codebee 0.1.6 → 0.1.8

package/app/core/automation.py

@@ -299,11 +299,36 @@ def _launch_run(t):
                                     time.strftime("%m-%d %H:%M"))).strip()[:40],
                "goal": t.get("prompt") or "",
                "workdir": (t.get("workdir") or "").strip()}
-    task = store.create_task(payload)
-    run = store.create_run("orchestration", task["title"], task_id=task["id"])
-    store.update_task_status(task["id"], "queued")
-    jobs.enqueue({"kind": "orchestration", "run_id": run["id"], "task_id": task["id"]})
-    return run["id"]
+    task = None
+    run = None
+    try:
+        task = store.create_task(payload)
+        run = store.create_run("orchestration", task["title"], task_id=task["id"])
+        store.update_task_status(task["id"], "queued")
+        jobs.enqueue({"kind": "orchestration", "run_id": run["id"],
+                      "task_id": task["id"]})
+        return run["id"]
+    except Exception:
+        # A scheduler failure must not leave a task/run that looks queued forever.
+        # Keep the public error generic; the detailed traceback stays in the
+        # service log and the run record remains useful for diagnostics.
+        log.exception("automation: 运行入队失败 task=%s run=%s",
+                      (task or {}).get("id"), (run or {}).get("id"))
+        if run:
+            try:
+                store.update_run(run["id"], status="failed",
+                                 error="任务入队失败，请稍后重试",
+                                 ended_at=time.strftime("%Y-%m-%d %H:%M:%S"))
+            except Exception:
+                log.exception("automation: 运行失败收口失败 run=%s", run.get("id"))
+        elif task:
+            # create_run may fail before a run is available. The task was
+            # already persisted by create_task, so mark it failed as well.
+            try:
+                store.update_task_status(task["id"], "failed")
+            except Exception:
+                log.exception("automation: 任务失败收口失败 task=%s", task.get("id"))
+        raise
 
 
 def _fire(snapshot, now):

package/app/core/catalog.py

@@ -10,6 +10,7 @@ from __future__ import annotations
 import copy
 import json
 import re
+import sys
 import threading
 
 from . import paths
@@ -74,12 +75,13 @@ DEFAULT_CATALOG = [
     },
     {
         "id": "aider", "name": "Aider", "cli_group": "installable",
-        "note": "Python 系结对编程 CLI；用 py -3.13 安装，模型经其配置/环境变量设置",
+        "note": "Python 系结对编程 CLI；用 uv tool 安装（自动备好 3.12 托管解释器），"
+                "模型经其配置/环境变量设置",
         "detect": {"cli": "aider"},
         "orch": {"kind": "aider", "command": "aider"},
         "config": {"path": "~/.aider.conf.yml", "format": None, "model_key": None},
-        "install": "py -3.13 -m pip install -U aider-chat",
-        "upgrade": "py -3.13 -m pip install -U aider-chat",
+        "install": "uv tool install --python 3.12 aider-chat",
+        "upgrade": "uv tool upgrade aider-chat",
         "default_enabled": False,
     },
     {
@@ -216,6 +218,16 @@ CONFIG_PATCH = {
 }
 
 
+# 安装命令平台修正（{id: (install, upgrade)}）：仅非 Windows 套用。winget 是
+# Windows 包管理器、py 启动器 Windows 独有，这两类命令在 macOS/Linux 上必失败
+# ——load() 幂等改写为等价的 npm/pip 渠道（与 CONFIG_PATCH 同策略：覆盖无风险，
+# 旧数据不修正的话每次点安装/升级都注定失败）。
+INSTALL_PATCH_NONWIN = {
+    "claude-code": ("npm install -g @anthropic-ai/claude-code",
+                    "npm install -g @anthropic-ai/claude-code@latest"),
+}
+
+
 def npm_pkg_name(cmd):
     """从 npm 安装命令里取包名（支持 @scope/name@latest）。
 
@@ -256,6 +268,13 @@ def derive_uninstall(cmd):
         if pkgs:
             head = m.group(1).replace("pip install", "pip uninstall")
             return "%s -y %s" % (head, pkgs[0])
+    m = re.search(r"uv\s+tool\s+install\b(.*)$", c)
+    if m:
+        # uv tool install 的包名在末尾（--python 3.12 这类 flag 的值不带横杠，
+        # 取最后一个非 flag token 才是包名）
+        pkgs = [t for t in m.group(1).split() if not t.startswith("-")]
+        if pkgs:
+            return "uv tool uninstall %s" % pkgs[-1]
     return None
 
 
@@ -289,6 +308,28 @@ def _apply_config_patch(entries):
             e["config"] = dict(patch)
 
 
+def _apply_install_patch(entries):
+    # Aider 渠道整体迁移到 uv tool（跨平台、含存量数据幂等修正）：anaconda 的
+    # py -3.13 装 aider 必挂在 numpy 源码构建（老 setuptools 引用 py3.12 已删除
+    # 的 pkgutil.ImpImporter），而多数机器又没有 3.9-3.12 的 pip 解释器；uv 能
+    # 自带托管解释器一条命令装好。只迁移仍旧 pip 形态的配置，已是 uv 或用户
+    # 自定义的其他命令不动（2026-09-18 本机实装失败定版）。
+    aider_uv = ("uv tool install --python 3.12 aider-chat",
+                "uv tool upgrade aider-chat")
+    for e in entries:
+        if e.get("id") != "aider":
+            continue
+        cur = (e.get("install") or "").strip()
+        if "pip" in cur and "install" in cur:
+            e["install"], e["upgrade"] = aider_uv
+    if sys.platform == "win32":
+        return
+    for e in entries:
+        patch = INSTALL_PATCH_NONWIN.get(e.get("id"))
+        if patch:
+            e["install"], e["upgrade"] = patch
+
+
 def _merge_new_defaults(entries):
     """把内置默认里「新增的」条目补进已加载清单（同 id 已存在则原样保留）。
 
@@ -321,6 +362,7 @@ def load(force=False):
         _apply_resume_patch(entries)
         _apply_launch_patch(entries)
         _apply_config_patch(entries)
+        _apply_install_patch(entries)
         _CACHE["entries"] = entries
         return entries
 

package/app/core/errorlog.py

@@ -0,0 +1,179 @@
+# -*- coding: utf-8 -*-
+"""错误台账：把分散在各处的失败（步骤失败/超时/看门狗击杀、CLI 启动失败）
+统一落成结构化记录——「客户遇到了问题我们根本不知道」的第一块拼图。
+
+落盘：data/errors/errors-YYYYMM.jsonl（按月分文件，append-only，一行一记录）。
+与 usage.py 同一套纪律：
+- record() 任何异常都吞掉——埋点失败绝不能影响任务执行；
+- 落盘前先脱敏（scrub_text）：密钥/令牌/绝对路径/超长文本一概不进台账。
+
+隐私红线（上传侧在 telemetry.py 再兜一道底）：
+1. API KEY / 凭据绝不入台账；
+2. 任务正文 / 章节内容（用户版权内容）绝不入台账——detail 只允许失败原因摘录；
+3. 绝对路径剥成相对片段。
+"""
+from __future__ import annotations
+
+import json
+import re
+import threading
+import time
+import uuid
+
+from . import paths
+
+LOCK = threading.RLock()
+
+FIELDS = ("id", "ts", "day", "category", "reason", "detail", "provider",
+          "model", "tool", "role", "run_id", "task_id", "step",
+          "exit_code", "app_version", "os")
+
+# detail 摘录上限（字符）：失败原因足够，不要变成日志搬运
+DETAIL_LIMIT = 600
+
+
+# ---------------------------------------------------------------- 脱敏
+
+# 常见密钥形态：OpenAI 系 sk- 前缀、Bearer 令牌、显式 key/secret/token 赋值
+_KEY_PATTERNS = (
+    (re.compile(r"\bsk-[A-Za-z0-9_-]{8,}"), "[key]"),
+    (re.compile(r"\b(?:Bearer|bearer)\s+\S+"), "Bearer [key]"),
+    # 赋值/JSON 两种形态都要吃：api_key=xxx、api_key: "xxx"（冒号前可有闭引号）
+    (re.compile(r"(?i)\b((?:api[_-]?|access[_-]?|secret[_-]?|auth[_-]?)(?:key|token|secret))"
+                r"""["']?\s*[:=，,]\s*["']?[A-Za-z0-9._~+/=-]{8,}"""), r"\1[key]"),
+    # 裸长十六进制/64 位串（可能是凭据指纹）
+    (re.compile(r"\b[0-9a-fA-F]{40,}\b"), "[token]"),
+)
+
+# 绝对路径：Windows 盘符路径、UNC、POSIX 家目录——剥掉盘符/用户名只留尾部结构
+_PATH_PATTERNS = (
+    (re.compile(r"(?i)\b[A-Z]:\\(?:[^\\/:*?\"<>|\r\n]+\\)*[^\\/:*?\"<>|\r\n]*"),
+     lambda m: "[path]" + m.group(0).split("\\")[-1]),
+    (re.compile(r"(?i)\b(?:\\\\[^\\\s]+\\[^\s]+)"), "[path]"),
+    (re.compile(r"(?:/Users/|/home/|~)[^\s\"':]+"), lambda m: "[path]" + m.group(0).rsplit("/", 1)[-1]),
+)
+
+
+def scrub_text(text, limit=DETAIL_LIMIT):
+    """自由文本 → 可安全落台账/上传的摘录：剥密钥、剥绝对路径、截断。
+
+    顺序有讲究：先剥密钥（可能出现在路径或赋值串里），再剥路径，最后截断。
+    任何输入（None/非字符串）都安全。
+    """
+    if not isinstance(text, str):
+        text = "" if text is None else str(text)
+    out = text
+    for pat, rep in _KEY_PATTERNS:
+        out = pat.sub(rep, out)
+    for pat, rep in _PATH_PATTERNS:
+        try:
+            out = pat.sub(rep, out)
+        except Exception:
+            pass
+    out = out.strip()
+    if len(out) > limit:
+        out = out[:limit] + "…"
+    return out
+
+
+def _coerce_str(v, limit):
+    return str(v or "")[:limit]
+
+
+def _os_tag():
+    import sys
+    return {"win32": "windows", "darwin": "macos"}.get(sys.platform, sys.platform)
+
+
+def _version():
+    try:
+        from . import selfupdate
+        return str(selfupdate.package_version() or "dev")
+    except Exception:
+        return "dev"
+
+
+def record(category="", reason="", detail="", provider="", model="", tool="",
+           role="", run_id="", task_id="", step=0, exit_code=None, source=""):
+    """追加一条错误记录。缺省字段留空；任何异常都吞掉（埋点不拖垮业务）。
+
+    category: step / run / launch / service（枚举放宽，未知值原样入库）
+    reason:   失败原因码（error_codes.ErrorCode 的字符串值优先，如 TIMEOUT）
+    detail:   失败摘录（内部先过 scrub_text，绝不存原文）
+    """
+    try:
+        rec = {
+            "id": "%s-%s" % (time.strftime("%Y-%m-%d %H:%M:%S"),
+                             uuid.uuid4().hex[:8]),
+            "ts": time.strftime("%Y-%m-%d %H:%M:%S"),
+            "day": time.strftime("%Y-%m-%d"),
+            "category": _coerce_str(category, 24) or "step",
+            "reason": _coerce_str(reason, 40) or "UNKNOWN",
+            "detail": scrub_text(detail),
+            "provider": _coerce_str(provider, 60),
+            "model": _coerce_str(model, 80),
+            "tool": _coerce_str(tool, 24),
+            "role": _coerce_str(role, 40),
+            "run_id": _coerce_str(run_id, 64),
+            "task_id": _coerce_str(task_id, 64),
+            "step": int(step or 0),
+            "exit_code": exit_code,
+            "app_version": _version()[:24],
+            "os": _os_tag(),
+        }
+        line = json.dumps(rec, ensure_ascii=False)
+        day = rec["day"]
+        with LOCK:
+            paths.ERRORS_DIR.mkdir(parents=True, exist_ok=True)
+            with open(paths.ERRORS_DIR / ("errors-%s.jsonl" % day[:7].replace("-", "")),
+                      "a", encoding="utf-8") as f:
+                f.write(line + "\n")
+    except Exception:
+        pass
+
+
+def iter_records(days=0):
+    """按时间范围读取台账（days=0 表示全部）。返回按写入顺序的记录列表。"""
+    out = []
+    try:
+        files = sorted(paths.ERRORS_DIR.glob("errors-*.jsonl")) if paths.ERRORS_DIR.is_dir() else []
+    except Exception:
+        return out
+    since_day = ""
+    if days:
+        import datetime
+        since_day = (datetime.date.today()
+                     - datetime.timedelta(days=int(days) - 1)).isoformat()
+    for p in files:
+        try:
+            for line in p.read_text(encoding="utf-8", errors="replace").splitlines():
+                line = line.strip()
+                if not line.startswith("{"):
+                    continue
+                try:
+                    r = json.loads(line)
+                except Exception:
+                    continue
+                if not isinstance(r, dict):
+                    continue
+                if since_day and str(r.get("day", "")) < since_day:
+                    continue
+                out.append(r)
+        except Exception:
+            continue
+    return out
+
+
+def pending_since(cursor, limit=200):
+    """取 cursor（上次上传到的记录 id）之后的记录，最多 limit 条。
+
+    id 以时间开头、uuid 尾巴保证唯一且字典序≈时间序；返回 (records, new_cursor)。
+    """
+    try:
+        recs = [r for r in iter_records(0) if str(r.get("id") or "") > str(cursor or "")]
+        recs.sort(key=lambda r: str(r.get("id") or ""))
+        batch = recs[:max(1, int(limit))]
+        new_cursor = str(batch[-1].get("id")) if batch else str(cursor or "")
+        return batch, new_cursor
+    except Exception:
+        return [], str(cursor or "")

package/app/core/flows.py

@@ -23,7 +23,7 @@ ENGINES = ("code", "review", "direct")
 # 引擎默认参数：自定义流程留空时的兜底
 ENGINE_DEFAULTS = {
     "review": {"manuscript": "output.md", "rubric": ["内容", "结构", "表达"],
-               "threshold": 7.0, "rounds": 2},
+               "threshold": 7.0, "rounds": 2, "best_of": 1},
     "code": {"verify_command": ""},
     # direct：无参数——目标+附件即全部输入，跑完即止
 }
@@ -118,7 +118,7 @@ _ID_RE = re.compile(r"^[a-z][a-z0-9_-]{0,31}$")
 # 自定义流程的 icon 在 upsert_flow 里直接落盘，不走 overrides。
 _EDITABLE = ("name", "goal_hint", "note", "manuscript", "rubric",
              "threshold", "rounds", "serial", "draft_prompt", "critique_prompt",
-             "verify_command")
+             "verify_command", "best_of")
 
 
 def _read():
@@ -279,6 +279,10 @@ def upsert_flow(payload):
             flow["rounds"] = max(1, min(5, int(payload.get("rounds") or dflt["rounds"])))
         except Exception:
             flow["rounds"] = dflt["rounds"]
+        try:
+            flow["best_of"] = max(1, min(3, int(payload.get("best_of") or dflt.get("best_of") or 1)))
+        except Exception:
+            flow["best_of"] = 1
         serial = _norm_serial(payload.get("serial"))
         if serial:
             flow["serial"] = serial

package/app/core/gitmod.py

@@ -33,6 +33,11 @@ def _git(workdir, *args, timeout=20):
     return r
 
 
+def _gh(workdir, *args, timeout=60):
+    """GitHub CLI 调用（pr_create 用）；与 _git 同构，缺失时由调用方给可读错误。"""
+    return runner.run_process(argv=["gh", *args], cwd=str(workdir), timeout=timeout)
+
+
 def _attach_entry(line):
     """status --porcelain 的一行是否为任务附件目录下的未跟踪项（不算脏改动）。"""
     if not line.startswith("?? "):
@@ -837,7 +842,7 @@ def file_diff(workdir, path):
 # 写操作白名单：action → 是否需要 confirm（不可恢复类）
 _WB_ACTIONS = {"checkout", "fetch", "pull", "push", "stage", "unstage",
                "discard", "delete", "commit", "stage_all", "unstage_all",
-               "stash_push", "stash_pop", "stash_drop"}
+               "stash_push", "stash_pop", "stash_drop", "pr_create"}
 
 
 def _res(r, extra=None):
@@ -938,6 +943,45 @@ def workbench_op(workdir, action, params=None):
             return _res(r)
         return _res(_git(wd, "reset", "-q", timeout=60))
 
+    if action == "pr_create":
+        # 「创建 PR」（借鉴 agent-orchestrator 的 planning→merge 闭环）：当前分支
+        # 推到 origin 后用 gh CLI 建 PR。gh 未安装/未登录给可读错误，不静默降级。
+        if info.get("detached") or not str(info.get("branch") or "").strip():
+            return False, "游离 HEAD 状态不能建 PR，请先切换到任务分支", {}
+        if not (info.get("remotes") or []):
+            return False, "该仓库没有配置远程（git remote），无法创建 PR", {}
+        if not _gh(wd, "--version", timeout=15)["ok"]:
+            return False, "未安装 gh CLI（https://cli.github.com/）：安装并 gh auth login 后可用", {}
+        br = info["branch"].strip()
+        base = str(params.get("base") or "").strip()
+        if not base:
+            bset = set(info.get("branches") or []) | set(info.get("remote_branches") or [])
+            for cand in ("main", "master"):
+                if cand in bset:
+                    base = cand
+                    break
+        if not base or base == br:
+            return False, "无法确定 PR 目标分支（base），请显式指定 base 参数", {}
+        title = str(params.get("title") or "").strip()[:120] or ("PR: %s" % br)
+        body_text = str(params.get("body") or "").strip()[:2000]
+        r_push = _git(wd, "push", "-u", "origin", br, timeout=180)
+        if not r_push["ok"]:
+            return _res(r_push)
+        argv = ["pr", "create", "--base", base, "--head", br, "--title", title]
+        if body_text:
+            argv += ["--body", body_text]
+        else:
+            argv += ["--fill"]   # 无描述时让 gh 用提交记录自动生成
+        r = _gh(wd, *argv, timeout=120)
+        ok, err, _d = _res(r)
+        if not ok:
+            low = (err or "").lower()
+            if "gh auth" in low or "authentic" in low or "log in" in low:
+                return False, "gh 未登录：请先在本机运行 gh auth login", {}
+            return False, (err or "gh pr create 失败")[:300], {}
+        lines = [ln.strip() for ln in (r.get("stdout") or "").splitlines() if ln.strip()]
+        return True, "", {"url": lines[-1] if lines else "", "branch": br, "base": base}
+
     if action == "commit":
         msg = str(params.get("message") or "").strip()[:500]
         if not msg:

package/app/core/health.py

@@ -59,6 +59,17 @@ def init(data_dir=None):
                     _PROVIDERS[name] = st
         except (OSError, json.JSONDecodeError):
             pass
+        # 「绑定链·」静态告警胶囊已从产品移除（2026-09-18：普通用户误伤，
+        # 用户拍板）——老版本持久化的条目启动即清，别让 0.1.6 的状态阴魂不散。
+        stale = [k for k in _PROVIDERS if k.startswith("绑定链·")]
+        if stale:
+            for k in stale:
+                _PROVIDERS.pop(k, None)
+            _persist()
+        # 供应商已删除的幽灵条目启动即清；运行期新增的幽灵由探针在探活时
+        # 发现（gone 标记）随手清——探针循环里不做全量扫描，避免和测试
+        # 夹具/运行期建条目赛跑。
+        _prune_missing()
         if not _STARTED:
             _STARTED = True
             t = threading.Thread(target=_probe_loop, name="health-probe", daemon=True)
@@ -159,56 +170,6 @@ def report_failure(provider: str, error: str = "", *, model: str = "",
 
 # ------------------------------------------------- 静态死链告警（绑定层）
 
-def report_binding_dead(key: str, error: str = ""):
-    """静态死链告警：某 CLI 的绑定链在起跑前就已全部失效（厂商停用/删除/无密钥、
-    模型停用、协议不匹配）。与运行期 report_failure 的区别：不计数、不进探针，
-    立即置 down+告警（尊重静默）；绑定恢复后由 report_binding_ok 自动解除。
-    伪供应商名带「绑定链·」前缀，不会与真实厂商名相撞。"""
-    if not key:
-        return
-    with _LOCK:
-        name = "绑定链·%s" % key
-        st = _PROVIDERS.get(name)
-        if st is None:
-            st = _PROVIDERS.setdefault(name, {
-                "name": name, "provider_id": "", "model": "",
-                "status": "ok", "consecutive_failures": 0,
-                "first_fail_at": 0, "last_fail_at": 0, "last_ok_at": 0,
-                "last_error": "", "alerted": False, "silenced": False,
-                "silence_until": 0, "probe_next_at": 0, "probe_backoff_idx": 0,
-                "recovered_at": 0,
-            })
-        st["static"] = True
-        st["binding_key"] = str(key)
-        st["name"] = name
-        st["status"] = "down"
-        st["last_fail_at"] = _now()
-        st["last_error"] = str(error or "")[:300]
-        st["probe_next_at"] = 0
-        if not _effective_silenced(st) and not st.get("alerted"):
-            st["alerted"] = True
-            log.warning("[health] ⚠️ 绑定链告警：%s 全部失效（%s）",
-                        name, st["last_error"][:120])
-        _persist()
-
-
-def report_binding_ok(key: str):
-    """绑定链恢复可用：自动解除该 CLI 的静态死链告警（无记录时零开销）。"""
-    if not key:
-        return
-    with _LOCK:
-        st = _PROVIDERS.get("绑定链·%s" % key)
-        if not st or not st.get("static"):
-            return
-        st["status"] = "recovered"
-        st["alerted"] = False
-        st["silenced"] = False
-        st["silence_until"] = 0
-        st["recovered_at"] = _now()
-        log.info("[health] 绑定链 %s 已恢复（静态告警解除）", st["name"])
-        _persist()
-
-
 # ---------------------------------------------------------------- 手动操作
 
 def silence(provider: str, minutes: int = 0):
@@ -294,16 +255,20 @@ def down_names() -> set:
 # ---------------------------------------------------------------- 探针
 
 def _probe_one(st):
-    """对单个故障 provider 发轻量探活。返回是否恢复。"""
+    """对单个故障 provider 发轻量探活。返回 (ok, gone)：
+    ok=是否恢复；gone=供应商已从配置里删除（健康记录应随之清掉）。"""
     try:
         from . import modelhub
         pid = st.get("provider_id") or ""
         model = st.get("model") or ""
-        provs = {p.get("name"): p for p in modelhub.providers()}
-        prov = provs.get(st.get("name")) or (modelhub.providers() and next(
-            (p for p in modelhub.providers() if p.get("id") == pid), None))
-        if not prov or not prov.get("api_key"):
-            return False  # 供应商被停用/删除：无法探测，保持状态
+        plist = modelhub.providers()
+        provs = {p.get("name"): p for p in plist}
+        prov = provs.get(st.get("name")) or next(
+            (p for p in plist if pid and p.get("id") == pid), None)
+        if not prov:
+            return False, True   # 供应商已删除：探不了也不再是「没恢复」
+        if not prov.get("api_key"):
+            return False, False  # 供应商在但没密钥：无法探测，保持状态
         if not model:
             try:
                 names = modelhub._enabled_models(prov)
@@ -311,12 +276,36 @@ def _probe_one(st):
             except Exception:
                 model = ""
         if not model:
-            return False
+            return False, False
         res = modelhub.chat(prov["id"], model, "1", max_tokens=8, timeout=20)
-        return bool(res.get("ok"))
+        return bool(res.get("ok")), False
     except Exception as e:
         log.debug("probe %s error: %s", st.get("name"), e)
-        return False
+        return False, False
+
+
+def _prune_missing():
+    """供应商被删除后清掉它的健康记录。
+
+    幽灵条目探不活也永远不会恢复，还会每轮把 last_fail_at 刷成「刚刚」，
+    健康页看起来像它一直在报错（Z.ai 删掉后仍显示 429 的误伤案，2026-09-18）。
+    """
+    try:
+        from . import modelhub
+        plist = modelhub.providers()
+    except Exception:
+        return
+    ids = {p.get("id") for p in plist}
+    names = {p.get("name") for p in plist}
+    with _LOCK:
+        stale = [n for n, st in _PROVIDERS.items()
+                 if not ((st.get("provider_id") and st.get("provider_id") in ids)
+                         or st.get("name") in names)]
+        if stale:
+            for n in stale:
+                _PROVIDERS.pop(n, None)
+            _persist()
+            log.info("[health] 清除已删除供应商的健康记录：%s", ", ".join(stale))
 
 
 def _probe_loop():
@@ -333,8 +322,13 @@ def _probe_loop():
                     st["probe_next_at"] = _now() + backoff
                     st["probe_backoff_idx"] = idx + 1
             for name, st in targets:
-                ok = _probe_one(st)
-                if ok:
+                ok, gone = _probe_one(st)
+                if gone:
+                    with _LOCK:
+                        if _PROVIDERS.pop(name, None) is not None:
+                            _persist()
+                    log.info("[health] %s 已删除，健康记录随之清除", name)
+                elif ok:
                     report_success(name)
                     log.info("[health] 探针确认 %s 已恢复", name)
                 else: