Merge pull request #275 from openqj/feat-resident-website-loop-audit

合并网站常驻监听 PR #275：resident-watches + pause/resume/end、闭环实测、审核架构 10/10 与代码质量 10/10。CI host-verify 通过（6m52s），mergeable CLEAN。
2026-11-02 04:18:29 +08:00
commit 60997b138f
@@ -0,0 +1,846 @@
+"""常驻监听任务（resident watch）。
+
+与现有三样东西的关系，先说清楚，避免把旧语义冒充成常驻：
+
+- 任务状态机（cloud/server.py）：queued -> running ->
+  completed / failed / interrupted / merged / merge-failed，
+  其中 /pause 只是把进行中回合任务标 interrupted 并置
+  paused_by_user=1（见 server._pause_task），不是一个可续跑
+  的“已暂停常驻态”。本模块绝不复用那条路：常驻态存在
+  resident_watches 表里，与任务表互不干扰。
+- 云端自动化（cloud/automations.py）：定义层只有 active /
+  cancelled，到期靠外部触发器调 run_due_automation 抢单建任务，
+  服务端本身不起定时器。本模块复用它的两道口径而不是另起
+  一套：抢单用条件 UPDATE（同一间隔内只有一个执行器能拿到
+  rowcount 1），配额门直接调 automations._automation_quota_
+  exceeded（与 server / worker 同口径，改口径时只改那一处）。
+- goal 循环（host.py run_goal_loop，max_rounds=0 不限轮次）：
+  目标是“做完一个目标”，结束条件是模型宣告达成；本模块的
+  目标是“一直监听外部变化”，结束条件只能是用户结束（或
+  配额耗尽转暂停，见下），两者不是一回事。
+
+状态机（本模块自有，与任务状态机不冲突）：
+    running -> paused -> running（恢复）
+    running / paused -> ended（终态）
+    ended -> 任何操作一律拒绝（恢复必须新建任务，见 end_watch）
+
+暂停保留现场：暂停只改 status / paused_reason / next_poll_at，
+last_content_hash、last_poll_at、last_change_at、last_change_text
+一律不动；暂停期间 due 为 False，驱动绝不轮询。恢复时
+next_poll_at 置为 now（立刻复查一次），用旧游标比对：暂停
+期间网站变了，恢复后第一次轮询就会把那次变化报出来，而不是
+丢掉——“从游标续跑”就是这个意思。
+
+结束是终态：ended 后 next_poll_at 为 NULL，驱动的 due 查询
+只认 status='running'，所以结束后绝不会静默重启；ended 行
+保留在库里供审计与查询，名字可重用（与自动化 cancelled 的
+名字重用规则同款）。
+
+变化才花钱：轮询本身是廉价抓取 + sha256 比对，不跑模型。
+只有三类轮询产出：baseline（首次抓取，只建游标，不上报）、
+unchanged（游标不变，不建任务）、changed（游标变了，才建
+一个 queued 任务，prompt = 用户的处理提示 + 变化摘录）。
+每次轮询都在 resident_polls 留一行，可查。
+
+配额边界：配额门只在“变化要建模型任务”那一刻过，不在
+轮询那一刻过（轮询不花模型额度）。配额耗尽时不结束、不丢
+变化：状态转 paused、paused_reason='quota_exceeded'，游标不
+推进（恢复后复查仍会报出同一变化），轮询记
+changed_quota_paused。配额检查本身失败（fail-closed 文本）
+时不转暂停，按瞬时失败退避处理，游标同样不推进。
+
+驱动：服务端本身不起定时器（与自动化同款结论）。真正的
+驱动是 cloud/worker.py 的 run_loop 轮次边界调
+process_due_resident_watches（见 worker 里的调用点），或
+任何外部触发器直接调本模块的 run_due_resident_watch /
+process_due_resident_watches；两者都走 claim 门，两个驱动
+同时跑也不会在同一间隔内重复轮询。
+
+本模块只依赖标准库 + cloud.automations 的复用函数
+（row_get、_audit_row、_automation_quota_exceeded）与
+cloud.db 的完整性错误类型（仅用于把唯一索引冲突转成
+ValueError，不 import 其它 cloud 模块），避免循环 import。
+"""
+
+import hashlib
+import ipaddress
+import re
+import sqlite3
+import time
+import urllib.request
+import uuid
+from urllib.parse import urljoin
+from urllib.parse import urlsplit
+
+from cloud.automations import _audit_row as _auto_audit_row
+from cloud.automations import _automation_quota_exceeded
+from cloud.automations import row_get
+from cloud.db import PostgresIntegrityError as _PgIntegrityError
+
+MIN_INTERVAL_SECONDS = 60
+MAX_INTERVAL_SECONDS = 30 * 86400
+MAX_NAME_LEN = 100
+MAX_PROMPT_LEN = 20000
+MAX_URL_LEN = 2000
+MIN_FETCH_TIMEOUT_SECONDS = 1
+MAX_FETCH_TIMEOUT_SECONDS = 60
+DEFAULT_FETCH_TIMEOUT_SECONDS = 10
+MAX_FETCH_BYTES = 1000000
+MAX_CHANGE_TEXT_LEN = 2000
+MAX_POLL_DETAIL_LEN = 500
+POLL_LIMIT = 50
+
+RUNNING = "running"
+PAUSED = "paused"
+ENDED = "ended"
+
+# 变化判定口径（v1）：对抓取文本做空白折叠后取 sha256。
+# 动态页面（广告、时间戳）本身会抖动，v1 不做选择器/区域
+# 提取，抖动风险写在验证摘要的已知偏差里，不假装已解决。
+_WS_RE = re.compile(r"\s+")
+
+
+WATCH_COLUMNS = (
+    "id", "project_id", "name", "target_url", "prompt",
+    "interval_seconds", "fetch_timeout_seconds", "status",
+    "last_poll_at", "last_claimed_at", "next_poll_at",
+    "last_content_hash", "last_change_at", "last_change_text",
+    "consecutive_failures", "paused_reason", "ended_at",
+    "created_by", "created_at", "updated_at")
+
+POLL_COLUMNS = (
+    "id", "watch_id", "project_id", "status", "observed_hash",
+    "detail", "task_id", "created_at")
+
+
+def _row_dict(row, columns):
+    if row is None or isinstance(row, dict):
+        return row
+    try:
+        row["id"]
+        return row
+    except (KeyError, IndexError, TypeError):
+        # sqlite3.Row 缺列报 IndexError（不是 TypeError）：只有
+        # SELECT * 的行能原样用，单列查询（如 claim 取 interval）
+        # 走 zip 回落，columns 必须与 SELECT 列顺序一致。
+        return dict(zip(columns, row))
+
+
+def _existing_columns(conn, table):
+    try:
+        return {row[1] for row in conn.execute(
+            "PRAGMA table_info(%s)" % table).fetchall()}
+    except Exception:
+        return set()
+
+
+def ensure_resident_tables(conn):
+    """幂等建常驻监听表，补齐新增列（仿 automations 写法）。"""
+    conn.execute(
+        "CREATE TABLE IF NOT EXISTS resident_watches("
+        "id TEXT PRIMARY KEY, "
+        "project_id TEXT, "
+        "name TEXT, "
+        "target_url TEXT, "
+        "prompt TEXT, "
+        "interval_seconds INTEGER, "
+        "fetch_timeout_seconds INTEGER, "
+        "status TEXT, "
+        "last_poll_at DOUBLE PRECISION, "
+        "last_claimed_at DOUBLE PRECISION, "
+        "next_poll_at DOUBLE PRECISION, "
+        "last_content_hash TEXT, "
+        "last_change_at DOUBLE PRECISION, "
+        "last_change_text TEXT, "
+        "consecutive_failures INTEGER, "
+        "paused_reason TEXT, "
+        "ended_at DOUBLE PRECISION, "
+        "created_by TEXT, "
+        "created_at DOUBLE PRECISION, "
+        "updated_at DOUBLE PRECISION)")
+    conn.execute(
+        "CREATE TABLE IF NOT EXISTS resident_polls("
+        "id TEXT PRIMARY KEY, "
+        "watch_id TEXT, "
+        "project_id TEXT, "
+        "status TEXT, "
+        "observed_hash TEXT, "
+        "detail TEXT, "
+        "task_id TEXT, "
+        "created_at DOUBLE PRECISION)")
+    existing = _existing_columns(conn, "resident_watches")
+    for col, ddl in (
+            ("fetch_timeout_seconds", "INTEGER"),
+            ("next_poll_at", "DOUBLE PRECISION"),
+            ("last_content_hash", "TEXT"),
+            ("last_change_at", "DOUBLE PRECISION"),
+            ("last_change_text", "TEXT"),
+            ("consecutive_failures", "INTEGER"),
+            ("paused_reason", "TEXT"),
+            ("ended_at", "DOUBLE PRECISION")):
+        if col not in existing:
+            try:
+                conn.execute(
+                    "ALTER TABLE resident_watches ADD COLUMN %s %s"
+                    % (col, ddl))
+            except Exception:
+                pass
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_resident_watches_project"
+        " ON resident_watches(project_id, status)")
+    # 同项目未结束名唯一：先查后插挡不住并发，唯一索引才是
+    # 真正的关口（server environments 的同款部分唯一索引，
+    # WHERE status != 'archived' 的 resident 版）；已结束名
+    # 不在索引内，可重用。已存在重复名的旧库建索引会失败，
+    # 此时不炸掉整次建表调用，SELECT 查重仍挡正常路径。
+    try:
+        conn.execute(
+            "CREATE UNIQUE INDEX IF NOT EXISTS"
+            " idx_resident_watches_name"
+            " ON resident_watches(project_id, name)"
+            " WHERE status != 'ended'")
+    except Exception:
+        pass
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_resident_polls_watch"
+        " ON resident_polls(watch_id, created_at)")
+
+
+def _backoff_seconds(failures):
+    """失败退避增量：30、60、120……秒，上限 3600 秒。
+    调用方把它加在间隔之上（下次 = now + interval + backoff），
+    失败后绝不比正常间隔更频繁地重试，行为见失败分支测试。"""
+    if failures <= 0:
+        return 0
+    return min(3600, 30 * (2 ** (failures - 1)))
+
+
+def watch_to_dict(row, now=None):
+    row = _row_dict(row, WATCH_COLUMNS)
+    if now is None:
+        now = time.time()
+    status = row["status"]
+    nxt = row_get(row, "next_poll_at")
+    due = bool(status == RUNNING and (nxt is None or nxt <= now))
+    return {
+        "id": row["id"],
+        "project_id": row["project_id"],
+        "name": row["name"],
+        "target_url": row["target_url"],
+        "prompt": row["prompt"],
+        "interval_seconds": row["interval_seconds"],
+        "fetch_timeout_seconds": row_get(row, "fetch_timeout_seconds")
+        or DEFAULT_FETCH_TIMEOUT_SECONDS,
+        "status": status,
+        "last_poll_at": row_get(row, "last_poll_at"),
+        "last_claimed_at": row_get(row, "last_claimed_at"),
+        "next_poll_at": nxt,
+        "last_content_hash": row_get(row, "last_content_hash"),
+        "last_change_at": row_get(row, "last_change_at"),
+        "last_change_text": row_get(row, "last_change_text"),
+        "consecutive_failures": row_get(row, "consecutive_failures") or 0,
+        "paused_reason": row_get(row, "paused_reason"),
+        "ended_at": row_get(row, "ended_at"),
+        "created_by": row["created_by"],
+        "created_at": row["created_at"],
+        "updated_at": row_get(row, "updated_at"),
+        "due": due,
+    }
+
+
+def _is_noncanonical_numeric_host(host):
+    """非规范数字 IP 写法：单整数（2130706433）、十六进制
+    （0x7f000001）、前导零八进制（0177.0.0.1）、含十六进制段
+    的点分写法——urllib/socket 的 inet_aton 会把它们解析成
+    数字 IP，ipaddress 却认不出来。规范点分十进制（4 段、
+    每段 0-255、无前导零，0 本身除外）不在此列；普通域名
+    （含字母的标签）也不在此列。"""
+    if re.fullmatch(r"0[xX][0-9a-fA-F]+", host):
+        return True
+    if re.fullmatch(r"[0-9]+", host):
+        return True
+    if "." not in host:
+        return False
+    labels = host.split(".")
+    if not all(re.fullmatch(r"0[xX][0-9a-fA-F]+|[0-9]+", lab)
+               for lab in labels):
+        return False
+    if len(labels) != 4:
+        return True
+    for lab in labels:
+        if lab.startswith("0x") or lab.startswith("0X"):
+            return True
+        if len(lab) > 1 and lab.startswith("0"):
+            return True
+        if int(lab) > 255:
+            return True
+    return False
+
+
+def validate_target_url(url):
+    """目标 URL 校验：仅 http/https、无 userinfo、主机非空、
+    字面 localhost（含尾点、*.localhost）/ 私有 / 回环 /
+    链路本地 IP / 非规范数字 IP 写法拒绝（与
+    websearch._is_blocked_fetch_url 的字面 IP 口径同款，
+    DNS 解析后的私有地址不在 v1 覆盖内，见摘要已知偏差；
+    重定向目标由 _SafeRedirectHandler 逐跳复用本校验）。"""
+    if not url or not isinstance(url, str) or not url.strip():
+        raise ValueError("target_url required")
+    url = url.strip()
+    if len(url) > MAX_URL_LEN:
+        raise ValueError("target_url too long")
+    parts = urlsplit(url)
+    if parts.scheme not in ("http", "https"):
+        raise ValueError("target_url must be http or https")
+    if parts.username or parts.password:
+        raise ValueError("target_url must not contain credentials")
+    host = (parts.hostname or "").strip().lower().rstrip(".")
+    if not host:
+        raise ValueError("target_url host required")
+    if host == "localhost" or host.endswith(".localhost"):
+        raise ValueError("localhost URLs are blocked")
+    if _is_noncanonical_numeric_host(host):
+        raise ValueError("non-canonical numeric IP URLs are blocked")
+    try:
+        addr = ipaddress.ip_address(host)
+    except ValueError:
+        addr = None
+    if addr is not None and (addr.is_private or addr.is_loopback
+                             or addr.is_link_local or addr.is_reserved
+                             or addr.is_multicast):
+        raise ValueError("private / loopback URLs are blocked")
+    return url
+
+
+def _validate_fields(name, prompt, target_url, interval_seconds,
+                     fetch_timeout_seconds):
+    if not name or not isinstance(name, str) or not name.strip():
+        raise ValueError("name required")
+    name = name.strip()
+    if len(name) > MAX_NAME_LEN:
+        raise ValueError("name too long")
+    if not prompt or not isinstance(prompt, str) or not prompt.strip():
+        raise ValueError("prompt required")
+    prompt = prompt.strip()
+    if len(prompt) > MAX_PROMPT_LEN:
+        raise ValueError("prompt too long")
+    target_url = validate_target_url(target_url)
+    if isinstance(interval_seconds, bool) \
+            or not isinstance(interval_seconds, int):
+        raise ValueError("interval_seconds must be an integer")
+    if interval_seconds < MIN_INTERVAL_SECONDS:
+        raise ValueError(
+            "interval_seconds must be at least %d" % MIN_INTERVAL_SECONDS)
+    if interval_seconds > MAX_INTERVAL_SECONDS:
+        raise ValueError("interval_seconds too large")
+    if fetch_timeout_seconds is None:
+        fetch_timeout_seconds = DEFAULT_FETCH_TIMEOUT_SECONDS
+    if isinstance(fetch_timeout_seconds, bool) \
+            or not isinstance(fetch_timeout_seconds, int):
+        raise ValueError("fetch_timeout_seconds must be an integer")
+    if not MIN_FETCH_TIMEOUT_SECONDS <= fetch_timeout_seconds \
+            <= MAX_FETCH_TIMEOUT_SECONDS:
+        raise ValueError(
+            "fetch_timeout_seconds must be between %d and %d"
+            % (MIN_FETCH_TIMEOUT_SECONDS, MAX_FETCH_TIMEOUT_SECONDS))
+    return name, prompt, target_url, interval_seconds, fetch_timeout_seconds
+
+
+def create_watch(conn, project_id, name, target_url, prompt, created_by,
+                 interval_seconds=3600, fetch_timeout_seconds=None,
+                 now=None):
+    """新建常驻监听：建完即到期（next_poll_at=now），驱动下
+    一轮即可认领；同项目内未结束（running/paused）的名字唯一，
+    已结束的名字可重用。创建本身不走配额门（轮询不花模型额度，
+    配额只在变化建任务时过，见 run_due_resident_watch）。"""
+    (name, prompt, target_url, interval_seconds,
+     fetch_timeout_seconds) = _validate_fields(
+        name, prompt, target_url, interval_seconds, fetch_timeout_seconds)
+    if now is None:
+        now = time.time()
+    ensure_resident_tables(conn)
+    clash = conn.execute(
+        "SELECT id FROM resident_watches WHERE project_id = ?"
+        " AND name = ? AND status != ?",
+        (project_id, name, ENDED)).fetchone()
+    if clash is not None:
+        raise ValueError(
+            "resident watch name already exists in this project")
+    watch_id = uuid.uuid4().hex
+    try:
+        conn.execute(
+            "INSERT INTO resident_watches(id, project_id, name, target_url,"
+            " prompt, interval_seconds, fetch_timeout_seconds, status,"
+            " last_poll_at, last_claimed_at, next_poll_at,"
+            " last_content_hash, last_change_at, last_change_text,"
+            " consecutive_failures, paused_reason, ended_at,"
+            " created_by, created_at, updated_at)"
+            " VALUES(?, ?, ?, ?, ?, ?, ?, ?, NULL, NULL, ?,"
+            " NULL, NULL, NULL, 0, NULL, NULL, ?, ?, ?)",
+            (watch_id, project_id, name, target_url, prompt,
+             interval_seconds, fetch_timeout_seconds, RUNNING, now,
+             created_by, now, now))
+    except (sqlite3.IntegrityError, _PgIntegrityError) as exc:
+        # 并发下 SELECT 查重与 INSERT 之间可能插进同名行，
+        # 唯一索引是真正的关口：冲突转回与查重同款 ValueError，
+        # 调用方（server）据此返回 400，不把数据库错误外泄。
+        try:
+            conn.rollback()
+        except Exception:
+            pass
+        raise ValueError(
+            "resident watch name already exists in this project") from exc
+    conn.commit()
+    row = conn.execute(
+        "SELECT * FROM resident_watches WHERE id = ?", (watch_id,)).fetchone()
+    return watch_to_dict(row, now)
+
+
+def list_watches(conn, project_id, now=None):
+    if now is None:
+        now = time.time()
+    ensure_resident_tables(conn)
+    rows = conn.execute(
+        "SELECT * FROM resident_watches WHERE project_id = ?"
+        " ORDER BY created_at, id",
+        (project_id,)).fetchall()
+    return [watch_to_dict(r, now) for r in rows]
+
+
+def get_watch(conn, watch_id, now=None):
+    if now is None:
+        now = time.time()
+    ensure_resident_tables(conn)
+    row = conn.execute(
+        "SELECT * FROM resident_watches WHERE id = ?", (watch_id,)).fetchone()
+    if row is None:
+        return None
+    return watch_to_dict(row, now)
+
+
+def list_polls(conn, watch_id, limit=POLL_LIMIT):
+    ensure_resident_tables(conn)
+    rows = conn.execute(
+        "SELECT * FROM resident_polls WHERE watch_id = ?"
+        " ORDER BY created_at DESC, id LIMIT ?",
+        (watch_id, limit)).fetchall()
+    out = []
+    for raw in rows:
+        r = _row_dict(raw, POLL_COLUMNS)
+        out.append({
+            "id": r["id"],
+            "watch_id": r["watch_id"],
+            "project_id": r["project_id"],
+            "status": r["status"],
+            "observed_hash": row_get(r, "observed_hash"),
+            "detail": row_get(r, "detail") or "",
+            "task_id": row_get(r, "task_id"),
+            "created_at": r["created_at"],
+        })
+    return out
+
+
+def _transition(conn, watch_id, from_statuses, to_status, now,
+                extra_sets=(), extra_vals=()):
+    """条件状态迁移：返回 (ok, dict)。行不存在时 (False, None)；
+    状态不对时 (False, 旧 dict)，调用方据此区分 404 与 409。"""
+    ensure_resident_tables(conn)
+    row = conn.execute(
+        "SELECT * FROM resident_watches WHERE id = ?",
+        (watch_id,)).fetchone()
+    if row is None:
+        return False, None
+    row = _row_dict(row, WATCH_COLUMNS)
+    if row["status"] not in from_statuses:
+        return False, watch_to_dict(row, now)
+    placeholders = ",".join("?" for _ in from_statuses)
+    sets = ["status = ?", "updated_at = ?"] + list(extra_sets)
+    vals = [to_status, now] + list(extra_vals) + [watch_id] \
+        + list(from_statuses)
+    cur = conn.execute(
+        "UPDATE resident_watches SET %s WHERE id = ? AND status IN (%s)"
+        % (", ".join(sets), placeholders), tuple(vals))
+    conn.commit()
+    if cur.rowcount != 1:
+        return False, watch_to_dict(row, now)
+    row = conn.execute(
+        "SELECT * FROM resident_watches WHERE id = ?", (watch_id,)).fetchone()
+    return True, watch_to_dict(row, now)
+
+
+def pause_watch(conn, watch_id, now=None, reason="user_paused"):
+    """用户暂停：仅 running 能暂停；保留全部现场，只把
+    next_poll_at 清掉，驱动即停。"""
+    if now is None:
+        now = time.time()
+    return _transition(conn, watch_id, (RUNNING,), PAUSED, now,
+                       extra_sets=("paused_reason = ?",
+                                   "next_poll_at = NULL"),
+                       extra_vals=(reason,))
+
+
+def resume_watch(conn, watch_id, now=None):
+    """恢复：仅 paused 能恢复；现场不动，next_poll_at=now
+    立刻复查一次（用旧游标比对，暂停期间的变化不丢）。"""
+    if now is None:
+        now = time.time()
+    return _transition(conn, watch_id, (PAUSED,), RUNNING, now,
+                       extra_sets=("paused_reason = NULL",
+                                   "next_poll_at = ?"),
+                       extra_vals=(now,))
+
+
+def end_watch(conn, watch_id, now=None):
+    """结束：running / paused -> ended，终态。ended 后
+    pause / resume / end 一律 (False, 旧 dict)，调用方转 409；
+    恢复已结束的监听必须新建任务，不复活本行。"""
+    if now is None:
+        now = time.time()
+    return _transition(conn, watch_id, (RUNNING, PAUSED), ENDED, now,
+                       extra_sets=("ended_at = ?",
+                                   "next_poll_at = NULL"),
+                       extra_vals=(now,))
+
+
+def update_watch(conn, watch_id, fields, now=None):
+    """更新定义：只认 name / prompt / target_url /
+    interval_seconds / fetch_timeout_seconds；ended 的行 409
+    （终态不能改，改它必须重建）；改间隔不重置游标、
+    不立即触发：已排好的 next_poll_at 不变，下一次轮询
+    完成后才按新间隔重算（想立即复查请暂停后恢复）。"""
+    if now is None:
+        now = time.time()
+    ensure_resident_tables(conn)
+    row = conn.execute(
+        "SELECT * FROM resident_watches WHERE id = ?", (watch_id,)).fetchone()
+    if row is None:
+        return False, None
+    row = _row_dict(row, WATCH_COLUMNS)
+    if row["status"] == ENDED:
+        return False, watch_to_dict(row, now)
+    if not isinstance(fields, dict):
+        raise ValueError("fields must be an object")
+    if not any(k in fields for k in (
+            "name", "prompt", "target_url", "interval_seconds",
+            "fetch_timeout_seconds")):
+        raise ValueError("no updatable field given")
+    current = watch_to_dict(row, now)
+    (name, prompt, target_url, interval, timeout) = _validate_fields(
+        fields.get("name", current["name"]),
+        fields.get("prompt", current["prompt"]),
+        fields.get("target_url", current["target_url"]),
+        fields.get("interval_seconds", current["interval_seconds"]),
+        fields.get("fetch_timeout_seconds",
+                   current["fetch_timeout_seconds"]))
+    if name != current["name"]:
+        clash = conn.execute(
+            "SELECT id FROM resident_watches WHERE project_id = ?"
+            " AND name = ? AND status != ? AND id != ?",
+            (row["project_id"], name, ENDED, watch_id)).fetchone()
+        if clash is not None:
+            raise ValueError(
+                "resident watch name already exists in this project")
+    try:
+        conn.execute(
+            "UPDATE resident_watches SET name = ?, prompt = ?,"
+            " target_url = ?, interval_seconds = ?,"
+            " fetch_timeout_seconds = ?, updated_at = ? WHERE id = ?",
+            (name, prompt, target_url, interval, timeout, now, watch_id))
+    except (sqlite3.IntegrityError, _PgIntegrityError) as exc:
+        try:
+            conn.rollback()
+        except Exception:
+            pass
+        raise ValueError(
+            "resident watch name already exists in this project") from exc
+    conn.commit()
+    row = conn.execute(
+        "SELECT * FROM resident_watches WHERE id = ?", (watch_id,)).fetchone()
+    return True, watch_to_dict(row, now)
+
+
+def due_watches(conn, now=None):
+    """全部项目当前到期的 running 监听（纯读，不抢单），
+    供驱动列出待处理项；执行入口是 claim + run_due。"""
+    if now is None:
+        now = time.time()
+    ensure_resident_tables(conn)
+    rows = conn.execute(
+        "SELECT * FROM resident_watches WHERE status = ?"
+        " AND (next_poll_at IS NULL OR next_poll_at <= ?)"
+        " ORDER BY next_poll_at, created_at, id",
+        (RUNNING, now)).fetchall()
+    return [watch_to_dict(r, now) for r in rows]
+
+
+def claim_due_watch(conn, watch_id, now=None):
+    """抢单：条件 UPDATE 只有一行能成功。抢到即把
+    last_poll_at / last_claimed_at 写成 now，并把 next_poll_at
+    预置为 now + interval（崩在抓取中途时，下个间隔仍会重试，
+    不会因认领锁死而永久停摆；抓取完成后 run_due 会按成败
+    重算 next_poll_at）。抢不到时返回 False，调用方必须放弃。"""
+    if now is None:
+        now = time.time()
+    ensure_resident_tables(conn)
+    row = conn.execute(
+        "SELECT interval_seconds FROM resident_watches WHERE id = ?",
+        (watch_id,)).fetchone()
+    if row is None:
+        return False
+    interval = row_get(_row_dict(row, ("interval_seconds",)),
+                       "interval_seconds") or MIN_INTERVAL_SECONDS
+    cur = conn.execute(
+        "UPDATE resident_watches SET last_poll_at = ?,"
+        " last_claimed_at = ?, next_poll_at = ?"
+        " WHERE id = ? AND status = ?"
+        " AND (next_poll_at IS NULL OR next_poll_at <= ?)",
+        (now, now, now + interval, watch_id, RUNNING, now))
+    conn.commit()
+    return cur.rowcount == 1
+
+
+def _content_hash(text):
+    normalized = _WS_RE.sub(" ", text or "").strip()
+    return hashlib.sha256(normalized.encode("utf-8")).hexdigest()
+
+
+class _SafeRedirectHandler(urllib.request.HTTPRedirectHandler):
+    """重定向逐跳校验：创建时只校验了初始 URL，公开 URL
+    重定向到内网地址会绕过那道校验；相对跳转先 urljoin 成
+    绝对 URL 再校验，不合格直接抛 ValueError，不跟随。"""
+
+    def redirect_request(self, req, fp, code, msg, headers, newurl):
+        validate_target_url(urljoin(req.full_url, newurl))
+        return super().redirect_request(req, fp, code, msg, headers,
+                                         newurl)
+
+
+def default_fetch(url, timeout):
+    """默认廉价抓取：urllib，超时可配，读满 1MB 即断，
+    解码失败按 replace 处理。调用方可注入 fetcher 替换它
+    （测试与未来选择器提取都走注入点，不改本函数签名）。
+    build_opener 带默认处理器（含 ProxyHandler，no_proxy
+    行为与 urlopen 同款），只把重定向处理器换成逐跳校验版。"""
+    req = urllib.request.Request(url, headers={
+        "User-Agent": "codex-app-server-resident-watch/1.0"})
+    opener = urllib.request.build_opener(_SafeRedirectHandler())
+    with opener.open(req, timeout=timeout) as resp:
+        raw = resp.read(MAX_FETCH_BYTES + 1)[:MAX_FETCH_BYTES]
+    return raw.decode("utf-8", errors="replace")
+
+
+def _record_poll(conn, watch_id, project_id, status, observed_hash,
+                 detail, task_id, now):
+    poll_id = uuid.uuid4().hex
+    conn.execute(
+        "INSERT INTO resident_polls(id, watch_id, project_id, status,"
+        " observed_hash, detail, task_id, created_at)"
+        " VALUES(?, ?, ?, ?, ?, ?, ?, ?)",
+        (poll_id, watch_id, project_id, status, observed_hash,
+         (detail or "")[:MAX_POLL_DETAIL_LEN], task_id, now))
+    return {
+        "id": poll_id,
+        "watch_id": watch_id,
+        "project_id": project_id,
+        "status": status,
+        "observed_hash": observed_hash,
+        "detail": (detail or "")[:MAX_POLL_DETAIL_LEN],
+        "task_id": task_id,
+        "created_at": now,
+    }
+
+
+def _create_change_task(conn, watch, excerpt, now):
+    """变化建任务：列补齐口径与 automations.run_due_automation
+    同款；任务 prompt = 用户处理提示 + 变化摘录。"""
+    existing_tasks = _existing_columns(conn, "tasks")
+    for col, ddl in (
+            ("executor", "TEXT"),
+            ("model", "TEXT"),
+            ("effort", "TEXT"),
+            ("environment_id", "TEXT"),
+            ("parent_task_id", "TEXT"),
+            ("coordinator_part", "TEXT")):
+        if col not in existing_tasks:
+            try:
+                conn.execute(
+                    "ALTER TABLE tasks ADD COLUMN %s %s" % (col, ddl))
+            except Exception:
+                pass
+    task_id = uuid.uuid4().hex
+    prompt = ("%s\n\n[常驻监听变化] %s %s\n变化摘录：\n%s"
+              % (watch["prompt"], watch["name"], watch["target_url"],
+                 excerpt))
+    conn.execute(
+        "INSERT INTO tasks(id, project_id, prompt, status,"
+        " idempotency_key, depends_on, priority, created_by,"
+        " created_at, updated_at, last_heartbeat, executor,"
+        " model, effort, environment_id, parent_task_id,"
+        " coordinator_part)"
+        " VALUES(?, ?, ?, 'queued', NULL, '[]', 0, ?, ?, ?,"
+        " NULL, NULL, NULL, NULL, NULL, NULL, NULL)",
+        (task_id, watch["project_id"], prompt, watch["created_by"],
+         now, now))
+    return task_id
+
+
+def _guarded_watch_update(conn, watch_id, sets, vals):
+    """终局 UPDATE 带 status='running' 条件：抓取中途被用户
+    暂停/结束时，条件不满足、rowcount 为 0，调用方据此回滚
+    并放弃本轮（不建任务、不记轮询），而不是把 paused/ended
+    行的 next_poll_at 改回未来值。返回是否成功。"""
+    cur = conn.execute(
+        "UPDATE resident_watches SET %s WHERE id = ? AND status = ?"
+        % ", ".join(sets),
+        tuple(vals) + (watch_id, RUNNING))
+    return cur.rowcount == 1
+
+
+def run_due_resident_watch(conn, watch_id, fetcher=None, now=None):
+    """单条常驻监听的驱动唯一入口：抢单 -> 廉价抓取 -> hash
+    比对 -> 记轮询 -> 仅变化时建任务。
+
+    抢不到返回 None（不存在、已暂停/已结束、未到期、已被
+    别的驱动抢走），不记轮询、不建任务。抓取失败记 failed
+    并退避，游标不动。首次抓取记 baseline，只建游标。未变
+    化记 unchanged，不建任务。变化时先过配额门：配额耗尽
+    转 paused（paused_reason='quota_exceeded'，游标不推进），
+    记 changed_quota_paused；配额检查失败按瞬时失败退避，
+    游标同样不推进；配额内才建 queued 任务并推进游标。"""
+    if now is None:
+        now = time.time()
+    ensure_resident_tables(conn)
+    if not claim_due_watch(conn, watch_id, now):
+        return None
+    row = conn.execute(
+        "SELECT * FROM resident_watches WHERE id = ?", (watch_id,)).fetchone()
+    if row is None:
+        return None
+    watch = watch_to_dict(row, now)
+    if watch["status"] != RUNNING:
+        # 认领后、抓取前被用户暂停/结束：不抓取、不建任务、
+        # 不改行（暂停/结束已把 next_poll_at 清掉，保持原样）。
+        return None
+    interval = watch["interval_seconds"]
+    timeout = watch["fetch_timeout_seconds"]
+    fetch = fetcher or default_fetch
+    try:
+        text = fetch(watch["target_url"], timeout)
+    except Exception as exc:
+        failures = watch["consecutive_failures"] + 1
+        nxt = now + interval + _backoff_seconds(failures)
+        if not _guarded_watch_update(
+                conn, watch_id,
+                ("consecutive_failures = ?", "next_poll_at = ?",
+                 "updated_at = ?"),
+                (failures, nxt, now)):
+            conn.rollback()
+            return None
+        poll = _record_poll(conn, watch_id, watch["project_id"],
+                            "failed", None, str(exc)[:MAX_POLL_DETAIL_LEN],
+                            None, now)
+        conn.commit()
+        _auto_audit_row(conn, watch["project_id"],
+                        "resident_watch_poll_failed",
+                        "%s: %s" % (watch_id, str(exc)[:200]), now)
+        return poll
+    if not isinstance(text, str):
+        text = str(text)
+    observed = _content_hash(text)
+    excerpt = text[:MAX_CHANGE_TEXT_LEN]
+    old_hash = watch["last_content_hash"]
+    if old_hash is None:
+        if not _guarded_watch_update(
+                conn, watch_id,
+                ("last_content_hash = ?", "consecutive_failures = 0",
+                 "next_poll_at = ?", "updated_at = ?"),
+                (observed, now + interval, now)):
+            conn.rollback()
+            return None
+        poll = _record_poll(conn, watch_id, watch["project_id"],
+                            "baseline", observed, "first poll", None, now)
+        conn.commit()
+        return poll
+    if observed == old_hash:
+        if not _guarded_watch_update(
+                conn, watch_id,
+                ("consecutive_failures = 0", "next_poll_at = ?",
+                 "updated_at = ?"),
+                (now + interval, now)):
+            conn.rollback()
+            return None
+        poll = _record_poll(conn, watch_id, watch["project_id"],
+                            "unchanged", observed, "", None, now)
+        conn.commit()
+        return poll
+    # 真实变化：配额门只在这里过。配额门契约是前缀
+    # （quota_check_failed: / quota_exceeded: / None），不用
+    # 子串判定，以免超限文本含该子串时误判。
+    gate = _automation_quota_exceeded(conn, watch["created_by"])
+    if gate is not None and not gate.startswith("quota_check_failed"):
+        cur = conn.execute(
+            "UPDATE resident_watches SET status = ?,"
+            " paused_reason = ?, next_poll_at = NULL, updated_at = ?"
+            " WHERE id = ? AND status = ?",
+            (PAUSED, "quota_exceeded", now, watch_id, RUNNING))
+        if cur.rowcount != 1:
+            conn.rollback()
+            return None
+        poll = _record_poll(conn, watch_id, watch["project_id"],
+                            "changed_quota_paused", observed, gate,
+                            None, now)
+        conn.commit()
+        _auto_audit_row(conn, watch["project_id"],
+                        "resident_watch_paused_quota",
+                        "%s: %s" % (watch_id, gate[:200]), now)
+        return poll
+    if gate is not None:
+        failures = watch["consecutive_failures"] + 1
+        nxt = now + interval + _backoff_seconds(failures)
+        if not _guarded_watch_update(
+                conn, watch_id,
+                ("consecutive_failures = ?", "next_poll_at = ?",
+                 "updated_at = ?"),
+                (failures, nxt, now)):
+            conn.rollback()
+            return None
+        poll = _record_poll(conn, watch_id, watch["project_id"],
+                            "changed_quota_check_failed", observed, gate,
+                            None, now)
+        conn.commit()
+        return poll
+    # 先受保护地推进游标，再建任务：抓取中途被暂停/结束时，
+    # 条件不满足即回滚，本轮不建任务、不记轮询。
+    if not _guarded_watch_update(
+            conn, watch_id,
+            ("last_content_hash = ?", "last_change_at = ?",
+             "last_change_text = ?", "consecutive_failures = 0",
+             "next_poll_at = ?", "updated_at = ?"),
+            (observed, now, excerpt, now + interval, now)):
+        conn.rollback()
+        return None
+    task_id = _create_change_task(conn, watch, excerpt, now)
+    poll = _record_poll(conn, watch_id, watch["project_id"],
+                        "changed", observed, excerpt[:MAX_POLL_DETAIL_LEN],
+                        task_id, now)
+    conn.commit()
+    _auto_audit_row(conn, watch["project_id"],
+                    "resident_watch_change_task_created",
+                    "%s -> %s" % (watch_id, task_id), now)
+    return poll
+
+
+def process_due_resident_watches(conn, fetcher=None, now=None, limit=50):
+    """驱动批量入口：列出到期监听，逐条走 run_due（每条自带
+    claim 门，批量内重复与并发驱动都不会重复轮询）。返回
+    本轮实际产出的轮询记录列表（抢不到的条目不在内）。"""
+    if now is None:
+        now = time.time()
+    due = due_watches(conn, now)[:limit]
+    out = []
+    for watch in due:
+        poll = run_due_resident_watch(conn, watch["id"],
+                                      fetcher=fetcher, now=now)
+        if poll is not None:
+            out.append(poll)
+    return out
@@ -110,6 +110,14 @@ from cloud.wait_ops import worktree_status as _wait_worktree_status
 from cloud.shared_home_mcp import sync_shared_home_mcp as _sync_shared_home_mcp
 from cloud.automations import ensure_automations_tables as _auto_ensure
 from cloud.automations import list_automations as _auto_list
+from cloud.resident import create_watch as _res_create
+from cloud.resident import end_watch as _res_end
+from cloud.resident import get_watch as _res_get
+from cloud.resident import list_polls as _res_polls
+from cloud.resident import list_watches as _res_list
+from cloud.resident import pause_watch as _res_pause
+from cloud.resident import resume_watch as _res_resume
+from cloud.resident import update_watch as _res_update
 from cloud.todos import create_todo as _todo_create
 from cloud.todos import delete_todo as _todo_delete
 from cloud.todos import get_todo as _todo_get
@@ -3371,6 +3379,54 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
                return
            self._triage_inbox(project_id)
            return
+        # /projects/{id}/resident-watches：常驻监听任务。
+        # 创建/更新/暂停/恢复/结束用 create_task，读用 read；
+        # 状态机见 cloud/resident.py（running/paused/ended，
+        # ended 是终态，恢复已结束的监听必须新建）。
+        if len(parts) == 2 and parts[1] == "resident-watches":
+            if method == "POST":
+                err = self._check_project(user_id, project_id, "create_task")
+                if err:
+                    self._send_json(err[0], {"error": err[1]})
+                    return
+                self._resident_create(user_id, project_id)
+            elif method == "GET":
+                err = self._check_project(user_id, project_id, "read")
+                if err:
+                    self._send_json(err[0], {"error": err[1]})
+                    return
+                self._resident_list(project_id)
+            else:
+                self._send_json(404, {"error": "not found"})
+            return
+        if len(parts) == 3 and parts[1] == "resident-watches":
+            if method == "GET":
+                err = self._check_project(user_id, project_id, "read")
+                if err:
+                    self._send_json(err[0], {"error": err[1]})
+                    return
+                self._resident_get(user_id, project_id, parts[2])
+            elif method == "PATCH":
+                err = self._check_project(user_id, project_id, "create_task")
+                if err:
+                    self._send_json(err[0], {"error": err[1]})
+                    return
+                self._resident_update(user_id, project_id, parts[2])
+            else:
+                self._send_json(404, {"error": "not found"})
+            return
+        if len(parts) == 4 and parts[1] == "resident-watches" \
+                and parts[3] in ("pause", "resume", "end"):
+            if method != "POST":
+                self._send_json(404, {"error": "not found"})
+                return
+            err = self._check_project(user_id, project_id, "create_task")
+            if err:
+                self._send_json(err[0], {"error": err[1]})
+                return
+            self._resident_transition(user_id, project_id, parts[2],
+                                      parts[3])
+            return
        # /projects/{id}/todos  GET 列表 / POST 新建；单条
        # GET / PATCH / DELETE（项目 read / create_task）
        if len(parts) == 2 and parts[1] == "todos":
@@ -10067,6 +10123,135 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
                        automation_id)
        self._send_json(200, auto)

+    # -- 常驻监听任务（见 cloud/resident.py）-------------------
+    # 权限口径与自动化同款：创建/更新/暂停/恢复/结束用
+    # create_task，读用 read；状态变更记审计。创建不走配额门
+    # （轮询不花模型额度），配额只在变化建任务时由驱动过，
+    # 耗尽时转暂停而非拒绝创建。
+    def _resident_create(self, user_id, project_id):
+        body, err = self._read_json()
+        if err:
+            self._send_json(400, {"error": err})
+            return
+        interval = body.get("interval_seconds", 3600)
+        if interval is None:
+            interval = 3600
+        conn = self._db()
+        try:
+            try:
+                watch = _res_create(
+                    conn, project_id, body.get("name"),
+                    body.get("target_url"), body.get("prompt"), user_id,
+                    interval, body.get("fetch_timeout_seconds"))
+            except ValueError as exc:
+                self._send_json(400, {"error": str(exc)})
+                return
+            except Exception as exc:
+                self._send_json(502, {"error": str(exc)[:500]})
+                return
+        finally:
+            conn.close()
+        with _auth_store(self.server.db_path) as store:
+            store.audit(user_id, project_id, "resident_watch_created",
+                        watch["id"])
+        self._send_json(201, watch)
+
+    def _resident_list(self, project_id):
+        conn = self._db()
+        try:
+            try:
+                watches = _res_list(conn, project_id)
+            except Exception as exc:
+                self._send_json(502, {"error": str(exc)[:500]})
+                return
+            conn.commit()
+        finally:
+            conn.close()
+        self._send_json(200, {
+            "project_id": project_id, "resident_watches": watches,
+            "due_count": sum(1 for w in watches if w["due"])})
+
+    def _resident_get(self, user_id, project_id, watch_id):
+        conn = self._db()
+        try:
+            watch = _res_get(conn, watch_id)
+            polls = []
+            polls_note = ""
+            if watch is not None:
+                try:
+                    polls = _res_polls(conn, watch_id)
+                except Exception:
+                    polls_note = "polls unavailable"
+        finally:
+            conn.close()
+        if watch is None or watch["project_id"] != project_id:
+            self._send_json(404, {"error": "resident watch not found"})
+            return
+        self._send_json(200, dict(watch, polls=polls,
+                                  polls_note=polls_note))
+
+    def _resident_update(self, user_id, project_id, watch_id):
+        body, jerr = self._read_json()
+        if jerr:
+            self._send_json(400, {"error": jerr})
+            return
+        conn = self._db()
+        try:
+            try:
+                ok, watch = _res_update(conn, watch_id, body)
+            except ValueError as exc:
+                self._send_json(400, {"error": str(exc)})
+                return
+            except Exception as exc:
+                self._send_json(502, {"error": str(exc)[:500]})
+                return
+        finally:
+            conn.close()
+        if watch is None or watch["project_id"] != project_id:
+            self._send_json(404, {"error": "resident watch not found"})
+            return
+        if not ok:
+            self._send_json(409, {
+                "error": "ended resident watch cannot be updated;"
+                         " create a new watch instead"})
+            return
+        with _auth_store(self.server.db_path) as store:
+            store.audit(user_id, project_id, "resident_watch_updated",
+                        watch_id)
+        self._send_json(200, watch)
+
+    def _resident_transition(self, user_id, project_id, watch_id, action):
+        fn = {"pause": _res_pause, "resume": _res_resume,
+              "end": _res_end}[action]
+        conn = self._db()
+        try:
+            try:
+                ok, watch = fn(conn, watch_id)
+            except Exception as exc:
+                self._send_json(502, {"error": str(exc)[:500]})
+                return
+        finally:
+            conn.close()
+        if watch is None or watch["project_id"] != project_id:
+            self._send_json(404, {"error": "resident watch not found"})
+            return
+        if not ok:
+            if watch["status"] == "ended":
+                self._send_json(409, {
+                    "error": "ended resident watch is terminal;"
+                             " create a new watch instead"})
+            else:
+                self._send_json(409, {
+                    "error": "resident watch in status %s cannot %s"
+                             % (watch["status"], action)})
+            return
+        audit_action = {"pause": "resident_watch_paused",
+                        "resume": "resident_watch_resumed",
+                        "end": "resident_watch_ended"}[action]
+        with _auth_store(self.server.db_path) as store:
+            store.audit(user_id, project_id, audit_action, watch_id)
+        self._send_json(200, watch)
+
    def _triage_inbox(self, project_id):
        """分诊收件箱(只读)：到期自动化、失败/中断任务、
        待处理后续消息、最近自动化产出。见 automations 模块
@@ -0,0 +1,493 @@
+"""常驻监听任务单测（cloud/resident.py）：不跑真实模型、不联网，
+抓取一律注入假 fetcher。覆盖创建校验、暂停保游标、恢复续跑、
+结束终态、防重复轮询、仅变化建任务、失败退避、配额耗尽转暂停、
+worker 驱动批量入口。"""
+import http.server
+import os
+import socket
+import sqlite3
+import sys
+import tempfile
+import threading
+import unittest
+from pathlib import Path
+from types import SimpleNamespace
+
+try:
+    import cloud.resident as resident_module
+    from cloud.resident import (
+        claim_due_watch,
+        create_watch,
+        default_fetch,
+        end_watch,
+        ensure_resident_tables,
+        get_watch,
+        list_polls,
+        pause_watch,
+        process_due_resident_watches,
+        resume_watch,
+        run_due_resident_watch,
+        update_watch,
+        validate_target_url,
+    )
+    from cloud.server import init_db
+    from cloud.worker import Worker
+except ImportError:  # discover -s cloud/tests 时把项目根目录补进 sys.path
+    sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
+    import cloud.resident as resident_module
+    from cloud.resident import (
+        claim_due_watch,
+        create_watch,
+        default_fetch,
+        end_watch,
+        ensure_resident_tables,
+        get_watch,
+        list_polls,
+        pause_watch,
+        process_due_resident_watches,
+        resume_watch,
+        run_due_resident_watch,
+        update_watch,
+        validate_target_url,
+    )
+    from cloud.server import init_db
+    from cloud.worker import Worker
+
+URL = "https://example.com/data"
+
+
+class _Fetcher:
+    """可变假抓取：value 改了，下一次轮询就看到新内容。"""
+
+    def __init__(self, value="v1"):
+        self.value = value
+        self.calls = []
+
+    def __call__(self, url, timeout):
+        self.calls.append((url, timeout))
+        return self.value
+
+
+def _failing_fetcher(url, timeout):
+    raise RuntimeError("boom")
+
+
+class ResidentTestCase(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.db_path = os.path.join(self.tmp.name, "control.db")
+        init_db(self.db_path)
+        self.conn = sqlite3.connect(self.db_path)
+        self.conn.row_factory = sqlite3.Row
+        ensure_resident_tables(self.conn)
+        self.now = 1700000000.0
+
+    def tearDown(self):
+        self.conn.close()
+        self.tmp.cleanup()
+
+    def _watch(self, name="watch1", **kw):
+        kw.setdefault("interval_seconds", 60)
+        return create_watch(self.conn, "proj", name, URL,
+                            "变化时汇报", "owner1", now=self.now, **kw)
+
+    def _tasks(self):
+        cur = self.conn.execute("SELECT id, prompt, status FROM tasks")
+        return cur.fetchall()
+
+    def _set_quota(self, user_id, micros):
+        self.conn.execute(
+            "CREATE TABLE IF NOT EXISTS user_quotas("
+            "user_id TEXT PRIMARY KEY, quota_usd_micros BIGINT, "
+            " updated_at DOUBLE PRECISION)")
+        self.conn.execute(
+            "INSERT OR REPLACE INTO user_quotas(user_id,"
+            " quota_usd_micros, updated_at) VALUES(?, ?, ?)",
+            (user_id, micros, self.now))
+        self.conn.commit()
+
+    def test_create_validation(self):
+        watch = self._watch()
+        self.assertEqual(watch["status"], "running")
+        self.assertTrue(watch["due"])
+        self.assertEqual(watch["next_poll_at"], self.now)
+        for kwargs, _ in (
+                ({"target_url": "ftp://example.com/x"}, None),
+                ({"target_url": "https://localhost/x"}, None),
+                ({"target_url": "http://127.0.0.1/x"}, None),
+                ({"target_url": "https://10.0.0.5/x"}, None),
+                ({"target_url": "https://user:pw@example.com/x"}, None),
+                ({"interval_seconds": 30}, None),
+                ({"name": "watch1"}, None),
+                ({"prompt": "  "}, None)):
+            with self.assertRaises(ValueError):
+                base = {"name": "other", "target_url": URL,
+                        "prompt": "p", "interval_seconds": 60}
+                base.update(kwargs)
+                create_watch(self.conn, "proj", base["name"],
+                             base["target_url"], base["prompt"], "owner1",
+                             interval_seconds=base["interval_seconds"],
+                             now=self.now)
+
+    def test_baseline_unchanged_changed_only_change_creates_task(self):
+        fetcher = _Fetcher("v1")
+        watch = self._watch()
+        poll = run_due_resident_watch(self.conn, watch["id"],
+                                      fetcher=fetcher, now=self.now)
+        self.assertEqual(poll["status"], "baseline")
+        self.assertEqual(len(self._tasks()), 0)
+        # 同一间隔内第二次：抢不到，不重复轮询。
+        self.assertIsNone(run_due_resident_watch(
+            self.conn, watch["id"], fetcher=fetcher, now=self.now))
+        self.assertEqual(len(fetcher.calls), 1)
+        self.now += 61
+        poll = run_due_resident_watch(self.conn, watch["id"],
+                                      fetcher=fetcher, now=self.now)
+        self.assertEqual(poll["status"], "unchanged")
+        self.assertEqual(len(self._tasks()), 0)
+        fetcher.value = "v2"
+        self.now += 61
+        poll = run_due_resident_watch(self.conn, watch["id"],
+                                      fetcher=fetcher, now=self.now)
+        self.assertEqual(poll["status"], "changed")
+        self.assertIsNotNone(poll["task_id"])
+        tasks = self._tasks()
+        self.assertEqual(len(tasks), 1)
+        self.assertIn("v2", tasks[0][1])
+        self.assertIn("变化时汇报", tasks[0][1])
+        after = get_watch(self.conn, watch["id"], now=self.now)
+        self.assertEqual(after["last_change_text"], "v2")
+        self.assertIsNotNone(after["last_change_at"])
+        self.assertEqual(
+            [p["status"] for p in list_polls(self.conn, watch["id"])],
+            ["changed", "unchanged", "baseline"])
+
+    def test_pause_preserves_cursor_and_resume_reports_pause_period_change(self):
+        fetcher = _Fetcher("v1")
+        watch = self._watch()
+        run_due_resident_watch(self.conn, watch["id"],
+                               fetcher=fetcher, now=self.now)
+        cursor = get_watch(self.conn, watch["id"], now=self.now)
+        self.assertIsNotNone(cursor["last_content_hash"])
+        ok, paused = pause_watch(self.conn, watch["id"], now=self.now)
+        self.assertTrue(ok)
+        self.assertEqual(paused["status"], "paused")
+        self.assertEqual(paused["last_content_hash"],
+                         cursor["last_content_hash"])
+        self.assertIsNone(paused["next_poll_at"])
+        self.assertFalse(paused["due"])
+        # 暂停期间即使时间过了间隔，驱动也不轮询。
+        self.assertIsNone(run_due_resident_watch(
+            self.conn, watch["id"], fetcher=fetcher, now=self.now + 3600))
+        self.assertEqual(len(fetcher.calls), 1)
+        # 暂停期间网站变了；恢复后第一次轮询用旧游标比对，报出来。
+        fetcher.value = "v2"
+        ok, resumed = resume_watch(self.conn, watch["id"], now=self.now)
+        self.assertTrue(ok)
+        self.assertEqual(resumed["status"], "running")
+        self.assertEqual(resumed["last_content_hash"],
+                         cursor["last_content_hash"])
+        poll = run_due_resident_watch(self.conn, watch["id"],
+                                      fetcher=fetcher, now=self.now)
+        self.assertEqual(poll["status"], "changed")
+        self.assertEqual(len(self._tasks()), 1)
+
+    def test_pause_resume_on_wrong_status_are_rejected(self):
+        watch = self._watch()
+        ok, _ = resume_watch(self.conn, watch["id"], now=self.now)
+        self.assertFalse(ok)
+        ok, _ = pause_watch(self.conn, watch["id"], now=self.now)
+        self.assertTrue(ok)
+        ok, _ = pause_watch(self.conn, watch["id"], now=self.now)
+        self.assertFalse(ok)
+
+    def test_end_is_terminal_and_name_reusable(self):
+        watch = self._watch()
+        ok, ended = end_watch(self.conn, watch["id"], now=self.now)
+        self.assertTrue(ok)
+        self.assertEqual(ended["status"], "ended")
+        self.assertIsNone(ended["next_poll_at"])
+        for fn in (pause_watch, resume_watch, end_watch):
+            ok, old = fn(self.conn, watch["id"], now=self.now)
+            self.assertFalse(ok)
+            self.assertEqual(old["status"], "ended")
+        self.assertIsNone(run_due_resident_watch(
+            self.conn, watch["id"], fetcher=_Fetcher(), now=self.now + 3600))
+        ok, _ = update_watch(self.conn, watch["id"],
+                             {"prompt": "新提示"}, now=self.now)
+        self.assertFalse(ok)
+        # 已结束的名字可重用。
+        again = self._watch()
+        self.assertEqual(again["status"], "running")
+
+    def test_claim_is_single_winner(self):
+        watch = self._watch()
+        self.assertTrue(claim_due_watch(self.conn, watch["id"], now=self.now))
+        self.assertFalse(claim_due_watch(self.conn, watch["id"],
+                                         now=self.now))
+        # 下个间隔到了才能再抢（claim 已把 next_poll_at 预置到下间隔）。
+        self.assertTrue(claim_due_watch(self.conn, watch["id"],
+                                        now=self.now + 61))
+
+    def test_fetch_failure_backoff_and_cursor_unchanged(self):
+        watch = self._watch()
+        run_due_resident_watch(self.conn, watch["id"],
+                               fetcher=_Fetcher("v1"), now=self.now)
+        cursor = get_watch(self.conn, watch["id"], now=self.now)
+        self.now += 61
+        poll = run_due_resident_watch(self.conn, watch["id"],
+                                      fetcher=_failing_fetcher, now=self.now)
+        self.assertEqual(poll["status"], "failed")
+        after = get_watch(self.conn, watch["id"], now=self.now)
+        self.assertEqual(after["consecutive_failures"], 1)
+        self.assertEqual(after["last_content_hash"],
+                         cursor["last_content_hash"])
+        # 失败退避：下次 = now + 间隔 + 30 秒，不是立刻重试。
+        self.assertEqual(after["next_poll_at"], self.now + 60 + 30)
+        self.now = after["next_poll_at"]
+        run_due_resident_watch(self.conn, watch["id"],
+                               fetcher=_failing_fetcher, now=self.now)
+        after2 = get_watch(self.conn, watch["id"], now=self.now)
+        self.assertEqual(after2["consecutive_failures"], 2)
+        self.assertEqual(after2["next_poll_at"], self.now + 60 + 60)
+
+    def test_quota_exceeded_pauses_not_ends_and_keeps_cursor(self):
+        fetcher = _Fetcher("v1")
+        watch = self._watch()
+        run_due_resident_watch(self.conn, watch["id"],
+                               fetcher=fetcher, now=self.now)
+        cursor_hash = get_watch(self.conn, watch["id"],
+                                now=self.now)["last_content_hash"]
+        self._set_quota("owner1", 0)
+        fetcher.value = "v2"
+        self.now += 61
+        poll = run_due_resident_watch(self.conn, watch["id"],
+                                      fetcher=fetcher, now=self.now)
+        self.assertEqual(poll["status"], "changed_quota_paused")
+        after = get_watch(self.conn, watch["id"], now=self.now)
+        self.assertEqual(after["status"], "paused")
+        self.assertEqual(after["paused_reason"], "quota_exceeded")
+        self.assertNotEqual(after["status"], "ended")
+        self.assertEqual(after["last_content_hash"], cursor_hash)
+        self.assertEqual(len(self._tasks()), 0)
+        # 恢复后配额仍耗尽：再次转暂停，不擅自结束、不丢变化。
+        resume_watch(self.conn, watch["id"], now=self.now)
+        poll = run_due_resident_watch(self.conn, watch["id"],
+                                      fetcher=fetcher, now=self.now)
+        self.assertEqual(poll["status"], "changed_quota_paused")
+        # 配额恢复后，同一变化仍能报出来（游标没被配额暂停推进）。
+        self._set_quota("owner1", 100000000)
+        resume_watch(self.conn, watch["id"], now=self.now)
+        poll = run_due_resident_watch(self.conn, watch["id"],
+                                      fetcher=fetcher, now=self.now)
+        self.assertEqual(poll["status"], "changed")
+        self.assertEqual(len(self._tasks()), 1)
+
+    def test_update_watch(self):
+        watch = self._watch()
+        ok, updated = update_watch(
+            self.conn, watch["id"],
+            {"interval_seconds": 3600, "prompt": "新处理提示"},
+            now=self.now)
+        self.assertTrue(ok)
+        self.assertEqual(updated["interval_seconds"], 3600)
+        self.assertEqual(updated["prompt"], "新处理提示")
+        with self.assertRaises(ValueError):
+            update_watch(self.conn, watch["id"], {}, now=self.now)
+        with self.assertRaises(ValueError):
+            update_watch(self.conn, watch["id"],
+                         {"target_url": "https://localhost/x"}, now=self.now)
+
+    def test_process_due_and_worker_driver(self):
+        fetcher = _Fetcher("v1")
+        self._watch(name="a")
+        self._watch(name="b")
+        polls = process_due_resident_watches(self.conn, fetcher=fetcher,
+                                             now=self.now)
+        self.assertEqual(len(polls), 2)
+        self.assertEqual([p["status"] for p in polls],
+                         ["baseline", "baseline"])
+        self.assertEqual(process_due_resident_watches(
+            self.conn, fetcher=fetcher, now=self.now), [])
+        # worker 驱动入口：注入假抓取，变化时产出 changed 轮询。
+        worker = Worker(self.db_path, "/fake/host.py",
+                        env={"XCM_API_KEY": "test-placeholder"},
+                        runner=lambda cmd: SimpleNamespace(
+                            returncode=0, stdout="{}", stderr=""))
+        fetcher.value = "v2"
+        self.now += 61
+        polls = worker.process_due_resident_watches(
+            now=self.now, fetcher=fetcher)
+        self.assertEqual(len(polls), 2)
+        self.assertTrue(all(p["status"] == "changed" for p in polls))
+        self.assertEqual(len(self._tasks()), 2)
+
+    def test_url_validation_blocks_localhost_and_numeric_forms(self):
+        for bad in ("http://localhost./x", "http://foo.localhost/x",
+                    "http://2130706433/x", "http://0x7f000001/x",
+                    "http://0177.0.0.1/x", "http://0x7f.0.0.1/x"):
+            with self.assertRaises(ValueError, msg=bad):
+                validate_target_url(bad)
+        # 规范点分十进制公网 IP 与带尾点的普通域名仍放行。
+        self.assertEqual(validate_target_url("https://8.8.8.8/x"),
+                         "https://8.8.8.8/x")
+        self.assertEqual(validate_target_url("https://example.com./x"),
+                         "https://example.com./x")
+
+    def test_unique_index_guards_concurrent_name_clash(self):
+        idx = self.conn.execute(
+            "SELECT name FROM sqlite_master WHERE type = 'index'"
+            " AND name = 'idx_resident_watches_name'").fetchone()
+        self.assertIsNotNone(idx)
+        self._watch()
+        # 正常路径仍先报 ValueError。
+        with self.assertRaises(ValueError):
+            self._watch()
+        # 绕过 SELECT 查重直接插：唯一索引本身必须拦住，
+        # 这才是并发下真正的关口。
+        with self.assertRaises(sqlite3.IntegrityError):
+            self.conn.execute(
+                "INSERT INTO resident_watches(id, project_id, name,"
+                " target_url, prompt, interval_seconds, status,"
+                " created_by, created_at, updated_at)"
+                " VALUES('dup', 'proj', 'watch1', ?, 'p', 60,"
+                " 'running', 'owner1', ?, ?)",
+                (URL, self.now, self.now))
+        other = self._watch(name="watch2")
+        with self.assertRaises(ValueError):
+            update_watch(self.conn, other["id"], {"name": "watch1"},
+                         now=self.now)
+
+    def test_quota_check_failed_is_transient_not_paused(self):
+        fetcher = _Fetcher("v1")
+        watch = self._watch()
+        run_due_resident_watch(self.conn, watch["id"],
+                               fetcher=fetcher, now=self.now)
+        cursor_hash = get_watch(self.conn, watch["id"],
+                                now=self.now)["last_content_hash"]
+        original = resident_module._automation_quota_exceeded
+        resident_module._automation_quota_exceeded = lambda conn, owner: (
+            "quota_check_failed: boom")
+        try:
+            fetcher.value = "v2"
+            self.now += 61
+            poll = run_due_resident_watch(
+                self.conn, watch["id"], fetcher=fetcher, now=self.now)
+        finally:
+            resident_module._automation_quota_exceeded = original
+        self.assertEqual(poll["status"], "changed_quota_check_failed")
+        after = get_watch(self.conn, watch["id"], now=self.now)
+        self.assertEqual(after["status"], "running")
+        self.assertEqual(after["last_content_hash"], cursor_hash)
+        self.assertEqual(after["consecutive_failures"], 1)
+        self.assertEqual(after["next_poll_at"], self.now + 60 + 30)
+        self.assertEqual(len(self._tasks()), 0)
+
+    def test_pause_during_fetch_aborts_without_task(self):
+        fetcher = _Fetcher("v1")
+        watch = self._watch()
+        run_due_resident_watch(self.conn, watch["id"],
+                               fetcher=fetcher, now=self.now)
+        cursor_hash = get_watch(self.conn, watch["id"],
+                                now=self.now)["last_content_hash"]
+        db_path = self.db_path
+
+        def pausing_fetcher(url, timeout):
+            conn2 = sqlite3.connect(db_path)
+            try:
+                conn2.row_factory = sqlite3.Row
+                pause_watch(conn2, watch["id"], now=self.now + 61)
+            finally:
+                conn2.close()
+            return "v2"
+
+        self.now += 61
+        poll = run_due_resident_watch(
+            self.conn, watch["id"], fetcher=pausing_fetcher, now=self.now)
+        self.assertIsNone(poll)
+        after = get_watch(self.conn, watch["id"], now=self.now)
+        self.assertEqual(after["status"], "paused")
+        self.assertIsNone(after["next_poll_at"])
+        self.assertEqual(after["last_content_hash"], cursor_hash)
+        self.assertEqual(len(self._tasks()), 0)
+        self.assertEqual(
+            [p["status"] for p in list_polls(self.conn, watch["id"])],
+            ["baseline"])
+
+
+class RedirectFetchTestCase(unittest.TestCase):
+    """重定向校验：只把 redirect-test.test 指到本地，其它主机
+    走原始解析；no_proxy 同款处理，tearDown 全部还原。"""
+
+    HOST = "redirect-test.test"
+
+    def setUp(self):
+        class Handler(http.server.BaseHTTPRequestHandler):
+            def do_GET(self):
+                if self.path == "/go":
+                    self.send_response(302)
+                    self.send_header("Location", "/final")
+                elif self.path == "/bad":
+                    self.send_response(302)
+                    self.send_header(
+                        "Location", "http://127.0.0.1:1/secret")
+                else:
+                    body = b"final-body"
+                    self.send_response(200)
+                    self.send_header("Content-Length", str(len(body)))
+                    self.end_headers()
+                    self.wfile.write(body)
+                    return
+                self.end_headers()
+
+            def log_message(self, *args):
+                pass
+
+        self.server = http.server.ThreadingHTTPServer(
+            ("127.0.0.1", 0), Handler)
+        self.port = self.server.server_address[1]
+        self.thread = threading.Thread(
+            target=self.server.serve_forever, daemon=True)
+        self.thread.start()
+        self._orig_getaddrinfo = socket.getaddrinfo
+
+        def patched(host, port, family=0, type=0, proto=0, flags=0):
+            if host == self.HOST:
+                return [(socket.AF_INET, socket.SOCK_STREAM, 6, "",
+                         ("127.0.0.1", port))]
+            return self._orig_getaddrinfo(
+                host, port, family, type, proto, flags)
+
+        socket.getaddrinfo = patched
+        self._old_no_proxy = os.environ.get("no_proxy")
+        self._old_NO_PROXY = os.environ.get("NO_PROXY")
+        os.environ["no_proxy"] = (
+            (self._old_no_proxy + ",") if self._old_no_proxy
+            else "") + self.HOST
+        os.environ["NO_PROXY"] = (
+            (self._old_NO_PROXY + ",") if self._old_NO_PROXY
+            else "") + self.HOST
+
+    def tearDown(self):
+        socket.getaddrinfo = self._orig_getaddrinfo
+        if self._old_no_proxy is None:
+            os.environ.pop("no_proxy", None)
+        else:
+            os.environ["no_proxy"] = self._old_no_proxy
+        if self._old_NO_PROXY is None:
+            os.environ.pop("NO_PROXY", None)
+        else:
+            os.environ["NO_PROXY"] = self._old_NO_PROXY
+        self.server.shutdown()
+        self.server.server_close()
+        self.thread.join(timeout=5)
+
+    def test_relative_redirect_followed_and_private_blocked(self):
+        base = "http://%s:%d" % (self.HOST, self.port)
+        self.assertEqual(default_fetch(base + "/go", 10), "final-body")
+        with self.assertRaises(ValueError):
+            default_fetch(base + "/bad", 10)
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,180 @@
+"""常驻监听服务端单测：POST/GET /projects/{id}/resident-watches、
+PATCH、POST pause / resume / end。不跑真实模型、不联网。
+骨架与 test_automations_server.py 同款（本地临时端口 + 临时
+SQLite）。"""
+import http.client
+import json
+import os
+import sqlite3
+import sys
+import tempfile
+import threading
+import unittest
+from pathlib import Path
+
+try:
+    from cloud.auth import AuthStore
+    from cloud.server import create_server
+except ImportError:  # discover -s cloud/tests 时把项目根目录补进 sys.path
+    sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
+    from cloud.auth import AuthStore
+    from cloud.server import create_server
+
+URL = "https://example.com/data"
+
+
+class ResidentServerTestCase(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.db_path = os.path.join(self.tmp.name, "control.db")
+        self.auth = AuthStore(self.db_path)
+        self.owner_id, self.owner_key = self.auth.create_user("owner")
+        self.outsider_id, self.outsider_key = self.auth.create_user(
+            "outsider")
+        self.server = create_server(self.db_path, "127.0.0.1", 0)
+        self.port = self.server.server_address[1]
+        self.thread = threading.Thread(
+            target=self.server.serve_forever, daemon=True)
+        self.thread.start()
+        status, self.project = self._req(
+            "POST", "/projects", {"name": "demo"}, self.owner_key)
+        self.assertEqual(status, 201)
+
+    def tearDown(self):
+        self.server.shutdown()
+        self.server.server_close()
+        self.thread.join(timeout=5)
+        try:
+            self.auth.conn.close()
+        except Exception:
+            pass
+        self.tmp.cleanup()
+
+    def _req(self, method, path, body=None, key=None):
+        conn = http.client.HTTPConnection(
+            "127.0.0.1", self.port, timeout=10)
+        hdrs = {}
+        payload = None
+        if body is not None:
+            payload = json.dumps(body).encode("utf-8")
+            hdrs["Content-Type"] = "application/json"
+        if key is not None:
+            hdrs["X-API-Key"] = key
+        try:
+            conn.request(method, path, body=payload, headers=hdrs)
+            resp = conn.getresponse()
+            data = resp.read()
+        finally:
+            conn.close()
+        try:
+            parsed = json.loads(data.decode("utf-8")) if data else None
+        except ValueError:
+            parsed = None
+        return resp.status, parsed
+
+    def _url(self, suffix=""):
+        return "/projects/%s/resident-watches%s" % (
+            self.project["id"], suffix)
+
+    def _create(self, **extra):
+        body = {"name": "watch1", "target_url": URL,
+                "prompt": "变化时汇报"}
+        body.update(extra)
+        return self._req("POST", self._url(), body, self.owner_key)
+
+    def _audit_actions(self):
+        conn = sqlite3.connect(self.db_path)
+        try:
+            rows = conn.execute(
+                "SELECT action FROM audit_log WHERE project_id = ?",
+                (self.project["id"],)).fetchall()
+        finally:
+            conn.close()
+        return [r[0] for r in rows]
+
+    def test_create_list_get_and_validation(self):
+        status, watch = self._create(interval_seconds=3600)
+        self.assertEqual(status, 201, watch)
+        self.assertEqual(watch["status"], "running")
+        self.assertEqual(watch["target_url"], URL)
+        self.assertTrue(watch["due"])
+        self.assertIsNotNone(watch["next_poll_at"])
+        self.assertIn("resident_watch_created", self._audit_actions())
+        status, body = self._req("GET", self._url(), None, self.owner_key)
+        self.assertEqual(status, 200)
+        self.assertEqual(len(body["resident_watches"]), 1)
+        self.assertEqual(body["due_count"], 1)
+        status, body = self._req(
+            "GET", self._url("/" + watch["id"]), None, self.owner_key)
+        self.assertEqual(status, 200)
+        self.assertEqual(body["id"], watch["id"])
+        self.assertEqual(body["polls"], [])
+        for bad in (
+                {"name": "x", "target_url": "ftp://example.com/x",
+                 "prompt": "p"},
+                {"name": "x", "target_url": "https://localhost/x",
+                 "prompt": "p"},
+                {"name": "x", "target_url": URL, "prompt": "p",
+                 "interval_seconds": 30},
+                {"name": "watch1", "target_url": URL, "prompt": "p"},
+                {"target_url": URL, "prompt": "p"},
+                {"name": "x", "prompt": "p"}):
+            status, _ = self._req("POST", self._url(), bad, self.owner_key)
+            self.assertEqual(status, 400, bad)
+        status, _ = self._req(
+            "GET", self._url("/nope"), None, self.owner_key)
+        self.assertEqual(status, 404)
+
+    def test_pause_resume_end_lifecycle_and_terminal_409(self):
+        status, watch = self._create()
+        self.assertEqual(status, 201)
+        wid = watch["id"]
+        status, body = self._req(
+            "POST", self._url("/%s/pause" % wid), {}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["status"], "paused")
+        self.assertEqual(body["paused_reason"], "user_paused")
+        self.assertIsNone(body["next_poll_at"])
+        self.assertFalse(body["due"])
+        status, body = self._req(
+            "POST", self._url("/%s/pause" % wid), {}, self.owner_key)
+        self.assertEqual(status, 409)
+        status, body = self._req(
+            "POST", self._url("/%s/resume" % wid), {}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["status"], "running")
+        self.assertIsNone(body["paused_reason"])
+        self.assertTrue(body["due"])
+        status, body = self._req(
+            "POST", self._url("/%s/end" % wid), {}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["status"], "ended")
+        for action in ("pause", "resume", "end"):
+            status, body = self._req(
+                "POST", self._url("/%s/%s" % (wid, action)), {},
+                self.owner_key)
+            self.assertEqual(status, 409, action)
+            self.assertIn("terminal", body["error"])
+        actions = self._audit_actions()
+        self.assertIn("resident_watch_paused", actions)
+        self.assertIn("resident_watch_resumed", actions)
+        self.assertIn("resident_watch_ended", actions)
+
+    def test_update_and_outsider_forbidden(self):
+        status, watch = self._create()
+        self.assertEqual(status, 201)
+        wid = watch["id"]
+        status, body = self._req(
+            "PATCH", self._url("/" + wid),
+            {"interval_seconds": 7200}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["interval_seconds"], 7200)
+        status, _ = self._req(
+            "POST", self._url(),
+            {"name": "x", "target_url": URL, "prompt": "p"},
+            self.outsider_key)
+        self.assertIn(status, (401, 403, 404))
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -0,0 +1,381 @@
+"""常驻监听网站闭环端到端测试。
+
+上一轮只验到注入假抓取的单测：真实抓取、真实 run_loop 轮次
+边界、经服务端 API 的全链路（创建 -> 基线 -> 变化建任务 ->
+暂停 -> 恢复 -> 失败退避 -> 结束）都没有跑过。本文件补这一环，
+不改任何生产默认值、不改生产校验口径：
+
+- 间隔仍是生产下限 60 秒；到期不是改间隔，而是给驱动传一个
+  更晚的 now（模拟时间流逝），与上一轮单测同款时间注入。
+- 目标 URL 仍走生产校验：主机名 website-loop.test 不是 IP 字
+  面、不是 localhost，所以校验通过；测试进程内只把这个主机
+  名的 DNS 解析指向本地测试 HTTP 服务（相当于 /etc/hosts 的
+  一条测试条目），抓取本身走真实 default_fetch（urllib、超时、
+  1MB 截断），不注入假 fetcher、不绕过生产代码。
+- 内容服务是线程内真实 http.server，可变内容、也可停服；
+  控制面是真实 ThreadingHTTPServer + 真实 SQLite 临时库；
+  变化建出的任务、轮询、审计都直接读库/接口核对。
+
+不联网、不跑真实模型、不碰生产。
+"""
+import http.client
+import http.server
+import json
+import os
+import socket
+import sqlite3
+import sys
+import tempfile
+import threading
+import time
+import unittest
+from pathlib import Path
+from types import SimpleNamespace
+
+try:
+    from cloud.auth import AuthStore
+    from cloud.resident import _content_hash
+    from cloud.server import create_server
+    from cloud.worker import Worker
+except ImportError:  # discover -s cloud/tests 时把项目根目录补进 sys.path
+    sys.path.insert(0, str(Path(__file__).resolve().parents[2]))
+    from cloud.auth import AuthStore
+    from cloud.resident import _content_hash
+    from cloud.server import create_server
+    from cloud.worker import Worker
+
+HOST = "website-loop.test"
+
+
+class WebsiteLoopTestCase(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.db_path = os.path.join(self.tmp.name, "control.db")
+        self.auth = AuthStore(self.db_path)
+        self.owner_id, self.owner_key = self.auth.create_user("owner")
+        self.server = create_server(self.db_path, "127.0.0.1", 0)
+        self.port = self.server.server_address[1]
+        self.thread = threading.Thread(
+            target=self.server.serve_forever, daemon=True)
+        self.thread.start()
+        status, self.project = self._req(
+            "POST", "/projects", {"name": "demo"}, self.owner_key)
+        self.assertEqual(status, 201, self.project)
+        # 可变内容服务：state 字典跨重启保留，停服/重启可复现。
+        self.state = {"content": "version one"}
+        state = self.state
+
+        class Handler(http.server.BaseHTTPRequestHandler):
+            def do_GET(self):
+                body = state["content"].encode("utf-8")
+                self.send_response(200)
+                self.send_header("Content-Type", "text/plain")
+                self.send_header("Content-Length", str(len(body)))
+                self.end_headers()
+                self.wfile.write(body)
+
+            def log_message(self, *args):
+                pass
+
+        self._handler = Handler
+        self.content_server = http.server.ThreadingHTTPServer(
+            ("127.0.0.1", 0), Handler)
+        self.content_port = self.content_server.server_address[1]
+        self.content_thread = threading.Thread(
+            target=self.content_server.serve_forever, daemon=True)
+        self.content_thread.start()
+        # 测试专用 DNS：只把 website-loop.test 指到本地服务，
+        # 其它主机一律走原始解析；生产校验与抓取代码一字不改。
+        self._orig_getaddrinfo = socket.getaddrinfo
+
+        def patched(host, port, family=0, type=0, proto=0, flags=0):
+            if host == HOST:
+                return [(socket.AF_INET, socket.SOCK_STREAM, 6, "",
+                         ("127.0.0.1", port))]
+            return self._orig_getaddrinfo(
+                host, port, family, type, proto, flags)
+
+        socket.getaddrinfo = patched
+        # 测试环境自带 egress 代理：urllib 默认会把这个测试主机名
+        # 也送代理并被断开。只给本测试主机名加 no_proxy（测试
+        # 进程内生效），生产抓取代码与代理行为一字不改。
+        self._old_no_proxy = os.environ.get("no_proxy")
+        self._old_NO_PROXY = os.environ.get("NO_PROXY")
+        os.environ["no_proxy"] = (
+            (self._old_no_proxy + ",") if self._old_no_proxy
+            else "") + HOST
+        os.environ["NO_PROXY"] = (
+            (self._old_NO_PROXY + ",") if self._old_NO_PROXY
+            else "") + HOST
+        self.worker = Worker(
+            self.db_path, "/fake/host.py",
+            env={"XCM_API_KEY": "test-placeholder"},
+            runner=lambda cmd: SimpleNamespace(
+                returncode=0, stdout="{}", stderr=""))
+
+    def tearDown(self):
+        # setUp 中途失败时属性可能未建：getattr 防御，避免
+        # tearDown 的 AttributeError 掩盖原始失败。
+        if getattr(self, "_orig_getaddrinfo", None) is not None:
+            socket.getaddrinfo = self._orig_getaddrinfo
+        if getattr(self, "_old_no_proxy", None) is None:
+            if "_old_no_proxy" in self.__dict__:
+                os.environ.pop("no_proxy", None)
+        else:
+            os.environ["no_proxy"] = self._old_no_proxy
+        if getattr(self, "_old_NO_PROXY", None) is None:
+            if "_old_NO_PROXY" in self.__dict__:
+                os.environ.pop("NO_PROXY", None)
+        else:
+            os.environ["NO_PROXY"] = self._old_NO_PROXY
+        try:
+            self.content_server.shutdown()
+            self.content_server.server_close()
+        except Exception:
+            pass
+        try:
+            self.content_thread.join(timeout=5)
+        except Exception:
+            pass
+        try:
+            self.server.shutdown()
+            self.server.server_close()
+        except Exception:
+            pass
+        try:
+            self.thread.join(timeout=5)
+        except Exception:
+            pass
+        try:
+            self.auth.conn.close()
+        except Exception:
+            pass
+        try:
+            self.tmp.cleanup()
+        except Exception:
+            pass
+
+    def _req(self, method, path, body=None, key=None):
+        conn = http.client.HTTPConnection("127.0.0.1", self.port,
+                                          timeout=10)
+        hdrs = {}
+        payload = None
+        if body is not None:
+            payload = json.dumps(body).encode("utf-8")
+            hdrs["Content-Type"] = "application/json"
+        if key is not None:
+            hdrs["X-API-Key"] = key
+        try:
+            conn.request(method, path, body=payload, headers=hdrs)
+            resp = conn.getresponse()
+            data = resp.read()
+        finally:
+            conn.close()
+        try:
+            parsed = json.loads(data.decode("utf-8")) if data else None
+        except ValueError:
+            parsed = None
+        return resp.status, parsed
+
+    def _url(self, suffix=""):
+        return "/projects/%s/resident-watches%s" % (
+            self.project["id"], suffix)
+
+    def _create_watch(self):
+        body = {"name": "site1",
+                "target_url": "http://%s:%d/data" % (
+                    HOST, self.content_port),
+                "prompt": "变化时汇报",
+                "interval_seconds": 60}
+        status, watch = self._req("POST", self._url(), body,
+                                  self.owner_key)
+        self.assertEqual(status, 201, watch)
+        return watch
+
+    def _get_watch(self, watch_id):
+        status, body = self._req(
+            "GET", self._url("/" + watch_id), None, self.owner_key)
+        self.assertEqual(status, 200, body)
+        return body
+
+    def _tasks(self):
+        conn = sqlite3.connect(self.db_path)
+        try:
+            return conn.execute(
+                "SELECT id, prompt, status FROM tasks"
+                " ORDER BY created_at, id").fetchall()
+        finally:
+            conn.close()
+
+    def _audit(self):
+        conn = sqlite3.connect(self.db_path)
+        try:
+            return [r[0] for r in conn.execute(
+                "SELECT action FROM audit_log WHERE project_id = ?",
+                (self.project["id"],)).fetchall()]
+        finally:
+            conn.close()
+
+    def test_full_website_loop_end_to_end(self):
+        # 1. 创建：经服务端 API，running、游标为空、下次轮询在位。
+        watch = self._create_watch()
+        wid = watch["id"]
+        self.assertEqual(watch["status"], "running")
+        self.assertEqual(watch["interval_seconds"], 60)
+        self.assertEqual(watch["fetch_timeout_seconds"], 10)
+        self.assertIsNone(watch["last_content_hash"])
+        self.assertIsNotNone(watch["next_poll_at"])
+        self.assertIn("resident_watch_created", self._audit())
+        t0 = time.time()
+
+        # 2. 基线：真实抓取 version one，只建游标，不建任务。
+        polls = self.worker.process_due_resident_watches(now=t0 + 120)
+        self.assertEqual(len(polls), 1, polls)
+        self.assertEqual(polls[0]["status"], "baseline")
+        self.assertEqual(
+            polls[0]["observed_hash"], _content_hash("version one"))
+        self.assertEqual(self._tasks(), [])
+        body = self._get_watch(wid)
+        self.assertEqual(body["last_content_hash"],
+                         _content_hash("version one"))
+        self.assertEqual(len(body["polls"]), 1)
+        self.assertEqual(body["polls"][0]["status"], "baseline")
+
+        # 3. 变化：内容改成 version two，到期轮询建出 queued 任务。
+        self.state["content"] = "version two"
+        polls = self.worker.process_due_resident_watches(now=t0 + 240)
+        self.assertEqual(len(polls), 1, polls)
+        self.assertEqual(polls[0]["status"], "changed")
+        self.assertIsNotNone(polls[0]["task_id"])
+        tasks = self._tasks()
+        self.assertEqual(len(tasks), 1)
+        self.assertEqual(tasks[0][0], polls[0]["task_id"])
+        self.assertEqual(tasks[0][2], "queued")
+        self.assertIn("变化时汇报", tasks[0][1])
+        self.assertIn("version two", tasks[0][1])
+        self.assertIn("resident_watch_change_task_created",
+                      self._audit())
+        body = self._get_watch(wid)
+        self.assertEqual(body["last_content_hash"],
+                         _content_hash("version two"))
+        self.assertIn("version two", body["last_change_text"])
+
+        # 4. 暂停：暂停期间再改内容，驱动不轮询、游标不动。
+        status, body = self._req(
+            "POST", self._url("/%s/pause" % wid), {}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["status"], "paused")
+        self.assertIsNone(body["next_poll_at"])
+        self.state["content"] = "version three"
+        polls = self.worker.process_due_resident_watches(now=t0 + 400)
+        self.assertEqual(polls, [])
+        body = self._get_watch(wid)
+        self.assertEqual(len(body["polls"]), 2)
+        self.assertEqual(body["last_content_hash"],
+                         _content_hash("version two"))
+
+        # 5. 恢复：用旧游标比对，报出暂停期间的变化并建任务，
+        #    不是从头建基线。
+        status, body = self._req(
+            "POST", self._url("/%s/resume" % wid), {}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["status"], "running")
+        polls = self.worker.process_due_resident_watches(now=t0 + 400)
+        self.assertEqual(len(polls), 1, polls)
+        self.assertEqual(polls[0]["status"], "changed")
+        tasks = self._tasks()
+        self.assertEqual(len(tasks), 2)
+        self.assertIn("version three", tasks[1][1])
+        body = self._get_watch(wid)
+        self.assertEqual(body["last_content_hash"],
+                         _content_hash("version three"))
+
+        # 6. 失败：停服后轮询记 failed、退避、不结束、游标不动；
+        #    恢复服务后内容未变，轮询记 unchanged、失败清零。
+        self.content_server.shutdown()
+        self.content_server.server_close()
+        self.content_thread.join(timeout=5)
+        polls = self.worker.process_due_resident_watches(now=t0 + 600)
+        self.assertEqual(len(polls), 1, polls)
+        self.assertEqual(polls[0]["status"], "failed")
+        body = self._get_watch(wid)
+        self.assertEqual(body["status"], "running")
+        self.assertEqual(body["consecutive_failures"], 1)
+        self.assertEqual(body["last_content_hash"],
+                         _content_hash("version three"))
+        self.assertGreater(body["next_poll_at"], t0 + 600 + 60)
+        self.content_server = http.server.ThreadingHTTPServer(
+            ("127.0.0.1", self.content_port), self._handler)
+        self.content_thread = threading.Thread(
+            target=self.content_server.serve_forever, daemon=True)
+        self.content_thread.start()
+        polls = self.worker.process_due_resident_watches(now=t0 + 800)
+        self.assertEqual(len(polls), 1, polls)
+        self.assertEqual(polls[0]["status"], "unchanged")
+        body = self._get_watch(wid)
+        self.assertEqual(body["consecutive_failures"], 0)
+        self.assertEqual(len(self._tasks()), 2)
+
+        # 7. 结束：终态，驱动不再轮询，改/停/恢复一律 409。
+        status, body = self._req(
+            "POST", self._url("/%s/end" % wid), {}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["status"], "ended")
+        self.assertIsNone(body["next_poll_at"])
+        self.assertEqual(body["last_content_hash"],
+                         _content_hash("version three"))
+        polls = self.worker.process_due_resident_watches(now=t0 + 1000)
+        self.assertEqual(polls, [])
+        for action in ("pause", "resume", "end"):
+            status, body = self._req(
+                "POST", self._url("/%s/%s" % (wid, action)), {},
+                self.owner_key)
+            self.assertEqual(status, 409, action)
+        status, _ = self._req(
+            "PATCH", self._url("/" + wid), {"interval_seconds": 120},
+            self.owner_key)
+        self.assertEqual(status, 409)
+        body = self._get_watch(wid)
+        self.assertEqual(body["status"], "ended")
+        self.assertEqual(len(body["polls"]), 5)
+
+    def test_run_loop_boundary_calls_resident_driver(self):
+        # 8. 驱动接线：run_loop 轮次边界确实调到常驻驱动，
+        #    不只是测试直接调 process_due。run_all 等无关边界
+        #    在本测试里打桩，只留 resident 驱动走真实抓取。
+        watch = self._create_watch()
+        wid = watch["id"]
+        worker = self.worker
+        calls = []
+        original = Worker.process_due_resident_watches
+
+        def recording():
+            # 实例属性上的函数不会自动绑定 self，必须闭包捕获
+            # worker，不能写成 def recording(self)。
+            calls.append(time.time())
+            return original(worker)
+
+        worker.process_due_resident_watches = recording
+        worker.run_all = lambda: []
+        worker.cleanup_stale_containers = lambda: None
+        worker._write_heartbeat = lambda: None
+        worker.register = lambda: None
+        worker._process_push_due = lambda: None
+        worker.process_auto_memories = lambda: None
+        worker._reap_kill_requests = lambda *a, **k: None
+        worker.auto_resume_interrupted = lambda: None
+        worker._heartbeat_running_tasks = lambda: None
+        worker._registry_heartbeat = lambda: None
+        worker._code_changed = lambda: False
+        worker._code_signature_valid = lambda: True
+        worker.run_loop(interval=0.05, max_seconds=0.4)
+        self.assertGreaterEqual(len(calls), 1)
+        body = self._get_watch(wid)
+        self.assertEqual(len(body["polls"]), 1)
+        self.assertEqual(body["polls"][0]["status"], "baseline")
+        self.assertEqual(body["last_content_hash"],
+                         _content_hash("version one"))
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -6582,6 +6582,33 @@ class Worker:
            log_event(self.logger, "worker_push_due_failed",
                      message="push due processing failed: %s" % exc)

+    def process_due_resident_watches(self, now=None, fetcher=None):
+        """常驻监听驱动（worker 轮次边界调用）：到期监听逐条
+        走 resident.process_due_resident_watches，抢单 + 廉价
+        抓取 + 仅变化建任务。全部异常只记日志，不炸循环；
+        fetcher 缺省用 resident 默认抓取，单测可经
+        self.resident_fetcher 或参数注入假抓取。"""
+        if now is None:
+            now = time.time()
+        if fetcher is None:
+            fetcher = getattr(self, "resident_fetcher", None)
+        try:
+            from cloud.resident import process_due_resident_watches as _run
+            conn = _db_connect(self.db_path)
+            try:
+                polls = _run(conn, fetcher=fetcher, now=now)
+            finally:
+                conn.close()
+        except Exception as exc:
+            log_event(self.logger, "worker_resident_watch_failed",
+                      message="resident watch processing failed: %s" % exc)
+            return []
+        for poll in polls:
+            log_event(self.logger, "worker_resident_watch_poll",
+                      message="resident watch poll: %s %s"
+                              % (poll["watch_id"], poll["status"]))
+        return polls
+
    def run_loop(self, interval=300, max_seconds=None, stop_event=None,
                 drain_timeout=60, sweep_threshold=300):
        """常驻循环：每 interval 秒跑一轮 run_all，空闲也不退出。
@@ -6664,6 +6691,10 @@ class Worker:
                # G11/G12 后台投递：摘要（每小时）与 outbox 重试，
                # 轮次边界做，不中断在跑任务，失败只记日志。
                self._process_push_due()
+                # 常驻监听驱动：同款轮次边界后台扫，到期监听
+                # 廉价抓取、仅变化建任务；服务端本身不起定时器，
+                # 这里就是常驻任务的实际驱动（见 resident 模块）。
+                self.process_due_resident_watches()
                # 自动记忆提取：同款轮次边界后台扫，失败只记日志
                # （process_auto_memories 内部已兜底，不炸循环）。
                self.process_auto_memories()
@@ -0,0 +1,201 @@
+# 常驻监听网站闭环：独立架构与代码质量审核报告
+
+审核员非实现者。审核对象：网站闭环 worktree
+`~/workspace/codex-app-server-resident-website-loop`，分支
+`feat-resident-website-loop`，头 `23725f9`（父 `561b389`，基底
+网站版 `0722ca3`，`origin/main 622c51e`）。本审核在独立 worktree
+`~/workspace/codex-app-server-resident-audit`（新分支
+`feat-resident-website-loop-audit`，从 `23725f9` 起）上做，闭环
+worktree、resident-task worktree、泛化 worktree、主 checkout
+（feat-repo-picker）一律未动。
+
+## 0. 评分标准（先定标准，再打分）
+
+只评架构与代码质量两项，满分各 10 分。控制台 UI、真实外网、
+生产 PG、双进程抢单实测、其它探针是其它维度/后续项，本轮不做，
+也不用它们冒充或无故扣这两项的分；但若某项确实卡住架构或代码
+质量本身，列为 blocker，不擅自定。
+
+### 架构 10 分标准
+
+- A1 分层干净：resident 模块自有状态机，与任务状态机、自动化
+  定义层互不干扰；server 只做路由/权限/审计，worker 只做驱动
+  边界调用；无循环依赖。
+- A2 状态机可推理：running/paused/ended、暂停保现场、恢复续跑、
+  结束终态 409、失败退避、配额暂停不结束的语义在所有路径一致，
+  且并发下不被竞态破坏（认领、改名、暂停/结束中途轮询）。
+- A3 驱动与认领正确：服务端不起定时器是明说的，实际驱动在
+  worker 轮次边界；claim 条件 UPDATE 是防重复的唯一关口；崩后
+  不锁死。
+- A4 快照/变化语义干净：轮询廉价、仅变化建任务、首次只建基线、
+  每次轮询可查；变化摘录与游标推进在同一事务边界内。
+- A5 安全边界在正确层：URL 校验、超时、1MB 截断、SSRF 防护不
+  只在创建时校验一次，抓取路径（包括重定向）也要守住；v1 的
+  DNS 解析后校验缺口必须明说，不得假装已解决。
+- A6 可扩展结构：网站是本轮唯一用例，但快照/抓取与状态机/认领
+  不应缠成一团；未来加探针时状态机与驱动骨架不应重写（本轮
+  不实现其它探针，只审结构）。
+
+### 代码质量 10 分标准
+
+- Q1 命名、函数长度与复杂度可读；长函数（run_due）分支清晰，
+  不靠注释掩盖逻辑。
+- Q2 无重复、无死代码、无测试值渗入生产代码；复用已有口径
+  （自动化 claim/配额/审计、环境名唯一索引先例）而不是另起一套。
+- Q3 错误处理诚实：失败、配额耗尽、配额检查失败三条路不混淆；
+  审计与轮询记录不静默丢失；异常不炸驱动循环。
+- Q4 测试可读、可隔离、可复现：本地 HTTP/DNS/no_proxy 等测试
+  手段干净还原、无环境泄漏；断言指向行为而非实现细节摆设。
+- Q5 静态检查：居民差分内 ruff E、F 干净（仓库惯用 `%` 格式、
+  宽泛 except 等不在仓库门禁内的规则不计入；基底已有的 E、F
+  不计入本轮）。
+
+## 1. 改动前分数
+
+- 架构：9.0 / 10
+- 代码质量：9.0 / 10
+
+基底是好的：分层、状态机、claim、配额复用、轮询可查、测试
+隔离（DNS/no_proxy 在 loop 测试里干净还原）都属实，定向 15 项
+基线全过（见第 4 节）。扣分只来自下面逐项 findings，不是整体
+推倒重来。
+
+## 2. Findings 清单（改动前）
+
+### 严重
+
+- F1 名称唯一性是先查后插、无数据库约束：create_watch /
+  update_watch 先 SELECT 查重再 INSERT/UPDATE，两个并发请求
+  可同时通过检查、插出同项目同名 running 监听；server 的
+  environments 已有先例：`CREATE UNIQUE INDEX … WHERE status
+  != 'archived'`。本轮用同款部分唯一索引
+  `WHERE status != 'ended'`（已结束名可重用）+ 完整性错误转
+  ValueError 修复。
+- F2 暂停/结束与进行中轮询竞态：run_due 认领后只读一次状态，
+  之后若用户暂停/结束，失败/基线/未变化/配额/变化分支的 UPDATE
+  均不带状态条件，会把 paused/ended 行的 next_poll_at 改回
+  未来值、并可能在用户已结束之后仍建出变化任务。修复：认领后
+  复查状态；全部终局 UPDATE 带 `status='running'` 条件，条件
+  不满足时回滚、不建任务、不记轮询。
+
+### 一般
+
+- F3 抓取路径 SSRF 只守了创建时：default_fetch 用默认 opener，
+  重定向目标不校验，公开 URL 重定向到内网地址可绕过创建校验；
+  另 localhost.、单整数/十六进制/前导零等非规范 IP 写法可绕过
+  字面 IP 检查。修复：重定向处理器逐跳校验（相对跳转先 urljoin
+  再校验）；校验时归一化主机（去尾点、*.localhost 挡掉）并拒绝
+  非规范数字 IP 写法。DNS 解析后校验仍属 v1 缺口，继续列为已知
+  偏差，不假装已解决。
+- F4 配额门判定靠子串：`\"quota_check_failed\" not in gate` 把
+  配额门返回文本当协议用，超限文本一旦含该子串即误判。配额门
+  的真实契约是前缀（`quota_check_failed:` / `quota_exceeded:` /
+  None），修复为 `startswith` 判定，并补 changed_quota_check_failed
+  的定向测试（该分支此前无测试）。
+- F5 row_get 重复造轮：resident 自带与 automations.row_get 逐字
+  同语义的 row_get，却已从 automations 导入另外两个函数。修复：
+  直接导入 automations.row_get，删本地定义。
+- F6 server 审计动作名嵌套三元：`_resident_transition` 里
+  `\"ended\" if … else action + \"d\" if … else \"resumed\"` 可读性差，
+  且改动时易错。修复为显式动作映射字典，审计字符串不变。
+
+### 建议
+
+- S1 test_resident.py 导入了未用的 `time`（ruff F401，居民差分
+  内唯一 E、F 问题）。删除。
+- S2 退避文档与代码不一致：`_backoff_seconds` 文档写“30、60、
+  120……秒”，代码实际下次轮询是 `now + interval + backoff`。
+  行为本身合理（不比间隔更频繁），文档误导。修复文档，行为不动。
+- S3 loop 测试 tearDown 不够防御：setUp 中途失败时 tearDown
+  访问未建属性会掩盖原始失败。修复为 getattr 防御。
+
+### 已核查、不是问题（不扣分）
+
+- 审计行“写在 commit 之后会丢”：核查 `_auto_audit_row` 本身
+  执行 INSERT 后自己 commit，失败/配额暂停/变化三处审计行在
+  模拟关闭连接后仍可查，不是缺陷。
+- server/worker 全量 ruff E、F 的其余项（如 worker.py:2781/2787
+  的 F821、server.py:10293 的 E501）经核对均为基底已有或由居民
+  插入导致的行号位移，非居民新增代码，不计入本轮。
+- 仓库默认 ruff 全规则的 39 项（如 UP031 `%` 格式）与仓库惯用
+  写法冲突，不在仓库门禁内，不计入。
+
+## 3. 修复（全部关闭，每项有对应测试或定向复核）
+
+- F1 名称唯一性：`ensure_resident_tables` 新增部分唯一索引
+  `idx_resident_watches_name ON resident_watches(project_id, name)
+  WHERE status != 'ended'`（server environments 同款口径）；
+  create/update 的 INSERT/UPDATE 包完整性错误转 ValueError
+  （sqlite3.IntegrityError / PostgresIntegrityError，调用方仍
+  返回 400）。测试：索引存在断言、绕过查重直接插报
+  IntegrityError、改名撞名报 ValueError。
+- F2 暂停/结束中途轮询：认领后复查状态（非 running 直接返回
+  None，不改行）；新增 `_guarded_watch_update`，失败/基线/
+  未变化/配额检查失败/变化推进游标的终局 UPDATE 全部带
+  `status='running'` 条件，条件不满足时回滚、不建任务、不记
+  轮询；变化分支先受保护推进游标、再建任务。测试：抓取回调
+  里经第二连接暂停监听，run_due 返回 None、无任务、游标与
+  next_poll_at 保持暂停态、轮询数不变。
+- F3 抓取路径 SSRF：新增 `_SafeRedirectHandler`，重定向逐跳
+  urljoin 后复用 `validate_target_url`，不合格抛 ValueError
+  不跟随；`default_fetch` 改走 `build_opener`（默认处理器含
+  ProxyHandler，no_proxy 行为与 urlopen 同款）；校验归一化
+  主机（去尾点、挡 *.localhost）并新增 `_is_noncanonical_
+  numeric_host` 拒绝单整数/十六进制/前导零八进制等非规范
+  数字 IP 写法。测试：localhost.、*.localhost、2130706433、
+  0x7f000001、0177.0.0.1、0x7f.0.0.1 均报 ValueError，规范
+  公网 IP 与带尾点域名仍放行；相对重定向跟随、私有绝对重
+  定向报 ValueError（DNS 映射测试主机，同 loop 测试口径）。
+- F4 配额门判定：`\"quota_check_failed\" not in gate` 改为
+  `gate.startswith(\"quota_check_failed\")`，与配额门前缀契约
+  对齐。测试：注入返回 `quota_check_failed: boom` 的配额门，
+  变化轮询记 changed_quota_check_failed、状态仍 running、
+  游标不动、失败数 1、下次 = now + 间隔 + 30、无任务。
+- F5 row_get 重复：删 resident 本地定义，直接导入
+  automations.row_get，内部调用点一字未动，模块依赖说明
+  同步更新。
+- F6 server 审计动作名：嵌套三元改为显式映射字典
+  {pause: paused, resume: resumed, end: ended}，审计字符串
+  不变，服务端测试的审计断言仍过。
+- S1 删 test_resident.py 未用 `time` 导入；S2 `_backoff_seconds`
+  文档改为“退避增量，下次 = now + 间隔 + backoff”，行为不动；
+  S3 loop 测试 tearDown 改为 getattr 防御，setUp 中途失败不
+  再掩盖原始失败。
+
+## 4. 测试数字（全部本轮实测，无删断言、无 skip、无改默认值）
+
+- 定向基线（改动前）：`cloud.tests.test_resident` +
+  `test_resident_server` + `test_resident_website_loop` = 15 项全过。
+- 定向改动后：同三文件 = 20 项全过（基线 15 + 新增 5：URL
+  变体、唯一索引、配额检查失败、中途暂停、重定向），4.450 秒；
+  ruff E、F 在居民差分 4 文件上全部通过。
+- 全量（CI 同款三条 discover，改动后、报告提交前跑出，之后
+  无任何代码或文案改动）：顶层 tests 105 项全过；cloud/tests
+  1551 项全过、9 跳过（384.252 秒）；deploy/tests 44 项全过
+  （单独重跑确认）；合计 1700 项、0 失败、9 跳过。对账：闭环
+  基线 1695（105 + 1546 + 44）+ 新增 5 = 1700，无测试丢失、
+  无新失败。
+- 覆盖率：本仓库 CI 无覆盖率门禁，不虚构覆盖率变化；新增
+  5 项定向测试即上述全量的一部分。
+
+## 5. 复评分数
+
+- 架构：10 / 10
+- 代码质量：10 / 10
+
+复评依据：F1/F2 两项严重 findings 已由数据库约束与状态条件
+UPDATE 从机制上关闭（非仅文档说明），并有定向测试锁定；
+F3 安全边界已从“创建时一次校验”补到“抓取重定向逐跳校验”
+与非规范 IP 写法；F4/F5/F6/S1/S2/S3 全部关闭且测试全过。
+剩余的已知偏差（控制台 UI、全文 hash 抖动、DNS 解析后 SSRF、
+真实外网/生产 PG/双进程抢单未验）属其它维度/后续项，已在
+第 6 节单独列出，不计入本轮两项分数，也不被用来冒充已完成。
+
+## 6. 剩余 blocker 与未做事项
+
+- 未做：未 push、未合并、未跑 deploy.py、未部署；不做其它探针、
+  不接控制台 UI、不做生产 PG 与双进程抢单实测。
+- 已知偏差（其它维度，不计入本轮两项分数）：控制台 UI 未接；
+  变化判定为全文 hash、无选择器提取，动态页面抖动可误报；
+  SSRF v1 不做 DNS 解析后校验；真实外网、生产 PG、双进程抢单
+  未验。
@@ -0,0 +1,69 @@
+# 常驻监听任务（resident watch）验证摘要
+
+## 基底与交付物（代码就绪，local only）
+
+- 工作区：`~/workspace/codex-app-server-resident-task`
+- 分支：`feat-resident-task`，基底 `origin/main` = `622c51e`（Merge PR #274，2026-11-02 fetch 到的头，与任务书一致）。
+- 主 checkout（`~/workspace/codex-app-server`，`feat-repo-picker`，有脏文件）与其它现存 worktree 一律未动。
+- 本轮提交前状态：`cloud/server.py` +187 行、`cloud/worker.py` +31 行，新增 `cloud/resident.py`、`cloud/tests/test_resident.py`、`cloud/tests/test_resident_server.py`、本摘要。
+- **未做**：未 push、未合并到 main、未跑 `deploy.py`、未部署、未建生产任务、未动生产／主机 env／系统 key／sub2api／生产网关／仓库可见性／GitHub App 权限、未杀生产容器。
+
+## 基底核实（先调查结论：main 上常驻监听不存在，不是重复建设）
+
+- 任务状态机仍是 `queued -> running -> completed / failed / interrupted / merged / merge-failed`；`/pause` 只是把进行中回合任务标 `interrupted` 并置 `paused_by_user=1`（`cloud/server.py` `_pause_task`），不是可续跑的常驻暂停态——本实现不复用它，另建常驻态。
+- 云端自动化（`cloud/automations.py`）定义层只有 `active / cancelled`，含 `automations / automation_runs` 表与 `run_due_automation` 的 claim 抢单、配额门；服务端本身不起定时器，到期触发靠外部触发器。本实现复用它的 claim 口径与配额函数，不另起冲突语义。
+- goal 循环（`host.py` `run_goal_loop`，`max_rounds=0` 不限轮次）结束条件是模型宣告达成，不是“一直监听外部变化”；定时自动化每次到期建一个新任务跑一轮就结束。两者都不是常驻监听，本摘要不把它们说成已实现。
+- main 上与暂停相关的已有机制（MCP `pause_task` 工具、长任务中断自动恢复）针对的都是回合任务，不是常驻监听；grep `pause|resident` 在 `cloud/*.py` 与 `host.py` 中未发现常驻监听实现。
+
+## 设计抉择
+
+- 新层不改任务表：`cloud/resident.py`（标准库 + 复用 `cloud.automations` 的 `_audit_row`、`_automation_quota_exceeded`）新增 `resident_watches` / `resident_polls` 两张表，避免改动任务状态机牵连 worker 认领、sweep、自动恢复。
+- 状态机：`running -> paused -> running`；`running / paused -> ended`（终态）。`ended` 后 pause / resume / end / update 一律 409，恢复已结束的监听必须新建任务（文档与测试双锁定，见下）。
+- 暂停保留现场：暂停只改 `status / paused_reason / next_poll_at`，`last_content_hash / last_poll_at / last_change_at / last_change_text` 不动；暂停期间 `due=False`，驱动不轮询。恢复时 `next_poll_at=now` 立刻复查，用旧游标比对——暂停期间的变化在恢复后第一次轮询即报出，不丢。
+- 驱动（明说）：服务端本身不起定时器，与自动化同款结论。实际驱动是 `cloud/worker.py` 的 `run_loop` 轮次边界调 `process_due_resident_watches()`（新增方法，异常只记日志、不炸循环；`self.resident_fetcher` 可注入假抓取供测试）。任何外部触发器也可直接调 `run_due_resident_watch / process_due_resident_watches`，两者都走 claim 门。
+- 防重复轮询：`claim_due_watch` 条件 UPDATE（`status='running'` 且 `next_poll_at <= now`），抢到才把 `last_poll_at / last_claimed_at` 写成 now 并预置 `next_poll_at=now+interval`（崩在抓取中途时下个间隔仍重试，不因认领锁死永久停摆；完成后按成败重算）。
+- 变化才花钱：轮询 = 廉价抓取 + 空白折叠后 sha256 比对，不跑模型。`baseline`（首次，只建游标）、`unchanged`（不建任务）、`changed`（才建一个 queued 任务，prompt = 用户处理提示 + 变化摘录）三种产出都记 `resident_polls`，可查。
+- 配额边界：配额门只在“变化要建模型任务”那一刻过（复用自动化同款 `_automation_quota_exceeded`）。耗尽时**不结束、不丢变化**：转 `paused`（`paused_reason='quota_exceeded'`）、游标不推进、记 `changed_quota_paused`；配额恢复后恢复，同一变化仍能报出。配额检查本身失败时按瞬时失败退避，不转暂停，游标同样不推进。
+- 安全边界：间隔下限 60 秒（与自动化同款）、上限 30 天；抓取超时 1–60 秒（默认 10 秒）、读满 1MB 即断；失败退避 30 / 60 / 120……秒、上限 3600 秒；目标 URL 仅 http/https、无 userinfo、字面 localhost／私有／回环／链路本地 IP 拒绝。
+- API（`cloud/server.py`，权限与自动化同款：写 `create_task`、读 `read`，全部记审计）：
+  - `POST /projects/{id}/resident-watches`（name / target_url / prompt 必填；interval_seconds 默认 3600；fetch_timeout_seconds 可选；创建不走配额门）
+  - `GET /projects/{id}/resident-watches`（附 `due_count`）、`GET /projects/{id}/resident-watches/{wid}`（附最近轮询 `polls`）
+  - `PATCH /projects/{id}/resident-watches/{wid}`（改名／提示／目标／间隔／超时；ended 409；改间隔不立即触发，下次轮询完成后按新间隔重算）
+  - `POST /projects/{id}/resident-watches/{wid}/pause|resume|end`
+  - 状态查询返回常驻态、游标（`last_content_hash`）、最后变化（`last_change_at / last_change_text`）、下次轮询（`next_poll_at`）、失败次数与暂停原因。
+
+## 验收清单逐项结果
+
+1. 能新建常驻任务（目标 URL、轮询间隔、变化处理提示）：✅ 模块与服务端测试覆盖；非空壳——新建即到期，驱动下一轮即可认领。
+2. 三态清晰（运行中／已暂停／已结束终态）：✅ 状态机与 HTTP 409 语义双覆盖。
+3. 暂停保留现场、暂停期间停轮询、恢复从游标续跑：✅ 测试 `test_pause_preserves_cursor_and_resume_reports_pause_period_change`（暂停期间变了，恢复后第一次轮询即报出）。
+4. 结束终态、结束后不再静默重启、结束后恢复被拒绝：✅ `ended` 后 due 查询只认 `running`，pause / resume / end / update 一律 409（服务端错误文本含 terminal 并提示新建）；测试覆盖。
+5. 真正有驱动、claim 防重复、心跳与 stale 回收：✅ 驱动为 worker `run_loop` 轮次边界（不是服务端定时器，明说）；claim 单胜者测试覆盖；stale 口径为“认领预置 next_poll_at，崩后下个间隔重试”+ worker 既有每轮心跳／sweep（常驻监听本身不另起心跳表，见已知偏差）。
+6. 变化才花钱、每次轮询与变化可查：✅ baseline / unchanged 不建任务、changed 才建任务；`resident_polls` 逐轮记录。
+7. 间隔下限、抓取超时、失败退避、除配额耗尽外不擅自结束、配额耗尽只转可恢复暂停并明示原因：✅ 全部测试覆盖；配额耗尽测试断言状态为 `paused` 且明确 `!= ended`。
+8. 权限与审计沿用任务／自动化同款、pause / resume / end 记审计：✅ 服务端测试断言 `resident_watch_created / paused / resumed / ended` 审计行；跨项目行一律 404（不泄露存在性）；局外人写操作被拒。
+9. API 含暂停、恢复、结束与状态查询：✅ 见上；控制台 UI 本轮未接，列为已知偏差，不写成已完成。
+
+## 真实测试数字
+
+- 定向（改动后）：`python3 -m unittest cloud.tests.test_resident cloud.tests.test_resident_server` —— 13 项（模块 10 + 服务端 3），全部通过。
+- 全量（CI 同款三条 discover，改动后、仅最后一次文案注释修订前跑出；该修订只改 `update_watch` 的 docstring，不改任何代码逻辑，修订后定向 13 项与自动化相关 31 项重跑全部通过）：
+  - `discover -s tests`：105 项，0 失败，0 跳过。
+  - `discover -s cloud/tests`：1544 项，0 失败，9 跳过。
+  - `discover -s deploy/tests`：44 项（单独重跑确认），0 失败。
+  - 合计 1693 项、0 失败、9 跳过。失败未靠删断言、加 skip、改默认值蒙混；新增测试 13 项全部计入上述 cloud 套件。
+- 新模块定向覆盖率（`coverage run --source=cloud.resident`，仅定向 13 项）：`cloud/resident.py` 312 语句、85% 覆盖。未覆盖主因：默认真实 HTTP 抓取（`default_fetch`，单元测试一律注入假抓取，不联网）与个别异常分支（PG 方言补列、配额检查失败分支）；本仓库 CI（`.github/workflows/host-verify.yml`）本身没有覆盖率门禁，故不报“覆盖率变化”——全量覆盖率基线未在本轮重算，不虚构数字。
+
+## 已知偏差与未做事项
+
+- 控制台 UI 未接：`cloud/console.html` 是打包产物，本轮没有常驻监听面板；用户现阶段只能走 API。接 UI 是后续一轮，不在本轮验收内。
+- 变化判定 v1 较粗：空白折叠后全文 sha256，没有选择器／区域提取；动态页面（广告、时间戳、随机推荐）本身会抖动成“变化”并建任务。缓解在后续：加 CSS 选择器或正文提取口径、或变化阈值。
+- SSRF 防护 v1 只挡字面 localhost／私有 IP，不做 DNS 解析后校验（解析后重定向到内网、DNS rebinding 不在覆盖内）；`websearch.fetch_url` 的同款字面口径已对齐，但不要把 v1 当成完整 SSRF 方案。
+- 常驻监听没有独立心跳表：stale 回收靠认领预置的 `next_poll_at` + worker 既有心跳／sweep；驱动进程整体死亡时，已认领未完成的当次轮询要等到下个间隔才补跑（与自动化 claim 的同款取舍）。
+- 未验证：真实外网抓取（单元测试不联网）、生产 PG 上的表迁移（`PRAGMA` 补列走自动化同款适配写法，但未在 PG 实测）、worker 常驻循环里的长时驱动（仅方法级测试，未跑真实 `run_loop` 长循环）、多 worker 并发抢单（仅条件 UPDATE 单胜者测试，未做双进程实测）。
+- 未做：未 push、未合并、未部署；UI、选择器提取、DNS 级 SSRF、PG 实测、双进程抢单实测留待后续轮次。
+
+## Blocker
+
+- 无本轮 blocker（全部工作均为本地代码与本地测试，未触及只有用户能给的账号／密钥／商户／商店账号／宿主／维护窗口，也未发生花真钱、删生产数据、改生产可见性的动作）。
+- 上线前仍需用户侧决策（不是本轮 blocker，不擅自定）：控制台 UI 是否随本批上、默认轮询间隔与抓取超时是否按本轮默认值（60 秒下限／10 秒超时）拍板为生产默认值、PG 实测与部署窗口。
@@ -0,0 +1,125 @@
+# 常驻监听网站闭环验证摘要
+
+## 起点与交付物（代码就绪，local only）
+
+- 工作区：`~/workspace/codex-app-server-resident-website-loop`
+- 分支：`feat-resident-website-loop`，起点为上一轮网站版
+  `feat-resident-task` 的 `0722ca3`（基底 `origin/main 622c51e`）。
+- 实现提交：`561b389`（仅新增端到端测试文件
+  `cloud/tests/test_resident_website_loop.py`，364 行）。
+  本摘要为其后提交，工作区提交后干净。
+- 范围：只做网站用例。不做 server_status、agent_status、
+  relay_cost 等其它探针；不把 resident-watches 改名/泛化成
+  通用任务；不改 API、表名与任何生产默认值。
+- 泛化 worktree（`~/workspace/codex-app-server-resident-generic`）
+  本轮一律未碰；resident-task worktree、主 checkout
+  （`feat-repo-picker`）及其它 worktree 一律未动。
+- **未做**：未 push、未合并到 main、未跑 `deploy.py`、未部署、
+  未建生产任务、未动生产／主机 env／系统 key／sub2api／
+  生产网关／仓库可见性／GitHub App 权限、未杀生产容器、
+  未连真实外部网站。
+
+## 闭环逐环证据（全部真实链路，不是注入假抓取）
+
+测试文件：`cloud/tests/test_resident_website_loop.py`，2 项。
+内容服务是线程内真实 `http.server`（可变内容、可停服重启）；
+控制面是真实 `ThreadingHTTPServer` + 真实 SQLite 临时库；
+驱动走真实 `default_fetch`（urllib、超时、1MB 截断）。
+生产 URL 校验一字不改：测试主机名 `website-loop.test` 不是
+IP 字面、不是 localhost，所以校验通过；测试进程内只把该
+主机名的 DNS 指到本地服务（相当于 /etc/hosts 的一条测试
+条目），并给它加 `no_proxy`（见“测试中发现的两处假断点”）。
+
+1. 创建（API）：`POST /projects/{id}/resident-watches`
+  （name=site1、target_url=`http://website-loop.test:<本地端口>/data`、
+   prompt=“变化时汇报”、interval_seconds=60）返回 201，
+   状态 running、interval 60、fetch_timeout 10（生产默认）、
+   last_content_hash 为空、next_poll_at 在位；审计含
+   `resident_watch_created`。状态查询 `GET` 单条可看到
+   running、游标、下次轮询。
+2. 基线轮询：worker 驱动 `process_due_resident_watches`
+  （fetcher=None，即真实 default_fetch）抓到 “version one”，
+   记 `baseline` 轮询、游标为其 sha256，任务数为 0；
+   状态查询的 polls 第 1 条就是 baseline。
+3. 变化建任务：内容改成 “version two”，到期轮询记
+   `changed`、建出 1 个 queued 任务，任务 prompt 含用户提示
+   “变化时汇报”与摘录 “version two”，任务行与审计行
+   `resident_watch_change_task_created` 均可查；游标推进到
+   version two，last_change_text 含 version two。
+4. 暂停：API pause 后状态 paused、next_poll_at 为空；暂停
+   期间内容再改成 “version three”，驱动返回空、轮询数仍为
+   2、游标仍为 version two——暂停期间的变化没有被轮询上报。
+5. 恢复：API resume 后状态 running；下一次驱动用旧游标比
+   对，记 `changed` 并建第 2 个任务（prompt 含 version three），
+   不是从头建基线；游标推进到 version three。
+6. 失败环：停掉内容服务后驱动记 `failed` 轮询、
+   consecutive_failures=1、next_poll_at 大于 now+60+30
+   （间隔 + 首次退避 30 秒）、状态仍 running、游标仍为
+   version three——驱动不崩、监听不擅自结束。恢复服务后
+   （内容仍为 version three）驱动记 `unchanged`、失败清零、
+   任务数仍为 2。
+7. 结束：API end 后状态 ended、next_poll_at 为空、游标仍为
+   version three；驱动返回空；pause / resume / end / PATCH
+   更新一律 409；状态查询显示 ended 与最后游标/变化，
+   polls 总数为 5（baseline、changed、changed、failed、
+   unchanged）。
+8. 驱动接线：单独一测试把 `run_all` 等无关边界打桩，只留
+   resident 驱动走真实抓取，调真实 `Worker.run_loop`
+  （interval=0.05、max_seconds=0.4）；记录到 run_loop 轮次
+   边界至少调了 1 次 `process_due_resident_watches`，且状态
+   查询里出现 1 条 baseline 轮询、游标为 version one。
+   这证明接线在 run_loop 里，不只是测试直接调驱动函数。
+
+## 断点结论：生产代码无断点，无生产改动
+
+本轮逐环均通，生产代码（`cloud/resident.py`、`cloud/server.py`、
+`cloud/worker.py`）一字未改，`git diff` 为空、新增文件只有
+端到端测试。第一次跑出的 2 个失败都是测试脚手架问题，
+不是生产断点，修法与证据如下，不许读成“修了生产代码”：
+
+- 假断点 1——首次真实抓取 failed：测试环境自带 egress 代理，
+  urllib 默认把测试主机名也送代理并被代理断开。实测对照：
+  不加 no_proxy 时 `default_fetch` 报 RemoteDisconnected，
+  加 no_proxy 后返回 “version one”。修法只在测试 setUp 里
+  给 `website-loop.test` 加 no_proxy（测试进程内生效，撕掉
+  时还原），生产抓取与代理行为一字不改。
+- 假断点 2——run_loop 接线测试 TypeError：实例属性上的
+  recording 函数不会自动绑定 self。修法改为闭包捕获 worker，
+  不改 worker 代码。
+
+## 真实测试数字
+
+- 定向（全量后重跑，含本轮 2 项）：`python3 -m unittest
+  cloud.tests.test_resident cloud.tests.test_resident_server
+  cloud.tests.test_resident_website_loop` —— 15 项全部通过。
+  其中上一轮 13 项在新 worktree 动工前已先复核全过。
+- 全量（CI 同款三条 discover，定向之后、全量之前无生产改动，
+  全量后又重跑定向 15 项全过，时间关系如实写明）：
+  - `discover -s tests`：105 项，0 失败，0 跳过。
+  - `discover -s cloud/tests`：1546 项，0 失败，9 跳过
+    （上一轮 1544 + 本轮新增 2，对账无差）。
+  - `discover -s deploy/tests`：44 项，0 失败。
+  - 合计 1695 项、0 失败、9 跳过。失败未靠删断言、加 skip、
+    改默认值蒙混；生产间隔下限 60 秒、抓取超时默认 10 秒、
+    URL 校验口径一律未动，到期只用驱动 now 注入模拟时间流逝。
+
+## 已知偏差与未做事项（未写成已完成）
+
+- 控制台 UI 不在本轮闭环内：只能走 API，无常驻面板。
+- 真实外网抓取未验：只验了本地 HTTP 服务 + 测试 DNS 映射，
+  不要把本轮说成真实外网验证；生产环境下的代理/DNS/证书
+  行为仍属后续验证项。
+- 生产 PG 表迁移未验（只验 SQLite 临时库）、双进程抢单未验
+  （只验条件 UPDATE 单胜者的上一轮单测与本轮单驱动链路）。
+- 其它探针本轮明确不做：server_status、agent_status、
+  relay_cost、通用任务改名一律不在本轮范围内。
+- 变化判定仍是上一轮 v1 全文 hash，无选择器/区域提取；
+  SSRF 仍是 v1 只挡字面私有 IP，未做 DNS 解析后校验。
+- 未做：未 push、未合并、未部署；架构与代码质量的 10 分
+  审核是父级后续独立步骤，本轮不自评、不给自己打 10 分。
+
+## Blocker
+
+- 本轮无 blocker（纯本地代码与本地测试，未触及只有用户能给
+  的账号／密钥／商户／商店账号／宿主／维护窗口，也未发生
+  花真钱、删生产数据、改生产可见性的动作）。