feat: improve conductor files and runtime handling

agentmain.py

@@ -110,6 +110,9 @@ class GenericAgent:
         if not self.is_running: return
         print('Abort current task...')
         self.stop_sig = True
+        for sess in getattr(self.llmclient.backend, '_sessions', [self.llmclient.backend]):
+            try: sess.active_response.close()
+            except Exception: pass
         if self.handler is not None: self.handler.code_stop_signal.append(1)
             
     def put_task(self, query, source="user", images=None):

frontends/conductor.html

@@ -218,6 +218,7 @@
 
     .composer { padding: 14px; border-top: 1px solid rgba(31,42,68,.075); display: flex; gap: 12px;
                 align-items: end; background: rgba(255,255,255,.40); }
+    .composer.dragging { background: rgba(59,130,246,.13); outline: 2px dashed #3b82f6; outline-offset: -6px; }
     textarea#input { flex: 1; height: 52px; max-height: 140px; resize: none; outline: none; color: var(--text);
                background: rgba(255,255,255,.92); border: 1px solid rgba(31,42,68,.12); border-radius: 18px;
                padding: 14px 15px; font: inherit; line-height: 1.35; }
@@ -225,6 +226,19 @@
     button { height: 52px; min-width: 92px; border: 0; border-radius: 18px; color: white; font-weight: 750;
              cursor: pointer; background: linear-gradient(135deg, #3b82f6, #10b981);
              box-shadow: 0 10px 26px rgba(59,130,246,.22); }
+    .file-open, .attach { min-width: 0; padding: 0 10px; box-shadow: none; }
+    .file-open { height: 28px; border-radius: 9px; }
+    .attach { width: 44px; height: 52px; font-size: 18px; border-radius: 18px; }
+    .file-link { color: inherit; text-decoration: underline; cursor: pointer; }
+    #fileDialog { width: min(1100px, 96vw); height: min(800px, 92vh); padding: 0; border: 0; border-radius: 18px; box-shadow: 0 24px 80px #0004; }
+    #fileDialog::backdrop { background: #0f172a66; backdrop-filter: blur(3px); }
+    .file-bar { display: flex; align-items: center; gap: 8px; padding: 12px; border-bottom: 1px solid var(--border); }
+    .file-bar code { flex: 1; overflow: hidden; text-overflow: ellipsis; white-space: nowrap; }
+    .file-bar button { min-width: 0; height: 34px; padding: 0 12px; border-radius: 10px; box-shadow: none; }
+    #fileBody { height: calc(100% - 59px); overflow: auto; padding: 12px; }
+    .file-row { display: flex; gap: 10px; padding: 9px 11px; border-radius: 9px; cursor: pointer; }
+    .file-row:hover { background: #eef2ff; } .file-row span:first-child { flex: 1; } .file-size { color: var(--muted); }
+    .file-preview { line-height: 1.55; } .file-preview pre { white-space: pre-wrap; overflow-wrap: anywhere; }
 
     /* ---------- conductor 输出(横跨底部的独立卡片,默认收起) ---------- */
     .console {
@@ -314,18 +328,29 @@
       <section class="pane chat">
         <div class="head">
           <div class="title">Conversation <select id="llm" aria-label="Conductor model" disabled></select></div>
-          <div class="hint" id="wsStatus">connecting...</div>
+          <div class="hint"><button class="file-open" id="fileOpen">📁 文件</button> · <span id="wsStatus">connecting...</span></div>
         </div>
         <div class="messages" id="messages"></div>
-        <div class="composer">
-          <textarea id="input" placeholder="告诉 Conductor 你想做什么……"></textarea>
-          <button id="send">发送</button>
+        <div class="composer" id="composer">
+          <textarea id="input" placeholder="告诉 Conductor 你想做什么……（可拖入文件）"></textarea>
+          <button class="attach" id="fileUpload" title="上传文件">📎</button>
+          <button id="send">发送</button><input id="fileInput" type="file" multiple hidden>
         </div>
       </section>
     </section>
     </div><!-- /app-grid -->
   </main>
 
+  <dialog id="fileDialog">
+    <div class="file-bar"><button id="fileBack">↑</button><code id="filePath"></code><button id="fileClose">关闭</button></div>
+    <div id="fileBody"></div>
+  </dialog>
+  <dialog id="downloadDialog" style="width:min(520px,90vw);border:0;border-radius:16px;padding:22px;box-shadow:0 24px 80px #0004">
+    <strong>下载文件</strong>
+    <code id="downloadPath" style="display:block;margin:16px 0;overflow-wrap:anywhere;color:var(--muted)"></code>
+    <div style="display:flex;justify-content:flex-end;gap:10px"><button id="downloadCancel">取消</button><button id="downloadConfirm">下载</button></div>
+  </dialog>
+
   <script src="https://cdn.jsdelivr.net/npm/marked/marked.min.js"></script>
   <script>
     const cardsEl    = document.getElementById('cards');
@@ -341,10 +366,48 @@
     const consoleBar = document.getElementById('consoleBar');
     const consoleBody= document.getElementById('consoleBody');
     const logCount   = document.getElementById('logCount');
-    let ws;
+    const fileDialog = document.getElementById('fileDialog');
+    const fileBody   = document.getElementById('fileBody');
+    const filePathEl = document.getElementById('filePath');
+    const fileInput  = document.getElementById('fileInput');
+    const composer   = document.getElementById('composer');
+    const downloadDialog = document.getElementById('downloadDialog');
+    const downloadPath = document.getElementById('downloadPath');
+    let ws, fileDir='', fileParent='';
 
     function esc(s){ return String(s||'').replace(/[&<>"]/g,c=>({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;'}[c])); }
     function renderMd(s){ try { return marked.parse(s||'', { breaks: true, gfm: true }); } catch(e){ return esc(s); } }
+    function fileUrl(path){ return '/file?path='+encodeURIComponent(path); }
+    function humanSize(n){ return n==null?'':n<1024?n+' B':n<1048576?(n/1024).toFixed(1)+' KB':(n/1048576).toFixed(1)+' MB'; }
+    async function browseFiles(path=''){
+      const data=await fetch(path?'/files?path='+encodeURIComponent(path):'/files').then(r=>{if(!r.ok) throw Error(r.statusText); return r.json()});
+      fileDir=data.path; fileParent=data.parent; filePathEl.textContent=fileDir;
+      fileBody.innerHTML=data.items.map(x=>`<div class="file-row" data-path="${esc(x.path)}" data-dir="${x.dir?'1':''}"><span>${x.dir?'📁':'📄'} ${esc(x.name)}</span><span class="file-size">${humanSize(x.size)}</span></div>`).join('')||'<div class="empty">空目录</div>';
+    }
+    async function openFile(path){
+      if(!/\.(md|markdown|txt|py|js|ts|json|ya?ml|css|html?|xml|csv|log|ini|cfg|sh|ps1|bat|c|cc|cpp|h|hpp|java|go|rs|sql)$/i.test(path)){
+        downloadPath.textContent=path; downloadDialog.showModal();
+        return;
+      }
+      const text=await fetch(fileUrl(path)).then(r=>{if(!r.ok) throw Error(r.statusText); return r.text()});
+      filePathEl.textContent=path;
+      fileBody.innerHTML=/\.(md|markdown)$/i.test(path)?`<div class="file-preview">${renderMd(text)}</div>`:'';
+      if(!/\.(md|markdown)$/i.test(path)){ const pre=document.createElement('pre'); pre.textContent=text; fileBody.replaceChildren(pre); }
+    }
+    document.getElementById('downloadCancel').onclick=()=>downloadDialog.close();
+    document.getElementById('downloadConfirm').onclick=()=>{
+      const a=document.createElement('a'); a.href=fileUrl(downloadPath.textContent); a.download=downloadPath.textContent.split(/[\\/]/).pop();
+      document.body.appendChild(a); a.click(); a.remove(); downloadDialog.close();
+    };
+    async function uploadFiles(files){
+      const paths=[];
+      for(const file of files){
+        const r=await fetch('/file?name='+encodeURIComponent(file.name),{method:'POST',body:file});
+        if(!r.ok) throw Error(await r.text());
+        paths.push('`'+(await r.json()).path+'`');
+      }
+      inputEl.value+=(inputEl.value?'\n':'')+paths.join('\n'); checkLimit(); inputEl.focus();
+    }
 
     /* ---------- subagent 卡片 ---------- */
     function renderCards(items){
@@ -414,11 +477,28 @@
       div.className = `msg ${cls}` + (role==='user' && !item.read ? ' unread' : '');
       if(item.id) div.dataset.id = item.id;
       div.innerHTML = renderMd(item.msg || '');
+      (item.files||[]).forEach(f=>{ const code=[...div.querySelectorAll('code')].find(x=>x.textContent===(f.ref||f.path)); if(code){ const a=document.createElement('a'); a.className='file-link'; a.href=fileUrl(f.path); a.dataset.path=f.path; code.replaceWith(a); a.append(code); } });
       messagesEl.appendChild(div);
       messagesEl.scrollTop = messagesEl.scrollHeight;
     }
     function markAllRead(){ messagesEl.querySelectorAll('.msg.me.unread').forEach(el=>el.classList.remove('unread')); }
 
+    document.getElementById('fileOpen').onclick=()=>{ fileDialog.showModal(); browseFiles().catch(e=>fileBody.textContent=e); };
+    document.getElementById('fileClose').onclick=()=>fileDialog.close();
+    document.getElementById('fileBack').onclick=()=>browseFiles(filePathEl.textContent===fileDir?fileParent:fileDir);
+    document.getElementById('fileUpload').onclick=()=>fileInput.click();
+    fileBody.addEventListener('click', e=>{ const row=e.target.closest('.file-row'); if(row) (row.dataset.dir?browseFiles:openFile)(row.dataset.path); });
+    messagesEl.addEventListener('click', async e=>{
+      const a=e.target.closest('.file-link'); if(!a) return;
+      e.preventDefault(); fileDialog.showModal();
+      const path=a.dataset.path, parent=path.replace(/[\\\/][^\\\/]*$/, '');
+      try { await browseFiles(parent); await openFile(path); } catch(err) { fileBody.textContent=err; }
+    });
+    fileInput.onchange=async()=>{ try { await uploadFiles(fileInput.files); } catch(e) { alert('上传失败：'+e); } fileInput.value=''; };
+    ['dragenter','dragover'].forEach(type=>composer.addEventListener(type,e=>{ e.preventDefault(); composer.classList.add('dragging'); }));
+    ['dragleave','drop'].forEach(type=>composer.addEventListener(type,e=>{ e.preventDefault(); composer.classList.remove('dragging'); }));
+    composer.addEventListener('drop',e=>{ if(e.dataTransfer.files.length) uploadFiles(e.dataTransfer.files).catch(err=>alert('上传失败：'+err)); });
+
     /* ---------- conductor 输出 ---------- */
     function fmtTs(ms){ try{ return new Date(ms).toLocaleTimeString('zh-CN',{hour12:false}); }catch(e){ return ''; } }
     function logItemHtml(r){

frontends/conductor.py

@@ -1,9 +1,9 @@
-import os, sys, re, time, json, uuid, queue, asyncio, threading
+import os, sys, re, time, json, uuid, queue, asyncio, threading, argparse, base64, secrets
 from dataclasses import dataclass, field
 from typing import Dict, Any, Optional, List
 from contextlib import asynccontextmanager, suppress
 
-from fastapi import FastAPI, WebSocket, WebSocketDisconnect
+from fastapi import FastAPI, WebSocket, WebSocketDisconnect, Request, HTTPException
 from fastapi.responses import FileResponse, PlainTextResponse, JSONResponse
 from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
@@ -27,6 +27,15 @@ HOST = "127.0.0.1"
 PORT = 8900
 HTML_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), "conductor.html")
 
+parser = argparse.ArgumentParser()
+parser.add_argument("--host", default=HOST)
+parser.add_argument("--port", type=int, default=PORT)
+parser.add_argument("--key")
+parser.add_argument("--no-browser", action="store_true")
+args = parser.parse_args()
+if args.host not in ("127.0.0.1", "localhost", "::1") and not args.key:
+    parser.error("--key is required for non-loopback listening")
+
 
 def _settings_doc() -> dict:
     try:
@@ -112,6 +121,19 @@ async def lifespan(app: FastAPI):
 app = FastAPI(title="Conductor", lifespan=lifespan)
 app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
 
+class RemoteAuth:
+    def __init__(self, app): self.app = app
+    async def __call__(self, scope, receive, send):
+        remote = (scope.get("client") or ("",))[0]
+        got = dict(scope.get("headers", [])).get(b"authorization", b"")
+        want = b"Basic " + base64.b64encode(f"conductor:{args.key}".encode())
+        if args.key and remote not in ("127.0.0.1", "::1") and not secrets.compare_digest(got, want):
+            if scope["type"] == "websocket": return await send({"type": "websocket.close", "code": 1008})
+            return await PlainTextResponse("Unauthorized", 401, {"WWW-Authenticate": 'Basic realm="Conductor"'})(scope, receive, send)
+        await self.app(scope, receive, send)
+
+app.add_middleware(RemoteAuth)
+
 class ChatIn(BaseModel):
     msg: str
     role: str = "conductor"  # conductor | system | user
@@ -199,8 +221,23 @@ async def broadcast(payload: dict):
 
 def push_cards(): schedule_broadcast({"type": "subagents", "items": pool.snapshot()})
 
+FILE_ROOT = os.path.dirname(ROOT)
+FILE_HOME = os.path.join(ROOT, "temp")
+UPLOAD_DIR = os.path.join(FILE_HOME, "uploads")
+
+def file_path(path: str):
+    path = os.path.abspath(os.path.expanduser(path))
+    if os.path.commonpath((FILE_ROOT, path)) != FILE_ROOT: raise ValueError("path is above file browser root")
+    return path
+
 def add_chat(msg: str, role: str = "conductor", files: list = None, images: list = None):
-    item = {"id": short_id(), "role": role, "msg": msg, "ts": now_ms(), "read": role != "user", "files": files or [], "images": images or []}
+    files = list(files or [])
+    known = {x.get("path") for x in files}
+    for ref in re.findall(r"`([^`\n]+)`", msg):
+        try: path = file_path(ref)
+        except ValueError: continue
+        if os.path.isfile(path) and path not in known: files.append({"ref": ref, "path": path, "name": os.path.basename(path)})
+    item = {"id": short_id(), "role": role, "msg": msg, "ts": now_ms(), "read": role != "user", "files": files, "images": images or []}
     chat_messages.append(item)
     if len(chat_messages) > 200: del chat_messages[:-200]
     schedule_broadcast({"type": "chat", "item": item})
@@ -318,6 +355,9 @@ Conductor API\tBase: http://{HOST}:{PORT}
 POST /chat\tbody: {{"msg": "..."}}\t给用户发消息
 POST /subagent\tbody: {{"prompt": "..."}}\t启动新subagent，返回 {{"id": "xxx"}}；指定模型加参数llm(数字/名称)
 POST /approval\tbody: {{"prompt": "...", "source": "..."}}\t推一条待批任务到前端(后端不存)，用户同意则直接派发为subagent
+GET /files?path=...\t列举目录（默认temp，最高到GA上一级）
+GET /file?path=...\t下载文件
+POST /file?name=...\t上传到temp/uploads
 POST /subagent/{{id}}\tbody: {{"action": "keyinfo", "msg": "..."}}\t注入key_info（agent下轮可见）
 POST /subagent/{{id}}\tbody: {{"action": "input", "msg": "..."}}\t对subagent输入下一条指令；指定模型加参数llm(数字/名称)
 POST /subagent/{{id}}\tbody: {{"action": "stop"}}\t中断执行但保留（可继续input）
@@ -381,7 +421,8 @@ API: {base}；requests，GET /readme查用法，GET /chat读未读对话，GET /
 - 改写prompt时严禁添加用户未提及的假设、工具、前提条件。只能精炼/结构化用户原意，不能脑补，只能做很小的改写
 
 原则：
-- 信任subagent足够聪明，不要写具体步骤和容易探测的信息；能自己判断的自己判断，只在真正需要用户决策时打扰。\n
+- 信任subagent足够聪明，不要写具体步骤和容易探测的信息；能自己判断的自己判断，只在真正需要用户决策时打扰。
+- 向用户交付文件时，路径用反引号包裹。\n
 需要处理：
 {summary}"""
 
@@ -492,6 +533,28 @@ def conductor_token_stats():
 @app.get("/")
 def index(): return FileResponse(HTML_PATH)
 
+@app.get("/files")
+def files(path: str = FILE_HOME):
+    try: path = file_path(path)
+    except ValueError as e: raise HTTPException(403, str(e))
+    if not os.path.isdir(path): raise HTTPException(404, "directory not found")
+    items = [{"name": x.name, "path": x.path, "dir": x.is_dir(), "size": x.stat().st_size if x.is_file() else None}
+             for x in sorted(os.scandir(path), key=lambda x: (not x.is_dir(), x.name.lower()))]
+    return {"path": path, "parent": path if path == FILE_ROOT else os.path.dirname(path), "items": items}
+
+@app.get("/file")
+def get_file(path: str):
+    try: path = file_path(path)
+    except ValueError as e: raise HTTPException(403, str(e))
+    return FileResponse(path, filename=os.path.basename(path))
+
+@app.post("/file")
+async def put_file(name: str, request: Request):
+    os.makedirs(UPLOAD_DIR, exist_ok=True)
+    path = os.path.join(UPLOAD_DIR, os.path.basename(name))
+    with open(path, "wb") as f: f.write(await request.body())
+    return {"name": os.path.basename(path), "path": path}
+
 @app.get("/readme")
 def readme(): return PlainTextResponse(READMES["api"])
 
@@ -591,7 +654,7 @@ if __name__ == "__main__":
     import uvicorn
     # bridge 自启 conductor 时传 --no-browser:不在用户浏览器里弹一个独立 conductor UI,
     # 用户从桌面版「指挥家」页直接连过来即可。手动跑 conductor.py(没带 flag)保持原行为。
-    if "--no-browser" not in sys.argv:
+    if not args.no_browser:
         import webbrowser, threading
-        threading.Timer(1.0, lambda: webbrowser.open(f"http://{HOST}:{PORT}")).start()
-    uvicorn.run("conductor:app", host=HOST, port=PORT, reload=False)
+        threading.Timer(1.0, lambda: webbrowser.open(f"http://127.0.0.1:{args.port}")).start()
+    uvicorn.run(app, host=args.host, port=args.port)

frontends/stapp.py

@@ -300,8 +300,25 @@ def _poll_main_task(max_items=256):
     return done
 
 
+def _render_stat_badge(is_running, target=None):
+    """Render task usage/time into one stable element, without clearing its container."""
+    if 'task_start_ts' not in st.session_state or not hasattr(llmcore, 'STATS'):
+        return
+    end_ts = time.time() if is_running else st.session_state.get('task_end_ts', time.time())
+    secs = max(0, int(end_ts - st.session_state.task_start_ts))
+    stats = dict(llmcore.STATS)
+    short = lambda n: f'{n / 1000:.0f}k' if n >= 1000 else str(n)
+    usage = ((f"{stats['session']} │ " if stats.get('session') else '') +
+             f"{short(stats['ctx'])} chars·{stats['msgs']}msgs │ "
+             f"in {short(stats.get('inp', 0))} toks·cached{short(stats.get('cached', 0))}·out{short(stats.get('out', 0))} │ "
+             if 'ctx' in stats else '')
+    renderer = target.markdown if target is not None else st.markdown
+    renderer(f'<div class="ga-stat-badge">{usage}{secs // 60}:{secs % 60:02d}</div>',
+             unsafe_allow_html=True)
+
+
 @st.fragment(run_every=timedelta(seconds=1))
-def render_main_stream(frozen_host, live_slot, render_state):
+def render_main_stream(frozen_host, live_slot, stats_slot, render_state):
     """Append completed turns outside the fragment; redraw only the active turn."""
     done = _poll_main_task()
     if done is not None:
@@ -312,7 +329,7 @@ def render_main_stream(frozen_host, live_slot, render_state):
             if st.session_state.get('loop_enabled'):
                 b = get_controller()
                 b['obj'] = st.session_state.get('loop_prompt_input', ''); b['ready'] = False; b['job'] = b['epoch']; b['ev'].set()
-        # The final response is rendered once from history on the full-app rerun.
+        # The final response and frozen stats are rendered on the full-app rerun.
         st.rerun(scope="app")
 
     response = st.session_state.get("partial_response", "")
@@ -324,17 +341,18 @@ def render_main_stream(frozen_host, live_slot, render_state):
         with frozen_host:
             render_segments([segments[render_state['frozen']]])
         render_state['frozen'] += 1
-    # This external slot keeps a stable delta path; only the growing final segment changes.
+    # These external slots keep stable delta paths; only their contents change.
     with live_slot.container():
         render_segments([segments[-1]], suffix=" ▌")
+    _render_stat_badge(is_running=True, target=stats_slot)
 
 
-def mount_main_stream():
+def mount_main_stream(stats_slot):
     """Create persistent targets captured by the fragment until the next full rerun."""
     with st.chat_message("assistant"):
         frozen_host = st.container()
         live_slot = st.empty()
-        render_main_stream(frozen_host, live_slot, {'frozen': 0})
+        render_main_stream(frozen_host, live_slot, stats_slot, {'frozen': 0})
 
 if not hasattr(agent, "_ui_messages"): agent._ui_messages = st.session_state.get("messages", [])
 if "messages" not in st.session_state: st.session_state.messages = agent._ui_messages
@@ -380,6 +398,10 @@ _js_ime_fix = ("" if os.name == 'nt' else
     "f();new MutationObserver(f).observe(d.body,{childList:1,subtree:1})}()")
 _embed_html(f'<script>{_js_ime_fix}</script>', height=0)
 
+# Fixed target shared by full-app rendering and the one-second stream fragment.
+# The badge is position:fixed, so this placeholder does not affect its visual position.
+stats_slot = st.empty()
+
 _injected = st.session_state.pop('_inject_prompt', None)
 prompt = st.chat_input("any task?") or _injected
 if prompt:
@@ -454,33 +476,16 @@ if prompt:
     if hasattr(agent, '_pet_req') and not prompt.startswith('/'): agent._pet_req('state=walk')
     with st.chat_message("user"): st.markdown(prompt)
     _start_main_task(prompt)
-    mount_main_stream()
+    mount_main_stream(stats_slot)
 elif st.session_state.get('display_queue') is not None:
     # No new prompt but a task is mid-flight (typically a /btw rerun) — resume drain.
-    mount_main_stream()
+    mount_main_stream(stats_slot)
 
-# Usage/time label remains after completion; while running only its fragment updates once/sec.
+# Usage/time label remains after completion. During a task render_main_stream refreshes it.
 _has_task_stats = 'task_start_ts' in st.session_state
 _is_running = st.session_state.get('display_queue') is not None
-if _has_task_stats:
-    def _render_stat_badge():
-        if not hasattr(llmcore, 'STATS'): return
-        end_ts = time.time() if _is_running else st.session_state.get('task_end_ts', time.time())
-        secs = max(0, int(end_ts - st.session_state.task_start_ts))
-        stats = dict(llmcore.STATS)
-        short = lambda n: f'{n / 1000:.0f}k' if n >= 1000 else str(n)
-        usage = ((f"{stats['session']} │ " if stats.get('session') else '') +
-                 f"{short(stats['ctx'])} chars·{stats['msgs']}msgs │ "
-                 f"in {short(stats.get('inp', 0))} toks·cached{short(stats.get('cached', 0))}·out{short(stats.get('out', 0))} │ "
-                 if 'ctx' in stats else '')
-        st.markdown(f'<div class="ga-stat-badge">{usage}{secs // 60}:{secs % 60:02d}</div>',
-                    unsafe_allow_html=True)
-    if _is_running:
-        @st.fragment(run_every=timedelta(seconds=1))
-        def _live_stat_badge(): _render_stat_badge()
-        _live_stat_badge()
-    else:
-        _render_stat_badge()
+if _has_task_stats and not _is_running:
+    _render_stat_badge(is_running=False, target=stats_slot)
 
 if _has_task_stats or _is_running:
     st.markdown(

ga.py

@@ -312,7 +312,7 @@ class GenericAgentHandler(BaseHandler):
         cwd = os.path.normpath(os.path.abspath(raw_path))
         code_cwd = os.path.normpath(self.cwd)
         maxlen = self._get_tool_maxlen(10000, args)
-        if timeout > 600: result = '[ERROR] Timeout must be <= 600 seconds, run time-consuming code in background subprocess and monitor it!'
+        if timeout > 600: result = '[ERROR] Timeout must be <= 600 seconds; code not executed. Run time-consuming code in the background instead of waiting for it to finish in the foreground, verify it started successfully, and monitor it until completion or failure.'
         elif code_type == 'python' and _arg(args, "inline_eval", False, bool):
             ns = {'handler':self, 'parent':self.parent, 'history':json.dumps(self.parent.llmclient.backend.history)}
             old_cwd = os.getcwd()

llmcore.py

@@ -174,7 +174,9 @@ def _parse_claude_sse(resp_lines):
                 if current_block and current_block.get("type") == "text": current_block["text"] += text
                 if text: yield text
             elif delta.get("type") == "thinking_delta":
-                if current_block and current_block.get("type") == "thinking": current_block["thinking"] += delta.get("thinking", "")
+                thinking = delta.get("thinking", "")
+                if current_block and current_block.get("type") == "thinking": current_block["thinking"] += thinking
+                if thinking: yield thinking
             elif delta.get("type") == "signature_delta":
                 if current_block and current_block.get("type") == "thinking":
                     current_block["signature"] = current_block.get("signature", "") + delta.get("signature", "")
@@ -396,6 +398,7 @@ def _stream_with_retry(sess, url, headers, payload, parse_fn):
         try:
             with requests.post(url, headers=headers, json=payload, stream=sess.stream, 
                                timeout=(sess.connect_timeout, sess.read_timeout), proxies=sess.proxies, verify=sess.verify) as r:
+                sess.active_response = r
                 if r.status_code >= 400:
                     #pathlib.Path(__file__).parent.joinpath('temp','bad_requests.json').write_text(json.dumps({"url":url,"headers":headers,"payload":payload,"t":time.time()},ensure_ascii=False),encoding='utf-8')
                     d = _delay(r, attempt) if r.status_code in _RETRYABLE and attempt < sess.max_retries else None