codebee 0.1.56 → 0.1.57

package/app/core/runner.py

@@ -22,15 +22,15 @@ import tempfile
 import threading
 import time
 
+from . import beekeeper
 from .env_scrub import scrub_env
 from .error_codes import ErrorCode
 
 # 非 Windows 必须置 0：POSIX 的 Popen 对非零 creationflags 直接抛 ValueError，
 # 置 0 则两边通用（remote.py 同款守卫）。
 CREATE_NO_WINDOW = 0x08000000 if os.name == "nt" else 0
-DEFAULT_TIMEOUT = 1200  # 单步 20 分钟
-MAX_ATTEMPT_TIMEOUT_S = 60  # 无进展时单模型最多等 1 分钟
-DEFAULT_STREAM_ACTIVITY_TIMEOUT_S = 180  # 流式 CLI 的活动延长窗口
+DEFAULT_TIMEOUT = 1200  # 单步 20 分钟
+DEFAULT_STREAM_ACTIVITY_TIMEOUT_S = 180  # 流式 CLI 的活动延长窗口
 
 _BASH_CANDIDATES = [
     r"D:\Git\usr\bin\bash.exe",
@@ -219,6 +219,16 @@ def _drain_streams(proc, t_out, t_err, *, timeout=10):
         remaining = max(0.0, deadline - time.time())
         if remaining > 0:
             t.join(timeout=remaining)
+    # Killing the tree does not necessarily reap the root process on Windows.
+    # Reap it with the remaining bounded budget before callers inspect
+    # ``returncode``; otherwise long-running timeout tests can leak a live
+    # Popen handle and emit ResourceWarning during interpreter shutdown.
+    remaining = max(0.0, deadline - time.time())
+    if remaining > 0 and proc.poll() is None:
+        try:
+            proc.wait(timeout=remaining)
+        except (subprocess.TimeoutExpired, OSError):
+            pass
 
 
 def _pipe_reader(stream, chunks, log_fh, stamp=None):
@@ -238,13 +248,13 @@ def _pipe_reader(stream, chunks, log_fh, stamp=None):
             b = stream.read1(65536)
             if not b:
                 break
-            chunks.append(b)
-            if stamp is not None:
-                stamp[0] = time.time()
-                if len(stamp) > 1:
-                    stamp[1] = True
-                if len(stamp) > 2:
-                    stamp[2] = time.monotonic()
+            chunks.append(b)
+            if stamp is not None:
+                stamp[0] = time.time()
+                if len(stamp) > 1:
+                    stamp[1] = True
+                if len(stamp) > 2:
+                    stamp[2] = time.monotonic()
             if log_fh:
                 try:
                     log_fh.write(b)
@@ -564,98 +574,98 @@ def pretty_cli_log(text, max_event_chars=4000):
     return "\n".join(out)
 
 
-def _normalize_abort_markers(value):
-    """把重复错误配置规范成 ``[(marker, count), ...]``。
-
-    旧调用方传的是单个 ``(marker, count)`` 元组，必须继续支持；新的
-    调用方可以传字典或元组列表，以便同一 CLI 同时对多种断流文案止损。
-    """
-    if not value:
-        return []
-    if isinstance(value, str):
-        return [(value, 1)]
-    if isinstance(value, dict):
-        raw = list(value.items())
-    elif (isinstance(value, (tuple, list)) and len(value) == 2
-          and isinstance(value[0], str)):
-        raw = [value]
-    else:
-        try:
-            raw = list(value)
-        except TypeError:
-            return []
-    out = []
-    for item in raw:
-        if isinstance(item, str):
-            marker, limit = item, 1
-        else:
-            try:
-                marker, limit = item
-            except (TypeError, ValueError):
-                continue
-        marker = str(marker or "")
-        try:
-            limit = max(1, int(limit))
-        except (TypeError, ValueError):
-            continue
-        if marker:
-            out.append((marker, limit))
-    return out
-
-
-def run_process(argv=None, shell_cmd=None, stdin_text=None, cwd=None, env=None,
-                timeout=DEFAULT_TIMEOUT, cancel_event=None, log_path=None,
-                stall_timeout=0, repeat_abort=None, deadline=None,
-                abort_markers=None, activity_timeout=0):
+def _normalize_abort_markers(value):
+    """把重复错误配置规范成 ``[(marker, count), ...]``。
+
+    旧调用方传的是单个 ``(marker, count)`` 元组，必须继续支持；新的
+    调用方可以传字典或元组列表，以便同一 CLI 同时对多种断流文案止损。
+    """
+    if not value:
+        return []
+    if isinstance(value, str):
+        return [(value, 1)]
+    if isinstance(value, dict):
+        raw = list(value.items())
+    elif (isinstance(value, (tuple, list)) and len(value) == 2
+          and isinstance(value[0], str)):
+        raw = [value]
+    else:
+        try:
+            raw = list(value)
+        except TypeError:
+            return []
+    out = []
+    for item in raw:
+        if isinstance(item, str):
+            marker, limit = item, 1
+        else:
+            try:
+                marker, limit = item
+            except (TypeError, ValueError):
+                continue
+        marker = str(marker or "")
+        try:
+            limit = max(1, int(limit))
+        except (TypeError, ValueError):
+            continue
+        if marker:
+            out.append((marker, limit))
+    return out
+
+
+def run_process(argv=None, shell_cmd=None, stdin_text=None, cwd=None, env=None,
+                timeout=DEFAULT_TIMEOUT, cancel_event=None, log_path=None,
+                stall_timeout=0, repeat_abort=None, deadline=None,
+                abort_markers=None, activity_timeout=0):
     """通用子进程执行：并发读管道防死锁；超时/取消杀整棵进程树。
 
-    stall_timeout：停滞看门狗（秒，0=关闭）——超过该时长 stdout/stderr 无任何
-    新输出即判卡死，提前杀树返回（timed_out=True + stalled=True）。只对输出
-    持续流动的 CLI 开（codex JSONL 事件流）；claude json 到结束才一次性输出，
-    开了会把正常长任务误杀。
-
-    activity_timeout：基础 timeout 到期后，最近仍有真实输出的流式 CLI 允许继续
-    运行的最长静默窗口（秒，0=关闭）。它独立于 stall_timeout，避免环境把
-    停滞看门狗设为 0 后，Claude/Codex 又在基础超时处被硬杀；总延长仍有上限。
-
-    repeat_abort：可选 (marker, count)，也接受多个 ``(marker, count)``；同一
-    致命错误达到次数即提前杀树，处理 CLI 自身不断打印重连消息、因此永远
-    触发不了静默看门狗的假运行。abort_markers 是兼容扩展，用于不改变旧
-    repeat_abort 参数形状的情况下追加多种一次性断流标记。
-
-    deadline：可选的 ``time.monotonic()`` 绝对截止时刻。它比 timeout 优先，
-    到点会杀掉进程树并返回 deadline_exceeded=True，调用方可把整个任务收口
-    为 timeout，而不是把它误记成普通供应商失败。
-
-    返回 {ok, exit_code, stdout, stderr, duration, cancelled, timed_out, stalled,
-    deadline_exceeded, abort_marker}。
-    """
+    stall_timeout：停滞看门狗（秒，0=关闭）——超过该时长 stdout/stderr 无任何
+    新输出即判卡死，提前杀树返回（timed_out=True + stalled=True）。只对输出
+    持续流动的 CLI 开（codex JSONL 事件流）；claude json 到结束才一次性输出，
+    开了会把正常长任务误杀。
+
+    activity_timeout：基础 timeout 到期后，最近仍有真实输出的流式 CLI 允许继续
+    运行的最长静默窗口（秒，0=关闭）。它独立于 stall_timeout，避免环境把
+    停滞看门狗设为 0 后，Claude/Codex 又在基础超时处被硬杀；总延长仍有上限。
+
+    repeat_abort：可选 (marker, count)，也接受多个 ``(marker, count)``；同一
+    致命错误达到次数即提前杀树，处理 CLI 自身不断打印重连消息、因此永远
+    触发不了静默看门狗的假运行。abort_markers 是兼容扩展，用于不改变旧
+    repeat_abort 参数形状的情况下追加多种一次性断流标记。
+
+    deadline：可选的 ``time.monotonic()`` 绝对截止时刻。它比 timeout 优先，
+    到点会杀掉进程树并返回 deadline_exceeded=True，调用方可把整个任务收口
+    为 timeout，而不是把它误记成普通供应商失败。
+
+    返回 {ok, exit_code, stdout, stderr, duration, cancelled, timed_out, stalled,
+    deadline_exceeded, abort_marker}。
+    """
     if shell_cmd:
         # shell 串的解析器随平台：重定向/引号语法两边通用，只是解释器不同
         argv = ["cmd", "/c", shell_cmd] if os.name == "nt" else ["/bin/sh", "-c", shell_cmd]
-    if argv is None:
-        return {"ok": False, "exit_code": None, "stdout": "", "stderr": "argv 为空",
-                "duration": 0.0, "cancelled": False, "timed_out": False,
-                "stalled": False, "deadline_exceeded": False, "abort_marker": None}
-    try:
-        deadline = float(deadline) if deadline is not None else None
-    except (TypeError, ValueError):
-        deadline = None
-    try:
-        timeout = max(0.0, float(timeout))
-    except (TypeError, ValueError):
-        timeout = float(DEFAULT_TIMEOUT)
-    try:
-        activity_timeout = max(0.0, float(activity_timeout))
-    except (TypeError, ValueError):
-        activity_timeout = 0.0
-    if deadline is not None and time.monotonic() >= deadline:
-        return {"ok": False, "exit_code": None, "stdout": "",
-                "stderr": "[任务总时限已到，未启动子进程]", "duration": 0.0,
-                "cancelled": False, "timed_out": True, "stalled": False,
-                "deadline_exceeded": True, "abort_marker": None}
-    aborts = _normalize_abort_markers(repeat_abort)
-    aborts.extend(_normalize_abort_markers(abort_markers))
+    if argv is None:
+        return {"ok": False, "exit_code": None, "stdout": "", "stderr": "argv 为空",
+                "duration": 0.0, "cancelled": False, "timed_out": False,
+                "stalled": False, "deadline_exceeded": False, "abort_marker": None}
+    try:
+        deadline = float(deadline) if deadline is not None else None
+    except (TypeError, ValueError):
+        deadline = None
+    try:
+        timeout = max(0.0, float(timeout))
+    except (TypeError, ValueError):
+        timeout = float(DEFAULT_TIMEOUT)
+    try:
+        activity_timeout = max(0.0, float(activity_timeout))
+    except (TypeError, ValueError):
+        activity_timeout = 0.0
+    if deadline is not None and time.monotonic() >= deadline:
+        return {"ok": False, "exit_code": None, "stdout": "",
+                "stderr": "[任务总时限已到，未启动子进程]", "duration": 0.0,
+                "cancelled": False, "timed_out": True, "stalled": False,
+                "deadline_exceeded": True, "abort_marker": None}
+    aborts = _normalize_abort_markers(repeat_abort)
+    aborts.extend(_normalize_abort_markers(abort_markers))
     full_env = scrub_env(os.environ.copy(), mode="drop")
     if env:
         # 5A：env 关键字环境变量注入用户传入的 env（属于有意注入，例如模型 API key）
@@ -686,13 +696,16 @@ def run_process(argv=None, shell_cmd=None, stdin_text=None, cwd=None, env=None,
                 stdout=subprocess.PIPE, stderr=subprocess.PIPE,
                 creationflags=CREATE_NO_WINDOW,
                 start_new_session=(os.name != "nt"))  # POSIX 需独立进程组供 killpg 杀树；Windows 忽略该参数
-        except Exception as e:
-            return {"ok": False, "exit_code": None, "stdout": "",
-                    "stderr": "启动失败: %r" % e, "duration": 0.0,
-                    "cancelled": False, "timed_out": False, "stalled": False,
-                    "deadline_exceeded": False, "abort_marker": None}
+        except Exception as e:
+            return {"ok": False, "exit_code": None, "stdout": "",
+                    "stderr": "启动失败: %r" % e, "duration": 0.0,
+                    "cancelled": False, "timed_out": False, "stalled": False,
+                    "deadline_exceeded": False, "abort_marker": None}
+        # 挂巢（beekeeper）：服务无论以何种方式退出（含被硬杀），蜜蜂 CLI
+        # 一起带走。挂巢失败只降级，绝不影响本步起跑。
+        beekeeper.adopt(proc)
         out_chunks, err_chunks = [], []
-        stamp = [time.time(), False, time.monotonic()]  # [墙钟、是否有输出、单调时钟]
+        stamp = [time.time(), False, time.monotonic()]  # [墙钟、是否有输出、单调时钟]
         t_out = threading.Thread(target=_pipe_reader, args=(proc.stdout, out_chunks, log_fh, stamp), daemon=True)
         t_err = threading.Thread(target=_pipe_reader, args=(proc.stderr, err_chunks, log_fh, stamp), daemon=True)
         t_out.start()
@@ -709,41 +722,41 @@ def run_process(argv=None, shell_cmd=None, stdin_text=None, cwd=None, env=None,
                     except Exception:
                         pass
             threading.Thread(target=_feed, daemon=True).start()
-        start = time.time()
-        start_mono = time.monotonic()
-        cancelled = timed_out = stalled = repeat_aborted = False
-        deadline_exceeded = False
-        abort_marker = None
-        while True:
+        start = time.time()
+        start_mono = time.monotonic()
+        cancelled = timed_out = stalled = repeat_aborted = False
+        deadline_exceeded = False
+        abort_marker = None
+        while True:
             try:
                 proc.wait(timeout=0.4)
                 break
             except subprocess.TimeoutExpired:
                 pass
-            if cancel_event is not None and cancel_event.is_set():
-                cancelled = True
-                _kill_tree(proc.pid)
-                _drain_streams(proc, t_out, t_err, timeout=10)
-                break
-            if deadline is not None and time.monotonic() >= deadline:
-                deadline_exceeded = True
-                timed_out = True
-                _kill_tree(proc.pid)
-                _drain_streams(proc, t_out, t_err, timeout=5)
-                break
-            elapsed = time.monotonic() - start_mono
-            if elapsed > timeout:
-                # Claude/Codex 的工具循环可能持续吐 stream-json 事件而迟迟
-                # 不给最终 result。活动延长与 stall 看门狗分离：无输出的挂死
-                # 仍按原单模型 timeout 快速换将，收到过输出则最多再给一个活动窗口。
-                active = (activity_timeout > 0 and stamp[1]
-                          and time.monotonic() - stamp[2] < activity_timeout
-                          and elapsed < timeout + activity_timeout)
-                if not active:
-                    timed_out = True
-                    _kill_tree(proc.pid)
-                    _drain_streams(proc, t_out, t_err, timeout=5)
-                    break
+            if cancel_event is not None and cancel_event.is_set():
+                cancelled = True
+                _kill_tree(proc.pid)
+                _drain_streams(proc, t_out, t_err, timeout=10)
+                break
+            if deadline is not None and time.monotonic() >= deadline:
+                deadline_exceeded = True
+                timed_out = True
+                _kill_tree(proc.pid)
+                _drain_streams(proc, t_out, t_err, timeout=5)
+                break
+            elapsed = time.monotonic() - start_mono
+            if elapsed > timeout:
+                # Claude/Codex 的工具循环可能持续吐 stream-json 事件而迟迟
+                # 不给最终 result。活动延长与 stall 看门狗分离：无输出的挂死
+                # 仍按原单模型 timeout 快速换将，收到过输出则最多再给一个活动窗口。
+                active = (activity_timeout > 0 and stamp[1]
+                          and time.monotonic() - stamp[2] < activity_timeout
+                          and elapsed < timeout + activity_timeout)
+                if not active:
+                    timed_out = True
+                    _kill_tree(proc.pid)
+                    _drain_streams(proc, t_out, t_err, timeout=5)
+                    break
             if stall_timeout and time.time() - stamp[0] > stall_timeout:
                 # 停滞看门狗：长静默多为卡死（网关挂起/CLI 假死），与其耗满总
                 # 超时不如提前杀——错误按超时归类，走既有的换模型/换将链路
@@ -752,22 +765,22 @@ def run_process(argv=None, shell_cmd=None, stdin_text=None, cwd=None, env=None,
                 _kill_tree(proc.pid)
                 _drain_streams(proc, t_out, t_err, timeout=5)
                 break
-            if aborts:
-                try:
-                    recent = decode_output(b"".join(out_chunks)[-8000:] +
-                                           b"\n" + b"".join(err_chunks)[-8000:])
-                    for marker, limit in aborts:
-                        if recent.count(marker) >= limit:
-                            repeat_aborted = True
-                            abort_marker = marker
-                            timed_out = True
-                            _kill_tree(proc.pid)
-                            _drain_streams(proc, t_out, t_err, timeout=5)
-                            break
-                    if repeat_aborted:
-                        break
-                except (TypeError, ValueError):
-                    pass
+            if aborts:
+                try:
+                    recent = decode_output(b"".join(out_chunks)[-8000:] +
+                                           b"\n" + b"".join(err_chunks)[-8000:])
+                    for marker, limit in aborts:
+                        if recent.count(marker) >= limit:
+                            repeat_aborted = True
+                            abort_marker = marker
+                            timed_out = True
+                            _kill_tree(proc.pid)
+                            _drain_streams(proc, t_out, t_err, timeout=5)
+                            break
+                    if repeat_aborted:
+                        break
+                except (TypeError, ValueError):
+                    pass
         duration = round(time.time() - start, 1)
         t_out.join(timeout=5)
         t_err.join(timeout=5)
@@ -784,32 +797,32 @@ def run_process(argv=None, shell_cmd=None, stdin_text=None, cwd=None, env=None,
             pass
         if cancelled:
             stderr += "\n[已被用户取消]"
-        elif repeat_aborted:
-            limit = next((n for m, n in aborts if m == abort_marker), "?")
-            if len(aborts) == 1 and abort_markers is None:
-                # 保留旧版 UI/测试已经依赖的文案；多标记配置使用更具体的
-                # 标记名，方便定位究竟是哪一种断流触发了止损。
-                stderr += "\n[同一网络错误重复 %s 次，已提前终止进程树]" % limit
-            else:
-                stderr += "\n[错误标记 %s 重复 %s 次，已提前终止进程树]" % (
-                    abort_marker, limit)
-        elif deadline_exceeded:
-            stderr += "\n[任务总时限已到，已终止进程树]"
+        elif repeat_aborted:
+            limit = next((n for m, n in aborts if m == abort_marker), "?")
+            if len(aborts) == 1 and abort_markers is None:
+                # 保留旧版 UI/测试已经依赖的文案；多标记配置使用更具体的
+                # 标记名，方便定位究竟是哪一种断流触发了止损。
+                stderr += "\n[同一网络错误重复 %s 次，已提前终止进程树]" % limit
+            else:
+                stderr += "\n[错误标记 %s 重复 %s 次，已提前终止进程树]" % (
+                    abort_marker, limit)
+        elif deadline_exceeded:
+            stderr += "\n[任务总时限已到，已终止进程树]"
         elif stalled:
             stderr += "\n[输出停滞 %ss，已终止进程树]" % stall_timeout
         elif timed_out:
             stderr += "\n[超时 %ss，已终止进程树]" % timeout
-        return {
-            "ok": exit_code == 0 and not cancelled and not timed_out,
-            "exit_code": exit_code, "stdout": stdout, "stderr": stderr,
-            "duration": duration, "cancelled": cancelled, "timed_out": timed_out,
-            "stalled": stalled, "deadline_exceeded": deadline_exceeded,
-            "abort_marker": abort_marker,
-            "activity_extended": bool(activity_timeout and stamp[1]
-                                       and duration > timeout),
-            "activity_timeout": activity_timeout,
-            "last_output_age": max(0.0, time.time() - stamp[0]),
-        }
+        return {
+            "ok": exit_code == 0 and not cancelled and not timed_out,
+            "exit_code": exit_code, "stdout": stdout, "stderr": stderr,
+            "duration": duration, "cancelled": cancelled, "timed_out": timed_out,
+            "stalled": stalled, "deadline_exceeded": deadline_exceeded,
+            "abort_marker": abort_marker,
+            "activity_extended": bool(activity_timeout and stamp[1]
+                                       and duration > timeout),
+            "activity_timeout": activity_timeout,
+            "last_output_age": max(0.0, time.time() - stamp[0]),
+        }
     finally:
         if log_fh:
             try:
@@ -886,8 +899,8 @@ def _codex_fail_msg(stdout):
     2026-09-16 实测：配额/限流只出现在 error 与 turn.failed 事件里，旧解析器
     两者都丢——进程退出码非 0 时错误串里只剩 "Reading prompt from stdin..."，
     _quota_error/_transient_error 判不出可降级，健康后继模型从未被尝试。
-    Reconnecting... 是 CLI 内部重试噪音（可能自愈），不取作终态错误；重复
-    重连由进程看门狗提前中断。turn.failed 优先于裸 error（后者取最后一条兜底）。
+    Reconnecting... 是 CLI 内部重试噪音（可能自愈），不取作终态错误；重复
+    重连由进程看门狗提前中断。turn.failed 优先于裸 error（后者取最后一条兜底）。
     """
     terminal, last_err = "", ""
     for line in (stdout or "").splitlines():
@@ -980,8 +993,8 @@ _TRANSIENT = ("503", "502", "529", "429", "no available channel", "temporarily",
               # 2026-09-20 连载评审实测：codex 网关断流（stream disconnected）
               # 与 opencode 服务端 500（Unexpected server error）都是「重试/换将
               # 就可能活」的瞬态病，旧表判成终态导致整链早死
-              "stream disconnected", "stream closed before response.completed",
-              "unexpected server error",
+              "stream disconnected", "stream closed before response.completed",
+              "unexpected server error",
               # 2026-09-22 mo-so 实案：claude 报 "Can't reach the API server —
               # check your internet or DNS (ENOTFOUND)"，一家端点解析不到不代表
               # 跨厂商链上别家也不通，旧表判成终态直接跳去换将别的 CLI
@@ -996,33 +1009,33 @@ def _transient_error(err):
 
 # 欠费/配额/限流耗尽：换 KEY 与换厂商都该继续（同厂商另一账号往往还能用）。
 # 与 modelhub._QUOTA_HINTS 同源，这里独立一份避免 core 模块间循环依赖。
-_QUOTA = ("insufficient", "quota", "balance", "credit", "billing", "arrears",
-          "payment required", "402", "欠费", "余额", "额度", "exceeded",
-          "too many requests", "rate limit", "rate_limit", "http 429",
-          "并发", "超过限")
-
-
-def _quota_error(err):
-    err = (err or "").lower()
-    return any(k in err for k in _QUOTA)
-
-
-_AUTH = ("invalid api key", "invalid_api_key", "api key is invalid",
-         "incorrect api key", "unauthorized", "unauthorised",
-         "authentication failed", "authentication error", "authentication_error",
-         "invalid token", "invalid_token", "http 401", "status code 401")
-
-
-def _auth_error(err):
-    """只识别明确的凭据拒绝，不把一般模型权限错误（403）当成 KEY 错。"""
-    err = (err or "").lower()
-    return any(k in err for k in _AUTH)
-
-
-def _attempt_credential(att):
-    """绑定链中凭据的非敏感身份；不把密钥值写入日志或尝试记录。"""
-    return (str((att or {}).get("provider_id") or ""),
-            str((att or {}).get("key_id") or ""))
+_QUOTA = ("insufficient", "quota", "balance", "credit", "billing", "arrears",
+          "payment required", "402", "欠费", "余额", "额度", "exceeded",
+          "too many requests", "rate limit", "rate_limit", "http 429",
+          "并发", "超过限")
+
+
+def _quota_error(err):
+    err = (err or "").lower()
+    return any(k in err for k in _QUOTA)
+
+
+_AUTH = ("invalid api key", "invalid_api_key", "api key is invalid",
+         "incorrect api key", "unauthorized", "unauthorised",
+         "authentication failed", "authentication error", "authentication_error",
+         "invalid token", "invalid_token", "http 401", "status code 401")
+
+
+def _auth_error(err):
+    """只识别明确的凭据拒绝，不把一般模型权限错误（403）当成 KEY 错。"""
+    err = (err or "").lower()
+    return any(k in err for k in _AUTH)
+
+
+def _attempt_credential(att):
+    """绑定链中凭据的非敏感身份；不把密钥值写入日志或尝试记录。"""
+    return (str((att or {}).get("provider_id") or ""),
+            str((att or {}).get("key_id") or ""))
 
 
 # 限流（429）专项：与一般瞬态不同，它是「等一个窗口就常能自愈」的病——
@@ -1048,15 +1061,15 @@ def _refusal_error(err):
     return any(k in err for k in _REFUSAL)
 
 
-def _attempt_upstream(att):
-    """返回尝试项的上游身份；同一 host 换模型/供应商别名不重复烧超时。"""
-    provider = (att or {}).get("provider") or {}
-    base_url = str(provider.get("base_url") or "").strip()
-    match = re.match(r"^[a-zA-Z][a-zA-Z0-9+.-]*://([^/?#]+)", base_url)
-    if match:
-        return match.group(1).lower()
-    provider_id = str((att or {}).get("provider_id") or "").strip()
-    return "provider:" + provider_id if provider_id else "unknown"
+def _attempt_upstream(att):
+    """返回尝试项的上游身份；同一 host 换模型/供应商别名不重复烧超时。"""
+    provider = (att or {}).get("provider") or {}
+    base_url = str(provider.get("base_url") or "").strip()
+    match = re.match(r"^[a-zA-Z][a-zA-Z0-9+.-]*://([^/?#]+)", base_url)
+    if match:
+        return match.group(1).lower()
+    provider_id = str((att or {}).get("provider_id") or "").strip()
+    return "provider:" + provider_id if provider_id else "unknown"
 
 
 def _claude_drift_note(att_model, parsed):
@@ -1083,11 +1096,11 @@ def _rate_limited(err):
 
 
 # 审计日志错误行探针：只认强信号，避免把回显提示词里的普通词当错误
-_LOG_ERR_HINT = re.compile(
-    r"(429|402|rate[_ ]?limit|quota|余额|欠费|insufficient|billing|arrears"
-    r"|401|unauthorized|unauthorised|invalid[_ ]api[_ ]key|authentication[_ ]error"
-    r"|ENOTFOUND|ECONNREFUSED|failed to run prompt|api_error|overloaded)",
-    re.I)
+_LOG_ERR_HINT = re.compile(
+    r"(429|402|rate[_ ]?limit|quota|余额|欠费|insufficient|billing|arrears"
+    r"|401|unauthorized|unauthorised|invalid[_ ]api[_ ]key|authentication[_ ]error"
+    r"|ENOTFOUND|ECONNREFUSED|failed to run prompt|api_error|overloaded)",
+    re.I)
 
 
 def _augment_error_from_log(out, log_path, limit=300):
@@ -1110,7 +1123,7 @@ def _augment_error_from_log(out, log_path, limit=300):
     out["error"] = (out["error"] + "；日志错误行: " + hits[-1][:limit])[-800:]
 
 
-def _grace_wait(cancel_event, seconds, log_path=None, why=""):
+def _grace_wait(cancel_event, seconds, log_path=None, why=""):
     """链尾限流宽限等待：小片睡眠随时响应取消。返回 False=已取消（别再重试）。"""
     if log_path:
         try:
@@ -1119,11 +1132,11 @@ def _grace_wait(cancel_event, seconds, log_path=None, why=""):
                          % (why or "上游限流（429）", seconds))
         except Exception:
             pass
-    end = time.monotonic() + seconds
-    while time.monotonic() < end:
-        if cancel_event is not None and cancel_event.is_set():
-            return False
-        time.sleep(min(3, max(0.5, end - time.monotonic())))
+    end = time.monotonic() + seconds
+    while time.monotonic() < end:
+        if cancel_event is not None and cancel_event.is_set():
+            return False
+        time.sleep(min(3, max(0.5, end - time.monotonic())))
     return cancel_event is None or not cancel_event.is_set()
 
 
@@ -1240,11 +1253,11 @@ def _prompt_to_file(prompt, workdir):
         return None
 
 
-# Windows 上不少 CLI 先经过 cmd.exe/.cmd 垫片，实际安全上限接近 8191，
-# 而不是 CreateProcess 的理论 32767。留出命令、路径和环境参数余量，超过
-# 该值统一落盘，避免 MiMo/Kimi 评审在启动前直接报「命令行太长」。
-_ARGV_PROMPT_SAFE = 20000
-_ARGV_PROMPT_SHIM_SAFE = 6000
+# Windows 上不少 CLI 先经过 cmd.exe/.cmd 垫片，实际安全上限接近 8191，
+# 而不是 CreateProcess 的理论 32767。留出命令、路径和环境参数余量，超过
+# 该值统一落盘，避免 MiMo/Kimi 评审在启动前直接报「命令行太长」。
+_ARGV_PROMPT_SAFE = 20000
+_ARGV_PROMPT_SHIM_SAFE = 6000
 
 
 def _build_call(agent, kind, sid, readonly, model, prompt, images=None, workdir=None):
@@ -1382,27 +1395,27 @@ def _build_call(agent, kind, sid, readonly, model, prompt, images=None, workdir=
     # 超长时落盘临时文件、参数位换成读文件指令——评审/作者 CLI 非交互模式均带
     # 读文件工具（kimi 0.43 实测可读），读不到的调用输出不可解析，由评审全挂
     # 防线兜底判失败，绝不静默降级。
-    prompt_safe = _ARGV_PROMPT_SAFE
-    if argv:
-        launcher = str(argv[0]).lower()
-        if (launcher.endswith((".cmd", ".bat")) or
-                os.path.basename(launcher) in ("npm", "npx", "pnpm", "yarn", "bun")):
-            prompt_safe = _ARGV_PROMPT_SHIM_SAFE
-    if argv is not None and stdin_text is None and len(prompt) >= prompt_safe \
-            and argv.count(prompt) == 1:
+    prompt_safe = _ARGV_PROMPT_SAFE
+    if argv:
+        launcher = str(argv[0]).lower()
+        if (launcher.endswith((".cmd", ".bat")) or
+                os.path.basename(launcher) in ("npm", "npx", "pnpm", "yarn", "bun")):
+            prompt_safe = _ARGV_PROMPT_SHIM_SAFE
+    if argv is not None and stdin_text is None and len(prompt) >= prompt_safe \
+            and argv.count(prompt) == 1:
         pf = _prompt_to_file(prompt, workdir)
-        if pf:
-            argv[argv.index(prompt)] = (
-                "[系统] 本次完整指令因命令行长度限制已写入文件：%s\n"
-                "请先用读文件工具完整读取该文件，然后把文件内容当作你的任务指令执行。" % pf)
-            # Keep the original payload on stdin as a compatibility path for
-            # simple generic adapters and test shims that read stdin. CLI tools
-            # that intentionally use argv still receive the bounded file hint.
-            if kind == "generic" and stdin_text is None and len(tmpl) > 1:
-                # 模板明确将提示词放进 argv 时，原文只保存在文件里；同时再
-                # 写入 stdin 会让部分包装器重复读取，也无法解决 Windows 命令行上限。
-                stdin_text = None
-            tmp_files.append(pf)
+        if pf:
+            argv[argv.index(prompt)] = (
+                "[系统] 本次完整指令因命令行长度限制已写入文件：%s\n"
+                "请先用读文件工具完整读取该文件，然后把文件内容当作你的任务指令执行。" % pf)
+            # Keep the original payload on stdin as a compatibility path for
+            # simple generic adapters and test shims that read stdin. CLI tools
+            # that intentionally use argv still receive the bounded file hint.
+            if kind == "generic" and stdin_text is None and len(tmpl) > 1:
+                # 模板明确将提示词放进 argv 时，原文只保存在文件里；同时再
+                # 写入 stdin 会让部分包装器重复读取，也无法解决 Windows 命令行上限。
+                stdin_text = None
+            tmp_files.append(pf)
     return argv, stdin_text, prompt, tmp_files
 
 
@@ -1426,50 +1439,50 @@ def _check_approval(agent):
     return False, "sensitive step requires policy=ask UI flow (Phase 5) — currently blocking"
 
 
-def _stall_timeout(env_name, default):
+def _stall_timeout(env_name, default):
     """停滞看门狗秒数：stdout/stderr 持续无新输出超过该时长 → 判卡死提前杀，
     错误按超时归类走既有换模型/换将链路。只对事件流持续流动的 CLI 启用
     （codex JSONL / claude stream-json）；qwen/opencode/generic 结束才一次性
     输出，开了会误杀正常长任务，仍靠总超时兜底。env 可调，0=关闭。"""
     try:
         return max(0, int(os.environ.get(env_name, str(default))))
-    except Exception:
-        return default
-
-
-def _activity_timeout(env_name, default=DEFAULT_STREAM_ACTIVITY_TIMEOUT_S):
-    """流式 CLI 的活动延长窗口，独立于停滞看门狗配置。"""
-    try:
-        return max(0, int(os.environ.get(env_name, str(default))))
-    except Exception:
-        return default
-
-
-def _max_model_attempts(has_deadline=False):
-    """返回一次 runner 调用允许消耗的候选数。
-
-    生产任务会传入统一 deadline，默认收敛到 3 个候选；没有 deadline 的
-    旧式直接调用保留最多 4 个候选，以兼容已有的「同一上游两模型后换
-    供应商」策略。运维可用环境变量进一步收紧/放宽，最少 1 个。
-    """
-    default = 3 if has_deadline else 4
-    try:
-        value = int(os.environ.get("TUTTI_MAX_MODEL_ATTEMPTS", str(default)))
-        return max(1, min(12, value))
-    except (TypeError, ValueError):
-        return default
-
-
-_INVALID_MODEL_NAMES = frozenset(("auto", "default", "none"))
-
-
-def _invalid_model(model):
-    return str(model or "").strip().lower() in _INVALID_MODEL_NAMES
-
-
-def run_agent(agent, prompt, workdir=None, readonly=True,
-              timeout=DEFAULT_TIMEOUT, cancel_event=None, log_path=None, resume=None,
-              images=None, require_tools=False, deadline=None):
+    except Exception:
+        return default
+
+
+def _activity_timeout(env_name, default=DEFAULT_STREAM_ACTIVITY_TIMEOUT_S):
+    """流式 CLI 的活动延长窗口，独立于停滞看门狗配置。"""
+    try:
+        return max(0, int(os.environ.get(env_name, str(default))))
+    except Exception:
+        return default
+
+
+def _max_model_attempts(has_deadline=False):
+    """返回一次 runner 调用允许消耗的候选数。
+
+    生产任务会传入统一 deadline，默认收敛到 3 个候选；没有 deadline 的
+    旧式直接调用保留最多 4 个候选，以兼容已有的「同一上游两模型后换
+    供应商」策略。运维可用环境变量进一步收紧/放宽，最少 1 个。
+    """
+    default = 3 if has_deadline else 4
+    try:
+        value = int(os.environ.get("TUTTI_MAX_MODEL_ATTEMPTS", str(default)))
+        return max(1, min(12, value))
+    except (TypeError, ValueError):
+        return default
+
+
+_INVALID_MODEL_NAMES = frozenset(("auto", "default", "none"))
+
+
+def _invalid_model(model):
+    return str(model or "").strip().lower() in _INVALID_MODEL_NAMES
+
+
+def run_agent(agent, prompt, workdir=None, readonly=True,
+              timeout=DEFAULT_TIMEOUT, cancel_event=None, log_path=None, resume=None,
+              images=None, require_tools=False, deadline=None):
     """执行一次智能体调用，返回统一结构
     {ok, text, json, cost_usd, tokens, error, error_code, raw}。
     agent 来自 registry.effective_agents()；resume 为已有会话 id，仅真实智能体生效
@@ -1483,66 +1496,65 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
 
     模型尝试顺序来自 _resolve_attempts：跨厂商链（每条独立 env）或
     主模型 + 降级备选；瞬态/配额/超时/上游拒答换下一条，取消与解析
-    失败不降级；同一上游超时两次即跳过该上游剩余模型。
-
-    5E：catalog `orch.timeout_ms`（毫秒）优先于 caller 传入的 timeout；
-    单模型无进展时最多等 60 秒，持续 stream 活动交由 stall_timeout 看门狗
-    收口。deadline 是
-    任务共享的 ``time.monotonic()`` 绝对截止时刻；传入后所有模型、空响应
-    重试和限流宽限共用同一剩余预算。
-    """
+    失败不降级；同一上游超时两次即跳过该上游剩余模型。
+
+    5E：catalog `orch.timeout_ms`（毫秒）优先于 caller 传入的 timeout；
+    单模型时限由该配置或 caller timeout 决定，任务 deadline 可进一步收紧；
+    持续 stream 活动交由 stall_timeout 看门狗收口。deadline 是
+    任务共享的 ``time.monotonic()`` 绝对截止时刻；传入后所有模型、空响应
+    重试和限流宽限共用同一剩余预算。
+    """
     # 5E：catalog orch.timeout_ms 优先
     orch_timeout_ms = (agent.get("orch") or {}).get("timeout_ms")
-    if orch_timeout_ms:
-        timeout = float(orch_timeout_ms) / 1000.0
-    try:
-        timeout = float(timeout)
-    except (TypeError, ValueError):
-        timeout = float(DEFAULT_TIMEOUT)
-    timeout = max(0.0, timeout)
-    try:
-        deadline = float(deadline) if deadline is not None else None
-    except (TypeError, ValueError):
-        deadline = None
-    if deadline is not None and time.monotonic() >= deadline:
-        return {"ok": False, "text": "", "json": None, "cost_usd": 0.0,
-                "tokens": 0, "usage": None, "error": "任务总时限已到，未启动模型",
-                "error_code": ErrorCode.TIMEOUT,
-                "raw": {"exit_code": None, "timed_out": True,
-                        "deadline_exceeded": True, "duration": 0.0},
-                "kind": agent.get("kind", "generic"), "model": None,
-                "attempts": []}
-    timeout = min(timeout, MAX_ATTEMPT_TIMEOUT_S)
+    if orch_timeout_ms:
+        timeout = float(orch_timeout_ms) / 1000.0
+    try:
+        timeout = float(timeout)
+    except (TypeError, ValueError):
+        timeout = float(DEFAULT_TIMEOUT)
+    timeout = max(0.0, timeout)
+    try:
+        deadline = float(deadline) if deadline is not None else None
+    except (TypeError, ValueError):
+        deadline = None
+    if deadline is not None and time.monotonic() >= deadline:
+        return {"ok": False, "text": "", "json": None, "cost_usd": 0.0,
+                "tokens": 0, "usage": None, "error": "任务总时限已到，未启动模型",
+                "error_code": ErrorCode.TIMEOUT,
+                "raw": {"exit_code": None, "timed_out": True,
+                        "deadline_exceeded": True, "duration": 0.0},
+                "kind": agent.get("kind", "generic"), "model": None,
+                "attempts": []}
     kind = agent.get("kind", "generic")
     if kind == "aider":
         repo_issue = _git_repo_issue(workdir)
         if repo_issue:
             return {"ok": False, "text": "", "json": None, "cost_usd": 0.0,
-                    "tokens": 0, "usage": None, "error": repo_issue,
-                    "error_code": ErrorCode.VENDOR_ERROR, "raw": None,
-                    "kind": kind, "model": agent.get("model"), "attempts": []}
-    if kind == "codex":
-        stall_t = _stall_timeout("TUTTI_CODEX_STALL_TIMEOUT", 600)
-        activity_t = _activity_timeout("TUTTI_CODEX_ACTIVITY_TIMEOUT")
-    elif kind == "claude":
-        stall_t = _stall_timeout("TUTTI_CLAUDE_STALL_TIMEOUT", 600)
-        activity_t = _activity_timeout("TUTTI_CLAUDE_ACTIVITY_TIMEOUT")
-    else:
+                    "tokens": 0, "usage": None, "error": repo_issue,
+                    "error_code": ErrorCode.VENDOR_ERROR, "raw": None,
+                    "kind": kind, "model": agent.get("model"), "attempts": []}
+    if kind == "codex":
+        stall_t = _stall_timeout("TUTTI_CODEX_STALL_TIMEOUT", 600)
+        activity_t = _activity_timeout("TUTTI_CODEX_ACTIVITY_TIMEOUT")
+    elif kind == "claude":
+        stall_t = _stall_timeout("TUTTI_CLAUDE_STALL_TIMEOUT", 600)
+        activity_t = _activity_timeout("TUTTI_CLAUDE_ACTIVITY_TIMEOUT")
+    else:
         # 数据驱动：catalog orch.stall_timeout_s——给 kimi/qwen 这类评审用
         # CLI 配置后，静默挂死即杀（qwen 900s：2026-09-22 MCP 收尾死锁案）；
         # mimo 等结束才一次性输出的仍留 0（开了必误杀），靠总超时兜底
         try:
             stall_t = max(0, int((agent.get("orch") or {}).get("stall_timeout_s") or 0))
-        except Exception:
-            stall_t = 0
-        activity_t = 0
+        except Exception:
+            stall_t = 0
+        activity_t = 0
     # 5G：approval NEVER 一线（无人值守不静默降级）
     ok, reason = _check_approval(agent)
     if not ok:
         return {"ok": False, "text": "", "json": None, "cost_usd": 0.0, "tokens": 0,
-                "usage": None, "error": reason,
-                "error_code": ErrorCode.ENV_BLOCK,
-                "raw": None, "kind": kind, "model": None, "attempts": []}
+                "usage": None, "error": reason,
+                "error_code": ErrorCode.ENV_BLOCK,
+                "raw": None, "kind": kind, "model": None, "attempts": []}
     base_env = dict(agent.get("env") or {})
     if kind == "claude":
         bash = find_git_bash()
@@ -1555,64 +1567,64 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
                 "usage": None,
                 "error": "该 CLI 未配置会话恢复（catalog orch.resume_argv_template），"
                          "无法在已有会话上继续",
-                "error_code": ErrorCode.VENDOR_ERROR,
-                "sid": "", "raw": None, "kind": kind, "model": None, "attempts": []}
-
-    attempts = _resolve_attempts(agent)
-    # 启动前过滤明显的占位模型名。它们常由某个 CLI 的会话标题/默认配置
-    # 泄漏进编排链；让子进程自己等待网关返回会把一个确定性配置错拖满
-    # 多轮超时。非法项不消耗模型调用预算，后面的真实候选仍可接手。
-    filtered_attempts = []
-    preflight_failures = []
-    for att in attempts:
-        if _invalid_model(att.get("model")):
-            preflight_failures.append({
-                "model": att.get("model"),
-                "provider_id": att.get("provider_id") or "",
-                "provider": (att.get("provider") or {}).get("name", "")
-                            if isinstance(att.get("provider"), dict) else "",
-                "ok": False, "skipped": True, "error":
-                    "无效模型名 %r，启动前跳过" % att.get("model"),
-            })
-            continue
-        filtered_attempts.append(att)
-    attempts = filtered_attempts[:_max_model_attempts(deadline is not None)]
-    out = None
-    attempt_history = list(preflight_failures)
-    if not attempts:
-        return {"ok": False, "text": "", "json": None, "cost_usd": 0.0,
-                "tokens": 0, "usage": None,
-                "error": (preflight_failures[-1].get("error")
-                           if preflight_failures else "没有可尝试的模型"),
-                "error_code": ErrorCode.VENDOR_ERROR, "raw": None,
-                "kind": kind, "model": None, "attempts": attempt_history}
-    grace_left = RATE_LIMIT_GRACE_N  # 链尾限流宽限预算：整链撞 429 时原地等一个窗口再试一次
-    timed_out_upstreams = {}  # 同一上游两次超时后跳过剩余模型
-    skipped_upstreams = set()
-    auth_failed_credentials = set()
-    for ai, att in enumerate(attempts):
-        credential = _attempt_credential(att)
-        if credential in auth_failed_credentials:
-            attempt_history.append({
-                "model": att.get("model") or "",
-                "provider_id": att.get("provider_id") or "",
-                "provider": ((att.get("provider") or {}).get("name", "")
-                             if isinstance(att.get("provider"), dict) else ""),
-                "kind": kind, "ok": False, "skipped": True,
-                "error_code": ErrorCode.VENDOR_ERROR,
-                "error": "同一凭据已被认证拒绝，跳过该凭据的其他模型",
-            })
-            if out is not None:
-                out["attempts"] = list(attempt_history)
-            continue
-        upstream = _attempt_upstream(att)
-        if upstream in skipped_upstreams:
-            continue
-        # 有任务 deadline 时，所有候选共享同一个截止时刻；没有 deadline
-        # 的历史直接调用保留「每个候选最多 60 秒」的旧语义。
-        model_deadline = (att.get("_deadline") or deadline or
-                          (time.monotonic() + timeout))
-        env = dict(base_env)
+                "error_code": ErrorCode.VENDOR_ERROR,
+                "sid": "", "raw": None, "kind": kind, "model": None, "attempts": []}
+
+    attempts = _resolve_attempts(agent)
+    # 启动前过滤明显的占位模型名。它们常由某个 CLI 的会话标题/默认配置
+    # 泄漏进编排链；让子进程自己等待网关返回会把一个确定性配置错拖满
+    # 多轮超时。非法项不消耗模型调用预算，后面的真实候选仍可接手。
+    filtered_attempts = []
+    preflight_failures = []
+    for att in attempts:
+        if _invalid_model(att.get("model")):
+            preflight_failures.append({
+                "model": att.get("model"),
+                "provider_id": att.get("provider_id") or "",
+                "provider": (att.get("provider") or {}).get("name", "")
+                            if isinstance(att.get("provider"), dict) else "",
+                "ok": False, "skipped": True, "error":
+                    "无效模型名 %r，启动前跳过" % att.get("model"),
+            })
+            continue
+        filtered_attempts.append(att)
+    attempts = filtered_attempts[:_max_model_attempts(deadline is not None)]
+    out = None
+    attempt_history = list(preflight_failures)
+    if not attempts:
+        return {"ok": False, "text": "", "json": None, "cost_usd": 0.0,
+                "tokens": 0, "usage": None,
+                "error": (preflight_failures[-1].get("error")
+                           if preflight_failures else "没有可尝试的模型"),
+                "error_code": ErrorCode.VENDOR_ERROR, "raw": None,
+                "kind": kind, "model": None, "attempts": attempt_history}
+    grace_left = RATE_LIMIT_GRACE_N  # 链尾限流宽限预算：整链撞 429 时原地等一个窗口再试一次
+    timed_out_upstreams = {}  # 同一上游两次超时后跳过剩余模型
+    skipped_upstreams = set()
+    auth_failed_credentials = set()
+    for ai, att in enumerate(attempts):
+        credential = _attempt_credential(att)
+        if credential in auth_failed_credentials:
+            attempt_history.append({
+                "model": att.get("model") or "",
+                "provider_id": att.get("provider_id") or "",
+                "provider": ((att.get("provider") or {}).get("name", "")
+                             if isinstance(att.get("provider"), dict) else ""),
+                "kind": kind, "ok": False, "skipped": True,
+                "error_code": ErrorCode.VENDOR_ERROR,
+                "error": "同一凭据已被认证拒绝，跳过该凭据的其他模型",
+            })
+            if out is not None:
+                out["attempts"] = list(attempt_history)
+            continue
+        upstream = _attempt_upstream(att)
+        if upstream in skipped_upstreams:
+            continue
+        # 有任务 deadline 时，所有候选共享同一个截止时刻；否则每个候选
+        # 使用 catalog 配置或 caller 提供的单模型时限。
+        model_deadline = (att.get("_deadline") or deadline or
+                          (time.monotonic() + timeout))
+        env = dict(base_env)
         env.update(att["env"])
         if kind == "codex":
             # 网关自定义模型名（如 [opencode]xxx）是合法 API 模型名不能改，但方括号
@@ -1626,40 +1638,40 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
         elif att["from_chain"] and "codex_provider" in eff_agent:
             # 链内条目未注入供应商时不能沿用上一条（可能是另一家厂商）的 -c 覆盖
             del eff_agent["codex_provider"]
-        for attempt in range(2):  # claude 偶发空响应（0 token）自动重试一次
-            remaining = model_deadline - time.monotonic()
-            if remaining <= 0:
-                res = {"ok": False, "exit_code": None, "stdout": "", "stderr": "",
-                       "duration": timeout, "cancelled": False, "timed_out": True,
-                       "stalled": False,
-                       "deadline_exceeded": deadline is not None}
-            else:
-                argv, stdin_text, prompt_eff, tmp_files = _build_call(
-                    eff_agent, kind, sid, readonly,
-                    att["model"], prompt,
-                    images=images if kind == "codex" else None,
-                    workdir=workdir)
-                try:
-                    repeat_guard = (("Reconnecting...", 2)
-                                    if kind == "codex" else None)
-                    stream_abort_markers = (
-                        (("stream disconnected", 1),
-                         ("stream closed before response.completed", 1))
-                        if kind == "codex" else None)
-                    res = run_process(argv=argv, stdin_text=stdin_text, cwd=workdir, env=env,
-                                      timeout=min(timeout, remaining), deadline=deadline,
-                                      cancel_event=cancel_event,
-                                      log_path=log_path, stall_timeout=stall_t,
-                                      activity_timeout=activity_t,
-                                      repeat_abort=repeat_guard,
-                                      abort_markers=stream_abort_markers)
-                finally:
-                    # 超长指令临时文件：CLI 进程已结束（管道已收），即刻清场不污染工作目录
-                    for tf in tmp_files:
-                        try:
-                            os.remove(tf)
-                        except OSError:
-                            pass
+        for attempt in range(2):  # claude 偶发空响应（0 token）自动重试一次
+            remaining = model_deadline - time.monotonic()
+            if remaining <= 0:
+                res = {"ok": False, "exit_code": None, "stdout": "", "stderr": "",
+                       "duration": timeout, "cancelled": False, "timed_out": True,
+                       "stalled": False,
+                       "deadline_exceeded": deadline is not None}
+            else:
+                argv, stdin_text, prompt_eff, tmp_files = _build_call(
+                    eff_agent, kind, sid, readonly,
+                    att["model"], prompt,
+                    images=images if kind == "codex" else None,
+                    workdir=workdir)
+                try:
+                    repeat_guard = (("Reconnecting...", 2)
+                                    if kind == "codex" else None)
+                    stream_abort_markers = (
+                        (("stream disconnected", 1),
+                         ("stream closed before response.completed", 1))
+                        if kind == "codex" else None)
+                    res = run_process(argv=argv, stdin_text=stdin_text, cwd=workdir, env=env,
+                                      timeout=min(timeout, remaining), deadline=deadline,
+                                      cancel_event=cancel_event,
+                                      log_path=log_path, stall_timeout=stall_t,
+                                      activity_timeout=activity_t,
+                                      repeat_abort=repeat_guard,
+                                      abort_markers=stream_abort_markers)
+                finally:
+                    # 超长指令临时文件：CLI 进程已结束（管道已收），即刻清场不污染工作目录
+                    for tf in tmp_files:
+                        try:
+                            os.remove(tf)
+                        except OSError:
+                            pass
             out = {"ok": res["ok"], "text": "", "json": None, "cost_usd": 0.0,
                    "tokens": 0, "usage": None, "error": "", "error_code": "",
                    "sid": "", "raw": res, "kind": kind, "model": att["model"],
@@ -1677,9 +1689,9 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
                         else "超时" if res["timed_out"]
                         else "取消" if res["cancelled"]
                         else "退出码 %s" % res["exit_code"])
-                if res.get("deadline_exceeded"):
-                    head = "任务总时限已到"
-                out["error"] = head + ("；stderr/stdout: " + tail if tail else "")
+                if res.get("deadline_exceeded"):
+                    head = "任务总时限已到"
+                out["error"] = head + ("；stderr/stdout: " + tail if tail else "")
                 if kind == "codex":
                     fm = _codex_fail_msg(res["stdout"])
                     if fm:
@@ -1789,44 +1801,44 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
                 if not out["text"]:
                     out["error_code"] = _classify_failure(res, kind=kind, empty_output=True)
             break
-        _report_key(att, out)
-        raw = out.get("raw") or {}
-        attempt_history.append({
-            "model": out.get("model") or att.get("model") or "",
-            "provider_id": out.get("provider_id") or att.get("provider_id") or "",
-            "provider": ((out.get("provider") or {}).get("name", "")
-                         if isinstance(out.get("provider"), dict) else ""),
-            "kind": kind,
-            "ok": bool(out.get("ok")),
-            "duration": raw.get("duration", 0.0),
-            "timed_out": bool(raw.get("timed_out")),
-            "deadline_exceeded": bool(raw.get("deadline_exceeded")),
-            "abort_marker": raw.get("abort_marker"),
-            "error_code": str(out.get("error_code") or ""),
-            "error": (out.get("error") or "")[:500],
-        })
-        out["attempts"] = list(attempt_history)
-        if out["ok"] or ai == len(attempts) - 1:
-            remaining = model_deadline - time.monotonic()
-            if (not out["ok"] and grace_left > 0 and _rate_limited(out.get("error"))
-                    and remaining > 2
-                    and not (cancel_event is not None and cancel_event.is_set())):
-                # 链尾限流宽限：enumerate 活列表——把链尾这条 append 回去，
-                # 下一轮迭代就是「原地重试一次」；再撞限流时预算已耗尽照常判死。
-                grace_left -= 1
-                grace_s = min(RATE_LIMIT_GRACE_S, remaining - 1)
-                if _grace_wait(cancel_event, grace_s, log_path=log_path):
-                    retry = dict(att, _deadline=model_deadline)
-                    attempts.append(retry)
-                    continue
-            return out
-        if _auth_error(out.get("error")):
-            auth_failed_credentials.add(credential)
-            if not any(_attempt_credential(candidate) not in auth_failed_credentials
-                       for candidate in attempts[ai + 1:]):
-                return out
-            continue
-        # 换将闸门：什么失败值得烧链上下一个候选。
+        _report_key(att, out)
+        raw = out.get("raw") or {}
+        attempt_history.append({
+            "model": out.get("model") or att.get("model") or "",
+            "provider_id": out.get("provider_id") or att.get("provider_id") or "",
+            "provider": ((out.get("provider") or {}).get("name", "")
+                         if isinstance(out.get("provider"), dict) else ""),
+            "kind": kind,
+            "ok": bool(out.get("ok")),
+            "duration": raw.get("duration", 0.0),
+            "timed_out": bool(raw.get("timed_out")),
+            "deadline_exceeded": bool(raw.get("deadline_exceeded")),
+            "abort_marker": raw.get("abort_marker"),
+            "error_code": str(out.get("error_code") or ""),
+            "error": (out.get("error") or "")[:500],
+        })
+        out["attempts"] = list(attempt_history)
+        if out["ok"] or ai == len(attempts) - 1:
+            remaining = model_deadline - time.monotonic()
+            if (not out["ok"] and grace_left > 0 and _rate_limited(out.get("error"))
+                    and remaining > 2
+                    and not (cancel_event is not None and cancel_event.is_set())):
+                # 链尾限流宽限：enumerate 活列表——把链尾这条 append 回去，
+                # 下一轮迭代就是「原地重试一次」；再撞限流时预算已耗尽照常判死。
+                grace_left -= 1
+                grace_s = min(RATE_LIMIT_GRACE_S, remaining - 1)
+                if _grace_wait(cancel_event, grace_s, log_path=log_path):
+                    retry = dict(att, _deadline=model_deadline)
+                    attempts.append(retry)
+                    continue
+            return out
+        if _auth_error(out.get("error")):
+            auth_failed_credentials.add(credential)
+            if not any(_attempt_credential(candidate) not in auth_failed_credentials
+                       for candidate in attempts[ai + 1:]):
+                return out
+            continue
+        # 换将闸门：什么失败值得烧链上下一个候选。
         # · 瞬态/配额（文本特征，含 CLI 内部 api_retry 的 429/5xx 字样）照旧；
         # · 超时：显式可换（看 res 标志位）——慢上游换快候选可能救回。
         #   2026-09-23 夜班实案：旧逻辑里超时能否换将全赌输出尾部碰巧含有
@@ -1835,20 +1847,20 @@ def run_agent(agent, prompt, workdir=None, readonly=True,
         #   换异上游候选常常就活；判死则整 run 白烧（同案：gemini 拒答后
         #   链上云知声候选从未被尝试）。
         raw = out.get("raw") or {}
-        if (raw.get("cancelled")   # 取消是用户意志，绝不降级（尾段可能带瞬态字样）
-                or not (_transient_error(out.get("error")) or _quota_error(out.get("error"))
-                        or _refusal_error(out.get("error"))
-                        or raw.get("timed_out"))):
+        if (raw.get("cancelled")   # 取消是用户意志，绝不降级（尾段可能带瞬态字样）
+                or not (_transient_error(out.get("error")) or _quota_error(out.get("error"))
+                        or _refusal_error(out.get("error"))
+                        or raw.get("timed_out"))):
             return out  # 取消/解析失败等真终态不降级
-        # 同一上游的模型连续两次超时就收手——
-        # 链上候选绑同一上游时换将=换壳不换命（2026-09-22 429 集群教训；
-        # 2026-09-23 夜班实案同一慢上游连烧 3×20 分钟全超时零产出），最多
-        # 两个模型预算后转向其他上游。其余瞬态/拒答仍按链长全部尝试：各家病根
+        # 同一上游的模型连续两次超时就收手——
+        # 链上候选绑同一上游时换将=换壳不换命（2026-09-22 429 集群教训；
+        # 2026-09-23 夜班实案同一慢上游连烧 3×20 分钟全超时零产出），最多
+        # 两个模型预算后转向其他上游。其余瞬态/拒答仍按链长全部尝试：各家病根
         # 不同，每个候选都值得一次机会。
-        if raw.get("timed_out"):
-            timed_out_upstreams[upstream] = timed_out_upstreams.get(upstream, 0) + 1
-            if timed_out_upstreams[upstream] >= 2:
-                skipped_upstreams.add(upstream)
+        if raw.get("timed_out"):
+            timed_out_upstreams[upstream] = timed_out_upstreams.get(upstream, 0) + 1
+            if timed_out_upstreams[upstream] >= 2:
+                skipped_upstreams.add(upstream)
     return out