fix: 多Agent协作 JSON 截断导致主管规划解析失败

- 根因: deepseek-v4-flash 规划输出超 max_tokens=2000 被截断，JSON 不完整解析失败 - agents.py: 新增 _chat_json() 统一处理 JSON 类调用，解析失败自动加大 max_tokens 重试(×1/×2/×3) - 主管规划/辩论质询/辩论裁决/评审全部改用 _chat_json 并提高初始上限(3000/1200/2500/2000) - 实测: 主管(计算器网页版4子任务)✅ 评审(92分)✅ 辩论(指定阵容)✅
智能体大模型接口全面切换 + 新增视觉智能体
2026-08-12 17:08:16 +08:00 · 2026-08-12 13:22:56 +08:00
7 changed files with 176 additions and 29 deletions
@@ -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(
-        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)
+    # 1) 主管拆解（JSON 解析失败自动加大 max_tokens 重试）
    try:
-        subtasks = _extract_json(plan_text)
-        if isinstance(subtasks, dict):
-            subtasks = subtasks.get('subtasks') or subtasks.get('tasks') or []
+        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')}],
+            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 ''
@@ -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:
@@ -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}

@@ -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']})
@@ -132,7 +134,20 @@ 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]))
-    messages.append({'role': 'user', 'content': user_text})
+    # 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


@@ -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()
@@ -623,6 +623,7 @@ 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 +637,27 @@ 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'});
@@ -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 = []