codebee 0.1.2 → 0.1.4

package/app/core/manager.py

@@ -1,1511 +1,1523 @@
-# -*- coding: utf-8 -*-
-"""智能体管理器：安装检测、版本、安装/升级、模型配置读写。
-
-安全约束：catalog 中的配置文件路径展开后必须落在用户主目录内，
-防止相对路径穿越到预期之外的系统位置。
-"""
-from __future__ import annotations
-
-import json
-import os
-import re
-import shutil
-import socket
-import subprocess
-import sys
-import threading
-import time
-import webbrowser
-import zlib
-from pathlib import Path
-
-from . import catalog, paths, runner
-
-CREATE_NO_WINDOW = 0x08000000
-VERSION_TTL = 300  # 版本缓存 5 分钟
-
-_LOCK = threading.RLock()
-_STATE = {"detected": {}, "versions": {}, "detect_ts": 0.0, "detect_ev": None}
-
-# 能自动写入默认模型的 config.format（其余格式只能手动编辑）
-_WRITABLE_FORMATS = ("toml-line", "toml-section", "json", "json-path", "jsonc",
-                     "yaml-line")
-
-
-def _expand(p):
-    return os.path.abspath(os.path.expanduser(os.path.expandvars(p)))
-
-
-def _safe_config_path(raw):
-    """展开并校验配置路径：必须在用户主目录内（防穿越）。"""
-    if not raw:
-        return None
-    full = _expand(raw)
-    home = os.path.abspath(os.path.expanduser("~"))
-    try:
-        if os.path.commonpath([full, home]) != home:
-            return None
-    except ValueError:
-        return None
-    return full
-
-
-# ---------------------------------------------------------------- 检测
-
-def detect_entry(entry):
-    d = entry.get("detect") or {}
-    if d.get("cli"):
-        path = shutil.which(d["cli"])
-        return {"installed": bool(path), "detail": path or ""}
-    if d.get("exe"):
-        full = _expand(d["exe"])
-        return {"installed": os.path.isfile(full), "detail": full if os.path.isfile(full) else ""}
-    if d.get("uwp"):
-        base = os.path.join(os.environ.get("LOCALAPPDATA", ""), "Packages", d["uwp"])
-        return {"installed": os.path.isdir(base), "detail": base if os.path.isdir(base) else ""}
-    if d.get("dir"):
-        full = _expand(d["dir"])
-        return {"installed": os.path.isdir(full), "detail": full if os.path.isdir(full) else ""}
-    return {"installed": False, "detail": ""}
-
-
-def detect_all(force=False):
-    """检测全部条目。检测（慢磁盘 IO）在锁外跑：shutil.which/isfile 在
-    Windows 上遇到断链的 PATH 项可能卡数秒，持锁会把所有并发请求堵死
-    （曾导致 SSE 多连接时服务假死）。等待方有界等待 30s 后拿旧结果。
-    """
-    with _LOCK:
-        if not force and _STATE["detected"] and time.time() - _STATE["detect_ts"] < 60:
-            return _STATE["detected"]
-        ev = _STATE["detect_ev"]
-        lead = ev is None  # 我是本次检测的执行者
-        if lead:
-            ev = _STATE["detect_ev"] = threading.Event()
-    if not lead:
-        ev.wait(30)  # 检测完成或超时；两种情况都拿当前最新快照
-        with _LOCK:
-            return dict(_STATE["detected"])
-    try:
-        detected = {}
-        for entry in catalog.load():
-            try:
-                detected[entry["id"]] = detect_entry(entry)
-            except Exception as e:
-                detected[entry["id"]] = {"installed": False, "detail": "检测出错: %r" % e}
-        if detected:
-            with _LOCK:
-                _STATE["detected"] = detected
-    finally:
-        with _LOCK:
-            _STATE["detect_ts"] = time.time()
-            _STATE["detect_ev"] = None
-        ev.set()
-    return detected
-
-
-def _uwp_version(package_dir):
-    import xml.etree.ElementTree as ET
-    mf = os.path.join(package_dir, "AppxManifest.xml")
-    if not os.path.isfile(mf):
-        return None
-    try:
-        tree = ET.parse(mf)
-        for el in tree.iter():
-            if el.tag.endswith("}Identity") or el.tag == "Identity":
-                return el.get("Version")
-    except Exception:
-        pass
-    return None
-
-
-def _exe_version(path):
-    try:
-        r = subprocess.run(
-            ["powershell", "-NoProfile", "-Command",
-             "(Get-Item -LiteralPath '%s').VersionInfo.ProductVersion" % path.replace("'", "''")],
-            capture_output=True, creationflags=CREATE_NO_WINDOW, timeout=25)
-        out = r.stdout.decode("utf-8", "replace").strip()
-        return out or None
-    except Exception:
-        return None
-
-
-def version_of(entry):
-    """版本探测：CLI 走 --version；UWP 读 AppxManifest；exe 走 PowerShell（惰性缓存）。"""
-    eid = entry["id"]
-    with _LOCK:
-        cached = _STATE["versions"].get(eid)
-        if cached and time.time() - cached[0] < VERSION_TTL:
-            return cached[1]
-    det = (detect_all() or {}).get(eid) or {}
-    version = None
-    cli = (entry.get("detect") or {}).get("cli")
-    if cli and det.get("installed"):
-        try:
-            r = subprocess.run(["cmd", "/c", cli, "--version"], capture_output=True,
-                               creationflags=CREATE_NO_WINDOW, timeout=20)
-            out = (r.stdout or b"").decode("utf-8", "replace").strip()
-            if not out:
-                out = (r.stderr or b"").decode("utf-8", "replace").strip()
-            version = out.splitlines()[0][:60] if out else None
-        except Exception:
-            version = None
-    elif det.get("detail") and os.path.isdir(det["detail"]):
-        version = _uwp_version(det["detail"])
-    elif det.get("detail") and os.path.isfile(det["detail"]):
-        version = _exe_version(det["detail"])
-    with _LOCK:
-        _STATE["versions"][eid] = (time.time(), version or "-")
-    return version or "-"
-
-
-# ---------------------------------------------------------------- TOML 表内键
-
-def _toml_span(lines, table):
-    """定位顶层表 `[table]` 的行区间 [start, end)；start 为表头行。
-    只认顶层表头（行首无空白），不误吞嵌套 `[[array]]` 之外的子表——
-    子表在 TOML 里也是 `[a.b]` 顶层写法，同样按表头截断。"""
-    header = re.compile(r"^\[([^\[\]]+)\]\s*$")
-    start = None
-    for i, ln in enumerate(lines):
-        if ln.lstrip().startswith("["):
-            m = header.match(ln.strip())
-            if not m:
-                continue
-            if start is not None:
-                return start, i
-            if m.group(1).strip() == table:
-                start = i
-    if start is None:
-        return None, None
-    return start, len(lines)
-
-
-def _toml_read_value(text, table, leaf):
-    m = re.search(r'(?m)^\s*%s\s*=\s*"([^"]*)"\s*(#.*)?$' % re.escape(leaf), text) \
-        if table is None else None
-    if table is None:
-        return m.group(1) if m else None
-    start, end = _toml_span(text.splitlines(), table)
-    if start is None:
-        return None
-    for ln in text.splitlines()[start + 1:end]:
-        m = re.match(r'^\s*%s\s*=\s*"([^"]*)"\s*(#.*)?$' % re.escape(leaf), ln)
-        if m:
-            return m.group(1)
-    return None
-
-
-def _toml_write_value(text, table, leaf, value):
-    """就地写入 TOML 的 [table] leaf（双引号标量），保留其余内容与换行风格。"""
-    eol = "\r\n" if "\r\n" in text else "\n"
-    lines = text.splitlines()
-    new_line = '%s = "%s"' % (leaf, value.replace("\\", "\\\\").replace('"', '\\"'))
-    # 换值不换行：保留行尾注释等其余内容
-    pat = re.compile(r'^(\s*%s\s*=\s*)"(?:[^"\\]|\\.)*"(.*)$' % re.escape(leaf))
-    if table is None:
-        for i, ln in enumerate(lines):
-            m = pat.match(ln)
-            if m:
-                lines[i] = "%s\"%s\"%s" % (m.group(1), value.replace("\\", "\\\\").replace('"', '\\"'), m.group(2))
-                break
-        else:
-            lines.append(new_line)
-        return eol.join(lines).rstrip("\r\n") + eol
-    start, end = _toml_span(lines, table)
-    if start is None:  # 表不存在：整段追加
-        if lines and lines[-1].strip():
-            lines.append("")
-        lines.append("[%s]" % table)
-        lines.append(new_line)
-        return eol.join(lines).rstrip("\r\n") + eol
-    for i in range(start + 1, end):
-        m = pat.match(lines[i])
-        if m:
-            lines[i] = "%s\"%s\"%s" % (m.group(1), value.replace("\\", "\\\\").replace('"', '\\"'), m.group(2))
-            return eol.join(lines).rstrip("\r\n") + eol
-    lines.insert(end, new_line)  # 表内末尾追加（表头区间终点即下一表头前）
-    return eol.join(lines).rstrip("\r\n") + eol
-
-
-# ---------------------------------------------------------------- 模型配置
-
-def _config_path(entry):
-    cfg = entry.get("config") or {}
-    return _safe_config_path(cfg.get("path"))
-
-
-def _yaml_model_path(cfg):
-    """把 config.model_key 的点号路径拆成 (段, 键)；无点号时段为 None（顶层键）。"""
-    return _dotted_key(cfg)
-
-
-def _dotted_key(cfg):
-    """model_key 点号路径拆 (表/段, 键)；无点号时第一元为 None（顶层键）。"""
-    key = (cfg.get("model_key") or "model").strip()
-    if "." in key:
-        section, leaf = key.split(".", 1)
-        return section.strip(), leaf.strip()
-    return None, key
-
-
-def _yaml_quote(value):
-    """YAML 双引号标量。必须加引号：模型名可能以 [ 开头（YAML 流序列）或含 #。"""
-    return '"%s"' % value.replace("\\", "\\\\").replace('"', '\\"')
-
-
-def _yaml_unquote(raw):
-    """取 YAML 标量的值：剥引号、丢行尾注释。引号内的 # 不算注释。"""
-    s = (raw or "").strip()
-    if not s:
-        return ""
-    if s[0] in ("'", '"'):
-        q = s[0]
-        i = 1
-        buf = []
-        while i < len(s):
-            ch = s[i]
-            if q == '"' and ch == "\\" and i + 1 < len(s):
-                nxt = s[i + 1]
-                buf.append('"' if nxt == '"' else ("\\" if nxt == "\\" else nxt))
-                i += 2
-                continue
-            if ch == q:
-                if q == "'" and i + 1 < len(s) and s[i + 1] == "'":  # '' 转义
-                    buf.append("'")
-                    i += 2
-                    continue
-                break
-            buf.append(ch)
-            i += 1
-        return "".join(buf).strip()
-    # 无引号：截断行尾注释（# 前的空白才算注释起始）
-    return re.split(r"\s+#", s, 1)[0].strip()
-
-
-def _yaml_span(lines, section):
-    """定位顶层段的行区间 [start, end)；start 为段名行，其子键在 start+1..end。"""
-    start = None
-    for i, ln in enumerate(lines):
-        m = re.match(r"^([^\s#][^:]*):\s*(.*)$", ln)
-        if not m:
-            continue
-        if m.group(1).strip() == section:
-            start = i
-            continue
-        if start is not None:
-            return start, i
-    if start is None:
-        return None, None
-    return start, len(lines)
-
-
-def _yaml_read_value(text, section, leaf):
-    lines = text.splitlines()
-    if section is None:
-        m = re.search(r"(?m)^%s\s*:\s*(.+?)\s*$" % re.escape(leaf), text)
-        return _yaml_unquote(m.group(1)) or None if m else None
-    start, end = _yaml_span(lines, section)
-    if start is None:
-        return None
-    for ln in lines[start + 1:end]:
-        m = re.match(r"^\s+%s\s*:\s*(.+?)\s*$" % re.escape(leaf), ln)
-        if m:
-            return _yaml_unquote(m.group(1)) or None
-    return None
-
-
-def _yaml_write_value(text, section, leaf, value):
-    """就地写入 YAML 的 section.leaf，保留其余内容、缩进与换行风格。"""
-    eol = "\r\n" if "\r\n" in text else "\n"
-    lines = text.splitlines()
-    quoted = _yaml_quote(value)
-    if section is None:
-        for i, ln in enumerate(lines):
-            if re.match(r"^%s\s*:" % re.escape(leaf), ln):
-                lines[i] = "%s: %s" % (leaf, quoted)
-                break
-        else:
-            lines.append("%s: %s" % (leaf, quoted))
-        return eol.join(lines).rstrip("\r\n") + eol
-    start, end = _yaml_span(lines, section)
-    if start is None:  # 段不存在：整段追加
-        if lines and lines[-1].strip():
-            lines.append("")
-        lines.append("%s:" % section)
-        lines.append("  %s: %s" % (leaf, quoted))
-        return eol.join(lines).rstrip("\r\n") + eol
-    for i in range(start + 1, end):
-        m = re.match(r"^(\s+)%s\s*:" % re.escape(leaf), lines[i])
-        if m:  # 键已存在：只换值，保留原缩进
-            lines[i] = "%s%s: %s" % (m.group(1), leaf, quoted)
-            return eol.join(lines).rstrip("\r\n") + eol
-    indent, last = "  ", start  # 段存在但无该键：跟随段内缩进、追加到段尾
-    for i in range(start + 1, end):
-        if not lines[i].strip():
-            continue
-        if indent == "  ":
-            m = re.match(r"^(\s+)\S", lines[i])
-            if m:
-                indent = m.group(1)
-        last = i
-    lines.insert(last + 1, "%s%s: %s" % (indent, leaf, quoted))
-    return eol.join(lines).rstrip("\r\n") + eol
-
-
-def _json_path_get(data, keys):
-    """沿点号路径下钻 JSON 嵌套；终点必须是字符串。"""
-    cur = data
-    for k in keys:
-        if isinstance(cur, dict) and k in cur:
-            cur = cur[k]
-        else:
-            return None
-    return cur if isinstance(cur, str) else None
-
-
-def _json_path_set(data, keys, value):
-    """沿点号路径写入 JSON 嵌套，缺中间对象就地创建；
-    中途遇到非 dict（如 string 简写形式）升级为对象。"""
-    cur = data
-    for k in keys[:-1]:
-        if not isinstance(cur.get(k), dict):
-            cur[k] = {}
-        cur = cur[k]
-    cur[keys[-1]] = value
-
-
-def read_model(entry):
-    path = _config_path(entry)
-    cfg = entry.get("config") or {}
-    if not path or not os.path.isfile(path) or not cfg.get("format"):
-        return None
-    try:
-        text = open(path, encoding="utf-8", errors="replace").read()
-    except Exception:
-        return None
-    if cfg["format"] == "toml-line":
-        m = re.search(r'(?m)^\s*model\s*=\s*"([^"]+)"', text)
-        return m.group(1) if m else None
-    if cfg["format"] == "toml-section":
-        table, leaf = _dotted_key(cfg)
-        return _toml_read_value(text, table, leaf)
-    if cfg["format"] == "json":
-        try:
-            v = json.loads(text).get("model")
-            if v:
-                return v
-        except Exception:
-            pass
-        m = re.search(r'"model"\s*:\s*"([^"]+)"', text)
-        return m.group(1) if m else None
-    if cfg["format"] == "jsonc":
-        keys = (cfg.get("model_key") or "model").split(".")
-        try:
-            data = json.loads(_jsonc_strip_comments(text) or "{}")
-        except Exception:
-            return None
-        return _json_path_get(data, keys)
-    if cfg["format"] == "json-path":
-        try:
-            data = json.loads(text or "{}")
-        except Exception:
-            return None
-        return _json_path_get(data, (cfg.get("model_key") or "model").split("."))
-    if cfg["format"] == "yaml-line":
-        return _yaml_read_value(text, *_yaml_model_path(cfg))
-    return None
-
-
-def write_model(entry, model):
-    """写入默认模型（改动前自动备份 .bak）。支持 toml-line / toml-section /
-    json / json-path / jsonc / yaml-line。"""
-    path = _config_path(entry)
-    cfg = entry.get("config") or {}
-    fmt = cfg.get("format")
-    if not path:
-        return {"ok": False,
-                "error": "配置路径无效或不在用户主目录内，已拒绝写入"}
-    if fmt not in _WRITABLE_FORMATS:
-        return {"ok": False, "error": "该工具的模型配置格式暂不支持自动写入，请手动编辑 %s" % path}
-    model = (model or "").strip()
-    if not model:
-        return {"ok": False, "error": "模型名不能为空"}
-    # 写入前二次校验：解析真实路径后必须仍在用户主目录内（防穿越/符号链接），
-    # 校验通过后统一改用解析路径写入
-    home = os.path.abspath(os.path.expanduser("~"))
-    path = os.path.realpath(path)
-    try:
-        if os.path.commonpath([path, home]) != home:
-            return {"ok": False, "error": "配置路径越出用户主目录，已拒绝写入"}
-    except ValueError:
-        return {"ok": False, "error": "配置路径越出用户主目录，已拒绝写入"}
-    try:
-        if os.path.isfile(path):
-            shutil.copyfile(path, path + ".bak")
-        else:
-            # 各 CLI 首次运行都未必建主配置（grok 不建 config.toml、pi 不建
-            # settings.json、openclaw 缺失即安全默认——官方文档明确「缺失即
-            # 内置默认」）；它正是用户层覆盖的落点，缺了按需创建
-            Path(path).parent.mkdir(parents=True, exist_ok=True)
-            Path(path).write_bytes(b"{}" if fmt in ("json", "json-path", "jsonc") else b"")
-        if fmt == "jsonc":
-            # mimo（mimocode.jsonc）有注释，整体重解析会丢注释——复用
-            # _jsonc_set 做就地片段改写（只动目标键，其余原样保留）
-            keys = (cfg.get("model_key") or "model").split(".")
-            text = open(path, encoding="utf-8", errors="replace", newline="").read()
-            new_text, ok = _jsonc_set(text, tuple(keys),
-                                      json.dumps(model, ensure_ascii=False))
-            if not ok:
-                return {"ok": False,
-                        "error": "jsonc 结构异常，未能就地写入 %s（已避免覆盖）" % path}
-            Path(path).write_bytes(new_text.encode("utf-8"))
-        elif fmt == "toml-line":
-            text = open(path, encoding="utf-8", errors="replace").read()
-            new_line = 'model = "%s"' % model
-            if re.search(r'(?m)^\s*model\s*=\s*"[^"]*"', text):
-                text = re.sub(r'(?m)^\s*model\s*=\s*"[^"]*"', new_line, text)
-            else:
-                text = text.rstrip("\n") + "\n" + new_line + "\n"
-            Path(path).write_bytes(text.encode("utf-8"))
-        elif fmt == "toml-section":
-            text = open(path, encoding="utf-8", errors="replace", newline="").read()
-            text = _toml_write_value(text, *_dotted_key(cfg), value=model)
-            Path(path).write_bytes(text.encode("utf-8"))
-        elif fmt == "yaml-line":
-            # newline="" 关掉通用换行转换：文本层面看不出 \r\n 就会被静默改写成 LF，
-            # 用户的 Windows 配置不该因为写个模型名而整篇换行符被替换
-            text = open(path, encoding="utf-8", errors="replace", newline="").read()
-            text = _yaml_write_value(text, *_yaml_model_path(cfg), value=model)
-            Path(path).write_bytes(text.encode("utf-8"))
-        elif fmt == "json-path":
-            # openclaw（agents.defaults.model.primary）：默认模型藏在嵌套对象里，
-            # 且 openclaw 对未知顶层键直接拒绝启动——绝不能写顶层 "model"
-            text = open(path, encoding="utf-8", errors="replace", newline="").read()
-            try:
-                data = json.loads(text) if text.strip() else {}
-            except Exception:
-                return {"ok": False, "error": "配置文件不是合法 JSON，已中止（避免覆盖）"}
-            _json_path_set(data, (cfg.get("model_key") or "model").split("."), model)
-            # pi 的 settings.json：defaultModel 必须配 defaultProvider 才能解析出
-            # (provider, model) 二元组；catalog 里声明了的伴随键一并落盘
-            for k, v in (cfg.get("model_extra_keys") or {}).items():
-                _json_path_set(data, k.split("."), v)
-            Path(path).write_bytes(
-                json.dumps(data, ensure_ascii=False, indent=2).encode("utf-8"))
-        else:
-            try:
-                data = json.loads(open(path, encoding="utf-8", errors="replace").read())
-            except FileNotFoundError:
-                return {"ok": False,
-                        "error": "配置文件尚未生成（%s 首次运行后才有），暂无法写入" % path}
-            except Exception:
-                return {"ok": False, "error": "配置文件不是合法 JSON，已中止（避免覆盖）"}
-            data["model"] = model
-            Path(path).write_bytes(
-                json.dumps(data, ensure_ascii=False, indent=2).encode("utf-8"))
-        return {"ok": True, "model": read_model(entry)}
-    except Exception as e:
-        return {"ok": False, "error": repr(e)}
-
-
-# ---------------------------------------------------------------- 一键打开
-
-def _port_open(port, timeout=0.5):
-    try:
-        with socket.create_connection(("127.0.0.1", port), timeout=timeout):
-            return True
-    except OSError:
-        return False
-
-
-def launch_env(entry):
-    """打开交互/网页版时注入的子进程环境变量：与编排同源的绑定凭据
-    （dsh=DEEPSEEK_*，claude=ANTHROPIC_*，codex=ORCH_API_KEY）。交互进程
-    脱离了编排链路，没有这层注入就拿不到 API key。无绑定时返回 {}。"""
-    try:
-        from . import modelhub  # 惰性导入：modelhub 体量大且避免潜在环
-        b = modelhub.resolve_binding(entry["id"]) or {}
-    except Exception:
-        return {}
-    return dict(b.get("env") or {})
-
-
-def _launch_log_path(entry):
-    """web 类启动日志的落点：id 白名单化后仅作文件名成分，最终路径必须仍围栏
-    在数据目录内（catalog.json 用户可编辑，id 不可信；非白名单 id 用 crc32
-    稳定代称——跨进程重启不变，「已在运行」回读上次日志才找得到）。"""
-    raw_id = str(entry.get("id") or "")
-    if re.fullmatch(r"[A-Za-z0-9_-]{1,64}", raw_id):
-        stem = raw_id
-    else:
-        stem = "agent-%d" % (zlib.crc32(raw_id.encode("utf-8")) & 0xFFFFFFFF)
-    data_root = os.path.abspath(str(paths.DATA_DIR))
-    p = Path(data_root, "launch", stem + ".log")
-    try:
-        if os.path.commonpath([os.path.abspath(str(p)), data_root]) != data_root:
-            return None
-    except ValueError:
-        return None
-    return p
-
-
-def _best_url(log_path, port):
-    """从启动日志提取该端口的信任 URL（含 token 优先）。无日志/未匹配返回 None。"""
-    try:
-        text = open(str(log_path), encoding="utf-8", errors="replace").read()
-    except Exception:
-        return None
-    urls = re.findall(r"https?://[^\s\"'<>]+", text)
-    same = [u for u in urls if ":%d" % port in u]
-    if not same:
-        return None
-    tokened = [u for u in same if "token=" in u]
-    return (tokened or same)[0]
-
-
-def _yaml_model_ids(text, section):
-    """收集 section.models 序列里的全部模型 id（保持顺序）。段/键缺失返回 []。"""
-    lines = text.splitlines()
-    start, end = _yaml_span(lines, section)
-    if start is None:
-        return []
-    m_indent = None
-    for i in range(start + 1, end):
-        m = re.match(r"^(\s+)models\s*:\s*(?:#.*)?$", lines[i])
-        if m:
-            m_indent = len(m.group(1))
-            start = i
-            break
-    if m_indent is None:
-        return []
-    ids = []
-    for ln in lines[start + 1:end]:
-        if not ln.strip():
-            continue
-        if len(ln) - len(ln.lstrip(" ")) <= m_indent:
-            break  # models 列表结束（遇到同级或更浅缩进的键）
-        m = re.match(r"^\s*-\s+id\s*:\s*(.+?)\s*$", ln)
-        if m:
-            ids.append(_yaml_unquote(m.group(1)))
-    return ids
-
-
-def _yaml_ensure_model_entry(text, section, model, context_window=1000000):
-    """确保 section.models 序列里有 id==model 的条目，缺则按现有条目形状追加
-    （id/name/contextWindow——dsh 的 catalog 校验要求 id 与 name 非空）。
-    已存在或段/列表结构不完整时原文返回。返回 (new_text, added)。"""
-    eol = "\r\n" if "\r\n" in text else "\n"
-    lines = text.splitlines()
-    start, end = _yaml_span(lines, section)
-    if start is None:
-        return text, False
-    m_idx = m_indent = None
-    for i in range(start + 1, end):
-        m = re.match(r"^(\s+)models\s*:\s*(?:#.*)?$", lines[i])
-        if m:
-            m_idx, m_indent = i, len(m.group(1))
-            break
-    if m_idx is None:
-        return text, False  # 段内没有 models 键：不凭空造结构（保守）
-    item_indent = None
-    last_item = m_idx
-    i = m_idx + 1
-    while i < end:
-        ln = lines[i]
-        if not ln.strip():
-            i += 1
-            continue
-        if len(ln) - len(ln.lstrip(" ")) <= m_indent:
-            break  # models 列表结束
-        m = re.match(r"^(\s*)-\s+id\s*:\s*(.+?)\s*$", ln)
-        if m:
-            item_indent = m.group(1)
-            last_item = i
-            if _yaml_unquote(m.group(2)) == model:
-                return text, False
-        elif re.match(r"^\s+\S", ln):
-            last_item = i  # 条目的续属性行（name/contextWindow…）
-        i += 1
-    item_indent = item_indent or (" " * (m_indent + 2))
-    block = ["%s- id: %s" % (item_indent, _yaml_quote(model)),
-             "%s  name: %s" % (item_indent, _yaml_quote(model)),
-             "%s  contextWindow: %d" % (item_indent, context_window)]
-    lines[last_item + 1:last_item + 1] = block
-    return eol.join(lines).rstrip("\r\n") + eol, True
-
-
-def _sync_dsh_settings(entry, model, base_url):
-    """dsh 专属：把绑定模型的端点与模型目录写进 ~/.dsh/settings.yaml 的
-    llm-deepseek 段（agent-default-model.model 由 write_model 负责）。
-    端点不同步不行——settings 优先级高于 env，密钥会发给旧端点；
-    models 列表不同步不行——dsh web 的模型下拉只列它，缺条目就选不中。
-    返回错误串或 None。"""
-    path = _config_path(entry)
-    if not path:
-        return "dsh 配置路径无效"
-    path = os.path.realpath(path)
-    home = os.path.abspath(os.path.expanduser("~"))
-    try:
-        if os.path.commonpath([path, home]) != home:
-            return "dsh 配置路径越出用户主目录，已拒绝"
-    except ValueError:
-        return "dsh 配置路径越出用户主目录，已拒绝"
-    try:
-        if os.path.isfile(path):
-            text = open(path, encoding="utf-8", errors="replace", newline="").read()
-        else:
-            Path(path).parent.mkdir(parents=True, exist_ok=True)
-            text = ""
-        text2 = _yaml_write_value(text, "llm-deepseek", "baseURL", base_url)
-        text2, _added = _yaml_ensure_model_entry(text2, "llm-deepseek", model)
-        if text2 != text:
-            if os.path.isfile(path):
-                shutil.copyfile(path, path + ".bak")
-            Path(path).write_bytes(text2.encode("utf-8"))
-        return None
-    except Exception as e:
-        return repr(e)
-
-
-def _dsh_selfcheck_model(entry):
-    """dsh 无绑定时自检：agent-default-model.model 必须在端点 models 列表内，
-    否则 web UI 打开就是一个选不中的模型（glm-5.3-flash vs V4 端点的实况）。
-    不在列表则改选列表第一个并写回。返回 (生效模型, note)。"""
-    path = _config_path(entry)
-    if not path or not os.path.isfile(path):
-        return None, ""
-    try:
-        text = open(path, encoding="utf-8", errors="replace", newline="").read()
-    except Exception:
-        return None, ""
-    cur = _yaml_read_value(text, "agent-default-model", "model")
-    ids = _yaml_model_ids(text, "llm-deepseek")
-    if not ids:
-        return cur, "dsh 端点未登记任何模型，请先在 dsh 侧配置模型目录"
-    if cur in ids:
-        return cur, ""
-    first = ids[0]
-    w = write_model(entry, first)
-    fixed = w.get("model") if w.get("ok") else None
-    note = "dsh 默认模型 %s 不在端点模型列表，已改选 %s" % (cur or "（空）", first)
-    if not fixed:
-        note += "（写入失败：%s）" % w.get("error")
-    return fixed, note
-
-
-def _dsh_key_present():
-    """dsh 的密钥是否有着落：进程 env 或它自己的 ~/.dsh/.env（credentials-local）。"""
-    if os.environ.get("DEEPSEEK_API_KEY"):
-        return True
-    try:
-        envfile = os.path.join(os.path.expanduser("~"), ".dsh", ".env")
-        return "DEEPSEEK_API_KEY" in open(envfile, encoding="utf-8", errors="replace").read()
-    except Exception:
-        return False
-
-
-def _toml_section_set(text, section, pairs):
-    """就地写 TOML 段（[section] 下多键）。段存在则逐键替换，缺则整段追加在文末。
-    只处理 codex config.toml 这种顶层简单段；返回 (new_text, changed)。"""
-    eol = "\r\n" if "\r\n" in text else "\n"
-    lines = text.splitlines()
-    start = end = None
-    header = "[%s]" % section
-    for i, ln in enumerate(lines):
-        if ln.strip() == header:
-            start = i
-        elif start is not None and ln.startswith("[") and ln.rstrip().endswith("]"):
-            end = i
-            break
-    if start is None:
-        block = [header] + ["%s = %s" % (k, v) for k, v in pairs]
-        if lines and lines[-1].strip():
-            lines.append("")
-        lines.extend(block)
-        return eol.join(lines).rstrip("\r\n") + eol, True
-    end = end if end is not None else len(lines)
-    changed = False
-    todo = dict(pairs)
-    for i in range(start + 1, end):
-        m = re.match(r"^(\s*)([A-Za-z0-9_.-]+)\s*=", lines[i])
-        if m and m.group(2) in todo:
-            lines[i] = "%s%s = %s" % (m.group(1), m.group(2), todo.pop(m.group(2)))
-            changed = True
-    if todo:
-        ins = start + 1
-        while ins < end and not lines[ins].strip():
-            ins += 1
-        for k, v in list(todo.items())[::-1]:
-            lines.insert(ins, "%s = %s" % (k, v))
-        changed = True
-    return eol.join(lines).rstrip("\r\n") + eol, changed
-
-
-def _toml_top_set(text, key, value):
-    """写 TOML 顶层键（第一个 [段] 之前的区域）。返回 (new_text, changed)。"""
-    eol = "\r\n" if "\r\n" in text else "\n"
-    lines = text.splitlines()
-    pat = re.compile(r"^(\s*)%s\s*=\s*.+$" % re.escape(key))
-    for i, ln in enumerate(lines):
-        if pat.match(ln):
-            lines[i] = "%s%s = %s" % (pat.match(ln).group(1), key, value)
-            return eol.join(lines).rstrip("\r\n") + eol, True
-    first_section = next((i for i, ln in enumerate(lines)
-                          if ln.startswith("[") and ln.rstrip().endswith("]")), len(lines))
-    lines.insert(first_section, "%s = %s" % (key, value))
-    return eol.join(lines).rstrip("\r\n") + eol, True
-
-
-def _sync_codex_settings(entry, model, cp):
-    """codex 专属：把绑定供应商与模型写进 ~/.codex/config.toml
-    （[model_providers.orch] 段 + 顶层 model_provider/model）。
-
-    codex 交互 TUI 不认编排的 -c 一次性覆盖，也不认 ORCH_API_KEY env——
-    没有 config.toml 里的 provider 段，绑定模型根本无处可用；而 model 单写
-    不写 provider 会指到 codex 自带 openai 官方端点上（401）。与编排的
-    _codex_provider_args 同构，但落 config 文件。返回错误串或 None。"""
-    path = _config_path(entry)
-    if not path:
-        return "codex 配置路径无效"
-    path = os.path.realpath(path)
-    home = os.path.abspath(os.path.expanduser("~"))
-    try:
-        if os.path.commonpath([path, home]) != home:
-            return "codex 配置路径越出用户主目录，已拒绝"
-    except ValueError:
-        return "codex 配置路径越出用户主目录，已拒绝"
-    name = cp.get("name", "orch")
-    def q(v):
-        return '"%s"' % str(v).replace("\\", "\\\\").replace('"', '\\"')
-    pairs = [("name", q(cp.get("name", name))),
-             ("base_url", q(cp.get("base_url", ""))),
-             ("env_key", q(cp.get("env_key", "ORCH_API_KEY"))),
-             ("wire_api", q(cp.get("wire_api", "responses")))]
-    try:
-        if os.path.isfile(path):
-            text = open(path, encoding="utf-8", errors="replace", newline="").read()
-        else:
-            Path(path).parent.mkdir(parents=True, exist_ok=True)
-            text = ""
-        text2, _ = _toml_section_set(text, "model_providers.%s" % name, pairs)
-        text2, _ = _toml_top_set(text2, "model_provider", q(name))
-        if model:
-            text2, _ = _toml_top_set(text2, "model", q(model))
-        if text2 != text:
-            if os.path.isfile(path):
-                shutil.copyfile(path, path + ".bak")
-            Path(path).write_bytes(text2.encode("utf-8"))
-        return None
-    except Exception as e:
-        return repr(e)
-
-
-# ---------------------------------------------------------------- 打开前凭据注入
-
-def _jsonc_scan_object(text, start):
-    """扫描 text[start]（须为 '{'）起的 JSON(C) 对象：返回 (闭合偏移, 键表)。
-    键表为 [key, key_start, key_end, value_start, value_end]（偏移相对整个 text，
-    value_end 指向值结束后一格）。跳过字符串转义、// 与 /* */ 注释、嵌套括号，
-    未闭合时容错返回文末。"""
-    n = len(text)
-    i = start + 1
-    keys = []
-    state = "key"
-    depth = 0
-
-    def skip_string(j):
-        while j < n:
-            if text[j] == "\\":
-                j += 2
-                continue
-            if text[j] == '"':
-                return j + 1
-            j += 1
-        return n
-
-    def skip_comment(j):
-        if j + 1 < n and text[j + 1] == "/":
-            e = text.find("\n", j)
-            return n if e < 0 else e
-        if j + 1 < n and text[j + 1] == "*":
-            e = text.find("*/", j + 2)
-            return n if e < 0 else e + 2
-        return j + 1  # 非注释的孤立斜杠：当普通字符
-
-    while i < n:
-        ch = text[i]
-        if ch in " \t\r\n":
-            i += 1
-            continue
-        if ch == "/":
-            i = skip_comment(i)
-            continue
-        if state == "key":
-            if ch == "}":
-                return i, keys
-            if ch == ",":
-                i += 1
-                continue
-            if ch == '"':
-                j = skip_string(i + 1)
-                keys.append([text[i + 1:j - 1], i, j, 0, 0])
-                state = "colon"
-                i = j
-                continue
-            i += 1
-            continue
-        if state == "colon":
-            if ch == ":":
-                state = "value"
-            i += 1
-            continue
-        # state == value
-        if not keys:
-            return n, keys  # 结构异常：放弃扫描
-        keys[-1][3] = i
-        if ch in "{[":
-            depth = 0
-            j = i
-            while j < n:
-                c = text[j]
-                if c == '"':
-                    j = skip_string(j + 1)
-                    continue
-                if c == "/":
-                    j = skip_comment(j)
-                    continue
-                if c in "{[":
-                    depth += 1
-                elif c in "}]":
-                    depth -= 1
-                    if depth == 0:
-                        break
-                j += 1
-            keys[-1][4] = j + 1
-            i = j + 1
-        elif ch == '"':
-            j = skip_string(i + 1)
-            keys[-1][4] = j
-            i = j
-        else:  # 数字 / true / false / null
-            j = i
-            while j < n and text[j] not in ",}\r\n":
-                j += 1
-            keys[-1][4] = j
-            i = j
-    state = "key"
-    return n, keys
-
-
-def _jsonc_strip_comments(text):
-    """剥掉 jsonc 的 // 行注释与 /* */ 块注释，供 json.loads 解析。
-    字符串字面量内部的 //（$schema 的 https:// 等）不剥——逐字符扫描；
-    简单的 re.sub(r"//[^\\n]*") 会把 URL 截断导致解析失败。"""
-    out = []
-    i, n = 0, len(text)
-    while i < n:
-        ch = text[i]
-        if ch == '"':
-            j = i + 1
-            while j < n:
-                if text[j] == "\\":
-                    j += 2
-                    continue
-                if text[j] == '"':
-                    break
-                j += 1
-            out.append(text[i:min(j + 1, n)])
-            i = j + 1
-            continue
-        if ch == "/" and i + 1 < n and text[i + 1] == "/":
-            e = text.find("\n", i)
-            i = n if e < 0 else e  # 行注释：换行符本身保留
-            continue
-        if ch == "/" and i + 1 < n and text[i + 1] == "*":
-            e = text.find("*/", i + 2)
-            i = n if e < 0 else e + 2
-            out.append(" ")
-            continue
-        out.append(ch)
-        i += 1
-    return "".join(out)
-
-
-def _jsonc_set(text, path, value_json):
-    """JSON(C) 顶层就地写键：path=("provider","orch") 或 ("model",)。
-    只动目标片段，其余文本（含注释与原格式）原样保留。返回 (new_text, ok)。
-    空文件按空对象起笔（write_model 缺文件按需创建的产物）。"""
-    brace = text.find("{")
-    if brace < 0 and not text.strip():
-        return ('{\n  "%s": %s\n}' % (path[0], value_json), True)
-    if brace < 0:
-        return text, False
-    end, keys = _jsonc_scan_object(text, brace)
-    head = path[0]
-    if len(path) == 1:
-        for _k, _ks, _ke, vs, ve in keys:
-            if _k == head:
-                return text[:vs] + value_json + text[ve:], True
-        return _jsonc_insert_entry(text, brace, end, head, value_json), True
-    # 二级路径：先定位一级键的值对象
-    tgt = None
-    for k, _ks, _ke, vs, ve in keys:
-        if k == head:
-            tgt = (vs, ve)
-            break
-    if tgt is None:  # 一级键不存在：整体插入
-        block = json.dumps({path[1]: json.loads(value_json)}, ensure_ascii=False)
-        return _jsonc_insert_entry(text, brace, end, head, block), True
-    vs, ve = tgt
-    if text[vs:ve].lstrip()[0:1] != "{":
-        return text, False  # 一级值不是对象：保守放弃（不覆盖用户的非标结构）
-    end2, keys2 = _jsonc_scan_object(text, vs)
-    for k, _ks, _ke, v2s, v2e in keys2:
-        if k == path[1]:
-            return text[:v2s] + value_json + text[v2e:], True
-    seg = text[vs:end2 + 1]
-    new_seg = _jsonc_insert_entry(seg, 0, end2 - vs, path[1], value_json)
-    if new_seg is None:
-        return text, False
-    return text[:vs] + new_seg + text[end2 + 1:], True
-
-
-def _jsonc_insert_entry(text, brace, end, key, value_json):
-    """在 {brace..end} 对象的开头插入 "key": value（带尾逗号，不依赖原文件的
-    逗号风格）；对象为空时去掉多余逗号。"""
-    m = re.match(r"\{([ \t\r\n]*)", text[brace:end + 1])
-    first = brace + (m.end(1) if m else 1)  # m 从 brace 起 match：end(1) 已含 '{'
-    if first >= end:  # 空对象 {}
-        return text[:brace + 1] + ' "%s": %s ' % (key, value_json) + text[end:]
-    return text[:first] + '"%s": %s,\n  ' % (key, value_json) + text[first:]
-
-
-def _sync_settings_env(path, updates, remove_keys=()):
-    """把键值对写进目标 settings.json 的 env 段（claude/qwen 等同构：交互 TUI
-    启动时把该段合并进进程环境）。只动 env 相关键，其余内容保留；坏 JSON 中止
-    不覆盖；改动前 .bak。返回错误串或 None。path 必须已过主目录围栏校验。"""
-    try:
-        if os.path.isfile(path):
-            text = open(path, encoding="utf-8", errors="replace").read()
-            try:
-                data = json.loads(text)
-            except Exception:
-                return "settings.json 不是合法 JSON，已中止（避免覆盖）"
-        else:
-            Path(path).parent.mkdir(parents=True, exist_ok=True)
-            data = {}
-        if not isinstance(data, dict):
-            return "settings.json 结构异常（顶层不是对象），已中止"
-        env = data.get("env")
-        if not isinstance(env, dict):
-            env = {}
-        for k in remove_keys:
-            env.pop(k, None)
-        env.update(updates)
-        data["env"] = env
-        if os.path.isfile(path):
-            shutil.copyfile(path, path + ".bak")
-        Path(path).write_bytes(json.dumps(data, ensure_ascii=False, indent=2).encode("utf-8"))
-        return None
-    except Exception as e:
-        return repr(e)
-
-
-def _guard_home(path):
-    """展开并校验配置路径必须在用户主目录内；通过则返回 realpath，否则 None。"""
-    if not path:
-        return None
-    path = os.path.realpath(path)
-    home = os.path.abspath(os.path.expanduser("~"))
-    try:
-        if os.path.commonpath([path, home]) != home:
-            return None
-    except ValueError:
-        return None
-    return path
-
-
-def _sync_claude_settings(entry, model, prov):
-    """claude-code 专属：把 anthropic 供应商的端点+密钥+模型写进
-    ~/.claude/settings.json 的 env 段（交互 TUI 与无头共用该文件，只认
-    ANTHROPIC_*；编排降级链给的 ORCH_API_KEY 对它等于没 key）。
-    返回错误串或 None。"""
-    path = _guard_home(_config_path(entry))
-    if not path:
-        return "claude 配置路径无效或越出用户主目录，已拒绝"
-    updates = {"ANTHROPIC_BASE_URL": prov.get("base_url") or "",
-               # AUTH_TOKEN 走 Bearer 头（Z.ai 等原生 anthropic 网关的用法）；
-               # 与 x-api-key 互斥，清掉可能残留的 ANTHROPIC_API_KEY 防止带错头
-               "ANTHROPIC_AUTH_TOKEN": prov.get("api_key") or ""}
-    if model:
-        updates["ANTHROPIC_MODEL"] = model
-    return _sync_settings_env(path, updates, remove_keys=("ANTHROPIC_API_KEY",))
-
-
-def _sync_qwen_settings(entry, model, prov):
-    """qwencode 专属：openai 兼容供应商写进 ~/.qwen/settings.json 的 env 段
-    （qwen-code 实测认 OPENAI_API_KEY/OPENAI_BASE_URL/OPENAI_MODEL，存在即
-    优先走 openai 兼容通道——2026-09-15 真机对维云端点实测请求到达并鉴权）。
-    anthropic 协议未实证，不开（协议不匹配时宁可提示）。返回错误串或 None。"""
-    path = _guard_home(_config_path(entry))
-    if not path:
-        return "qwen 配置路径无效或越出用户主目录，已拒绝"
-    updates = {"OPENAI_API_KEY": prov.get("api_key") or "",
-               "OPENAI_BASE_URL": prov.get("base_url") or ""}
-    if model:
-        updates["OPENAI_MODEL"] = model
-    return _sync_settings_env(path, updates)
-
-
-def _opencode_config_candidates(entry):
-    """opencode 配置的候选路径：catalog 登记的 jsonc 优先，其次同目录的
-    opencode.json（opencode 两种文件名都认，用户现有安装多用 json）。"""
-    out = []
-    primary = _config_path(entry)
-    if primary:
-        out.append(primary)
-    alt = os.path.join(os.path.abspath(os.path.expanduser("~/.config/opencode")),
-                       "opencode.json")
-    if alt not in out:
-        out.append(alt)
-    return out
-
-
-def _sync_opencode_settings(entry, model, prov):
-    """opencode 专属：把绑定供应商写进其配置的 provider.orch 段 + 顶层 model
-    （opencode 交互 TUI 只认自家配置文件里的凭据，ORCH_API_KEY env 对它等于
-    没 key）。写入所有已存在的候选文件（避免新文件遮蔽旧文件的读取优先级），
-    全不存在时建 catalog 登记的那个。纯 JSON 走整体读改写；带注释的 JSONC 走
-    文本级就地 patch（保留注释）。返回错误串或 None。"""
-    npm = "@ai-sdk/anthropic" if prov.get("protocol") == "anthropic" else "@ai-sdk/openai-compatible"
-    block = {"npm": npm, "name": prov.get("name") or "CodeBee 绑定",
-             "options": {"baseURL": prov.get("base_url") or "",
-                         "apiKey": prov.get("api_key") or ""},
-             "models": {model: {"name": model}} if model else {}}
-    top_model = ("orch/" + model) if model else ""
-    targets = [p for p in _opencode_config_candidates(entry) if os.path.isfile(p)] \
-        or [_opencode_config_candidates(entry)[0]]
-    errs = []
-    for path in targets:
-        try:
-            path = os.path.realpath(path)
-            home = os.path.abspath(os.path.expanduser("~"))
-            try:
-                if os.path.commonpath([path, home]) != home:
-                    errs.append("路径越出主目录：%s" % path)
-                    continue
-            except ValueError:
-                errs.append("路径越出主目录：%s" % path)
-                continue
-            if os.path.isfile(path):
-                text = open(path, encoding="utf-8", errors="replace", newline="").read()
-            else:
-                Path(path).parent.mkdir(parents=True, exist_ok=True)
-                text = ""
-            try:
-                data = json.loads(text) if text.strip() else {}
-            except Exception:
-                data = None
-            if data is not None:  # 纯 JSON：整体读改写（保序）
-                if not isinstance(data, dict):
-                    errs.append("结构异常（顶层不是对象）：%s" % path)
-                    continue
-                provs = data.get("provider")
-                if not isinstance(provs, dict):
-                    provs = {}
-                provs["orch"] = block
-                data["provider"] = provs
-                if top_model:
-                    data["model"] = top_model
-                new_text = json.dumps(data, ensure_ascii=False, indent=2) + "\n"
-            else:  # JSONC（带注释）：文本级 patch
-                block_json = json.dumps(block, ensure_ascii=False)
-                new_text, ok = _jsonc_set(text, ("provider", "orch"), block_json)
-                if not ok:
-                    errs.append("JSONC 就地改写失败：%s" % path)
-                    continue
-                if top_model:
-                    new_text, _ = _jsonc_set(new_text, ("model",),
-                                             json.dumps(top_model))
-            if os.path.isfile(path):
-                shutil.copyfile(path, path + ".bak")
-            Path(path).write_bytes(new_text.encode("utf-8"))
-        except Exception as e:
-            errs.append("%s: %r" % (path, e))
-    return "；".join(errs) or None
-
-
-# 打开前专属注入通道：{agent_id: (可注入协议, 注入器)}。交互 TUI 脱离编排链路，
-# 只认自家配置文件里的凭据，编排降级给的 env（ORCH_API_KEY 等）对它们无效。
-_AGENT_INJECTORS = {
-    "claude-code": (("anthropic",), _sync_claude_settings),
-    "opencode": (("anthropic", "openai"), _sync_opencode_settings),
-    "qwencode": (("openai",), _sync_qwen_settings),
-}
-
-# 无专属注入通道的专有协议 CLI：env 注入大概率无效，打开时明确告知而非静默废
-# （mimo 系 opencode 衍生但配置路径未实证，先按提示类；grok 吃 XAI_API_KEY 但
-# 无端点 env 可指中转，openai 协议供应商也用不上）
-_NO_CHANNEL_HINT = ("grok-build", "pi", "mimo-code")
-
-
-def _sync_agent_injection(entry, binding):
-    """打开前把绑定链里第一个可注入供应商落进 CLI 自家配置。与编排降级链解耦：
-    协议不匹配不降级（claude 拿 openai 的 key 等于没 key）。返回给用户看的提示
-    （成功注入 / 不可用原因），None=该 CLI 无需处理。"""
-    from . import modelhub
-    spec = _AGENT_INJECTORS.get(entry["id"])
-    if spec:
-        protocols, injector = spec
-        pick, note = modelhub.launch_pick(entry["id"], protocols)
-        if pick:
-            prov = pick["provider"]
-            err = injector(entry, pick["model"], prov)
-            if err:
-                return "%s 凭据同步失败：%s（打开后可能需在其自带界面登录）" % (
-                    entry.get("name", entry["id"]), err)
-            return "已注入 %s（%s · %s）" % (prov.get("name") or "供应商",
-                                            prov.get("protocol"), prov.get("base_url", ""))
-        return note
-    if entry["id"] in _NO_CHANNEL_HINT:
-        if binding.get("env"):
-            return "已按绑定注入 env，但该 CLI 未必认 CodeBee 的凭据通道，打开后若要求登录请在其界面内登录"
-        return "未绑定可用供应商：打开后需在其自带界面登录；要打开即用请到「CLI 绑定」页绑定"
-    return None
-
-
-def _sync_launch_model(entry, binding):
-    """打开前把「CLI 绑定」页选中的模型落到该 CLI 自己的配置文件——保证
-    交互/网页版启动即选中绑定模型（绑定页是运行时模型唯一真源，目录页的
-    「默认模型」只是手动快照，会滞后）。dsh 额外同步端点与 models 目录；
-    codex 额外落 provider 段（交互 TUI 只认 config.toml，不认编排的 -c 覆盖
-    与 ORCH_API_KEY env）。返回给用户看的同步笔记列表。"""
-    notes = []
-    model = (binding.get("model") or "").strip()
-    prov = binding.get("provider") or {}
-    fmt = (entry.get("config") or {}).get("format")
-    is_dsh = entry["id"] in ("deepseek-harness", "dsh")
-    is_codex = entry["id"] in ("codex-cli", "codex")
-    cp = binding.get("codex_provider")
-    if model and fmt in _WRITABLE_FORMATS:
-        w = write_model(entry, model)
-        notes.append("模型已同步为 %s" % w["model"] if w.get("ok")
-                     else "模型同步失败：%s" % w.get("error"))
-    elif model and not is_dsh and entry["id"] not in _AGENT_INJECTORS:
-        # 有专属注入通道的 CLI 由 _sync_agent_injection 负责落模型（含 jsonc），
-        # 不再报「只读配置」以免与注入成功的提示互相矛盾
-        notes.append("该工具模型为只读配置，按其现有配置打开（绑定模型 %s 未自动写入）" % model)
-    if is_dsh and model and prov.get("base_url"):
-        err = _sync_dsh_settings(entry, model, prov["base_url"])
-        notes.append("dsh 端点已同步为 %s" % prov["base_url"] if not err
-                     else "dsh 端点同步失败：%s" % err)
-    elif is_dsh and not model:
-        _fixed, note = _dsh_selfcheck_model(entry)
-        if note:
-            notes.append(note)
-    if is_codex and cp:
-        err = _sync_codex_settings(entry, model, cp)
-        notes.append("codex 供应商已同步为 %s" % cp.get("base_url", "") if not err
-                     else "codex 供应商同步失败：%s" % err)
-    inj = _sync_agent_injection(entry, binding)
-    if inj:
-        notes.append(inj)
-    return notes
-
-
-def launch(entry, open_browser=True):
-    """一键打开：web 类后台起服务并自动开浏览器；console 类新开终端窗口跑交互 TUI。
-
-    打开前把「CLI 绑定」页的模型落盘到该 CLI 配置文件（dsh 连端点与 models
-    目录一起同步），并注入绑定密钥 env——保证打开即选中可用模型。"""
-    if not detect_entry(entry).get("installed"):
-        return {"ok": False, "error": "未安装，无法打开"}
-    launch = entry.get("launch") or {}
-    cmd = (launch.get("command") or "").strip()
-    if not cmd:
-        return {"ok": False, "error": "未配置打开命令（可在 data/catalog.json 补 launch 字段）"}
-    is_dsh = entry["id"] in ("deepseek-harness", "dsh")
-    try:
-        from . import modelhub
-        binding = modelhub.resolve_binding(entry["id"]) or {}
-    except Exception:
-        binding = {}
-    env = os.environ.copy()
-    env.update(binding.get("env") or {})
-    notes = _sync_launch_model(entry, binding)
-    if is_dsh and not (binding.get("env") or {}) and not _dsh_key_present():
-        notes.append("未发现 dsh 密钥：打开后可能需在 dsh 内登录配置；"
-                     "要打开即用，请到「CLI 绑定」页给 DeepSeek Harness 绑定 openai 协议供应商")
-    name = entry.get("name", entry["id"])
-    kind = (launch.get("kind") or "console").lower()
-    extra = ("；".join(notes)) if notes else ""
-
-    if kind == "web":
-        try:
-            port = int(launch.get("port") or 0)
-        except (TypeError, ValueError):
-            port = 0
-        if port <= 0:
-            return {"ok": False, "error": "web 类打开必须配置固定端口（launch.port）"}
-        bare = "http://127.0.0.1:%d" % port
-        log_path = _launch_log_path(entry)
-        if not log_path:
-            return {"ok": False, "error": "启动日志路径不可信，已拒绝打开"}
-        if _port_open(port):
-            # 已在运行：新实例抢不到端口、新 token 也拿不到——从上次启动日志
-            # 恢复带 token 的信任 URL（dsh web 有 /?token=... 围栏，裸开是 401）。
-            # 模型同步照做：正在跑的实例不重启，改的是它下次生效的配置
-            url = _best_url(log_path, port) or bare
-            if open_browser:
-                webbrowser.open(url)
-            msg = "服务已在运行，已打开 " + url
-            if extra:
-                msg += "（" + extra + "）"
-            return {"ok": True, "kind": "web", "url": url, "message": msg}
-        # 后台起服务：无窗口、不阻塞请求；子进程输出重定向到启动日志（cmd 层
-        # 重定向，路径含空格才加引号）——web 类普遍会在启动行打印带 token 的
-        # 信任 URL（如 dsh web），就绪线程从日志提取后再开浏览器
-        ls = str(log_path)
-        spawn_cmd = "%s > %s 2>&1" % (cmd, ('"%s"' % ls) if " " in ls else ls)
-        try:
-            log_path.parent.mkdir(parents=True, exist_ok=True)
-            subprocess.Popen(["cmd", "/c", spawn_cmd], cwd=str(paths.ROOT), env=env,
-                             stdin=subprocess.DEVNULL, stdout=subprocess.DEVNULL,
-                             stderr=subprocess.DEVNULL, creationflags=CREATE_NO_WINDOW)
-        except Exception as e:
-            return {"ok": False, "error": "无法启动服务: %r" % e}
-        if open_browser:
-            threading.Thread(target=_open_when_ready, args=(port, bare, log_path),
-                             name="launch-wait-%d" % port, daemon=True).start()
-        return {"ok": True, "kind": "web", "url": bare,
-                "message": "%s 正在启动，就绪后浏览器会自动打开（%s）%s"
-                           % (name, bare, ("；" + extra) if extra else "")}
-
-    if sys.platform != "win32":
-        return {"ok": False, "error": "终端窗口拉起暂仅支持 Windows"}
-    # start 为目标命令新开一个可见终端窗口；cmd /k 让 CLI 退出后窗口保留，
-    # 报错不至于一闪而过。外层 cmd 用 CREATE_NO_WINDOW 隐藏。
-    argv = ["cmd", "/c", "start", "CodeBee %s" % name, "/D", str(paths.ROOT),
-            "cmd", "/k", cmd]
-    try:
-        subprocess.Popen(argv, cwd=str(paths.ROOT), env=env,
-                         creationflags=CREATE_NO_WINDOW)
-    except Exception as e:
-        return {"ok": False, "error": "无法打开终端窗口: %r" % e}
-    return {"ok": True, "kind": "console", "message": "已在新的终端窗口打开 %s%s"
-            % (name, ("（" + extra + "）") if extra else "")}
-
-
-def _open_when_ready(port, url, log_path, timeout=30):
-    """等 web 服务端口就绪后开浏览器。优先从启动日志提取带 token 的信任 URL
-    （端口通了 token 行可能还差几十毫秒才落盘，故就绪后最多再等 6 秒）；
-    超时兜底也开——服务可能只是慢，用户手动刷新即可。"""
-    deadline = time.time() + timeout
-    while time.time() < deadline:
-        if _port_open(port):
-            break
-        time.sleep(0.5)
-    token_deadline = time.time() + 6
-    best = None
-    while time.time() < token_deadline:
-        best = _best_url(log_path, port)
-        if best and "token=" in best:
-            break
-        time.sleep(0.5)
-    try:
-        webbrowser.open(best or url)
-    except Exception:
-        pass
-
-
-# ---------------------------------------------------------------- 安装/升级
-
-def run_mgmt_command(entry, op, cancel_event=None, log_path=None):
-    """执行 install/upgrade/uninstall 命令（在任务队列里跑，日志实时落盘）。"""
-    if op == "uninstall":
-        cmd = catalog.uninstall_command(entry)
-        if not cmd:
-            return {"ok": False,
-                    "error": "无法推导卸载命令：请在 data/catalog.json 的 \"%s\" 里配置 uninstall 字段"
-                             % entry["id"]}
-    else:
-        cmd = entry.get(op)
-        if not cmd:
-            return {"ok": False,
-                    "error": "未配置 %s 命令：请在 data/catalog.json 的 \"%s\" 里补充，或用官方渠道安装"
-                             % (op, entry["id"])}
-    res = runner.run_process(shell_cmd=cmd, cwd=str(paths.ROOT),
-                             timeout=1800, cancel_event=cancel_event, log_path=log_path)
-    detect_all(force=True)
-    with _LOCK:
-        _STATE["versions"].pop(entry["id"], None)
-    return {"ok": res["ok"], "exit_code": res["exit_code"], "command": cmd,
-            "error": "" if res["ok"] else (res["stderr"][-800:] or "退出码 %s" % res["exit_code"])}
-
-
-# ---------------------------------------------------------------- 版本检查
-
-_UPDATE_CACHE = {}   # agent_id → (ts, {current, latest, updatable, note})
-UPDATE_TTL = 600
-
-
-def _npm_pkg_name(cmd):
-    """从 npm 安装命令里取包名（实现已统一到 catalog.npm_pkg_name）。"""
-    return catalog.npm_pkg_name(cmd)
-
-
-def _ver_tuple(s):
-    return [int(x) for x in re.findall(r"\d+", str(s or ""))[:4]]
-
-
-def check_update(entry, force=False):
-    """检查是否有新版本可用。npm 走 npm view；winget 走 winget upgrade 列表；
-    其他渠道标记为不支持。结果缓存 10 分钟。"""
-    eid = entry["id"]
-    if not force:
-        cached = _UPDATE_CACHE.get(eid)
-        if cached and time.time() - cached[0] < UPDATE_TTL:
-            return cached[1]
-    current = version_of(entry)
-    cur_num = ".".join(str(x) for x in _ver_tuple(current)) or "-"
-    result = {"current": current, "latest": None, "updatable": None, "note": ""}
-
-    cmd = entry.get("install") or entry.get("upgrade") or ""
-    pkg = _npm_pkg_name(cmd)
-    if pkg:
-        r = runner.run_process(argv=["cmd", "/c", "npm", "view", pkg, "version"], timeout=90)
-        latest = ""
-        if r["ok"]:
-            for line in (r["stdout"] or "").splitlines():
-                line = line.strip()
-                if line and re.match(r"^\d", line):
-                    latest = line
-        if not latest:
-            result["note"] = "查询 npm 失败（网络或 registry 问题）：" + (r["stderr"] or "")[-200:]
-        else:
-            result["latest"] = latest
-            result["updatable"] = bool(_ver_tuple(latest) > _ver_tuple(cur_num))
-            if not result["updatable"]:
-                result["note"] = "已是最新版本"
-    elif "winget" in cmd:
-        m = re.search(r"--id\s+([A-Za-z0-9._-]+)", cmd)
-        wid = m.group(1) if m else None
-        if not wid:
-            result["note"] = "无法从命令中解析 winget 包 ID"
-        else:
-            r = runner.run_process(argv=["cmd", "/c", "winget", "upgrade"], timeout=180)
-            if not r["ok"]:
-                result["note"] = "winget upgrade 查询失败：" + (r["stderr"] or "")[-200:]
-            else:
-                hit = [l for l in (r["stdout"] or "").splitlines() if wid.lower() in l.lower()]
-                result["updatable"] = bool(hit)
-                if hit:
-                    result["latest"] = "（winget 有可用更新）"
-                else:
-                    result["latest"] = cur_num
-                    result["note"] = "已是最新版本"
-    else:
-        result["note"] = "该渠道暂不支持自动检查更新，可直接点升级尝试"
-
-    with _LOCK:
-        _UPDATE_CACHE[eid] = (time.time(), result)
-    return result
-
-
-# 「进入智能体目录页自动检查更新」的后台任务状态
-_UPDATE_CHECK = {"running": False, "total": 0, "done": 0}
-
-
-def updates_checking():
-    with _LOCK:
-        return bool(_UPDATE_CHECK["running"])
-
-
-def update_info(entry):
-    """单个 CLI 的更新检查结果（供管理页卡片展示）。没查过时为 unknown。"""
-    with _LOCK:
-        cached = _UPDATE_CACHE.get(entry["id"])
-    if not cached:
-        return {"status": "unknown", "latest": None, "updatable": None,
-                "note": "", "checked_at": ""}
-    res = cached[1] or {}
-    up = res.get("updatable")
-    status = "updatable" if up is True else ("current" if up is False else "unsupported")
-    return {"status": status, "latest": res.get("latest"), "updatable": up,
-            "note": res.get("note") or "",
-            "checked_at": time.strftime("%H:%M", time.localtime(cached[0]))}
-
-
-def _checkable_entries():
-    """已安装且配了 install/upgrade 的条目——只有这些才谈得上「是否有新版本」。"""
-    detected = detect_all()
-    out = []
-    for e in catalog.load():
-        det = (detected or {}).get(e["id"]) or {}
-        if det.get("installed") and (e.get("upgrade") or e.get("install")):
-            out.append(e)
-    return out
-
-
-def check_updates_async(force=False):
-    """后台逐个检查已安装 CLI 的远端最新版本，返回本次要检查的条目数。
-
-    进入「智能体目录」页时自动触发。单条结果复用 check_update 的 10 分钟缓存，
-    所以反复进出页面几乎不产生额外子进程/网络开销；已在跑时不重复起线程。
-    """
-    with _LOCK:
-        if _UPDATE_CHECK["running"]:
-            return 0
-        entries = _checkable_entries()
-        _UPDATE_CHECK.update(running=True, total=len(entries), done=0)
-
-    def _worker():
-        try:
-            for e in entries:
-                try:
-                    check_update(e, force=force)
-                except Exception:
-                    pass
-                finally:
-                    with _LOCK:
-                        _UPDATE_CHECK["done"] += 1
-        finally:
-            with _LOCK:
-                _UPDATE_CHECK["running"] = False
-
-    threading.Thread(target=_worker, name="catalog-update-check", daemon=True).start()
-    return len(entries)
-
-
-def catalog_view():
-    """管理页数据：catalog + 检测 + 版本 + 模型 + 编排启用状态。"""
-    from . import registry
-    entries = catalog.load()
-    detected = detect_all(force=False)
-    enabled = registry.load_enabled()
-    view = []
-    for e in entries:
-        det = detected.get(e["id"]) or {}
-        pref = enabled.get(e["id"]) or {}
-        orch_enabled = bool(pref.get("enabled", e.get("default_enabled", False))) if e.get("orch") else False
-        view.append({
-            "id": e["id"], "name": e.get("name", e["id"]), "note": e.get("note", ""),
-            "group": "installed" if det.get("installed") else "installable",
-            "installed": det.get("installed", False),
-            "detail": det.get("detail", ""),
-            "version": version_of(e),
-            "config_path": _config_path(e),
-            "config_writable": (e.get("config") or {}).get("format") in _WRITABLE_FORMATS,
-            "model": read_model(e),
-            "orch_kind": (e.get("orch") or {}).get("kind"),
-            "orch_enabled": orch_enabled,
-            "update": update_info(e),
-            "has_install": bool(e.get("install")),
-            "has_upgrade": bool(e.get("upgrade")),
-            # 卸载命令由 install/upgrade 推导（或 catalog 显式配置），供 UI 确认框展示
-            "uninstall_cmd": catalog.uninstall_command(e) if det.get("installed") else None,
-            # 一键打开配置（kind=web/console + command）；没配的条目 UI 不出「打开」按钮
-            "launch": e.get("launch"),
-        })
-    return view
+# -*- coding: utf-8 -*-
+"""智能体管理器：安装检测、版本、安装/升级、模型配置读写。
+
+安全约束：catalog 中的配置文件路径展开后必须落在用户主目录内，
+防止相对路径穿越到预期之外的系统位置。
+"""
+from __future__ import annotations
+
+import json
+import os
+import re
+import shutil
+import socket
+import subprocess
+import sys
+import threading
+import time
+import webbrowser
+import zlib
+from pathlib import Path
+
+from . import catalog, paths, runner
+
+CREATE_NO_WINDOW = 0x08000000
+VERSION_TTL = 300  # 版本缓存 5 分钟
+
+_LOCK = threading.RLock()
+_STATE = {"detected": {}, "versions": {}, "detect_ts": 0.0, "detect_ev": None}
+
+# 能自动写入默认模型的 config.format（其余格式只能手动编辑）
+_WRITABLE_FORMATS = ("toml-line", "toml-section", "json", "json-path", "jsonc",
+                     "yaml-line")
+
+
+def _expand(p):
+    return os.path.abspath(os.path.expanduser(os.path.expandvars(p)))
+
+
+def _safe_config_path(raw):
+    """展开并校验配置路径：必须在用户主目录内（防穿越）。"""
+    if not raw:
+        return None
+    full = _expand(raw)
+    home = os.path.abspath(os.path.expanduser("~"))
+    try:
+        if os.path.commonpath([full, home]) != home:
+            return None
+    except ValueError:
+        return None
+    return full
+
+
+# ---------------------------------------------------------------- 检测
+
+def detect_entry(entry):
+    d = entry.get("detect") or {}
+    if d.get("cli"):
+        path = shutil.which(d["cli"])
+        return {"installed": bool(path), "detail": path or ""}
+    if d.get("exe"):
+        full = _expand(d["exe"])
+        return {"installed": os.path.isfile(full), "detail": full if os.path.isfile(full) else ""}
+    if d.get("uwp"):
+        base = os.path.join(os.environ.get("LOCALAPPDATA", ""), "Packages", d["uwp"])
+        return {"installed": os.path.isdir(base), "detail": base if os.path.isdir(base) else ""}
+    if d.get("dir"):
+        full = _expand(d["dir"])
+        return {"installed": os.path.isdir(full), "detail": full if os.path.isdir(full) else ""}
+    return {"installed": False, "detail": ""}
+
+
+def detect_all(force=False):
+    """检测全部条目。检测（慢磁盘 IO）在锁外跑：shutil.which/isfile 在
+    Windows 上遇到断链的 PATH 项可能卡数秒，持锁会把所有并发请求堵死
+    （曾导致 SSE 多连接时服务假死）。等待方有界等待 30s 后拿旧结果。
+    """
+    with _LOCK:
+        if not force and _STATE["detected"] and time.time() - _STATE["detect_ts"] < 60:
+            return _STATE["detected"]
+        ev = _STATE["detect_ev"]
+        lead = ev is None  # 我是本次检测的执行者
+        if lead:
+            ev = _STATE["detect_ev"] = threading.Event()
+    if not lead:
+        ev.wait(30)  # 检测完成或超时；两种情况都拿当前最新快照
+        with _LOCK:
+            return dict(_STATE["detected"])
+    try:
+        detected = {}
+        for entry in catalog.load():
+            try:
+                detected[entry["id"]] = detect_entry(entry)
+            except Exception as e:
+                detected[entry["id"]] = {"installed": False, "detail": "检测出错: %r" % e}
+        if detected:
+            with _LOCK:
+                _STATE["detected"] = detected
+    finally:
+        with _LOCK:
+            _STATE["detect_ts"] = time.time()
+            _STATE["detect_ev"] = None
+        ev.set()
+    return detected
+
+
+def _uwp_version(package_dir):
+    import xml.etree.ElementTree as ET
+    mf = os.path.join(package_dir, "AppxManifest.xml")
+    if not os.path.isfile(mf):
+        return None
+    try:
+        tree = ET.parse(mf)
+        for el in tree.iter():
+            if el.tag.endswith("}Identity") or el.tag == "Identity":
+                return el.get("Version")
+    except Exception:
+        pass
+    return None
+
+
+def _exe_version(path):
+    try:
+        r = subprocess.run(
+            ["powershell", "-NoProfile", "-Command",
+             "(Get-Item -LiteralPath '%s').VersionInfo.ProductVersion" % path.replace("'", "''")],
+            capture_output=True, creationflags=CREATE_NO_WINDOW, timeout=25)
+        out = r.stdout.decode("utf-8", "replace").strip()
+        return out or None
+    except Exception:
+        return None
+
+
+def version_of(entry):
+    """版本探测：CLI 走 --version；UWP 读 AppxManifest；exe 走 PowerShell（惰性缓存）。"""
+    eid = entry["id"]
+    with _LOCK:
+        cached = _STATE["versions"].get(eid)
+        if cached and time.time() - cached[0] < VERSION_TTL:
+            return cached[1]
+    det = (detect_all() or {}).get(eid) or {}
+    version = None
+    cli = (entry.get("detect") or {}).get("cli")
+    if cli and det.get("installed"):
+        try:
+            r = subprocess.run(["cmd", "/c", cli, "--version"], capture_output=True,
+                               creationflags=CREATE_NO_WINDOW, timeout=20)
+            out = (r.stdout or b"").decode("utf-8", "replace").strip()
+            if not out:
+                out = (r.stderr or b"").decode("utf-8", "replace").strip()
+            version = out.splitlines()[0][:60] if out else None
+        except Exception:
+            version = None
+    elif det.get("detail") and os.path.isdir(det["detail"]):
+        version = _uwp_version(det["detail"])
+    elif det.get("detail") and os.path.isfile(det["detail"]):
+        version = _exe_version(det["detail"])
+    with _LOCK:
+        _STATE["versions"][eid] = (time.time(), version or "-")
+    return version or "-"
+
+
+# ---------------------------------------------------------------- TOML 表内键
+
+def _toml_span(lines, table):
+    """定位顶层表 `[table]` 的行区间 [start, end)；start 为表头行。
+    只认顶层表头（行首无空白），不误吞嵌套 `[[array]]` 之外的子表——
+    子表在 TOML 里也是 `[a.b]` 顶层写法，同样按表头截断。"""
+    header = re.compile(r"^\[([^\[\]]+)\]\s*$")
+    start = None
+    for i, ln in enumerate(lines):
+        if ln.lstrip().startswith("["):
+            m = header.match(ln.strip())
+            if not m:
+                continue
+            if start is not None:
+                return start, i
+            if m.group(1).strip() == table:
+                start = i
+    if start is None:
+        return None, None
+    return start, len(lines)
+
+
+def _toml_read_value(text, table, leaf):
+    m = re.search(r'(?m)^\s*%s\s*=\s*"([^"]*)"\s*(#.*)?$' % re.escape(leaf), text) \
+        if table is None else None
+    if table is None:
+        return m.group(1) if m else None
+    start, end = _toml_span(text.splitlines(), table)
+    if start is None:
+        return None
+    for ln in text.splitlines()[start + 1:end]:
+        m = re.match(r'^\s*%s\s*=\s*"([^"]*)"\s*(#.*)?$' % re.escape(leaf), ln)
+        if m:
+            return m.group(1)
+    return None
+
+
+def _toml_write_value(text, table, leaf, value):
+    """就地写入 TOML 的 [table] leaf（双引号标量），保留其余内容与换行风格。"""
+    eol = "\r\n" if "\r\n" in text else "\n"
+    lines = text.splitlines()
+    new_line = '%s = "%s"' % (leaf, value.replace("\\", "\\\\").replace('"', '\\"'))
+    # 换值不换行：保留行尾注释等其余内容
+    pat = re.compile(r'^(\s*%s\s*=\s*)"(?:[^"\\]|\\.)*"(.*)$' % re.escape(leaf))
+    if table is None:
+        for i, ln in enumerate(lines):
+            m = pat.match(ln)
+            if m:
+                lines[i] = "%s\"%s\"%s" % (m.group(1), value.replace("\\", "\\\\").replace('"', '\\"'), m.group(2))
+                break
+        else:
+            lines.append(new_line)
+        return eol.join(lines).rstrip("\r\n") + eol
+    start, end = _toml_span(lines, table)
+    if start is None:  # 表不存在：整段追加
+        if lines and lines[-1].strip():
+            lines.append("")
+        lines.append("[%s]" % table)
+        lines.append(new_line)
+        return eol.join(lines).rstrip("\r\n") + eol
+    for i in range(start + 1, end):
+        m = pat.match(lines[i])
+        if m:
+            lines[i] = "%s\"%s\"%s" % (m.group(1), value.replace("\\", "\\\\").replace('"', '\\"'), m.group(2))
+            return eol.join(lines).rstrip("\r\n") + eol
+    lines.insert(end, new_line)  # 表内末尾追加（表头区间终点即下一表头前）
+    return eol.join(lines).rstrip("\r\n") + eol
+
+
+# ---------------------------------------------------------------- 模型配置
+
+def _config_path(entry):
+    cfg = entry.get("config") or {}
+    return _safe_config_path(cfg.get("path"))
+
+
+def _yaml_model_path(cfg):
+    """把 config.model_key 的点号路径拆成 (段, 键)；无点号时段为 None（顶层键）。"""
+    return _dotted_key(cfg)
+
+
+def _dotted_key(cfg):
+    """model_key 点号路径拆 (表/段, 键)；无点号时第一元为 None（顶层键）。"""
+    key = (cfg.get("model_key") or "model").strip()
+    if "." in key:
+        section, leaf = key.split(".", 1)
+        return section.strip(), leaf.strip()
+    return None, key
+
+
+def _yaml_quote(value):
+    """YAML 双引号标量。必须加引号：模型名可能以 [ 开头（YAML 流序列）或含 #。"""
+    return '"%s"' % value.replace("\\", "\\\\").replace('"', '\\"')
+
+
+def _yaml_unquote(raw):
+    """取 YAML 标量的值：剥引号、丢行尾注释。引号内的 # 不算注释。"""
+    s = (raw or "").strip()
+    if not s:
+        return ""
+    if s[0] in ("'", '"'):
+        q = s[0]
+        i = 1
+        buf = []
+        while i < len(s):
+            ch = s[i]
+            if q == '"' and ch == "\\" and i + 1 < len(s):
+                nxt = s[i + 1]
+                buf.append('"' if nxt == '"' else ("\\" if nxt == "\\" else nxt))
+                i += 2
+                continue
+            if ch == q:
+                if q == "'" and i + 1 < len(s) and s[i + 1] == "'":  # '' 转义
+                    buf.append("'")
+                    i += 2
+                    continue
+                break
+            buf.append(ch)
+            i += 1
+        return "".join(buf).strip()
+    # 无引号：截断行尾注释（# 前的空白才算注释起始）
+    return re.split(r"\s+#", s, 1)[0].strip()
+
+
+def _yaml_span(lines, section):
+    """定位顶层段的行区间 [start, end)；start 为段名行，其子键在 start+1..end。"""
+    start = None
+    for i, ln in enumerate(lines):
+        m = re.match(r"^([^\s#][^:]*):\s*(.*)$", ln)
+        if not m:
+            continue
+        if m.group(1).strip() == section:
+            start = i
+            continue
+        if start is not None:
+            return start, i
+    if start is None:
+        return None, None
+    return start, len(lines)
+
+
+def _yaml_read_value(text, section, leaf):
+    lines = text.splitlines()
+    if section is None:
+        m = re.search(r"(?m)^%s\s*:\s*(.+?)\s*$" % re.escape(leaf), text)
+        return _yaml_unquote(m.group(1)) or None if m else None
+    start, end = _yaml_span(lines, section)
+    if start is None:
+        return None
+    for ln in lines[start + 1:end]:
+        m = re.match(r"^\s+%s\s*:\s*(.+?)\s*$" % re.escape(leaf), ln)
+        if m:
+            return _yaml_unquote(m.group(1)) or None
+    return None
+
+
+def _yaml_write_value(text, section, leaf, value):
+    """就地写入 YAML 的 section.leaf，保留其余内容、缩进与换行风格。"""
+    eol = "\r\n" if "\r\n" in text else "\n"
+    lines = text.splitlines()
+    quoted = _yaml_quote(value)
+    if section is None:
+        for i, ln in enumerate(lines):
+            if re.match(r"^%s\s*:" % re.escape(leaf), ln):
+                lines[i] = "%s: %s" % (leaf, quoted)
+                break
+        else:
+            lines.append("%s: %s" % (leaf, quoted))
+        return eol.join(lines).rstrip("\r\n") + eol
+    start, end = _yaml_span(lines, section)
+    if start is None:  # 段不存在：整段追加
+        if lines and lines[-1].strip():
+            lines.append("")
+        lines.append("%s:" % section)
+        lines.append("  %s: %s" % (leaf, quoted))
+        return eol.join(lines).rstrip("\r\n") + eol
+    for i in range(start + 1, end):
+        m = re.match(r"^(\s+)%s\s*:" % re.escape(leaf), lines[i])
+        if m:  # 键已存在：只换值，保留原缩进
+            lines[i] = "%s%s: %s" % (m.group(1), leaf, quoted)
+            return eol.join(lines).rstrip("\r\n") + eol
+    indent, last = "  ", start  # 段存在但无该键：跟随段内缩进、追加到段尾
+    for i in range(start + 1, end):
+        if not lines[i].strip():
+            continue
+        if indent == "  ":
+            m = re.match(r"^(\s+)\S", lines[i])
+            if m:
+                indent = m.group(1)
+        last = i
+    lines.insert(last + 1, "%s%s: %s" % (indent, leaf, quoted))
+    return eol.join(lines).rstrip("\r\n") + eol
+
+
+def _json_path_get(data, keys):
+    """沿点号路径下钻 JSON 嵌套；终点必须是字符串。"""
+    cur = data
+    for k in keys:
+        if isinstance(cur, dict) and k in cur:
+            cur = cur[k]
+        else:
+            return None
+    return cur if isinstance(cur, str) else None
+
+
+def _json_path_set(data, keys, value):
+    """沿点号路径写入 JSON 嵌套，缺中间对象就地创建；
+    中途遇到非 dict（如 string 简写形式）升级为对象。"""
+    cur = data
+    for k in keys[:-1]:
+        if not isinstance(cur.get(k), dict):
+            cur[k] = {}
+        cur = cur[k]
+    cur[keys[-1]] = value
+
+
+def read_model(entry):
+    path = _config_path(entry)
+    cfg = entry.get("config") or {}
+    if not path or not os.path.isfile(path) or not cfg.get("format"):
+        return None
+    try:
+        text = open(path, encoding="utf-8", errors="replace").read()
+    except Exception:
+        return None
+    if cfg["format"] == "toml-line":
+        m = re.search(r'(?m)^\s*model\s*=\s*"([^"]+)"', text)
+        return m.group(1) if m else None
+    if cfg["format"] == "toml-section":
+        table, leaf = _dotted_key(cfg)
+        return _toml_read_value(text, table, leaf)
+    if cfg["format"] == "json":
+        try:
+            v = json.loads(text).get("model")
+            if v:
+                return v
+        except Exception:
+            pass
+        m = re.search(r'"model"\s*:\s*"([^"]+)"', text)
+        return m.group(1) if m else None
+    if cfg["format"] == "jsonc":
+        keys = (cfg.get("model_key") or "model").split(".")
+        try:
+            data = json.loads(_jsonc_strip_comments(text) or "{}")
+        except Exception:
+            return None
+        return _json_path_get(data, keys)
+    if cfg["format"] == "json-path":
+        try:
+            data = json.loads(text or "{}")
+        except Exception:
+            return None
+        return _json_path_get(data, (cfg.get("model_key") or "model").split("."))
+    if cfg["format"] == "yaml-line":
+        return _yaml_read_value(text, *_yaml_model_path(cfg))
+    return None
+
+
+def write_model(entry, model):
+    """写入默认模型（改动前自动备份 .bak）。支持 toml-line / toml-section /
+    json / json-path / jsonc / yaml-line。"""
+    path = _config_path(entry)
+    cfg = entry.get("config") or {}
+    fmt = cfg.get("format")
+    if not path:
+        return {"ok": False,
+                "error": "配置路径无效或不在用户主目录内，已拒绝写入"}
+    if fmt not in _WRITABLE_FORMATS:
+        return {"ok": False, "error": "该工具的模型配置格式暂不支持自动写入，请手动编辑 %s" % path}
+    model = (model or "").strip()
+    if not model:
+        return {"ok": False, "error": "模型名不能为空"}
+    # 写入前二次校验：解析真实路径后必须仍在用户主目录内（防穿越/符号链接），
+    # 校验通过后统一改用解析路径写入
+    home = os.path.abspath(os.path.expanduser("~"))
+    path = os.path.realpath(path)
+    try:
+        if os.path.commonpath([path, home]) != home:
+            return {"ok": False, "error": "配置路径越出用户主目录，已拒绝写入"}
+    except ValueError:
+        return {"ok": False, "error": "配置路径越出用户主目录，已拒绝写入"}
+    try:
+        if os.path.isfile(path):
+            shutil.copyfile(path, path + ".bak")
+        else:
+            # 各 CLI 首次运行都未必建主配置（grok 不建 config.toml、pi 不建
+            # settings.json、openclaw 缺失即安全默认——官方文档明确「缺失即
+            # 内置默认」）；它正是用户层覆盖的落点，缺了按需创建
+            Path(path).parent.mkdir(parents=True, exist_ok=True)
+            Path(path).write_bytes(b"{}" if fmt in ("json", "json-path", "jsonc") else b"")
+        if fmt == "jsonc":
+            # mimo（mimocode.jsonc）有注释，整体重解析会丢注释——复用
+            # _jsonc_set 做就地片段改写（只动目标键，其余原样保留）
+            keys = (cfg.get("model_key") or "model").split(".")
+            text = open(path, encoding="utf-8", errors="replace", newline="").read()
+            new_text, ok = _jsonc_set(text, tuple(keys),
+                                      json.dumps(model, ensure_ascii=False))
+            if not ok:
+                return {"ok": False,
+                        "error": "jsonc 结构异常，未能就地写入 %s（已避免覆盖）" % path}
+            Path(path).write_bytes(new_text.encode("utf-8"))
+        elif fmt == "toml-line":
+            text = open(path, encoding="utf-8", errors="replace").read()
+            new_line = 'model = "%s"' % model
+            if re.search(r'(?m)^\s*model\s*=\s*"[^"]*"', text):
+                text = re.sub(r'(?m)^\s*model\s*=\s*"[^"]*"', new_line, text)
+            else:
+                text = text.rstrip("\n") + "\n" + new_line + "\n"
+            Path(path).write_bytes(text.encode("utf-8"))
+        elif fmt == "toml-section":
+            text = open(path, encoding="utf-8", errors="replace", newline="").read()
+            text = _toml_write_value(text, *_dotted_key(cfg), value=model)
+            Path(path).write_bytes(text.encode("utf-8"))
+        elif fmt == "yaml-line":
+            # newline="" 关掉通用换行转换：文本层面看不出 \r\n 就会被静默改写成 LF，
+            # 用户的 Windows 配置不该因为写个模型名而整篇换行符被替换
+            text = open(path, encoding="utf-8", errors="replace", newline="").read()
+            text = _yaml_write_value(text, *_yaml_model_path(cfg), value=model)
+            Path(path).write_bytes(text.encode("utf-8"))
+        elif fmt == "json-path":
+            # openclaw（agents.defaults.model.primary）：默认模型藏在嵌套对象里，
+            # 且 openclaw 对未知顶层键直接拒绝启动——绝不能写顶层 "model"
+            text = open(path, encoding="utf-8", errors="replace", newline="").read()
+            try:
+                data = json.loads(text) if text.strip() else {}
+            except Exception:
+                return {"ok": False, "error": "配置文件不是合法 JSON，已中止（避免覆盖）"}
+            _json_path_set(data, (cfg.get("model_key") or "model").split("."), model)
+            # pi 的 settings.json：defaultModel 必须配 defaultProvider 才能解析出
+            # (provider, model) 二元组；catalog 里声明了的伴随键一并落盘。
+            # setdefault 语义：用户已设的值（如 defaultProvider: anthropic）不被空串覆盖
+            for k, v in (cfg.get("model_extra_keys") or {}).items():
+                cur = data
+                ks = k.split(".")
+                for kk in ks[:-1]:
+                    if not isinstance(cur.get(kk), dict):
+                        cur[kk] = {}
+                    cur = cur[kk]
+                if cur.get(ks[-1]) in (None, ""):
+                    cur[ks[-1]] = v
+            Path(path).write_bytes(
+                json.dumps(data, ensure_ascii=False, indent=2).encode("utf-8"))
+        else:
+            try:
+                data = json.loads(open(path, encoding="utf-8", errors="replace").read())
+            except FileNotFoundError:
+                return {"ok": False,
+                        "error": "配置文件尚未生成（%s 首次运行后才有），暂无法写入" % path}
+            except Exception:
+                return {"ok": False, "error": "配置文件不是合法 JSON，已中止（避免覆盖）"}
+            data["model"] = model
+            Path(path).write_bytes(
+                json.dumps(data, ensure_ascii=False, indent=2).encode("utf-8"))
+        return {"ok": True, "model": read_model(entry)}
+    except Exception as e:
+        return {"ok": False, "error": repr(e)}
+
+
+# ---------------------------------------------------------------- 一键打开
+
+def _port_open(port, timeout=0.5):
+    try:
+        with socket.create_connection(("127.0.0.1", port), timeout=timeout):
+            return True
+    except OSError:
+        return False
+
+
+def launch_env(entry):
+    """打开交互/网页版时注入的子进程环境变量：与编排同源的绑定凭据
+    （dsh=DEEPSEEK_*，claude=ANTHROPIC_*，codex=ORCH_API_KEY）。交互进程
+    脱离了编排链路，没有这层注入就拿不到 API key。无绑定时返回 {}。"""
+    try:
+        from . import modelhub  # 惰性导入：modelhub 体量大且避免潜在环
+        b = modelhub.resolve_binding(entry["id"]) or {}
+    except Exception:
+        return {}
+    return dict(b.get("env") or {})
+
+
+def _launch_log_path(entry):
+    """web 类启动日志的落点：id 白名单化后仅作文件名成分，最终路径必须仍围栏
+    在数据目录内（catalog.json 用户可编辑，id 不可信；非白名单 id 用 crc32
+    稳定代称——跨进程重启不变，「已在运行」回读上次日志才找得到）。"""
+    raw_id = str(entry.get("id") or "")
+    if re.fullmatch(r"[A-Za-z0-9_-]{1,64}", raw_id):
+        stem = raw_id
+    else:
+        stem = "agent-%d" % (zlib.crc32(raw_id.encode("utf-8")) & 0xFFFFFFFF)
+    data_root = os.path.abspath(str(paths.DATA_DIR))
+    p = Path(data_root, "launch", stem + ".log")
+    try:
+        if os.path.commonpath([os.path.abspath(str(p)), data_root]) != data_root:
+            return None
+    except ValueError:
+        return None
+    return p
+
+
+def _best_url(log_path, port):
+    """从启动日志提取该端口的信任 URL（含 token 优先）。无日志/未匹配返回 None。"""
+    try:
+        text = open(str(log_path), encoding="utf-8", errors="replace").read()
+    except Exception:
+        return None
+    urls = re.findall(r"https?://[^\s\"'<>]+", text)
+    same = [u for u in urls if ":%d" % port in u]
+    if not same:
+        return None
+    tokened = [u for u in same if "token=" in u]
+    return (tokened or same)[0]
+
+
+def _yaml_model_ids(text, section):
+    """收集 section.models 序列里的全部模型 id（保持顺序）。段/键缺失返回 []。"""
+    lines = text.splitlines()
+    start, end = _yaml_span(lines, section)
+    if start is None:
+        return []
+    m_indent = None
+    for i in range(start + 1, end):
+        m = re.match(r"^(\s+)models\s*:\s*(?:#.*)?$", lines[i])
+        if m:
+            m_indent = len(m.group(1))
+            start = i
+            break
+    if m_indent is None:
+        return []
+    ids = []
+    for ln in lines[start + 1:end]:
+        if not ln.strip():
+            continue
+        if len(ln) - len(ln.lstrip(" ")) <= m_indent:
+            break  # models 列表结束（遇到同级或更浅缩进的键）
+        m = re.match(r"^\s*-\s+id\s*:\s*(.+?)\s*$", ln)
+        if m:
+            ids.append(_yaml_unquote(m.group(1)))
+    return ids
+
+
+def _yaml_ensure_model_entry(text, section, model, context_window=1000000):
+    """确保 section.models 序列里有 id==model 的条目，缺则按现有条目形状追加
+    （id/name/contextWindow——dsh 的 catalog 校验要求 id 与 name 非空）。
+    已存在或段/列表结构不完整时原文返回。返回 (new_text, added)。"""
+    eol = "\r\n" if "\r\n" in text else "\n"
+    lines = text.splitlines()
+    start, end = _yaml_span(lines, section)
+    if start is None:
+        return text, False
+    m_idx = m_indent = None
+    for i in range(start + 1, end):
+        m = re.match(r"^(\s+)models\s*:\s*(?:#.*)?$", lines[i])
+        if m:
+            m_idx, m_indent = i, len(m.group(1))
+            break
+    if m_idx is None:
+        return text, False  # 段内没有 models 键：不凭空造结构（保守）
+    item_indent = None
+    last_item = m_idx
+    i = m_idx + 1
+    while i < end:
+        ln = lines[i]
+        if not ln.strip():
+            i += 1
+            continue
+        if len(ln) - len(ln.lstrip(" ")) <= m_indent:
+            break  # models 列表结束
+        m = re.match(r"^(\s*)-\s+id\s*:\s*(.+?)\s*$", ln)
+        if m:
+            item_indent = m.group(1)
+            last_item = i
+            if _yaml_unquote(m.group(2)) == model:
+                return text, False
+        elif re.match(r"^\s+\S", ln):
+            last_item = i  # 条目的续属性行（name/contextWindow…）
+        i += 1
+    item_indent = item_indent or (" " * (m_indent + 2))
+    block = ["%s- id: %s" % (item_indent, _yaml_quote(model)),
+             "%s  name: %s" % (item_indent, _yaml_quote(model)),
+             "%s  contextWindow: %d" % (item_indent, context_window)]
+    lines[last_item + 1:last_item + 1] = block
+    return eol.join(lines).rstrip("\r\n") + eol, True
+
+
+def _sync_dsh_settings(entry, model, base_url):
+    """dsh 专属：把绑定模型的端点与模型目录写进 ~/.dsh/settings.yaml 的
+    llm-deepseek 段（agent-default-model.model 由 write_model 负责）。
+    端点不同步不行——settings 优先级高于 env，密钥会发给旧端点；
+    models 列表不同步不行——dsh web 的模型下拉只列它，缺条目就选不中。
+    返回错误串或 None。"""
+    path = _config_path(entry)
+    if not path:
+        return "dsh 配置路径无效"
+    path = os.path.realpath(path)
+    home = os.path.abspath(os.path.expanduser("~"))
+    try:
+        if os.path.commonpath([path, home]) != home:
+            return "dsh 配置路径越出用户主目录，已拒绝"
+    except ValueError:
+        return "dsh 配置路径越出用户主目录，已拒绝"
+    try:
+        if os.path.isfile(path):
+            text = open(path, encoding="utf-8", errors="replace", newline="").read()
+        else:
+            Path(path).parent.mkdir(parents=True, exist_ok=True)
+            text = ""
+        text2 = _yaml_write_value(text, "llm-deepseek", "baseURL", base_url)
+        text2, _added = _yaml_ensure_model_entry(text2, "llm-deepseek", model)
+        if text2 != text:
+            if os.path.isfile(path):
+                shutil.copyfile(path, path + ".bak")
+            Path(path).write_bytes(text2.encode("utf-8"))
+        return None
+    except Exception as e:
+        return repr(e)
+
+
+def _dsh_selfcheck_model(entry):
+    """dsh 无绑定时自检：agent-default-model.model 必须在端点 models 列表内，
+    否则 web UI 打开就是一个选不中的模型（glm-5.3-flash vs V4 端点的实况）。
+    不在列表则改选列表第一个并写回。返回 (生效模型, note)。"""
+    path = _config_path(entry)
+    if not path or not os.path.isfile(path):
+        return None, ""
+    try:
+        text = open(path, encoding="utf-8", errors="replace", newline="").read()
+    except Exception:
+        return None, ""
+    cur = _yaml_read_value(text, "agent-default-model", "model")
+    ids = _yaml_model_ids(text, "llm-deepseek")
+    if not ids:
+        return cur, "dsh 端点未登记任何模型，请先在 dsh 侧配置模型目录"
+    if cur in ids:
+        return cur, ""
+    first = ids[0]
+    w = write_model(entry, first)
+    fixed = w.get("model") if w.get("ok") else None
+    note = "dsh 默认模型 %s 不在端点模型列表，已改选 %s" % (cur or "（空）", first)
+    if not fixed:
+        note += "（写入失败：%s）" % w.get("error")
+    return fixed, note
+
+
+def _dsh_key_present():
+    """dsh 的密钥是否有着落：进程 env 或它自己的 ~/.dsh/.env（credentials-local）。"""
+    if os.environ.get("DEEPSEEK_API_KEY"):
+        return True
+    try:
+        envfile = os.path.join(os.path.expanduser("~"), ".dsh", ".env")
+        return "DEEPSEEK_API_KEY" in open(envfile, encoding="utf-8", errors="replace").read()
+    except Exception:
+        return False
+
+
+def _toml_section_set(text, section, pairs):
+    """就地写 TOML 段（[section] 下多键）。段存在则逐键替换，缺则整段追加在文末。
+    只处理 codex config.toml 这种顶层简单段；返回 (new_text, changed)。"""
+    eol = "\r\n" if "\r\n" in text else "\n"
+    lines = text.splitlines()
+    start = end = None
+    header = "[%s]" % section
+    for i, ln in enumerate(lines):
+        if ln.strip() == header:
+            start = i
+        elif start is not None and ln.startswith("[") and ln.rstrip().endswith("]"):
+            end = i
+            break
+    if start is None:
+        block = [header] + ["%s = %s" % (k, v) for k, v in pairs]
+        if lines and lines[-1].strip():
+            lines.append("")
+        lines.extend(block)
+        return eol.join(lines).rstrip("\r\n") + eol, True
+    end = end if end is not None else len(lines)
+    changed = False
+    todo = dict(pairs)
+    for i in range(start + 1, end):
+        m = re.match(r"^(\s*)([A-Za-z0-9_.-]+)\s*=", lines[i])
+        if m and m.group(2) in todo:
+            lines[i] = "%s%s = %s" % (m.group(1), m.group(2), todo.pop(m.group(2)))
+            changed = True
+    if todo:
+        ins = start + 1
+        while ins < end and not lines[ins].strip():
+            ins += 1
+        for k, v in list(todo.items())[::-1]:
+            lines.insert(ins, "%s = %s" % (k, v))
+        changed = True
+    return eol.join(lines).rstrip("\r\n") + eol, changed
+
+
+def _toml_top_set(text, key, value):
+    """写 TOML 顶层键（第一个 [段] 之前的区域）。返回 (new_text, changed)。"""
+    eol = "\r\n" if "\r\n" in text else "\n"
+    lines = text.splitlines()
+    pat = re.compile(r"^(\s*)%s\s*=\s*.+$" % re.escape(key))
+    for i, ln in enumerate(lines):
+        if pat.match(ln):
+            lines[i] = "%s%s = %s" % (pat.match(ln).group(1), key, value)
+            return eol.join(lines).rstrip("\r\n") + eol, True
+    first_section = next((i for i, ln in enumerate(lines)
+                          if ln.startswith("[") and ln.rstrip().endswith("]")), len(lines))
+    lines.insert(first_section, "%s = %s" % (key, value))
+    return eol.join(lines).rstrip("\r\n") + eol, True
+
+
+def _sync_codex_settings(entry, model, cp):
+    """codex 专属：把绑定供应商与模型写进 ~/.codex/config.toml
+    （[model_providers.orch] 段 + 顶层 model_provider/model）。
+
+    codex 交互 TUI 不认编排的 -c 一次性覆盖，也不认 ORCH_API_KEY env——
+    没有 config.toml 里的 provider 段，绑定模型根本无处可用；而 model 单写
+    不写 provider 会指到 codex 自带 openai 官方端点上（401）。与编排的
+    _codex_provider_args 同构，但落 config 文件。返回错误串或 None。"""
+    path = _config_path(entry)
+    if not path:
+        return "codex 配置路径无效"
+    path = os.path.realpath(path)
+    home = os.path.abspath(os.path.expanduser("~"))
+    try:
+        if os.path.commonpath([path, home]) != home:
+            return "codex 配置路径越出用户主目录，已拒绝"
+    except ValueError:
+        return "codex 配置路径越出用户主目录，已拒绝"
+    name = cp.get("name", "orch")
+    def q(v):
+        return '"%s"' % str(v).replace("\\", "\\\\").replace('"', '\\"')
+    pairs = [("name", q(cp.get("name", name))),
+             ("base_url", q(cp.get("base_url", ""))),
+             ("env_key", q(cp.get("env_key", "ORCH_API_KEY"))),
+             ("wire_api", q(cp.get("wire_api", "responses")))]
+    try:
+        if os.path.isfile(path):
+            text = open(path, encoding="utf-8", errors="replace", newline="").read()
+        else:
+            Path(path).parent.mkdir(parents=True, exist_ok=True)
+            text = ""
+        text2, _ = _toml_section_set(text, "model_providers.%s" % name, pairs)
+        text2, _ = _toml_top_set(text2, "model_provider", q(name))
+        if model:
+            text2, _ = _toml_top_set(text2, "model", q(model))
+        if text2 != text:
+            if os.path.isfile(path):
+                shutil.copyfile(path, path + ".bak")
+            Path(path).write_bytes(text2.encode("utf-8"))
+        return None
+    except Exception as e:
+        return repr(e)
+
+
+# ---------------------------------------------------------------- 打开前凭据注入
+
+def _jsonc_scan_object(text, start):
+    """扫描 text[start]（须为 '{'）起的 JSON(C) 对象：返回 (闭合偏移, 键表)。
+    键表为 [key, key_start, key_end, value_start, value_end]（偏移相对整个 text，
+    value_end 指向值结束后一格）。跳过字符串转义、// 与 /* */ 注释、嵌套括号，
+    未闭合时容错返回文末。"""
+    n = len(text)
+    i = start + 1
+    keys = []
+    state = "key"
+    depth = 0
+
+    def skip_string(j):
+        while j < n:
+            if text[j] == "\\":
+                j += 2
+                continue
+            if text[j] == '"':
+                return j + 1
+            j += 1
+        return n
+
+    def skip_comment(j):
+        if j + 1 < n and text[j + 1] == "/":
+            e = text.find("\n", j)
+            return n if e < 0 else e
+        if j + 1 < n and text[j + 1] == "*":
+            e = text.find("*/", j + 2)
+            return n if e < 0 else e + 2
+        return j + 1  # 非注释的孤立斜杠：当普通字符
+
+    while i < n:
+        ch = text[i]
+        if ch in " \t\r\n":
+            i += 1
+            continue
+        if ch == "/":
+            i = skip_comment(i)
+            continue
+        if state == "key":
+            if ch == "}":
+                return i, keys
+            if ch == ",":
+                i += 1
+                continue
+            if ch == '"':
+                j = skip_string(i + 1)
+                keys.append([text[i + 1:j - 1], i, j, 0, 0])
+                state = "colon"
+                i = j
+                continue
+            i += 1
+            continue
+        if state == "colon":
+            if ch == ":":
+                state = "value"
+            i += 1
+            continue
+        # state == value
+        if not keys:
+            return n, keys  # 结构异常：放弃扫描
+        keys[-1][3] = i
+        if ch in "{[":
+            depth = 0
+            j = i
+            while j < n:
+                c = text[j]
+                if c == '"':
+                    j = skip_string(j + 1)
+                    continue
+                if c == "/":
+                    j = skip_comment(j)
+                    continue
+                if c in "{[":
+                    depth += 1
+                elif c in "}]":
+                    depth -= 1
+                    if depth == 0:
+                        break
+                j += 1
+            keys[-1][4] = j + 1
+            i = j + 1
+            state = "key"
+        elif ch == '"':
+            j = skip_string(i + 1)
+            keys[-1][4] = j
+            i = j
+            state = "key"
+        else:  # 数字 / true / false / null（或值后直接撞上 , } 换行的容错）
+            j = i
+            while j < n and text[j] not in ",}\r\n":
+                j += 1
+            keys[-1][4] = j
+            # j==i 说明值位置直接是分隔符/换行（空值或状态残留）——必须前进一格，
+            # 否则 while i < n 永远停在原地（死循环）
+            i = j if j > i else j + 1
+            state = "key"
+    return n, keys
+
+
+def _jsonc_strip_comments(text):
+    """剥掉 jsonc 的 // 行注释与 /* */ 块注释，供 json.loads 解析。
+    字符串字面量内部的 //（$schema 的 https:// 等）不剥——逐字符扫描；
+    简单的 re.sub(r"//[^\\n]*") 会把 URL 截断导致解析失败。"""
+    out = []
+    i, n = 0, len(text)
+    while i < n:
+        ch = text[i]
+        if ch == '"':
+            j = i + 1
+            while j < n:
+                if text[j] == "\\":
+                    j += 2
+                    continue
+                if text[j] == '"':
+                    break
+                j += 1
+            out.append(text[i:min(j + 1, n)])
+            i = j + 1
+            continue
+        if ch == "/" and i + 1 < n and text[i + 1] == "/":
+            e = text.find("\n", i)
+            i = n if e < 0 else e  # 行注释：换行符本身保留
+            continue
+        if ch == "/" and i + 1 < n and text[i + 1] == "*":
+            e = text.find("*/", i + 2)
+            i = n if e < 0 else e + 2
+            out.append(" ")
+            continue
+        out.append(ch)
+        i += 1
+    return "".join(out)
+
+
+def _jsonc_set(text, path, value_json):
+    """JSON(C) 顶层就地写键：path=("provider","orch") 或 ("model",)。
+    只动目标片段，其余文本（含注释与原格式）原样保留。返回 (new_text, ok)。
+    空文件按空对象起笔（write_model 缺文件按需创建的产物）。"""
+    brace = text.find("{")
+    if brace < 0 and not text.strip():
+        return ('{\n  "%s": %s\n}' % (path[0], value_json), True)
+    if brace < 0:
+        return text, False
+    end, keys = _jsonc_scan_object(text, brace)
+    head = path[0]
+    if len(path) == 1:
+        for _k, _ks, _ke, vs, ve in keys:
+            if _k == head:
+                return text[:vs] + value_json + text[ve:], True
+        return _jsonc_insert_entry(text, brace, end, head, value_json), True
+    # 二级路径：先定位一级键的值对象
+    tgt = None
+    for k, _ks, _ke, vs, ve in keys:
+        if k == head:
+            tgt = (vs, ve)
+            break
+    if tgt is None:  # 一级键不存在：整体插入
+        block = json.dumps({path[1]: json.loads(value_json)}, ensure_ascii=False)
+        return _jsonc_insert_entry(text, brace, end, head, block), True
+    vs, ve = tgt
+    if text[vs:ve].lstrip()[0:1] != "{":
+        return text, False  # 一级值不是对象：保守放弃（不覆盖用户的非标结构）
+    end2, keys2 = _jsonc_scan_object(text, vs)
+    for k, _ks, _ke, v2s, v2e in keys2:
+        if k == path[1]:
+            return text[:v2s] + value_json + text[v2e:], True
+    seg = text[vs:end2 + 1]
+    new_seg = _jsonc_insert_entry(seg, 0, end2 - vs, path[1], value_json)
+    if new_seg is None:
+        return text, False
+    return text[:vs] + new_seg + text[end2 + 1:], True
+
+
+def _jsonc_insert_entry(text, brace, end, key, value_json):
+    """在 {brace..end} 对象的开头插入 "key": value（带尾逗号，不依赖原文件的
+    逗号风格）；对象为空时去掉多余逗号。"""
+    m = re.match(r"\{([ \t\r\n]*)", text[brace:end + 1])
+    first = brace + (m.end(1) if m else 1)  # m 从 brace 起 match：end(1) 已含 '{'
+    if first >= end:  # 空对象 {}
+        return text[:brace + 1] + ' "%s": %s ' % (key, value_json) + text[end:]
+    return text[:first] + '"%s": %s,\n  ' % (key, value_json) + text[first:]
+
+
+def _sync_settings_env(path, updates, remove_keys=()):
+    """把键值对写进目标 settings.json 的 env 段（claude/qwen 等同构：交互 TUI
+    启动时把该段合并进进程环境）。只动 env 相关键，其余内容保留；坏 JSON 中止
+    不覆盖；改动前 .bak。返回错误串或 None。path 必须已过主目录围栏校验。"""
+    try:
+        if os.path.isfile(path):
+            text = open(path, encoding="utf-8", errors="replace").read()
+            try:
+                data = json.loads(text)
+            except Exception:
+                return "settings.json 不是合法 JSON，已中止（避免覆盖）"
+        else:
+            Path(path).parent.mkdir(parents=True, exist_ok=True)
+            data = {}
+        if not isinstance(data, dict):
+            return "settings.json 结构异常（顶层不是对象），已中止"
+        env = data.get("env")
+        if not isinstance(env, dict):
+            env = {}
+        for k in remove_keys:
+            env.pop(k, None)
+        env.update(updates)
+        data["env"] = env
+        if os.path.isfile(path):
+            shutil.copyfile(path, path + ".bak")
+        Path(path).write_bytes(json.dumps(data, ensure_ascii=False, indent=2).encode("utf-8"))
+        return None
+    except Exception as e:
+        return repr(e)
+
+
+def _guard_home(path):
+    """展开并校验配置路径必须在用户主目录内；通过则返回 realpath，否则 None。"""
+    if not path:
+        return None
+    path = os.path.realpath(path)
+    home = os.path.abspath(os.path.expanduser("~"))
+    try:
+        if os.path.commonpath([path, home]) != home:
+            return None
+    except ValueError:
+        return None
+    return path
+
+
+def _sync_claude_settings(entry, model, prov):
+    """claude-code 专属：把 anthropic 供应商的端点+密钥+模型写进
+    ~/.claude/settings.json 的 env 段（交互 TUI 与无头共用该文件，只认
+    ANTHROPIC_*；编排降级链给的 ORCH_API_KEY 对它等于没 key）。
+    返回错误串或 None。"""
+    path = _guard_home(_config_path(entry))
+    if not path:
+        return "claude 配置路径无效或越出用户主目录，已拒绝"
+    updates = {"ANTHROPIC_BASE_URL": prov.get("base_url") or "",
+               # AUTH_TOKEN 走 Bearer 头（Z.ai 等原生 anthropic 网关的用法）；
+               # 与 x-api-key 互斥，清掉可能残留的 ANTHROPIC_API_KEY 防止带错头
+               "ANTHROPIC_AUTH_TOKEN": prov.get("api_key") or ""}
+    if model:
+        updates["ANTHROPIC_MODEL"] = model
+    return _sync_settings_env(path, updates, remove_keys=("ANTHROPIC_API_KEY",))
+
+
+def _sync_qwen_settings(entry, model, prov):
+    """qwencode 专属：openai 兼容供应商写进 ~/.qwen/settings.json 的 env 段
+    （qwen-code 实测认 OPENAI_API_KEY/OPENAI_BASE_URL/OPENAI_MODEL，存在即
+    优先走 openai 兼容通道——2026-09-15 真机对维云端点实测请求到达并鉴权）。
+    anthropic 协议未实证，不开（协议不匹配时宁可提示）。返回错误串或 None。"""
+    path = _guard_home(_config_path(entry))
+    if not path:
+        return "qwen 配置路径无效或越出用户主目录，已拒绝"
+    updates = {"OPENAI_API_KEY": prov.get("api_key") or "",
+               "OPENAI_BASE_URL": prov.get("base_url") or ""}
+    if model:
+        updates["OPENAI_MODEL"] = model
+    return _sync_settings_env(path, updates)
+
+
+def _opencode_config_candidates(entry):
+    """opencode 配置的候选路径：catalog 登记的 jsonc 优先，其次同目录的
+    opencode.json（opencode 两种文件名都认，用户现有安装多用 json）。"""
+    out = []
+    primary = _config_path(entry)
+    if primary:
+        out.append(primary)
+    alt = os.path.join(os.path.abspath(os.path.expanduser("~/.config/opencode")),
+                       "opencode.json")
+    if alt not in out:
+        out.append(alt)
+    return out
+
+
+def _sync_opencode_settings(entry, model, prov):
+    """opencode 专属：把绑定供应商写进其配置的 provider.orch 段 + 顶层 model
+    （opencode 交互 TUI 只认自家配置文件里的凭据，ORCH_API_KEY env 对它等于
+    没 key）。写入所有已存在的候选文件（避免新文件遮蔽旧文件的读取优先级），
+    全不存在时建 catalog 登记的那个。纯 JSON 走整体读改写；带注释的 JSONC 走
+    文本级就地 patch（保留注释）。返回错误串或 None。"""
+    npm = "@ai-sdk/anthropic" if prov.get("protocol") == "anthropic" else "@ai-sdk/openai-compatible"
+    block = {"npm": npm, "name": prov.get("name") or "CodeBee 绑定",
+             "options": {"baseURL": prov.get("base_url") or "",
+                         "apiKey": prov.get("api_key") or ""},
+             "models": {model: {"name": model}} if model else {}}
+    top_model = ("orch/" + model) if model else ""
+    targets = [p for p in _opencode_config_candidates(entry) if os.path.isfile(p)] \
+        or [_opencode_config_candidates(entry)[0]]
+    errs = []
+    for path in targets:
+        try:
+            path = os.path.realpath(path)
+            home = os.path.abspath(os.path.expanduser("~"))
+            try:
+                if os.path.commonpath([path, home]) != home:
+                    errs.append("路径越出主目录：%s" % path)
+                    continue
+            except ValueError:
+                errs.append("路径越出主目录：%s" % path)
+                continue
+            if os.path.isfile(path):
+                text = open(path, encoding="utf-8", errors="replace", newline="").read()
+            else:
+                Path(path).parent.mkdir(parents=True, exist_ok=True)
+                text = ""
+            try:
+                data = json.loads(text) if text.strip() else {}
+            except Exception:
+                data = None
+            if data is not None:  # 纯 JSON：整体读改写（保序）
+                if not isinstance(data, dict):
+                    errs.append("结构异常（顶层不是对象）：%s" % path)
+                    continue
+                provs = data.get("provider")
+                if not isinstance(provs, dict):
+                    provs = {}
+                provs["orch"] = block
+                data["provider"] = provs
+                if top_model:
+                    data["model"] = top_model
+                new_text = json.dumps(data, ensure_ascii=False, indent=2) + "\n"
+            else:  # JSONC（带注释）：文本级 patch
+                block_json = json.dumps(block, ensure_ascii=False)
+                new_text, ok = _jsonc_set(text, ("provider", "orch"), block_json)
+                if not ok:
+                    errs.append("JSONC 就地改写失败：%s" % path)
+                    continue
+                if top_model:
+                    new_text, _ = _jsonc_set(new_text, ("model",),
+                                             json.dumps(top_model))
+            if os.path.isfile(path):
+                shutil.copyfile(path, path + ".bak")
+            Path(path).write_bytes(new_text.encode("utf-8"))
+        except Exception as e:
+            errs.append("%s: %r" % (path, e))
+    return "；".join(errs) or None
+
+
+# 打开前专属注入通道：{agent_id: (可注入协议, 注入器)}。交互 TUI 脱离编排链路，
+# 只认自家配置文件里的凭据，编排降级给的 env（ORCH_API_KEY 等）对它们无效。
+_AGENT_INJECTORS = {
+    "claude-code": (("anthropic",), _sync_claude_settings),
+    "opencode": (("anthropic", "openai"), _sync_opencode_settings),
+    "qwencode": (("openai",), _sync_qwen_settings),
+}
+
+# 无专属注入通道的专有协议 CLI：env 注入大概率无效，打开时明确告知而非静默废
+# （mimo 系 opencode 衍生但配置路径未实证，先按提示类；grok 吃 XAI_API_KEY 但
+# 无端点 env 可指中转，openai 协议供应商也用不上）
+_NO_CHANNEL_HINT = ("grok-build", "pi", "mimo-code")
+
+
+def _sync_agent_injection(entry, binding):
+    """打开前把绑定链里第一个可注入供应商落进 CLI 自家配置。与编排降级链解耦：
+    协议不匹配不降级（claude 拿 openai 的 key 等于没 key）。返回给用户看的提示
+    （成功注入 / 不可用原因），None=该 CLI 无需处理。"""
+    from . import modelhub
+    spec = _AGENT_INJECTORS.get(entry["id"])
+    if spec:
+        protocols, injector = spec
+        pick, note = modelhub.launch_pick(entry["id"], protocols)
+        if pick:
+            prov = pick["provider"]
+            err = injector(entry, pick["model"], prov)
+            if err:
+                return "%s 凭据同步失败：%s（打开后可能需在其自带界面登录）" % (
+                    entry.get("name", entry["id"]), err)
+            return "已注入 %s（%s · %s）" % (prov.get("name") or "供应商",
+                                            prov.get("protocol"), prov.get("base_url", ""))
+        return note
+    if entry["id"] in _NO_CHANNEL_HINT:
+        if binding.get("env"):
+            return "已按绑定注入 env，但该 CLI 未必认 CodeBee 的凭据通道，打开后若要求登录请在其界面内登录"
+        return "未绑定可用供应商：打开后需在其自带界面登录；要打开即用请到「CLI 绑定」页绑定"
+    return None
+
+
+def _sync_launch_model(entry, binding):
+    """打开前把「CLI 绑定」页选中的模型落到该 CLI 自己的配置文件——保证
+    交互/网页版启动即选中绑定模型（绑定页是运行时模型唯一真源，目录页的
+    「默认模型」只是手动快照，会滞后）。dsh 额外同步端点与 models 目录；
+    codex 额外落 provider 段（交互 TUI 只认 config.toml，不认编排的 -c 覆盖
+    与 ORCH_API_KEY env）。返回给用户看的同步笔记列表。"""
+    notes = []
+    model = (binding.get("model") or "").strip()
+    prov = binding.get("provider") or {}
+    fmt = (entry.get("config") or {}).get("format")
+    is_dsh = entry["id"] in ("deepseek-harness", "dsh")
+    is_codex = entry["id"] in ("codex-cli", "codex")
+    cp = binding.get("codex_provider")
+    if model and fmt in _WRITABLE_FORMATS:
+        w = write_model(entry, model)
+        notes.append("模型已同步为 %s" % w["model"] if w.get("ok")
+                     else "模型同步失败：%s" % w.get("error"))
+    elif model and not is_dsh and entry["id"] not in _AGENT_INJECTORS:
+        # 有专属注入通道的 CLI 由 _sync_agent_injection 负责落模型（含 jsonc），
+        # 不再报「只读配置」以免与注入成功的提示互相矛盾
+        notes.append("该工具模型为只读配置，按其现有配置打开（绑定模型 %s 未自动写入）" % model)
+    if is_dsh and model and prov.get("base_url"):
+        err = _sync_dsh_settings(entry, model, prov["base_url"])
+        notes.append("dsh 端点已同步为 %s" % prov["base_url"] if not err
+                     else "dsh 端点同步失败：%s" % err)
+    elif is_dsh and not model:
+        _fixed, note = _dsh_selfcheck_model(entry)
+        if note:
+            notes.append(note)
+    if is_codex and cp:
+        err = _sync_codex_settings(entry, model, cp)
+        notes.append("codex 供应商已同步为 %s" % cp.get("base_url", "") if not err
+                     else "codex 供应商同步失败：%s" % err)
+    inj = _sync_agent_injection(entry, binding)
+    if inj:
+        notes.append(inj)
+    return notes
+
+
+def launch(entry, open_browser=True):
+    """一键打开：web 类后台起服务并自动开浏览器；console 类新开终端窗口跑交互 TUI。
+
+    打开前把「CLI 绑定」页的模型落盘到该 CLI 配置文件（dsh 连端点与 models
+    目录一起同步），并注入绑定密钥 env——保证打开即选中可用模型。"""
+    if not detect_entry(entry).get("installed"):
+        return {"ok": False, "error": "未安装，无法打开"}
+    launch = entry.get("launch") or {}
+    cmd = (launch.get("command") or "").strip()
+    if not cmd:
+        return {"ok": False, "error": "未配置打开命令（可在 data/catalog.json 补 launch 字段）"}
+    is_dsh = entry["id"] in ("deepseek-harness", "dsh")
+    try:
+        from . import modelhub
+        binding = modelhub.resolve_binding(entry["id"]) or {}
+    except Exception:
+        binding = {}
+    env = os.environ.copy()
+    env.update(binding.get("env") or {})
+    notes = _sync_launch_model(entry, binding)
+    if is_dsh and not (binding.get("env") or {}) and not _dsh_key_present():
+        notes.append("未发现 dsh 密钥：打开后可能需在 dsh 内登录配置；"
+                     "要打开即用，请到「CLI 绑定」页给 DeepSeek Harness 绑定 openai 协议供应商")
+    name = entry.get("name", entry["id"])
+    kind = (launch.get("kind") or "console").lower()
+    extra = ("；".join(notes)) if notes else ""
+
+    if kind == "web":
+        try:
+            port = int(launch.get("port") or 0)
+        except (TypeError, ValueError):
+            port = 0
+        if port <= 0:
+            return {"ok": False, "error": "web 类打开必须配置固定端口（launch.port）"}
+        bare = "http://127.0.0.1:%d" % port
+        log_path = _launch_log_path(entry)
+        if not log_path:
+            return {"ok": False, "error": "启动日志路径不可信，已拒绝打开"}
+        if _port_open(port):
+            # 已在运行：新实例抢不到端口、新 token 也拿不到——从上次启动日志
+            # 恢复带 token 的信任 URL（dsh web 有 /?token=... 围栏，裸开是 401）。
+            # 模型同步照做：正在跑的实例不重启，改的是它下次生效的配置
+            url = _best_url(log_path, port) or bare
+            if open_browser:
+                webbrowser.open(url)
+            msg = "服务已在运行，已打开 " + url
+            if extra:
+                msg += "（" + extra + "）"
+            return {"ok": True, "kind": "web", "url": url, "message": msg}
+        # 后台起服务：无窗口、不阻塞请求；子进程输出重定向到启动日志（cmd 层
+        # 重定向，路径含空格才加引号）——web 类普遍会在启动行打印带 token 的
+        # 信任 URL（如 dsh web），就绪线程从日志提取后再开浏览器
+        ls = str(log_path)
+        spawn_cmd = "%s > %s 2>&1" % (cmd, ('"%s"' % ls) if " " in ls else ls)
+        try:
+            log_path.parent.mkdir(parents=True, exist_ok=True)
+            subprocess.Popen(["cmd", "/c", spawn_cmd], cwd=str(paths.ROOT), env=env,
+                             stdin=subprocess.DEVNULL, stdout=subprocess.DEVNULL,
+                             stderr=subprocess.DEVNULL, creationflags=CREATE_NO_WINDOW)
+        except Exception as e:
+            return {"ok": False, "error": "无法启动服务: %r" % e}
+        if open_browser:
+            threading.Thread(target=_open_when_ready, args=(port, bare, log_path),
+                             name="launch-wait-%d" % port, daemon=True).start()
+        return {"ok": True, "kind": "web", "url": bare,
+                "message": "%s 正在启动，就绪后浏览器会自动打开（%s）%s"
+                           % (name, bare, ("；" + extra) if extra else "")}
+
+    if sys.platform != "win32":
+        return {"ok": False, "error": "终端窗口拉起暂仅支持 Windows"}
+    # start 为目标命令新开一个可见终端窗口；cmd /k 让 CLI 退出后窗口保留，
+    # 报错不至于一闪而过。外层 cmd 用 CREATE_NO_WINDOW 隐藏。
+    argv = ["cmd", "/c", "start", "CodeBee %s" % name, "/D", str(paths.ROOT),
+            "cmd", "/k", cmd]
+    try:
+        subprocess.Popen(argv, cwd=str(paths.ROOT), env=env,
+                         creationflags=CREATE_NO_WINDOW)
+    except Exception as e:
+        return {"ok": False, "error": "无法打开终端窗口: %r" % e}
+    return {"ok": True, "kind": "console", "message": "已在新的终端窗口打开 %s%s"
+            % (name, ("（" + extra + "）") if extra else "")}
+
+
+def _open_when_ready(port, url, log_path, timeout=30):
+    """等 web 服务端口就绪后开浏览器。优先从启动日志提取带 token 的信任 URL
+    （端口通了 token 行可能还差几十毫秒才落盘，故就绪后最多再等 6 秒）；
+    超时兜底也开——服务可能只是慢，用户手动刷新即可。"""
+    deadline = time.time() + timeout
+    while time.time() < deadline:
+        if _port_open(port):
+            break
+        time.sleep(0.5)
+    token_deadline = time.time() + 6
+    best = None
+    while time.time() < token_deadline:
+        best = _best_url(log_path, port)
+        if best and "token=" in best:
+            break
+        time.sleep(0.5)
+    try:
+        webbrowser.open(best or url)
+    except Exception:
+        pass
+
+
+# ---------------------------------------------------------------- 安装/升级
+
+def run_mgmt_command(entry, op, cancel_event=None, log_path=None):
+    """执行 install/upgrade/uninstall 命令（在任务队列里跑，日志实时落盘）。"""
+    if op == "uninstall":
+        cmd = catalog.uninstall_command(entry)
+        if not cmd:
+            return {"ok": False,
+                    "error": "无法推导卸载命令：请在 data/catalog.json 的 \"%s\" 里配置 uninstall 字段"
+                             % entry["id"]}
+    else:
+        cmd = entry.get(op)
+        if not cmd:
+            return {"ok": False,
+                    "error": "未配置 %s 命令：请在 data/catalog.json 的 \"%s\" 里补充，或用官方渠道安装"
+                             % (op, entry["id"])}
+    res = runner.run_process(shell_cmd=cmd, cwd=str(paths.ROOT),
+                             timeout=1800, cancel_event=cancel_event, log_path=log_path)
+    detect_all(force=True)
+    with _LOCK:
+        _STATE["versions"].pop(entry["id"], None)
+    return {"ok": res["ok"], "exit_code": res["exit_code"], "command": cmd,
+            "error": "" if res["ok"] else (res["stderr"][-800:] or "退出码 %s" % res["exit_code"])}
+
+
+# ---------------------------------------------------------------- 版本检查
+
+_UPDATE_CACHE = {}   # agent_id → (ts, {current, latest, updatable, note})
+UPDATE_TTL = 600
+
+
+def _npm_pkg_name(cmd):
+    """从 npm 安装命令里取包名（实现已统一到 catalog.npm_pkg_name）。"""
+    return catalog.npm_pkg_name(cmd)
+
+
+def _ver_tuple(s):
+    return [int(x) for x in re.findall(r"\d+", str(s or ""))[:4]]
+
+
+def check_update(entry, force=False):
+    """检查是否有新版本可用。npm 走 npm view；winget 走 winget upgrade 列表；
+    其他渠道标记为不支持。结果缓存 10 分钟。"""
+    eid = entry["id"]
+    if not force:
+        cached = _UPDATE_CACHE.get(eid)
+        if cached and time.time() - cached[0] < UPDATE_TTL:
+            return cached[1]
+    current = version_of(entry)
+    cur_num = ".".join(str(x) for x in _ver_tuple(current)) or "-"
+    result = {"current": current, "latest": None, "updatable": None, "note": ""}
+
+    cmd = entry.get("install") or entry.get("upgrade") or ""
+    pkg = _npm_pkg_name(cmd)
+    if pkg:
+        r = runner.run_process(argv=["cmd", "/c", "npm", "view", pkg, "version"], timeout=90)
+        latest = ""
+        if r["ok"]:
+            for line in (r["stdout"] or "").splitlines():
+                line = line.strip()
+                if line and re.match(r"^\d", line):
+                    latest = line
+        if not latest:
+            result["note"] = "查询 npm 失败（网络或 registry 问题）：" + (r["stderr"] or "")[-200:]
+        else:
+            result["latest"] = latest
+            result["updatable"] = bool(_ver_tuple(latest) > _ver_tuple(cur_num))
+            if not result["updatable"]:
+                result["note"] = "已是最新版本"
+    elif "winget" in cmd:
+        m = re.search(r"--id\s+([A-Za-z0-9._-]+)", cmd)
+        wid = m.group(1) if m else None
+        if not wid:
+            result["note"] = "无法从命令中解析 winget 包 ID"
+        else:
+            r = runner.run_process(argv=["cmd", "/c", "winget", "upgrade"], timeout=180)
+            if not r["ok"]:
+                result["note"] = "winget upgrade 查询失败：" + (r["stderr"] or "")[-200:]
+            else:
+                hit = [l for l in (r["stdout"] or "").splitlines() if wid.lower() in l.lower()]
+                result["updatable"] = bool(hit)
+                if hit:
+                    result["latest"] = "（winget 有可用更新）"
+                else:
+                    result["latest"] = cur_num
+                    result["note"] = "已是最新版本"
+    else:
+        result["note"] = "该渠道暂不支持自动检查更新，可直接点升级尝试"
+
+    with _LOCK:
+        _UPDATE_CACHE[eid] = (time.time(), result)
+    return result
+
+
+# 「进入智能体目录页自动检查更新」的后台任务状态
+_UPDATE_CHECK = {"running": False, "total": 0, "done": 0}
+
+
+def updates_checking():
+    with _LOCK:
+        return bool(_UPDATE_CHECK["running"])
+
+
+def update_info(entry):
+    """单个 CLI 的更新检查结果（供管理页卡片展示）。没查过时为 unknown。"""
+    with _LOCK:
+        cached = _UPDATE_CACHE.get(entry["id"])
+    if not cached:
+        return {"status": "unknown", "latest": None, "updatable": None,
+                "note": "", "checked_at": ""}
+    res = cached[1] or {}
+    up = res.get("updatable")
+    status = "updatable" if up is True else ("current" if up is False else "unsupported")
+    return {"status": status, "latest": res.get("latest"), "updatable": up,
+            "note": res.get("note") or "",
+            "checked_at": time.strftime("%H:%M", time.localtime(cached[0]))}
+
+
+def _checkable_entries():
+    """已安装且配了 install/upgrade 的条目——只有这些才谈得上「是否有新版本」。"""
+    detected = detect_all()
+    out = []
+    for e in catalog.load():
+        det = (detected or {}).get(e["id"]) or {}
+        if det.get("installed") and (e.get("upgrade") or e.get("install")):
+            out.append(e)
+    return out
+
+
+def check_updates_async(force=False):
+    """后台逐个检查已安装 CLI 的远端最新版本，返回本次要检查的条目数。
+
+    进入「智能体目录」页时自动触发。单条结果复用 check_update 的 10 分钟缓存，
+    所以反复进出页面几乎不产生额外子进程/网络开销；已在跑时不重复起线程。
+    """
+    with _LOCK:
+        if _UPDATE_CHECK["running"]:
+            return 0
+        entries = _checkable_entries()
+        _UPDATE_CHECK.update(running=True, total=len(entries), done=0)
+
+    def _worker():
+        try:
+            for e in entries:
+                try:
+                    check_update(e, force=force)
+                except Exception:
+                    pass
+                finally:
+                    with _LOCK:
+                        _UPDATE_CHECK["done"] += 1
+        finally:
+            with _LOCK:
+                _UPDATE_CHECK["running"] = False
+
+    threading.Thread(target=_worker, name="catalog-update-check", daemon=True).start()
+    return len(entries)
+
+
+def catalog_view():
+    """管理页数据：catalog + 检测 + 版本 + 模型 + 编排启用状态。"""
+    from . import registry
+    entries = catalog.load()
+    detected = detect_all(force=False)
+    enabled = registry.load_enabled()
+    view = []
+    for e in entries:
+        det = detected.get(e["id"]) or {}
+        pref = enabled.get(e["id"]) or {}
+        orch_enabled = bool(pref.get("enabled", e.get("default_enabled", False))) if e.get("orch") else False
+        view.append({
+            "id": e["id"], "name": e.get("name", e["id"]), "note": e.get("note", ""),
+            "group": "installed" if det.get("installed") else "installable",
+            "installed": det.get("installed", False),
+            "detail": det.get("detail", ""),
+            "version": version_of(e),
+            "config_path": _config_path(e),
+            "config_writable": (e.get("config") or {}).get("format") in _WRITABLE_FORMATS,
+            "model": read_model(e),
+            "orch_kind": (e.get("orch") or {}).get("kind"),
+            "orch_enabled": orch_enabled,
+            "update": update_info(e),
+            "has_install": bool(e.get("install")),
+            "has_upgrade": bool(e.get("upgrade")),
+            # 卸载命令由 install/upgrade 推导（或 catalog 显式配置），供 UI 确认框展示
+            "uninstall_cmd": catalog.uninstall_command(e) if det.get("installed") else None,
+            # 一键打开配置（kind=web/console + command）；没配的条目 UI 不出「打开」按钮
+            "launch": e.get("launch"),
+        })
+    return view