V3.5.2 真流式对话+思考折叠/历史会话/能力标签/模型库/语音/知识库/按千次计费

1. 修复假流式：对话接口边收边吐(stream_chat 生成器直驱 _chat_stream_raw)；思考模型先流式思考内容(折叠)再流式回答；回答块显示 tok/s
2. 历史会话列表：重命名/置顶/删除(pinned/use_kb 列)；📝Markdown 一键开关，默认 Markdown 渲染
3. 模型能力标签(chat/thinking/vision/audio_in/audio_out/image_gen/video_gen/embedding/rerank)：接口库每模型勾选；
   对话按能力适配：视觉→传图、语音入→🎤录音(默认ASR转写)、语音出→🔊朗读(默认TTS)、思考→思考折叠；不匹配提前提示
4. 模型库标签：能力矩阵 + 系统默认模型(语音识别/合成/生图/生视频/embedding/rerank)
5. 计费修正：按次=元/千次；逐模型定价=模型名 输入价 缓存输入价 输出价(缓存价计缓存命中token)
6. 弹窗未保存提醒(点外部/X 时 confirm，全弹窗通用)
7. 知识库导航：#/kb 全局文档 CRUD + txt/md/pdf/docx 上传解析 + jieba BM25 检索；对话可一键注入知识库

README.md

@@ -3,7 +3,32 @@
 > 以「项目」为中心、以「AI Worker」为执行单元的项目管理平台。
 > 把大模型团队变成一支可指挥、可审计、可控成本的"虚拟团队"。
 
-**当前版本：V3.5**（精细化用量统计 / 大模型接口库 / AI Worker 团队 / 对话融合仪表盘 / 系统工作目录 / 流式单token超时）
+**当前版本：V3.5.2**（真流式对话+思考折叠 / 历史会话管理 / 模型能力标签 / 模型库+系统默认模型 / 语音输入输出 / 知识库导航 / 按千次计费）
+
+---
+
+## 🚀 V3.5.2 优化（本轮）
+
+| 能力 | 说明 |
+|---|---|
+| ⚡ 真流式对话 | 修复“一次性输出全部回答”的假流式 bug：服务端改为**边收边吐**（原先先等完整响应再一次性转发）；**思考模型先流式输出思考内容（自动折叠成「💭 思考过程」可展开）再流式输出回答**；回答块下方显示 **输出速度（tok/s）**、tokens/缓存命中/成本/首字延迟 |
+| 🕘 历史会话管理 | 对话页左侧新增「历史会话」列表：每个会话可 **✏️重命名 / 📌置顶 / 🗑删除**（置顶排前）；顶部「📝 Markdown」一键开关，消息默认按 Markdown 渲染 |
+| 🧠 模型能力标签 | 大模型接口的每个模型可勾选能力：对话/思考/视觉输入/语音输入/语音输出/图片生成/视频生成/Embedding/Rerank；对话按所选模型能力自动适配：**视觉模型→🖼️传图、语音输入→🎤录音转写、语音输出→🔊回答朗读、思考模型→展示思考过程**；不匹配时提前提示不可用 |
+| 🗂️ 模型库 + 系统默认模型 | AI Worker 新增「🧠 模型库」标签：全能力模型矩阵 + **系统默认模型配置**（语音识别/语音合成/图片生成/视频生成/embedding/rerank…），配置语音识别模型后对话即支持语音输入，配置语音合成模型后回答可生成语音 |
+| 💰 计费单位修正 | 按调用次数计费改为 **元/千次**（2 元/千次 = 单次 0.002 元）；逐模型定价格式改为 **模型名 输入价 缓存输入价 输出价**（缓存命中 token 按缓存价计） |
+| 🛡️ 弹窗未保存提醒 | 编辑弹窗有未保存修改时，点击弹窗外部或 ✕ 会提醒“是否不保存退出”（所有弹窗通用） |
+| 📚 知识库导航 | 侧边栏新增「📚 知识库」：全局文档增删改查、txt/md/pdf/docx 上传解析、全文检索；对话中可一键开启「📚 知识库」按钮，发送时自动注入相关知识 |
+
+---
+
+## 🚀 V3.5.1 优化（上一轮）
+
+| 能力 | 说明 |
+|---|---|
+| ⭐ 项目一键存为参考 | 项目卡片新增「⭐ 存为参考」按钮，把任意已创建项目一键保存到「从参考项目中新建」列表（复制项目+任务，原项目不受影响）；参考列表里显示「来自「原项目」」来源标记，非内置参考可删除，内置 3 个受保护 |
+| 🐛 修复标签切换 | AI Worker 页「大模型接口库 / 团队」标签点击无反应的 bug（onclick 里页面状态被重置）已修复，三标签正常切换 |
+| 💬 对话独立首页 | 左侧导航改为「💬 对话 / 📊 仪表盘」并列；首页为全新对话页：**上部=仪表盘关键信息摘要（项目/任务/Worker/成本/调用次数，点击即跳转仪表盘）**，中下部=主体对话界面（更高更宽，流式输出带闪烁光标，每条记录 tokens/缓存命中/成本/首字延迟） |
+| 📅 成本报表按日期 | 报表新增「按日期」维度：按天统计调用次数/输入/输出/缓存命中/总Tokens/成本，支持 近7天/30天/全部 筛选 |
 
 ---
 

app.py

@@ -7,6 +7,7 @@ import os
 import secrets
 import functools
 import time
+import base64 as _b64
 import json as _json
 from flask import Flask, request, jsonify, session, send_from_directory, redirect, Response
 
@@ -15,6 +16,7 @@ import db
 import engine
 import llm_gateway
 import rag
+import kb
 import notify
 import agents
 import eval as evalmod
@@ -25,6 +27,8 @@ import autostart
 
 app = Flask(__name__, static_folder='static', static_url_path='')
 app.secret_key = config.SECRET_KEY
+CHAT_UPLOAD_DIR = os.path.join(config.DATA_DIR, 'chat_uploads')
+os.makedirs(CHAT_UPLOAD_DIR, exist_ok=True)
 db.init_db()
 db.recover_stale_runs()
 evalmod.create_builtin_datasets()
@@ -232,16 +236,18 @@ def require_admin(fn):
 
 @app.route('/api/health')
 def health():
-    return jsonify({'ok': True, 'service': 'ai-worker-platform', 'version': 'v3.5.0'})
+    return jsonify({'ok': True, 'service': 'ai-worker-platform', 'version': 'v3.5.2'})
 
 
 # ---------------------------------------------------------------------------
 # V3.5 内置参考测试项目（从参考项目新建用，不进普通项目列表）
 # ---------------------------------------------------------------------------
 def seed_reference_projects():
-    """幂等：内置 3 个不同维度的简单参考测试项目，供「从参考项目中新建」快速复制。"""
-    if db.q('SELECT COUNT(*) c FROM projects WHERE is_reference=1')[0]['c'] > 0:
-        return
+    """幂等且自愈：内置 3 个不同维度的简单参考测试项目（ref_builtin=1，受保护），
+    供「从参考项目中新建」快速复制；历史数据自动补标记，缺失自动重建。"""
+    # 兼容历史：已有同名内置参考项目补 ref_builtin 标记
+    for name in ('产品文案速写（参考）', 'Python 小工具（参考）', '市场调研简报（参考）'):
+        db.w('UPDATE projects SET ref_builtin=1 WHERE name=? AND is_reference=1 AND ref_builtin=0', (name,))
     ts = db.now()
     refs = [
         {
@@ -287,10 +293,14 @@ def seed_reference_projects():
         },
     ]
     for ref in refs:
+        exist = db.q('SELECT id FROM projects WHERE name=? AND is_reference=1 AND ref_builtin=1',
+                     (ref['name'],), one=True)
+        if exist:
+            continue
         pid = db.w(
             'INSERT INTO projects (name, description, objective, acceptance_criteria, status, '
-            'budget_limit, deliver_type, workspace_dir, is_reference, auto_status, created_at, updated_at) '
-            'VALUES (?,?,?,?,?,?,?,?,1,?,?,?)',
+            'budget_limit, deliver_type, workspace_dir, is_reference, ref_builtin, auto_status, created_at, updated_at) '
+            'VALUES (?,?,?,?,?,?,?,?,1,1,?,?,?)',
             (ref['name'], '内置参考测试项目，可「从参考项目新建」快速复制', ref['objective'],
              ref['acceptance_criteria'], 'active', 0, 'web', '', 'none', ts, ts))
         db.w('UPDATE projects SET workspace_dir=? WHERE id=?', (f'ref_{pid}', pid))
@@ -860,12 +870,13 @@ def endpoints_api():
             return jsonify({'ok': False, 'error': 'Base URL 必填'}), 400
         models = [m.strip() for m in (d.get('models') or []) if m and m.strip()]
         eid = db.w(
-            'INSERT INTO llm_endpoints (name, provider, base_url, api_key, models, pricing, '
+            'INSERT INTO llm_endpoints (name, provider, base_url, api_key, models, pricing, capabilities, '
             'input_price, output_price, price_per_call, billing, description, status, created_at, updated_at) '
-            'VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?)',
+            'VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)',
             (d.get('name', '').strip(), d.get('provider', 'custom') or 'custom',
              d.get('base_url', '').strip(), d.get('api_key', ''), _json.dumps(models),
-             _json.dumps(d.get('pricing') or {}), float(d.get('input_price') or 0),
+             _json.dumps(d.get('pricing') or {}), _json.dumps(d.get('capabilities') or {}),
+             float(d.get('input_price') or 0),
              float(d.get('output_price') or 0), float(d.get('price_per_call') or 0),
              d.get('billing', 'token'), d.get('description', ''), d.get('status', 'enabled'),
              db.now(), db.now()))
@@ -878,6 +889,10 @@ def endpoints_api():
     for r in rows:
         r['models'] = _json.loads(r.get('models') or '[]')
         r['pricing'] = _json.loads(r.get('pricing') or '{}')
+        try:
+            r['capabilities'] = _json.loads(r.get('capabilities') or '{}') or {}
+        except Exception:
+            r['capabilities'] = {}
         r['worker_count'] = sum(1 for w in wc.values() if w.get('endpoint_id') == r['id'])
         r['workers'] = [w['name'] for w in wc.values() if w.get('endpoint_id') == r['id']][:20]
     return jsonify({'ok': True, 'data': rows})
@@ -913,6 +928,9 @@ def endpoint_detail(eid):
     if 'pricing' in d:
         sets.append('pricing=?')
         args.append(_json.dumps(d['pricing'] or {}))
+    if 'capabilities' in d:
+        sets.append('capabilities=?')
+        args.append(_json.dumps(d['capabilities'] or {}))
     if sets:
         args.append(db.now())
         db.w(f'UPDATE llm_endpoints SET {", ".join(sets)}, updated_at=? WHERE id=?', (*args, eid))
@@ -1005,20 +1023,72 @@ def team_detail(tid):
 @app.route('/api/reference_projects')
 @require_auth
 def reference_projects():
-    """内置参考测试项目（3 个，含任务列表），供「从参考项目中新建」快速复制"""
+    """参考测试项目（内置 + 用户保存的），供「从参考项目中新建」快速复制"""
     err = _check_perm_point('project.view')
     if err:
         return err
-    rows = db.q('SELECT * FROM projects WHERE is_reference=1 ORDER BY id')
+    rows = db.q('SELECT * FROM projects WHERE is_reference=1 ORDER BY ref_builtin DESC, id')
     out = []
     for r in rows:
         r['tasks'] = db.q('SELECT * FROM tasks WHERE project_id=? AND deleted=0 ORDER BY id', (r['id'],))
         for t in r['tasks']:
             t['depends_on'] = _json.loads(t.get('depends_on') or '[]')
+        src = db.q('SELECT id, name FROM projects WHERE id=?', (r.get('ref_source_id') or 0,), one=True)
+        r['source_project'] = dict(src) if src else None
         out.append(r)
     return jsonify({'ok': True, 'data': out})
 
 
+@app.route('/api/projects/<int:pid>/save_reference', methods=['POST'])
+@require_auth
+def project_save_reference(pid):
+    """把已创建的项目一键保存为参考项目（复制项目+任务，原项目不受影响），供「从参考项目中新建」使用"""
+    err = _check_project_perm(pid, 'view')
+    if err:
+        return err
+    proj = db.q('SELECT * FROM projects WHERE id=?', (pid,), one=True)
+    if not proj:
+        return jsonify({'ok': False, 'error': '项目不存在'}), 404
+    if proj.get('is_reference'):
+        return jsonify({'ok': False, 'error': '该项目本身已是参考项目'}), 400
+    ts = db.now()
+    new_name = (proj['name'].strip() or '参考项目') + '（参考）'
+    rpid = db.w(
+        'INSERT INTO projects (name, description, objective, acceptance_criteria, status, '
+        'budget_limit, deliver_type, deliver_note, workspace_dir, is_reference, ref_builtin, '
+        'ref_source_id, auto_status, created_at, updated_at) VALUES (?,?,?,?,?,?,?,?,?,1,0,?,?,?,?)',
+        (new_name, (proj.get('description') or '') + '\n（由项目「' + proj['name'] + '」一键保存为参考）',
+         proj.get('objective') or '', proj.get('acceptance_criteria') or '', 'active',
+         proj.get('budget_limit') or 0, proj.get('deliver_type') or 'web', proj.get('deliver_note') or '',
+         '', pid, 'none', ts, ts))
+    db.w('UPDATE projects SET workspace_dir=? WHERE id=?', (f'ref_{rpid}', rpid))
+    # 复制任务（depends_on 重新映射）
+    n = _copy_tasks_from_reference(proj, rpid)
+    db.w("INSERT INTO project_logs (project_id, level, message, created_at) VALUES (?,?,?,?)",
+         (rpid, 'info', f'由项目「{proj["name"]}」一键保存为参考项目，复制 {n} 个任务', ts))
+    enterprise.audit(enterprise.current_actor(), 'project.save_reference', f'project#{pid}',
+                     f'「{proj["name"]}」保存为参考项目 #{rpid}（{n} 任务）', request.remote_addr or '')
+    return jsonify({'ok': True, 'id': rpid, 'copied_tasks': n})
+
+
+@app.route('/api/reference_projects/<int:rid>', methods=['DELETE'])
+@require_auth
+def reference_project_delete(rid):
+    """删除参考项目（内置参考项目受保护）"""
+    err = _check_perm_point('project.manage')
+    if err:
+        return err
+    r = db.q('SELECT * FROM projects WHERE id=? AND is_reference=1', (rid,), one=True)
+    if not r:
+        return jsonify({'ok': False, 'error': '参考项目不存在'}), 404
+    if r.get('ref_builtin'):
+        return jsonify({'ok': False, 'error': '内置参考项目受保护，不可删除'}), 400
+    db.w('DELETE FROM task_logs WHERE task_id IN (SELECT id FROM tasks WHERE project_id=?)', (rid,))
+    db.w('DELETE FROM tasks WHERE project_id=?', (rid,))
+    db.w('DELETE FROM projects WHERE id=?', (rid,))
+    return jsonify({'ok': True})
+
+
 @app.route('/api/workspace/probe')
 @require_auth
 def workspace_probe():
@@ -1067,8 +1137,56 @@ def settings_workspace():
 # ---------------------------------------------------------------------------
 # V3.5 · 对话（融合在仪表盘顶部）：可选 大模型接口 / AI Worker / 团队，默认主力 AI Worker
 # ---------------------------------------------------------------------------
+# ---------------------------------------------------------------------------
+# V3.5.2 模型能力 / 系统默认模型 / 对话工具
+# ---------------------------------------------------------------------------
+DEFAULT_CAP_KEYS = {'chat': 'default_chat_model', 'asr': 'default_asr_model', 'tts': 'default_tts_model',
+                    'image': 'default_image_model', 'video': 'default_video_model',
+                    'embedding': 'default_embedding_model', 'rerank': 'default_rerank_model'}
+
+
+def default_model_for(cap):
+    """系统默认模型：设置存 "endpoint_id:model"，返回 (endpoint_row, model) 或 None"""
+    key = DEFAULT_CAP_KEYS.get(cap)
+    if not key:
+        return None
+    v = (db.get_setting(key, '') or '').strip()
+    if not v or ':' not in v:
+        return None
+    try:
+        eid, model = v.split(':', 1)
+        ep = db.q('SELECT * FROM llm_endpoints WHERE id=? AND status="enabled"', (int(eid),), one=True)
+        if ep and model:
+            return ep, model
+    except Exception:
+        pass
+    return None
+
+
+def _resolve_image(image):
+    """把聊天图片引用解析为模型可用的 URL / data URL"""
+    if not image:
+        return None
+    if image.startswith('data:'):
+        return image
+    if image.startswith('http'):
+        return image
+    if image.startswith('/api/chat/attachments/'):
+        fn = os.path.basename(image)
+        p = os.path.join(CHAT_UPLOAD_DIR, fn)
+        if os.path.isfile(p):
+            with open(p, 'rb') as f:
+                raw = f.read()
+            ext = os.path.splitext(fn)[1].lower().lstrip('.')
+            mime = {'jpg': 'image/jpeg', 'jpeg': 'image/jpeg', 'png': 'image/png',
+                    'gif': 'image/gif', 'webp': 'image/webp'}.get(ext, 'image/png')
+            return f'data:{mime};base64,{_b64.b64encode(raw).decode()}'
+    return image
+
+
 def _chat_target_cfg(session_row):
-    """解析对话目标，返回 (label, worker_or_none, llm_cfg, system_prompt, model)"""
+    """解析对话目标，返回 (label, worker_or_none, cfg, system_prompt, model, caps)
+    cfg 供 llm_gateway.stream_chat 使用（worker 与接口目标统一）"""
     ttype = session_row.get('target_type') or 'worker'
     tid = session_row.get('target_id') or 0
     model = session_row.get('model') or ''
@@ -1079,10 +1197,11 @@ def _chat_target_cfg(session_row):
         models = _json.loads(ep.get('models') or '[]') or ['']
         if not model:
             model = models[0]
+        caps = llm_gateway.endpoint_model_caps(ep, model)
         cfg = {'provider': ep.get('provider') or 'custom', 'model': model,
                'base_url': ep.get('base_url') or '', 'api_key': ep.get('api_key') or '',
                'endpoint': ep}
-        return f'接口「{ep["name"]}」/{model}', None, cfg, '', model
+        return f'接口「{ep["name"]}」/{model}', None, cfg, '', model, caps
     if ttype == 'team':
         team = db.q('SELECT * FROM worker_teams WHERE id=?', (tid,), one=True)
         if not team:
@@ -1092,48 +1211,187 @@ def _chat_target_cfg(session_row):
                     tuple(wids)) if wids else []
         if not rows:
             raise ValueError(f'团队「{team["name"]}」没有可用的 AI Worker 成员')
-        # 轮询分配成员：按该会话消息数取模
         n = db.q('SELECT COUNT(*) c FROM chat_messages WHERE session_id=? AND role="assistant"',
                  (session_row['id'],))[0]['c']
         worker = rows[n % len(rows)]
-        return f'团队「{team["name"]}」成员「{worker["name"]}」', worker, None, worker.get('system_prompt') or '', ''
-    # worker
+        cfg = llm_gateway.worker_llm_cfg(worker)
+        cfg['endpoint'] = cfg['endpoint'] or None
+        caps = llm_gateway.endpoint_model_caps(cfg.get('endpoint'), cfg['model']) if cfg.get('endpoint') else llm_gateway.guess_model_caps(cfg['model'])
+        return f'团队「{team["name"]}」成员「{worker["name"]}」', worker, cfg, worker.get('system_prompt') or '', cfg['model'], caps
     worker = db.q('SELECT * FROM workers WHERE id=? AND status="enabled"', (tid,), one=True)
     if not worker:
         raise ValueError('对话目标：AI Worker 不存在或已停用')
-    return f'Worker「{worker["name"]}」', worker, None, worker.get('system_prompt') or '', ''
+    cfg = llm_gateway.worker_llm_cfg(worker)
+    caps = llm_gateway.endpoint_model_caps(cfg.get('endpoint'), cfg['model']) if cfg.get('endpoint') else llm_gateway.guess_model_caps(cfg['model'])
+    return f'Worker「{worker["name"]}」', worker, cfg, worker.get('system_prompt') or '', cfg['model'], caps
 
 
 def _chat_history(sid, limit=24):
-    rows = db.q('SELECT role, content FROM chat_messages WHERE session_id=? AND error="" '
+    """取会话历史（含图片消息），返回 [{role, content, image}]"""
+    rows = db.q('SELECT role, content, image FROM chat_messages WHERE session_id=? AND error="" '
                 'ORDER BY id DESC LIMIT ?', (sid, limit))
     rows.reverse()
     out = []
     for r in rows:
-        if (r['content'] or '').strip():
-            out.append({'role': r['role'], 'content': r['content'][:8000]})
+        c = (r['content'] or '').strip()
+        img = _resolve_image(r.get('image') or '')
+        if c or img:
+            out.append({'role': r['role'], 'content': c[:8000], 'image': img})
     return out
 
 
 @app.route('/api/chat/options')
 @require_auth
 def chat_options():
-    """对话配置：可选 大模型接口/Worker/团队 + 主力 AI Worker"""
+    """对话配置：可选 大模型接口/Worker/团队 + 主力 AI Worker + 各模型能力标签 + 系统默认模型"""
     eps = db.q('SELECT id, name, provider, base_url, models, billing, input_price, output_price, '
-               'price_per_call, status FROM llm_endpoints WHERE status="enabled" ORDER BY id')
+               'price_per_call, capabilities, status FROM llm_endpoints WHERE status="enabled" ORDER BY id')
+    caps_map_all = {}
     for e in eps:
-        e['models'] = _json.loads(e.get('models') or '[]')
+        models = _json.loads(e.get('models') or '[]')
+        e['models'] = models
+        try:
+            caps_map = _json.loads(e.get('capabilities') or '{}') or {}
+        except Exception:
+            caps_map = {}
+        caps_map_all[e['id']] = caps_map
+        e['caps'] = {m: (caps_map.get(m) or llm_gateway.guess_model_caps(m)) for m in models}
     workers = db.q('SELECT id, name, provider, model, status, endpoint_id FROM workers WHERE status="enabled" ORDER BY id')
+    for w in workers:
+        ep = next((e for e in eps if e['id'] == w.get('endpoint_id')), None)
+        w['caps'] = llm_gateway.endpoint_model_caps(ep, w['model']) if ep else llm_gateway.guess_model_caps(w['model'])
     teams = db.q('SELECT * FROM worker_teams ORDER BY id DESC')
     for t in teams:
         t['worker_ids'] = _json.loads(t.get('worker_ids') or '[]')
         t['worker_count'] = len(t['worker_ids'])
+    defaults = {cap: default_model_for(cap) for cap in DEFAULT_CAP_KEYS}
+    dflt = {}
+    for cap, val in defaults.items():
+        if val:
+            ep, m = val
+            dflt[cap] = {'endpoint_id': ep['id'], 'endpoint_name': ep['name'], 'model': m}
     return jsonify({'ok': True, 'data': {
         'endpoints': eps, 'workers': workers, 'teams': teams,
         'main_worker_id': db.main_worker_id(),
+        'defaults': dflt,
+        'cap_labels': llm_gateway.CAP_LABELS,
     }})
 
 
+@app.route('/api/models/defaults', methods=['GET', 'PUT'])
+@require_auth
+@require_perm('setting.manage')
+def models_defaults():
+    """系统默认模型：语音识别/语音合成/图片生成/视频生成/embedding/rerank 等，配置后全局生效"""
+    caps = ['chat', 'asr', 'tts', 'image', 'video', 'embedding', 'rerank']
+    if request.method == 'PUT':
+        d = request.get_json(force=True) or {}
+        for cap in caps:
+            val = (d.get(cap) or '').strip()
+            if val and ':' not in val:
+                return jsonify({'ok': False, 'error': f'{cap} 默认模型格式应为 endpoint_id:model'}), 400
+            db.set_setting(DEFAULT_CAP_KEYS[cap], val)
+        return jsonify({'ok': True})
+    eps = db.q('SELECT id, name, models, capabilities FROM llm_endpoints WHERE status="enabled" ORDER BY id')
+    for e in eps:
+        e['models'] = _json.loads(e.get('models') or '[]')
+        try:
+            e['caps'] = _json.loads(e.get('capabilities') or '{}') or {}
+        except Exception:
+            e['caps'] = {}
+    out = {}
+    for cap in caps:
+        dm = default_model_for(cap)
+        out[cap] = {'value': f'{dm[0]["id"]}:{dm[1]}' if dm else '',
+                    'endpoint_id': dm[0]['id'] if dm else None, 'model': dm[1] if dm else ''}
+    return jsonify({'ok': True, 'data': {'defaults': out, 'endpoints': eps}})
+
+
+@app.route('/api/chat/upload_image', methods=['POST'])
+@require_auth
+def chat_upload_image():
+    """对话图片上传：保存到 data/chat_uploads/，返回可访问 URL（供视觉模型对话使用）"""
+    if 'file' not in request.files:
+        return jsonify({'ok': False, 'error': '缺少文件字段 file'}), 400
+    fs = request.files['file']
+    if not fs or not fs.filename:
+        return jsonify({'ok': False, 'error': '文件名为空'}), 400
+    ext = os.path.splitext(fs.filename)[1].lower()
+    if ext not in ('.png', '.jpg', '.jpeg', '.gif', '.webp'):
+        return jsonify({'ok': False, 'error': '仅支持 png/jpg/gif/webp 图片'}), 400
+    name = f'{db.now()}_{secrets.token_hex(4)}{ext}'
+    fs.save(os.path.join(CHAT_UPLOAD_DIR, name))
+    return jsonify({'ok': True, 'url': f'/api/chat/attachments/{name}'})
+
+
+@app.route('/api/chat/attachments/<path:name>')
+@require_auth
+def chat_attachment(name):
+    return send_from_directory(CHAT_UPLOAD_DIR, os.path.basename(name))
+
+
+@app.route('/api/chat/transcribe', methods=['POST'])
+@require_auth
+def chat_transcribe():
+    """语音输入：用系统默认 语音识别模型 转写（OpenAI 兼容 /audio/transcriptions）"""
+    dm = default_model_for('asr')
+    if not dm:
+        return jsonify({'ok': False, 'error': '未配置系统默认语音识别模型（模型库 → 系统默认模型）'}), 400
+    ep, model = dm
+    if 'file' not in request.files:
+        return jsonify({'ok': False, 'error': '缺少音频文件'}), 400
+    fs = request.files['file']
+    raw = fs.read()
+    if not raw:
+        return jsonify({'ok': False, 'error': '音频内容为空'}), 400
+    url = (ep.get('base_url') or '').rstrip('/') + '/audio/transcriptions'
+    headers = {'Authorization': f'Bearer {ep.get("api_key") or ""}'}
+    try:
+        import requests
+        resp = requests.post(url, headers=headers, timeout=120,
+                             files={'file': (fs.filename or 'audio.wav', raw, 'audio/wav')},
+                             data={'model': model})
+        if resp.status_code != 200:
+            return jsonify({'ok': False, 'error': f'转写失败({resp.status_code}): {resp.text[:200]}'}), 400
+        data = resp.json()
+        text = (data.get('text') or '').strip()
+        if not text:
+            return jsonify({'ok': False, 'error': '转写结果为空'}), 400
+        return jsonify({'ok': True, 'text': text})
+    except Exception as e:
+        return jsonify({'ok': False, 'error': f'语音识别失败：{str(e)[:200]}'}), 400
+
+
+@app.route('/api/chat/tts', methods=['POST'])
+@require_auth
+def chat_tts():
+    """语音输出：用系统默认 语音合成模型 把文本转语音（OpenAI 兼容 /audio/speech）"""
+    dm = default_model_for('tts')
+    if not dm:
+        return jsonify({'ok': False, 'error': '未配置系统默认语音合成模型（模型库 → 系统默认模型）'}), 400
+    ep, model = dm
+    d = request.get_json(force=True) or {}
+    text = (d.get('text') or '').strip()[:4000]
+    if not text:
+        return jsonify({'ok': False, 'error': '文本为空'}), 400
+    url = (ep.get('base_url') or '').rstrip('/') + '/audio/speech'
+    headers = {'Authorization': f'Bearer {ep.get("api_key") or ""}', 'Content-Type': 'application/json'}
+    try:
+        import requests
+        resp = requests.post(url, headers=headers, timeout=120,
+                             json={'model': model, 'input': text, 'voice': d.get('voice') or 'alloy'})
+        if resp.status_code != 200:
+            return jsonify({'ok': False, 'error': f'合成失败({resp.status_code}): {resp.text[:200]}'}), 400
+        audio = resp.content
+        ctype = resp.headers.get('Content-Type', 'audio/mpeg')
+        from flask import make_response
+        r = make_response(audio)
+        r.headers['Content-Type'] = ctype
+        return r
+    except Exception as e:
+        return jsonify({'ok': False, 'error': f'语音合成失败：{str(e)[:200]}'}), 400
+
+
 @app.route('/api/chat/sessions', methods=['GET', 'POST'])
 @require_auth
 def chat_sessions():
@@ -1149,7 +1407,7 @@ def chat_sessions():
             'VALUES (?,?,?,?,?,?)',
             ((d.get('title') or '新对话').strip(), ttype, tid, model, db.now(), db.now()))
         return jsonify({'ok': True, 'id': sid})
-    rows = db.q('SELECT * FROM chat_sessions ORDER BY id DESC LIMIT 50')
+    rows = db.q('SELECT * FROM chat_sessions ORDER BY pinned DESC, updated_at DESC LIMIT 100')
     for r in rows:
         last = db.q('SELECT role, content, created_at FROM chat_messages WHERE session_id=? '
                     'ORDER BY id DESC LIMIT 1', (r['id'],), one=True)
@@ -1158,30 +1416,55 @@ def chat_sessions():
     return jsonify({'ok': True, 'data': rows})
 
 
-@app.route('/api/chat/sessions/<int:sid>', methods=['GET', 'DELETE'])
+@app.route('/api/chat/sessions/<int:sid>', methods=['GET', 'PUT', 'DELETE'])
 @require_auth
 def chat_session_detail(sid):
     s = db.q('SELECT * FROM chat_sessions WHERE id=?', (sid,), one=True)
     if not s:
         return jsonify({'ok': False, 'error': '会话不存在'}), 404
+    if request.method == 'PUT':
+        """更新会话：重命名 / 置顶 / 切换知识库开关 / 切换目标"""
+        d = request.get_json(force=True) or {}
+        sets, args = [], []
+        if 'title' in d:
+            title = (d.get('title') or '').strip()
+            if title:
+                sets.append('title=?')
+                args.append(title[:100])
+        if 'pinned' in d:
+            sets.append('pinned=?')
+            args.append(1 if d.get('pinned') else 0)
+        if 'use_kb' in d:
+            sets.append('use_kb=?')
+            args.append(1 if d.get('use_kb') else 0)
+        if 'target_type' in d and 'target_id' in d:
+            sets.append('target_type=?'); args.append(d.get('target_type') or 'worker')
+            sets.append('target_id=?'); args.append(int(d.get('target_id') or 0))
+            sets.append('model=?'); args.append(d.get('model') or '')
+        if sets:
+            args.append(db.now())
+            db.w(f'UPDATE chat_sessions SET {", ".join(sets)}, updated_at=? WHERE id=?', (*args, sid))
+        return jsonify({'ok': True})
     if request.method == 'DELETE':
         db.w('DELETE FROM chat_messages WHERE session_id=?', (sid,))
         db.w('DELETE FROM chat_sessions WHERE id=?', (sid,))
         return jsonify({'ok': True})
     msgs = db.q('SELECT * FROM chat_messages WHERE session_id=? ORDER BY id', (sid,))
     try:
-        label, worker, cfg, sys_prompt, model = _chat_target_cfg(s)
+        label, worker, cfg, sys_prompt, model, caps = _chat_target_cfg(s)
         s['target_label'] = label
+        s['caps'] = caps
     except Exception as e:
         s['target_label'] = f'（{str(e)}）'
+        s['caps'] = []
     return jsonify({'ok': True, 'data': {'session': s, 'messages': msgs}})
 
 
-def _chat_gen(sid, user_content):
-    """SSE 生成器：流式转发模型输出（单 token 返回超时 / 首字延迟超时 由 llm_gateway 处理）"""
-    import json as _j
+def _chat_gen(sid, user_content, image=None, use_kb=False):
+    """SSE 生成器（V3.5.2 真流式）：边收边吐；思考模型先流式思考内容再流式回答；
+    支持知识库注入与图片消息。事件：start/reasoning/delta/done/error"""
     def sse(obj):
-        return f'data: {_j.dumps(obj, ensure_ascii=False)}\n\n'
+        return f'data: {_json.dumps(obj, ensure_ascii=False)}\n\n'
     s = db.q('SELECT * FROM chat_sessions WHERE id=?', (sid,), one=True)
     if not s:
         yield sse({'type': 'error', 'message': '会话不存在'}); return
@@ -1190,81 +1473,96 @@ def _chat_gen(sid, user_content):
     cfg = None
     sys_prompt = ''
     model = ''
-    chunks = []
-    worker_id = None
+    caps = []
     try:
-        label, worker, cfg, sys_prompt, model = _chat_target_cfg(s)
+        label, worker, cfg, sys_prompt, model, caps = _chat_target_cfg(s)
     except Exception as e:
         yield sse({'type': 'error', 'message': str(e)}); return
-    # 组装消息
+    yield sse({'type': 'start', 'label': label, 'caps': caps, 'model': model})
+    # 图片：仅视觉模型可带图
+    img_url = _resolve_image(image) if image else None
+    if img_url and 'vision' not in caps:
+        yield sse({'type': 'error', 'message': f'当前模型「{model}」不支持视觉输入（能力标签：{"、".join(caps) or "无"}）'}); return
+    # 组装消息（历史 + 知识库 + 本条）
     history = _chat_history(sid, 24)
-    history.append({'role': 'user', 'content': user_content})
+    if use_kb:
+        ctx, hits = kb.build_context(user_content, 3)
+        if ctx:
+            history.insert(0, {'role': 'system', 'content': ctx})
+            yield sse({'type': 'note', 'message': f'📚 已注入知识库参考 {len(hits)} 段'})
     if sys_prompt and not any(m['role'] == 'system' for m in history):
         history.insert(0, {'role': 'system', 'content': sys_prompt})
-    chunks = []
+    if img_url:
+        history.append({'role': 'user', 'content': [{'type': 'text', 'text': user_content},
+                                                    {'type': 'image_url', 'image_url': {'url': img_url}}]})
+    else:
+        history.append({'role': 'user', 'content': user_content})
+    chunks, thinking = [], []
     t0 = time.time()
     try:
-        if worker is not None:
-            def on_chunk(piece):
-                chunks.append(piece)
-            r = llm_gateway.chat_worker(worker, history, on_chunk=on_chunk)
-            for piece in chunks:
-                yield sse({'type': 'delta', 'content': piece})
-            worker_id = worker['id']
-            model = r['model']
-        else:
-            # 大模型接口：流式
-            r = llm_gateway.chat_stream(cfg['provider'], cfg['model'], history,
-                                        base_url=cfg['base_url'] or None, api_key=cfg['api_key'] or None,
-                                        on_chunk=lambda p: chunks.append(p))
-            for piece in chunks:
-                yield sse({'type': 'delta', 'content': piece})
-            r['cost'] = llm_gateway.calc_cost_ex(cfg['model'], r['prompt_tokens'],
-                                                 r['completion_tokens'], 1, cfg)
-            worker_id = None
-            model = r['model']
-        db.w('INSERT INTO chat_messages (session_id, role, content, model, worker_id, prompt_tokens, '
-             'completion_tokens, cached_tokens, cost, latency_ms, first_token_ms, created_at) '
-             'VALUES (?,?,?,?,?,?,?,?,?,?,?,?)',
-             (sid, 'assistant', ''.join(chunks), model, worker_id,
+        for evt, val in llm_gateway.stream_chat(cfg, history,
+                                                temperature=worker.get('temperature', 0.7) if worker else 0.7,
+                                                max_tokens=worker.get('max_tokens') if worker else None):
+            if evt == 'reasoning':
+                thinking.append(val)
+                yield sse({'type': 'reasoning', 'content': val})
+            elif evt == 'delta':
+                chunks.append(val)
+                yield sse({'type': 'delta', 'content': val})
+            elif evt == 'error':
+                raise llm_gateway.LLMError(val)
+            elif evt == 'done':
+                r = val
+        worker_id = worker['id'] if worker else None
+        db.w('INSERT INTO chat_messages (session_id, role, content, thinking, image, model, worker_id, '
+             'prompt_tokens, completion_tokens, cached_tokens, cost, latency_ms, first_token_ms, created_at) '
+             'VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?)',
+             (sid, 'assistant', ''.join(chunks), ''.join(thinking), '', r['model'], worker_id,
               r['prompt_tokens'], r['completion_tokens'], r.get('cached_tokens', 0),
               r['cost'], r.get('elapsed_ms', 0), r.get('first_token_ms') or 0, db.now()))
         yield sse({'type': 'done', 'usage': {
             'prompt_tokens': r['prompt_tokens'], 'completion_tokens': r['completion_tokens'],
             'cached_tokens': r.get('cached_tokens', 0), 'total_tokens': r['total_tokens'],
             'cost': r['cost'], 'first_token_ms': r.get('first_token_ms'),
-            'elapsed_ms': r.get('elapsed_ms', 0), 'model': model, 'label': label}})
+            'elapsed_ms': r.get('elapsed_ms', 0), 'model': r['model'], 'label': label,
+            'has_thinking': bool(thinking)}})
     except llm_gateway.LLMError as e:
         partial = ''.join(chunks)
-        db.w('INSERT INTO chat_messages (session_id, role, content, model, worker_id, error, created_at) '
-             'VALUES (?,?,?,?,?,?,?)',
-             (sid, 'assistant', partial, model or '', worker_id,
-              str(e)[:500], db.now()))
-        yield sse({'type': 'error', 'message': str(e), 'partial': partial})
+        db.w('INSERT INTO chat_messages (session_id, role, content, thinking, image, model, worker_id, error, created_at) '
+             'VALUES (?,?,?,?,?,?,?,?,?)',
+             (sid, 'assistant', partial, ''.join(thinking), '', model or '',
+              worker['id'] if worker else None, str(e)[:500], db.now()))
+        yield sse({'type': 'error', 'message': str(e), 'partial': partial,
+                   'has_thinking': bool(thinking)})
     except Exception as e:
-        db.w('INSERT INTO chat_messages (session_id, role, content, model, worker_id, error, created_at) '
-             'VALUES (?,?,?,?,?,?,?)',
-             (sid, 'assistant', ''.join(chunks), model or '', worker_id, str(e)[:500], db.now()))
-        yield sse({'type': 'error', 'message': str(e)})
+        db.w('INSERT INTO chat_messages (session_id, role, content, thinking, image, model, worker_id, error, created_at) '
+             'VALUES (?,?,?,?,?,?,?,?,?)',
+             (sid, 'assistant', ''.join(chunks), ''.join(thinking), '', model or '',
+              worker['id'] if worker else None, str(e)[:500], db.now()))
+        yield sse({'type': 'error', 'message': str(e), 'has_thinking': bool(thinking)})
 
 
 @app.route('/api/chat/sessions/<int:sid>/messages', methods=['POST'])
 @require_auth
 def chat_send(sid):
-    """发送对话消息，服务端按 token 流式返回（SSE）"""
+    """发送对话消息，服务端按 token 实时流式返回（SSE）：思考内容先行，回答随后"""
     s = db.q('SELECT id FROM chat_sessions WHERE id=?', (sid,), one=True)
     if not s:
         return jsonify({'ok': False, 'error': '会话不存在'}), 404
     d = request.get_json(force=True)
     content = (d.get('content') or '').strip()
-    if not content:
+    if not content and not d.get('image'):
         return jsonify({'ok': False, 'error': '消息内容为空'}), 400
     if len(content) > 60000:
         content = content[:60000]
-    db.w('INSERT INTO chat_messages (session_id, role, content, created_at) VALUES (?,?,?,?)',
-         (sid, 'user', content, db.now()))
+    image = (d.get('image') or '').strip()
+    use_kb = bool(d.get('use_kb'))
+    if use_kb:
+        db.w('UPDATE chat_sessions SET use_kb=1 WHERE id=?', (sid,))
+    db.w('INSERT INTO chat_messages (session_id, role, content, image, created_at) VALUES (?,?,?,?,?)',
+         (sid, 'user', content, image, db.now()))
     db.w('UPDATE chat_sessions SET updated_at=? WHERE id=?', (db.now(), sid))
-    return Response(_chat_gen(sid, content), mimetype='text/event-stream',
+    return Response(_chat_gen(sid, content, image, use_kb), mimetype='text/event-stream',
                     headers={'Cache-Control': 'no-cache', 'X-Accel-Buffering': 'no'})
 
 
@@ -1573,7 +1871,26 @@ def report_cost():
         scope_sql = 'WHERE project_id IN (%s)' % ','.join('?' * len(visible))
         scope_args = list(visible)
     group = request.args.get('group', 'project')
-    if group == 'worker':
+    if group == 'date':
+        # 按日期统计（V3.5）：date(created_at) 以本地时区分天；days 限制最近 N 天
+        days = 0
+        try:
+            days = int(request.args.get('days') or 0)
+        except Exception:
+            days = 0
+        extra_where = ''
+        extra_args = []
+        if days > 0:
+            extra_where = (scope_sql + ' AND' if scope_sql else 'WHERE') + ' created_at>=?'
+            extra_args = [db.now() - days * 86400]
+        rows = db.q(
+            f'SELECT date(created_at, "unixepoch", "localtime") day, COUNT(*) runs, '
+            f'COUNT(DISTINCT task_id) task_calls, '
+            f'SUM(prompt_tokens) prompt_tokens, SUM(completion_tokens) completion_tokens, '
+            f'SUM(cached_tokens) cached_tokens, SUM(total_tokens) tokens, SUM(cost) cost '
+            f'FROM cost_records {scope_sql}{extra_where} GROUP BY day ORDER BY day DESC',
+            scope_args + extra_args)
+    elif group == 'worker':
         rows = db.q(
             'SELECT worker_id, provider, model, COUNT(*) runs, COUNT(DISTINCT task_id) task_calls, '
             'SUM(prompt_tokens) prompt_tokens, SUM(completion_tokens) completion_tokens, '
@@ -2102,6 +2419,105 @@ def kb_search(pid):
     return jsonify({'ok': True, 'data': hits if hit else [], 'hit': hit})
 
 
+# ---------------------------------------------------------------------------
+# V3.5.2 · 全局知识库（导航「📚 知识库」：增删改查 / 上传 / 检索）
+# ---------------------------------------------------------------------------
+@app.route('/api/kb/docs', methods=['GET', 'POST'])
+@require_auth
+def kb_docs():
+    err = _check_perm_point('project.view')
+    if err:
+        return err
+    if request.method == 'POST':
+        d = request.get_json(force=True)
+        title = (d.get('title') or '').strip()
+        if not title:
+            return jsonify({'ok': False, 'error': '标题必填'}), 400
+        did = db.w(
+            'INSERT INTO kb_documents (title, content, tags, source, created_at, updated_at) VALUES (?,?,?,?,?,?)',
+            (title, d.get('content') or '', _json.dumps(d.get('tags') or []), d.get('source') or 'manual',
+             db.now(), db.now()))
+        n = kb.rebuild_chunks(did)
+        enterprise.audit(enterprise.current_actor(), 'kb.create', f'kb#{did}', f'「{title}」({n} 块)',
+                         request.remote_addr or '')
+        return jsonify({'ok': True, 'id': did, 'chunks': n})
+    q = (request.args.get('q') or '').strip()
+    page = max(int(request.args.get('page', 1)), 1)
+    page_size = min(int(request.args.get('page_size', 20)), 100)
+    where, args = [], []
+    if q:
+        where.append('(title LIKE ? OR content LIKE ?)')
+        args += [f'%{q}%', f'%{q}%']
+    where_sql = (' WHERE ' + ' AND '.join(where)) if where else ''
+    total = db.q(f'SELECT COUNT(*) c FROM kb_documents{where_sql}', args)[0]['c']
+    rows = db.q(f'SELECT * FROM kb_documents{where_sql} ORDER BY id DESC LIMIT ? OFFSET ?',
+                (*args, page_size, (page - 1) * page_size))
+    for r in rows:
+        r['tags'] = _json.loads(r.get('tags') or '[]')
+        r['chunks'] = db.q('SELECT COUNT(*) c FROM kb_chunks WHERE doc_id=?', (r['id'],))[0]['c']
+        r['preview'] = (r.get('content') or '')[:200]
+    return jsonify({'ok': True, 'data': rows, 'total': total, 'page': page, 'page_size': page_size})
+
+
+@app.route('/api/kb/docs/<int:did>', methods=['GET', 'PUT', 'DELETE'])
+@require_auth
+def kb_doc_detail(did):
+    d = db.q('SELECT * FROM kb_documents WHERE id=?', (did,), one=True)
+    if not d:
+        return jsonify({'ok': False, 'error': '文档不存在'}), 404
+    if request.method == 'GET':
+        d['tags'] = _json.loads(d.get('tags') or '[]')
+        return jsonify({'ok': True, 'data': d})
+    if request.method == 'DELETE':
+        db.w('DELETE FROM kb_chunks WHERE doc_id=?', (did,))
+        db.w('DELETE FROM kb_documents WHERE id=?', (did,))
+        return jsonify({'ok': True})
+    body = request.get_json(force=True)
+    if 'title' in body and (body.get('title') or '').strip():
+        db.w('UPDATE kb_documents SET title=? WHERE id=?', (body['title'].strip(), did))
+    if 'content' in body:
+        db.w('UPDATE kb_documents SET content=? WHERE id=?', (body.get('content') or '', did))
+    if 'tags' in body:
+        db.w('UPDATE kb_documents SET tags=? WHERE id=?', (_json.dumps(body['tags'] or []), did))
+    db.w('UPDATE kb_documents SET updated_at=? WHERE id=?', (db.now(), did))
+    kb.rebuild_chunks(did)
+    return jsonify({'ok': True})
+
+
+@app.route('/api/kb/search')
+@require_auth
+def kb_search_api():
+    q = request.args.get('q', '').strip()
+    if not q:
+        return jsonify({'ok': True, 'data': []})
+    hits = kb.search(q, top_k=8)
+    return jsonify({'ok': True, 'data': hits})
+
+
+@app.route('/api/kb/docs/upload', methods=['POST'])
+@require_auth
+def kb_upload():
+    """知识库上传：txt/md/html/csv/json/pdf/docx 抽取文本入库"""
+    if 'files' not in request.files:
+        return jsonify({'ok': False, 'error': '缺少文件字段 files'}), 400
+    saved = []
+    for fs in request.files.getlist('files'):
+        if not fs or not fs.filename:
+            continue
+        raw = fs.read()
+        text, ok = kb.extract_text(fs.filename, raw)
+        title = os.path.splitext(os.path.basename(fs.filename))[0].strip() or fs.filename
+        if not ok or not text.strip():
+            saved.append({'file': fs.filename, 'ok': False, 'error': '无法解析文本内容'})
+            continue
+        did = db.w(
+            'INSERT INTO kb_documents (title, content, tags, source, created_at, updated_at) VALUES (?,?,?,?,?,?)',
+            (title, text[:200000], _json.dumps(['上传']), 'file', db.now(), db.now()))
+        n = kb.rebuild_chunks(did)
+        saved.append({'file': fs.filename, 'ok': True, 'id': did, 'chunks': n, 'len': len(text)})
+    return jsonify({'ok': True, 'saved': saved})
+
+
 # ---------------------------------------------------------------------------
 # V3 · 交付体系：工作目录 / Demo 部署 / 打包 / 邮件送达
 # ---------------------------------------------------------------------------

db.py

@@ -401,10 +401,11 @@ CREATE TABLE IF NOT EXISTS llm_endpoints (
     base_url TEXT DEFAULT '',
     api_key TEXT DEFAULT '',
     models TEXT DEFAULT '[]',             -- JSON: 可用模型名列表
-    pricing TEXT DEFAULT '{}',            -- JSON: {model: {input: 元/1M, output: 元/1M}}
+    pricing TEXT DEFAULT '{}',            -- JSON: {model: {input: 元/1M, input_cache: 元/1M, output: 元/1M}}
+    capabilities TEXT DEFAULT '{}',       -- JSON: {model: [能力列表 chat/thinking/vision/audio_in/audio_out/image_gen/video_gen/embedding/rerank]}
     input_price REAL DEFAULT 0,           -- 兜底输入价（元/1M tokens）
     output_price REAL DEFAULT 0,          -- 兜底输出价（元/1M tokens）
-    price_per_call REAL DEFAULT 0,        -- 按调用次数计费单价（元/次）
+    price_per_call REAL DEFAULT 0,        -- 按调用次数计费单价（元/千次）
     billing TEXT DEFAULT 'token',         -- token=按token数计费 / call=按调用次数计费
     description TEXT DEFAULT '',
     status TEXT DEFAULT 'enabled',        -- enabled/disabled
@@ -427,6 +428,8 @@ CREATE TABLE IF NOT EXISTS chat_sessions (
     target_type TEXT DEFAULT 'worker',    -- model=大模型接口 / worker=AI Worker / team=团队
     target_id INTEGER DEFAULT 0,
     model TEXT DEFAULT '',                -- target_type=model 时选定的模型名
+    pinned INTEGER DEFAULT 0,             -- 置顶
+    use_kb INTEGER DEFAULT 0,             -- 是否注入知识库上下文
     created_at INTEGER,
     updated_at INTEGER
 );
@@ -436,6 +439,8 @@ CREATE TABLE IF NOT EXISTS chat_messages (
     session_id INTEGER NOT NULL,
     role TEXT DEFAULT 'user',             -- user/assistant
     content TEXT DEFAULT '',
+    thinking TEXT DEFAULT '',             -- 思考模型的过程内容
+    image TEXT DEFAULT '',                -- 用户消息附带的图片（路径/数据URL）
     model TEXT DEFAULT '',
     worker_id INTEGER,
     prompt_tokens INTEGER DEFAULT 0,
@@ -448,6 +453,26 @@ CREATE TABLE IF NOT EXISTS chat_messages (
     created_at INTEGER
 );
 
+CREATE TABLE IF NOT EXISTS kb_documents (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    title TEXT NOT NULL,
+    content TEXT DEFAULT '',
+    tags TEXT DEFAULT '[]',
+    source TEXT DEFAULT 'manual',
+    created_at INTEGER,
+    updated_at INTEGER
+);
+
+CREATE TABLE IF NOT EXISTS kb_chunks (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    doc_id INTEGER NOT NULL,
+    idx INTEGER DEFAULT 0,
+    content TEXT DEFAULT '',
+    tokens TEXT DEFAULT '[]'
+);
+
+CREATE INDEX IF NOT EXISTS idx_kb_chunks_doc ON kb_chunks(doc_id);
+
 CREATE INDEX IF NOT EXISTS idx_chat_msg_session ON chat_messages(session_id);
 CREATE INDEX IF NOT EXISTS idx_cost_worker ON cost_records(worker_id);
 """
@@ -549,12 +574,28 @@ def _migrate():
     pcols3 = {r['name'] for r in conn.execute('PRAGMA table_info(projects)')}
     if 'is_reference' not in pcols3:
         conn.execute('ALTER TABLE projects ADD COLUMN is_reference INTEGER DEFAULT 0')
+    if 'ref_builtin' not in pcols3:
+        conn.execute('ALTER TABLE projects ADD COLUMN ref_builtin INTEGER DEFAULT 0')
+    if 'ref_source_id' not in pcols3:
+        conn.execute('ALTER TABLE projects ADD COLUMN ref_source_id INTEGER DEFAULT 0')
     acols = {r['name'] for r in conn.execute('PRAGMA table_info(agent_runs)')}
     if 'workspace_dir' not in acols:
         conn.execute("ALTER TABLE agent_runs ADD COLUMN workspace_dir TEXT DEFAULT ''")
     scols = {r['name'] for r in conn.execute('PRAGMA table_info(chat_sessions)')}
     if 'model' not in scols:
         conn.execute("ALTER TABLE chat_sessions ADD COLUMN model TEXT DEFAULT ''")
+    if 'pinned' not in scols:
+        conn.execute('ALTER TABLE chat_sessions ADD COLUMN pinned INTEGER DEFAULT 0')
+    if 'use_kb' not in scols:
+        conn.execute('ALTER TABLE chat_sessions ADD COLUMN use_kb INTEGER DEFAULT 0')
+    mcols = {r['name'] for r in conn.execute('PRAGMA table_info(chat_messages)')}
+    if 'thinking' not in mcols:
+        conn.execute("ALTER TABLE chat_messages ADD COLUMN thinking TEXT DEFAULT ''")
+    if 'image' not in mcols:
+        conn.execute("ALTER TABLE chat_messages ADD COLUMN image TEXT DEFAULT ''")
+    ecols = {r['name'] for r in conn.execute('PRAGMA table_info(llm_endpoints)')}
+    if 'capabilities' not in ecols:
+        conn.execute("ALTER TABLE llm_endpoints ADD COLUMN capabilities TEXT DEFAULT '{}'")
     # V3.5 新表（幂等）
     conn.execute('''CREATE TABLE IF NOT EXISTS llm_endpoints (
         id INTEGER PRIMARY KEY AUTOINCREMENT,
@@ -606,8 +647,43 @@ def _migrate():
         error TEXT DEFAULT '',
         created_at INTEGER
     )''')
-    conn.execute('CREATE INDEX IF NOT EXISTS idx_chat_msg_session ON chat_messages(session_id)')
-    conn.execute('CREATE INDEX IF NOT EXISTS idx_cost_worker ON cost_records(worker_id)')
+    # V3.5.2：chat_messages 补 thinking/image；chat_sessions 补 pinned/use_kb；endpoints 补 capabilities
+    mcols = {r['name'] for r in conn.execute('PRAGMA table_info(chat_messages)')}
+    for col, ddl in (
+        ('thinking', "ALTER TABLE chat_messages ADD COLUMN thinking TEXT DEFAULT ''"),
+        ('image', "ALTER TABLE chat_messages ADD COLUMN image TEXT DEFAULT ''"),
+    ):
+        if col not in mcols:
+            conn.execute(ddl)
+    scols = {r['name'] for r in conn.execute('PRAGMA table_info(chat_sessions)')}
+    for col, ddl in (
+        ('pinned', 'ALTER TABLE chat_sessions ADD COLUMN pinned INTEGER DEFAULT 0'),
+        ('use_kb', 'ALTER TABLE chat_sessions ADD COLUMN use_kb INTEGER DEFAULT 0'),
+    ):
+        if col not in scols:
+            conn.execute(ddl)
+    ecols = {r['name'] for r in conn.execute('PRAGMA table_info(llm_endpoints)')}
+    if 'capabilities' not in ecols:
+        conn.execute("ALTER TABLE llm_endpoints ADD COLUMN capabilities TEXT DEFAULT '{}'")
+    # V3.5.2 知识库
+    conn.execute('''CREATE TABLE IF NOT EXISTS kb_documents (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        title TEXT NOT NULL,
+        content TEXT DEFAULT '',
+        tags TEXT DEFAULT '[]',
+        source TEXT DEFAULT 'manual',
+        created_at INTEGER,
+        updated_at INTEGER
+    )''')
+    conn.execute('''CREATE TABLE IF NOT EXISTS kb_chunks (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        doc_id INTEGER NOT NULL,
+        idx INTEGER DEFAULT 0,
+        content TEXT DEFAULT '',
+        tokens TEXT DEFAULT '[]'
+    )''')
+    conn.execute('CREATE INDEX IF NOT EXISTS idx_kb_chunks_doc ON kb_chunks(doc_id)')
+    conn.execute('CREATE INDEX IF NOT EXISTS idx_kb_title ON kb_documents(title)')
     conn.commit()
     conn.close()
 

kb.py

@@ -0,0 +1,140 @@
+# -*- coding: utf-8 -*-
+"""
+V3.5.2 全局知识库
+=================
+- kb_documents / kb_chunks：文档 + 分块（jieba 分词，BM25 风格检索）
+- 基本功能：增删改查、上传（txt/md/pdf）、全文检索、上下文注入（对话可选）
+"""
+import json
+import re
+import os
+
+import db
+import config
+
+KB_UPLOAD_DIR = os.path.join(config.DATA_DIR, 'kb_uploads')
+
+
+def _tok(text):
+    """jieba 分词（去停用字、只留长度>=2 的 token）"""
+    try:
+        import jieba
+        toks = []
+        for t in jieba.cut_for_search((text or '').lower()):
+            t = t.strip()
+            if len(t) >= 2 and not t.isdigit():
+                toks.append(t)
+        return toks
+    except Exception:
+        return [w for w in re.findall(r'[\u4e00-\u9fff]{2,}|[a-zA-Z0-9_]{2,}', (text or '').lower())]
+
+
+def _chunks(content, size=400, overlap=60):
+    """把文档切成小块（按段落聚合 + 超长硬切 + 前后重叠）"""
+    content = content or ''
+    paras = [p for p in re.split(r'\n+', content) if p.strip()]
+    blocks, buf = [], ''
+    for p in paras:
+        if buf and len(buf) + len(p) > size:
+            blocks.append(buf)
+            buf = ''
+        buf = (buf + '\n' + p) if buf else p
+    if buf:
+        blocks.append(buf)
+    out = []
+    for b in blocks:
+        while len(b) > size:
+            out.append(b[:size])
+            b = b[size - overlap:]
+        if b:
+            out.append(b)
+    return out or ['']
+
+
+def rebuild_chunks(doc_id):
+    doc = db.q('SELECT * FROM kb_documents WHERE id=?', (doc_id,), one=True)
+    if not doc:
+        return 0
+    db.w('DELETE FROM kb_chunks WHERE doc_id=?', (doc_id,))
+    n = 0
+    for i, c in enumerate(_chunks(doc.get('content') or '')):
+        db.w('INSERT INTO kb_chunks (doc_id, idx, content, tokens) VALUES (?,?,?,?)',
+             (doc_id, i, c, json.dumps(_tok(c))))
+        n += 1
+    return n
+
+
+def search(q, top_k=6):
+    """BM25 风格检索：返回 [{doc_id,title,content,score}]（按命中 token 数 + IDF 加权）"""
+    q_tokens = _tok(q)
+    if not q_tokens:
+        return []
+    rows = db.q('SELECT * FROM kb_chunks ORDER BY doc_id, idx')
+    if not rows:
+        return []
+    docs = {d['id']: d for d in db.q('SELECT id,title FROM kb_documents')}
+    df = {}
+    for c in rows:
+        for t in set(json.loads(c['tokens'] or '[]')):
+            df[t] = df.get(t, 0) + 1
+    n_docs = max(1, len(set(r['doc_id'] for r in rows)))
+    scored = []
+    for c in rows:
+        toks = json.loads(c['tokens'] or '[]')
+        tf = {}
+        for t in toks:
+            tf[t] = tf.get(t, 0) + 1
+        score = 0.0
+        for t in q_tokens:
+            if t in tf:
+                score += (1 + tf[t]) * max(0.1, (n_docs - df.get(t, 0) + 0.5) / (df.get(t, 0) + 0.5))
+        if score > 0:
+            scored.append({'doc_id': c['doc_id'], 'idx': c['idx'],
+                           'content': c['content'], 'score': round(score, 3),
+                           'title': docs.get(c['doc_id'], {}).get('title', '')})
+    scored.sort(key=lambda x: -x['score'])
+    return scored[:top_k]
+
+
+def build_context(q, top_k=4):
+    """把检索结果拼成可注入的上下文（用于对话/任务），返回 (ctx, hits)"""
+    hits = search(q, top_k)
+    if not hits:
+        return '', []
+    parts = []
+    for i, h in enumerate(hits):
+        parts.append(f"[{i + 1}]《{h['title']}》\n{h['content'][:900]}")
+    ctx = ('以下是与你问题相关的【知识库参考】资料（可据此回答）：\n' + '\n\n'.join(parts) + '\n\n----\n')
+    return ctx, hits
+
+
+def extract_text(filename, raw):
+    """按扩展名抽取文本：txt/md/html/csv/json；pdf 用 pypdf（有则装）。返回 (text, ok)"""
+    ext = os.path.splitext(filename)[1].lower()
+    name = filename or 'doc'
+    if ext in ('.txt', '.md', '.markdown', '.html', '.htm', '.csv', '.json', '.log', '.py', '.js', '.css'):
+        for enc in ('utf-8', 'gbk', 'utf-8-sig'):
+            try:
+                return raw.decode(enc), True
+            except Exception:
+                continue
+        return raw.decode('utf-8', errors='ignore'), True
+    if ext == '.pdf':
+        try:
+            from pypdf import PdfReader
+            import io
+            reader = PdfReader(io.BytesIO(raw))
+            text = '\n'.join((pg.extract_text() or '') for pg in reader.pages)
+            return text, bool(text.strip())
+        except Exception:
+            return '', False
+    if ext in ('.docx',):
+        try:
+            import io
+            from docx import Document
+            doc = Document(io.BytesIO(raw))
+            text = '\n'.join(p.text for p in doc.paragraphs)
+            return text, bool(text.strip())
+        except Exception:
+            return '', False
+    return '', False

llm_gateway.py

@@ -63,6 +63,58 @@ def _endpoint_pricing_map(ep):
         return {}
 
 
+# ---------------------------------------------------------------------------
+# 模型能力标签（V3.5.2）：chat/thinking/vision/audio_in/audio_out/image_gen/video_gen/embedding/rerank
+# ---------------------------------------------------------------------------
+CAP_LABELS = {
+    'chat': '💬 对话', 'thinking': '🧠 思考', 'vision': '👁️ 视觉输入',
+    'audio_in': '🎤 语音输入', 'audio_out': '🔊 语音输出', 'image_gen': '🎨 图片生成',
+    'video_gen': '🎬 视频生成', 'embedding': '🔢 Embedding', 'rerank': '🔀 Rerank',
+}
+ALL_CAPS = list(CAP_LABELS.keys())
+
+
+def endpoint_model_caps(endpoint, model):
+    """某接口下某模型的 能力标签 列表（缺省按模型名启发式推断）"""
+    if endpoint:
+        try:
+            caps_map = json.loads(endpoint.get('capabilities') or '{}') or {}
+        except Exception:
+            caps_map = {}
+        caps = list(caps_map.get(model) or [])
+        if caps:
+            return caps
+        # 兜底：模型名启发式
+        return guess_model_caps(model)
+    return guess_model_caps(model)
+
+
+def guess_model_caps(model):
+    """按模型名启发式推断能力（未显式配置时）"""
+    m = (model or '').lower()
+    caps = {'chat'}
+    if any(k in m for k in ('think', 'reason', 'r1', 'deepseek-reasoner', 'o1', 'o3', 'glm-4.6')):
+        caps.add('thinking')
+    if any(k in m for k in ('vision', 'vl', 'omni', 'qwen2.5-vl', 'glm-4v', 'llava', 'internvl', 'mini-omni')):
+        caps.add('vision')
+    if any(k in m for k in ('audio', 'voice', 'tts', 'speech', 'asr', 'whisper', 'transcri', 's2t', 'funasr', 'sensevoice')):
+        caps.add('audio_in')
+        caps.add('audio_out')
+    if any(k in m for k in ('image', 'dall', 'stable', 'flux', 'sd3', 'sora', 'wan', 'cogview', 'draw')):
+        caps.add('image_gen')
+    if any(k in m for k in ('video', 'sora', 'wan', 'kling', 'runway')):
+        caps.add('video_gen')
+    if any(k in m for k in ('embedding', 'text-embedding', 'bge')):
+        caps.add('embedding')
+    if any(k in m for k in ('rerank', 'bge-rerank')):
+        caps.add('rerank')
+    return sorted(caps)
+
+
+def model_has_cap(endpoint, model, cap):
+    return cap in endpoint_model_caps(endpoint, model)
+
+
 def worker_llm_cfg(worker):
     """解析 Worker 的大模型接口配置（V3.5）：
     优先取绑定的接口库 endpoint_id（统一鉴权/计价），worker 自带 base_url/api_key 可覆盖。
@@ -107,17 +159,23 @@ def calc_cost(model, prompt_tokens, completion_tokens):
     return round(prompt_tokens / 1e6 * pin + completion_tokens / 1e6 * pout, 6)
 
 
-def calc_cost_ex(model, prompt_tokens, completion_tokens, calls=1, cfg=None):
-    """按 Worker/接口库配置计价。cfg = worker_llm_cfg() 结果。"""
+def calc_cost_ex(model, prompt_tokens, completion_tokens, calls=1, cached_tokens=0, cfg=None):
+    """按 Worker/接口库配置计价。cfg = worker_llm_cfg() 结果。
+    - 按调用次数计费：price_per_call 单位为 元/千次，cost = price_per_call/1000 * calls
+    - 按 token 计费：逐模型 {input, input_cache, output}，缓存命中 token 按 input_cache 计价"""
     if cfg and cfg.get('endpoint'):
         ep = cfg['endpoint']
         if ep.get('billing') == 'call':
-            return round(float(ep.get('price_per_call') or 0) * max(1, calls), 6)
+            return round(float(ep.get('price_per_call') or 0) / 1000.0 * max(1, calls), 6)
         p = (cfg.get('pricing_map') or {}).get(model)
         if p:
-            pin, pout = float(p.get('input') or 0), float(p.get('output') or 0)
+            pin = float(p.get('input') or 0)
+            pcache = float(p.get('input_cache') if p.get('input_cache') is not None else p.get('cache') or 0)
+            pout = float(p.get('output') or 0)
             if pin or pout:
-                return round(prompt_tokens / 1e6 * pin + completion_tokens / 1e6 * pout, 6)
+                cached = max(0, min(cached_tokens or 0, prompt_tokens))
+                return round((prompt_tokens - cached) / 1e6 * pin + cached / 1e6 * pcache
+                             + completion_tokens / 1e6 * pout, 6)
         pin, pout = cfg.get('in_price') or 0, cfg.get('out_price') or 0
         if pin or pout:
             return round(prompt_tokens / 1e6 * pin + completion_tokens / 1e6 * pout, 6)
@@ -126,10 +184,10 @@ def calc_cost_ex(model, prompt_tokens, completion_tokens, calls=1, cfg=None):
 
 
 def worker_unit_price(worker):
-    """自动路由用：估算 Worker 单次调用成本（按 token 计费 = 输入价+0.5*输出价；按次计费 = 单价）"""
+    """自动路由用：估算 Worker 单次调用成本（按 token 计费 = 输入价+0.5*输出价；按次计费 = 单价/千次）"""
     cfg = worker_llm_cfg(worker)
     if cfg.get('endpoint') and cfg.get('billing') == 'call':
-        return float(cfg.get('price_per_call') or 0)
+        return float(cfg.get('price_per_call') or 0) / 1000.0
     p = (cfg.get('pricing_map') or {}).get(cfg['model'])
     if p:
         return float(p.get('input') or 0) + float(p.get('output') or 0) * 0.5
@@ -256,12 +314,16 @@ def _chat_stream_raw(provider, model, messages, temperature=0.7, max_tokens=None
             ch = choices[0]
             delta = ch.get('delta') or {}
             piece = delta.get('content') or ''
-            if not piece:
-                piece = delta.get('reasoning_content') or ''
             if piece:
                 if first_token_at is None:
                     first_token_at = time.time()
                 yield ('delta', piece)
+            else:
+                rp = delta.get('reasoning_content') or ''
+                if rp:
+                    if first_token_at is None:
+                        first_token_at = time.time()
+                    yield ('reasoning', rp)
             if ch.get('finish_reason'):
                 yield ('finish', ch.get('finish_reason'))
         return
@@ -316,6 +378,7 @@ def chat_stream(provider, model, messages, temperature=0.7, max_tokens=None,
     last_err = None
     for attempt in range(retries + 1):
         parts, usage = [], None
+        thinking = []
         finish_reason = None
         first_token_at = None
         t0 = time.time()
@@ -334,6 +397,10 @@ def chat_stream(provider, model, messages, temperature=0.7, max_tokens=None,
                             on_chunk(val)
                         except Exception:
                             pass
+                elif evt == 'reasoning':
+                    if first_token_at is None:
+                        first_token_at = time.time()
+                    thinking.append(val)
                 elif evt == 'usage':
                     usage = val
                 elif evt == 'finish':
@@ -350,6 +417,7 @@ def chat_stream(provider, model, messages, temperature=0.7, max_tokens=None,
             first_ms = int((first_token_at - t0) * 1000) if first_token_at else None
             return {
                 'text': text,
+                'thinking': ''.join(thinking),
                 'model': model,
                 'prompt_tokens': pt,
                 'completion_tokens': ct,
@@ -402,13 +470,74 @@ def chat_worker(worker, messages, temperature=None, max_tokens=None, on_chunk=No
                     max_tokens=max_tokens or worker.get('max_tokens') or 2000,
                     base_url=cfg['base_url'] or None, api_key=cfg['api_key'] or None,
                     on_chunk=on_chunk)
-    r['cost'] = calc_cost_ex(cfg['model'], r['prompt_tokens'], r['completion_tokens'], 1, cfg)
+    r['cost'] = calc_cost_ex(cfg['model'], r['prompt_tokens'], r['completion_tokens'], 1,
+                             r.get('cached_tokens', 0), cfg)
     r['worker_id'] = worker['id']
     r['provider'] = cfg['provider']
     r['model'] = cfg['model']
     return r
 
 
+def stream_chat(cfg, messages, temperature=0.7, max_tokens=None, extra_payload=None):
+    """对话专用的 实时流式生成器（V3.5.2）——真正边收边吐，思考与回答分开。
+    cfg: {'provider','model','base_url','api_key','endpoint'}（worker 或 接口 目标解析结果）
+    yield ('reasoning', piece) | ('delta', piece) | ('done', result) | ('error', msg)
+    result 含 thinking/text/usage/cost(含缓存命中价与按次计费)/first_token_ms/elapsed_ms。"""
+    provider = cfg['provider']
+    model = cfg['model']
+    if not model:
+        yield ('error', '未选择模型')
+        return
+    tk, fk, rk = db.get_llm_timeouts()
+    parts, thinking, usage = [], [], None
+    finish_reason = None
+    first_token_at = None
+    t0 = time.time()
+    tried = False
+    while True:
+        try:
+            for evt, val in _chat_stream_raw(
+                    provider, model, messages, temperature=temperature, max_tokens=max_tokens,
+                    base_url=cfg.get('base_url') or None, api_key=cfg.get('api_key') or None,
+                    token_timeout=tk, first_token_timeout=fk, extra_payload=extra_payload):
+                if evt == 'reasoning':
+                    if first_token_at is None:
+                        first_token_at = time.time()
+                    thinking.append(val)
+                    yield ('reasoning', val)
+                elif evt == 'delta':
+                    if first_token_at is None:
+                        first_token_at = time.time()
+                    parts.append(val)
+                    yield ('delta', val)
+                elif evt == 'usage':
+                    usage = val
+                elif evt == 'finish':
+                    finish_reason = val
+            break
+        except LLMError as e:
+            if parts or tried:
+                yield ('error', str(e))
+                return
+            tried = True  # 无任何输出时允许重试一次
+            continue
+    text = ''.join(parts)
+    pt, ct, cached = _usage_fields(usage)
+    usage_estimated = usage is None
+    if usage is None:
+        pt, ct = _estimate_prompt_tokens(messages), _estimate_completion_tokens(text)
+    elapsed_ms = int((time.time() - t0) * 1000)
+    first_ms = int((first_token_at - t0) * 1000) if first_token_at else None
+    cost = calc_cost_ex(model, pt, ct, 1, cached, cfg) if cfg.get('endpoint') else calc_cost(model, pt, ct)
+    yield ('done', {
+        'text': text, 'thinking': ''.join(thinking), 'model': model,
+        'prompt_tokens': pt, 'completion_tokens': ct, 'total_tokens': pt + ct,
+        'cached_tokens': cached, 'cost': cost, 'calls': 1,
+        'first_token_ms': first_ms, 'elapsed_ms': elapsed_ms,
+        'usage_estimated': usage_estimated, 'finish_reason': finish_reason,
+    })
+
+
 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):

static/index.html

@@ -11,8 +11,10 @@
   <aside id="sidebar">
     <div class="logo">🤖 <span>AI Worker</span><small>项目管理平台</small></div>
     <nav>
-      <a href="#/dashboard" data-route="dashboard" data-perm="dashboard.view" class="nav-chat">💬 对话 · 仪表盘</a>
+      <a href="#/chat" data-route="chat" data-perm="dashboard.view" class="nav-chat">💬 对话</a>
+      <a href="#/dashboard" data-route="dashboard" data-perm="dashboard.view">📊 仪表盘</a>
       <a href="#/projects" data-route="projects" data-perm="project.view">📁 项目</a>
+      <a href="#/kb" data-route="kb" data-perm="project.view">📚 知识库</a>
       <a href="#/workers" data-route="workers" data-perm="worker.view">🧑‍💻 AI Worker</a>
       <a href="#/agents" data-route="agents" data-perm="agent.view">🤝 多 Agent 协作</a>
       <a href="#/eval" data-route="eval" data-perm="eval.view">🎯 自动评估</a>

static/style.css

@@ -277,6 +277,7 @@ code{background:var(--panel2);border:1px solid var(--border);border-radius:6px;p
 
 /* ---------- V3.5 对话 ---------- */
 .chat-card{display:flex;flex-direction:column;max-height:560px;padding:14px}
+.chat-card.chat-tall{max-height:72vh;margin-top:14px}
 .chat-top{display:flex;justify-content:space-between;align-items:center;gap:10px;flex-wrap:wrap;margin-bottom:10px}
 .chat-title{font-size:16px;font-weight:700}
 .chat-sess{display:flex;gap:6px;align-items:center}
@@ -285,6 +286,7 @@ code{background:var(--panel2);border:1px solid var(--border);border-radius:6px;p
 .chat-target select{width:auto;max-width:240px}
 .chat-target-label{font-size:12px;color:var(--accent);font-weight:600}
 .chat-body{flex:1;min-height:180px;max-height:330px;overflow-y:auto;background:var(--bg);border:1px solid var(--border);border-radius:10px;padding:12px}
+.chat-card.chat-tall .chat-body{max-height:none}
 .chat-msg{display:flex;margin-bottom:10px}
 .chat-msg.user{justify-content:flex-end}
 .chat-msg.ai{justify-content:flex-start;flex-direction:column;align-items:flex-start}
@@ -292,8 +294,42 @@ code{background:var(--panel2);border:1px solid var(--border);border-radius:6px;p
 .chat-bubble.user{background:var(--accent);color:#fff;border-bottom-right-radius:3px}
 .chat-bubble.ai{background:var(--panel2);border:1px solid var(--border);border-bottom-left-radius:3px}
 .chat-bubble.ai.err{border-color:var(--danger)}
+.chat-cursor{display:inline-block;width:8px;height:15px;background:var(--accent);margin-left:2px;vertical-align:-2px;animation:chatblink .8s steps(1) infinite}
+@keyframes chatblink{50%{opacity:0}}
 .chat-usage{font-size:11px;color:var(--muted);margin-top:4px;font-family:ui-monospace,Consolas,monospace}
 .chat-input{display:flex;gap:10px;margin-top:10px;align-items:flex-end}
 .chat-input textarea{flex:1;resize:vertical}
 .ref-card{border-left:3px solid var(--accent)}
 .ref-card:hover{border-color:var(--accent)}
+.chat-stats{display:grid;grid-template-columns:repeat(5,1fr);gap:10px}
+.chat-stat-link{cursor:pointer;transition:border .15s, transform .1s}
+.chat-stat-link:hover{border-color:var(--accent);transform:translateY(-2px)}
+@media(max-width:1100px){.chat-stats{grid-template-columns:repeat(3,1fr)}}
+
+/* V3.5.2 对话：历史列表 + 主体 */
+.chat-wrap{display:grid;grid-template-columns:250px 1fr;gap:12px;min-height:0}
+.chat-hist{background:var(--bg);border:1px solid var(--border);border-radius:10px;padding:10px;overflow-y:auto;max-height:340px}
+.chat-card.chat-tall .chat-hist{max-height:56vh}
+.chat-hist-title{font-size:12px;color:var(--muted);font-weight:600;padding:2px 4px 8px;border-bottom:1px solid var(--border);margin-bottom:8px}
+.chat-hist-item{position:relative;padding:8px 10px;border-radius:8px;cursor:pointer;margin-bottom:4px;border:1px solid transparent}
+.chat-hist-item:hover{background:var(--panel2)}
+.chat-hist-item.on{background:var(--panel2);border-color:var(--accent)}
+.chat-hist-item .t{font-size:13px;font-weight:600;padding-right:56px;white-space:nowrap;overflow:hidden;text-overflow:ellipsis}
+.chat-hist-item .m{font-size:11px;color:var(--muted);margin-top:3px;white-space:nowrap;overflow:hidden;text-overflow:ellipsis}
+.chat-hist-item .ops{position:absolute;top:6px;right:6px;display:none;gap:2px}
+.chat-hist-item:hover .ops{display:flex}
+.chat-hist-item .op{background:var(--panel2);border:1px solid var(--border);border-radius:6px;color:var(--muted);font-size:12px;padding:2px 5px;cursor:pointer}
+.chat-hist-item .op:hover{color:var(--accent);border-color:var(--accent)}
+.chat-main{display:flex;flex-direction:column;min-width:0}
+.chat-caps{display:inline-flex;gap:4px;flex-wrap:wrap}
+.chat-attach{background:var(--panel2);border:1px solid var(--accent);border-radius:8px;padding:4px 10px;font-size:12px;align-self:center;white-space:nowrap}
+.chat-thinking{background:var(--panel2);border:1px solid var(--border);border-left:3px solid var(--purple);border-radius:8px;padding:6px 10px;margin-bottom:8px;max-width:78%}
+.chat-thinking summary{cursor:pointer;font-size:12px;color:var(--purple);font-weight:600}
+.chat-thinking-body{font-size:12px;color:var(--muted);margin-top:6px;max-height:140px;overflow-y:auto;white-space:pre-wrap;word-break:break-word}
+
+/* V3.5.2 大模型接口：模型编辑器 */
+.ep-models{border:1px solid var(--border);border-radius:10px;padding:8px;max-height:300px;overflow-y:auto;margin-bottom:8px}
+.ep-mrow{display:grid;grid-template-columns:1.2fr 0.8fr 1fr 0.8fr 2.6fr;gap:6px;align-items:center;padding:4px 2px;border-bottom:1px dashed var(--border)}
+.ep-mrow.ep-head{position:sticky;top:0;background:var(--panel);font-size:12px;color:var(--muted);z-index:2}
+.ep-mrow input{font-size:12px;padding:5px 7px}
+.ep-mrow .c-caps{display:flex;flex-wrap:wrap;gap:2px}