 'st-cancelled'}">${w.status === 'enabled' ? '启用' : '停用'}</span></td>
         <td><button class="btn sm" onclick="testWorker(${w.id})">测试</button>
+            <button class="btn sm" onclick="visionTest(${w.id})">🖼️ 视觉测试</button>
             <button class="btn sm" onclick="openWorkerModal(${JSON.stringify(w).replace(/"/g,'&quot;')}, ${JSON.stringify(ps.data).replace(/"/g,'&quot;')})">编辑</button>
             <button class="btn sm danger" onclick="delWorker(${w.id})">删除</button></td>
       </tr>`).join('') || '<tr><td colspan="9" class="empty">还没有 Worker，注册第一个虚拟员工吧</td></tr>'}
@@ -636,6 +821,27 @@ async function testWorker(wid) {
   catch (e) { toast('❌ ' + e.message, 'err'); }
 }
 
+async function visionTest(wid) {
+  openModal(`
+    <h3>🖼️ 视觉能力测试</h3>
+    <div class="tab-note">多模态智能体：输入图片 URL 与问题，验证图像理解能力。示例：<br>
+    https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/640px-Cat03.jpg</div>
+    <label>图片 URL</label><input id="vt-url" placeholder="https://...">
+    <label>问题</label><textarea id="vt-q" rows="2">请描述这张图片的内容，并给出专业分析。</textarea>
+    <button class="btn primary" id="vt-go">🚀 开始分析</button>
+    <div id="vt-out" style="margin-top:10px"></div>`);
+  $('#vt-go').addEventListener('click', async () => {
+    const url = $('#vt-url').value.trim();
+    if (!url) return toast('请填写图片 URL', 'err');
+    $('#vt-out').innerHTML = '<div class="tab-note"><span class="spin"></span> 视觉分析中（多模态调用约 10-60 秒）…</div>';
+    try {
+      const r = await api(`/api/workers/${wid}/vision_test`, {method:'POST', body: {image_url: url, question: $('#vt-q').value}});
+      $('#vt-out').innerHTML = `<pre style="white-space:pre-wrap;background:#f6f8fa;padding:10px;border-radius:8px;max-height:320px;overflow:auto">${esc(r.data.text)}</pre>
+        <div class="tab-note">tokens ${r.data.total_tokens} · 成本 ${fmtMoney(r.data.cost)} · 延迟 ${r.data.latency_ms ?? '—'}ms</div>`;
+    } catch (e) { $('#vt-out').innerHTML = `<div class="err">${esc(e.message)}</div>`; }
+  });
+}
+
 async function delWorker(wid) {
   if (!confirm('确认删除该 Worker？（历史任务将解除指派）')) return;
   await api(`/api/workers/${wid}`, {method:'DELETE'});

static/style.css

@@ -131,7 +131,8 @@ tr:hover td{background:var(--panel2)}
 .tabs{display:flex;gap:4px;margin-bottom:16px;border-bottom:1px solid var(--border)}
 .tabs a{padding:9px 16px;color:var(--muted);text-decoration:none;border-bottom:2px solid transparent;font-size:13px}
 .tabs a.active{color:var(--accent);border-bottom-color:var(--accent)}
-.dag-box{background:var(--panel);border:1px solid var(--border);border-radius:12px;padding:14px;overflow:auto}
+.dag-box{background:var(--panel);border:1px solid var(--border);border-radius:12px;overflow:hidden;height:520px;position:relative;touch-action:none;user-select:none;cursor:grab}
+.dag-box svg{display:block;transition:transform .05s linear}
 .dag-node{fill:var(--panel2);stroke:var(--border);stroke-width:1.5;rx:10;cursor:pointer}
 .dag-node:hover{stroke:var(--accent)}
 .dag-label{font-size:12px;fill:var(--text);text-anchor:middle}

templates.py

@@ -180,7 +180,7 @@ def apply_project_template(tpl, variables, worker_id=None):
     return pid, created
 
 
-def apply_team_template(tpl, variables, provider='doubao', model='doubao-seed-evolving'):
+def apply_team_template(tpl, variables, provider='deepseek', model='deepseek-v4-flash'):
     """应用团队模板 → 批量注册 Worker"""
     content = render(json.loads(tpl['content']), variables)
     created = []