diff --git a/cloud/approvals.py b/cloud/approvals.py
new file mode 100644
index 0000000..5396cab
--- /dev/null
+++ b/cloud/approvals.py
@@ -0,0 +1,302 @@
+"""运行时审批队列(细粒度审批，P1 第 4 项)。
+
+选型：只打通 elicitation 一类，其他审批类全部记为遗留。
+
+选 elicitation 的理由：
+- 它是 host.py 里唯一已经特判的 server_request 类，未来 hook
+  的落点就是 _answer_server_request 的 elicitation 分支，加法
+  可接，不用先修一条错的通用回包。
+- 决定语义最干净：二值，批准 = accept(content 为空对象，
+  与今天 approval=auto 时的行为完全一致)，拒绝 = decline，
+  不需要像 permissions 那样现拼权限 profile，也不需要像
+  file_change 的 cancel 那样附带“中断整轮”的副作用。
+- 触发方是 MCP 服务器向用户要表单输入，正是“第三方工具向
+  用户要东西”的场景，用户在控制台点允许或拒绝最符合直觉。
+
+关键语义(先定清楚，再写代码)：
+- 待审批期间原流程挂起等待，不先拒绝后补。队列里的 pending
+  项本身不做任何放行或拦截动作，它只是等一个决定。
+- 等待有超时：默认 3600 秒，可在创建时改(服务端入口只接受
+  60 到 86400 秒)。超时默认是拒绝，绝不是放行：过期扫荡把
+  pending 置为 expired，同时把 effective decision 记为 denied，
+  decided_by 记为 system-timeout，回执里看得见。
+- 批准放行、拒绝拦截：决定一旦写入就是终态，再次决定返回
+  already_decided，原决定绝不被覆盖。回执字段：decided_by、
+  decided_at、decision，批准与拒绝、过期都留痕。
+- 决定是 CAS：只有 pending 行能被改状态，并发下两个决定只
+  有一个能赢，输的那个拿到 already_decided 而不是假成功。
+
+诚实边界(本片做到哪、没做到哪)：
+- 本片交付的是队列服务本身：建表迁移、创建、查询、批准、
+  拒绝、过期扫荡、回执、服务端端点、控制台 UI、单测。
+- 本片没有把实时 elicitation 接到队列上，host.py 完全没动。
+  原因不是协议不让挂起：JSON-RPC 请求本来就可以晚回包，
+  挂起在协议上是成立的。真正缺的是通道：host.py 跑在 worker
+  拉起的子进程里，手头没有云 DB 连接，也没有到控制面的
+  带权请求通道(worker.py 本片禁动，没法给它递 DB 路径或
+  token)。在没有通道的情况下往 _answer_server_request 里
+  加“等决定”只会把回合无限挂死，所以不做假拦截。
+- 在通道落地之前，实时 elicitation 的行为与今天完全一致：
+  approval=auto 时 accept 空 content，其他一律 decline。
+  队列里的决定目前是记录加事后审计口径，供通道落地后的
+  hook 消费，绝不要对外宣称“实时请求已被事前拦截”。
+- 通道落地时 host 侧要做的事只有两件，都已在本模块留好
+  映射函数：elicitation 到达时先建 pending 项再挂起等决定，
+  决定到了按 elicitation_response 回包；超时按 expired 的
+  denied 处理。通道本身(worker 递 DB 路径还是控制面 HTTP
+  带权轮询，生产是 Postgres 所以 DB 直连只够 SQLite 开发
+  环境用)是合流阶段的单独设计，不在本片里猜一个。
+
+遗留的审批类(明确不塞进本片)：
+- item/commandExecution/requestApproval：命令审批。现 host
+  通用回包写的是 approve/deny，但新协议的决定词表是
+  accept/decline/acceptForSession 等，通用回包本身就是错的，
+  要先修词表再接队列；kind 里已预留 command，映射函数
+  command_response 已按新协议写好，但不接实时流。
+- item/fileChange/requestApproval：文件改动审批。决定里有
+  cancel(拒绝并立即中断整轮)，不是纯二值，接队列要先定
+  cancel 的轮次语义，暂不接。
+- item/permissions/requestApproval：权限审批。回包不是二值，
+  而是一份 GrantedPermissionProfile，要现拼 profile，暂不接。
+- item/tool/requestUserInput：这根本不是审批，是向用户要
+  输入内容(问题列表)，blocking 与否另有语义，应另起输入
+  队列，暂不接。
+- item/tool/call 与 account/chatgptAuthTokens/refresh：前者是
+  动态工具执行，后者是凭证刷新，都不是用户审批，排除。
+
+本模块只依赖标准库，不 import cloud 其他模块，避免循环
+import；服务端与未来 worker/hook 共用。占位符用 ?，建表与
+列补齐写法与 peer_messages 同款(db 适配层负责翻译)。
+"""
+
+import json
+import time
+import uuid
+
+KINDS = ("elicitation", "command", "file_change", "permissions",
+         "user_input")
+DECISIONS = ("approved", "denied")
+DEFAULT_TIMEOUT_SECONDS = 3600
+MAX_SUMMARY_CHARS = 2000
+MAX_DETAIL_CHARS = 20000
+
+
+def row_get(row, name):
+    """行取值容错：sqlite3.Row 缺列报 IndexError、dict 缺键报
+    KeyError、空行报 TypeError，一律回 None(与 peer_messages
+    的同名函数同款，模块独立所以这里自带一份)。"""
+    try:
+        return row[name]
+    except (KeyError, IndexError, TypeError):
+        return None
+
+
+def ensure_approvals_table(conn):
+    """幂等建审批队列表。调用方负责 commit(服务端入口在写
+    操作后统一 commit；本函数自己不 commit，避免半截事务)。"""
+    conn.execute(
+        "CREATE TABLE IF NOT EXISTS task_approvals("
+        "id TEXT PRIMARY KEY, "
+        "task_id TEXT, "
+        "project_id TEXT, "
+        "kind TEXT, "
+        "summary TEXT, "
+        "detail TEXT, "
+        "status TEXT, "
+        "requested_by TEXT, "
+        "decided_by TEXT, "
+        "decided_at DOUBLE PRECISION, "
+        "decision TEXT, "
+        "created_at DOUBLE PRECISION, "
+        "updated_at DOUBLE PRECISION, "
+        "expires_at DOUBLE PRECISION)"
+    )
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_approvals_task"
+        " ON task_approvals(task_id, status)")
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_approvals_project"
+        " ON task_approvals(project_id, created_at)")
+
+
+def _detail_to_str(detail):
+    if detail is None:
+        return None
+    if isinstance(detail, str):
+        text = detail
+    else:
+        try:
+            text = json.dumps(detail, ensure_ascii=False)
+        except (TypeError, ValueError):
+            text = str(detail)
+    if len(text) > MAX_DETAIL_CHARS:
+        raise ValueError("detail too long")
+    return text
+
+
+def _detail_from_str(text):
+    if text is None:
+        return None
+    try:
+        return json.loads(text)
+    except (TypeError, ValueError):
+        return text
+
+
+def approval_to_dict(row):
+    """行转 dict：回执字段原样带出，detail 尽量还原成对象。"""
+    if row is None:
+        return None
+    return {
+        "id": row_get(row, "id"),
+        "task_id": row_get(row, "task_id"),
+        "project_id": row_get(row, "project_id"),
+        "kind": row_get(row, "kind"),
+        "summary": row_get(row, "summary"),
+        "detail": _detail_from_str(row_get(row, "detail")),
+        "status": row_get(row, "status"),
+        "requested_by": row_get(row, "requested_by"),
+        "decided_by": row_get(row, "decided_by"),
+        "decided_at": row_get(row, "decided_at"),
+        "decision": row_get(row, "decision"),
+        "created_at": row_get(row, "created_at"),
+        "updated_at": row_get(row, "updated_at"),
+        "expires_at": row_get(row, "expires_at"),
+    }
+
+
+def create_approval(conn, task_id, project_id, kind="elicitation",
+                    summary="", detail=None, requested_by=None,
+                    timeout_seconds=DEFAULT_TIMEOUT_SECONDS, now=None):
+    """创建一条待审批项，返回 dict。非法输入抛 ValueError，
+    调用方据此回 400；成功时 commit。
+
+    timeout_seconds：None 用默认值；0 表示不过期(只允许
+    模块级调用，服务端入口另有限制)；负数非法。"""
+    if kind not in KINDS:
+        raise ValueError("kind must be one of: " + ", ".join(KINDS))
+    if not isinstance(summary, str) or not summary.strip():
+        raise ValueError("summary required")
+    summary = summary.strip()
+    if len(summary) > MAX_SUMMARY_CHARS:
+        raise ValueError("summary too long")
+    if timeout_seconds is None:
+        timeout_seconds = DEFAULT_TIMEOUT_SECONDS
+    if not isinstance(timeout_seconds, (int, float)) \
+            or isinstance(timeout_seconds, bool) \
+            or timeout_seconds < 0:
+        raise ValueError("timeout_seconds invalid")
+    detail_str = _detail_to_str(detail)
+    now = time.time() if now is None else now
+    expires_at = None
+    if timeout_seconds:
+        expires_at = now + timeout_seconds
+    item_id = uuid.uuid4().hex
+    ensure_approvals_table(conn)
+    conn.execute(
+        "INSERT INTO task_approvals(id, task_id, project_id,"
+        " kind, summary, detail, status, requested_by,"
+        " decided_by, decided_at, decision,"
+        " created_at, updated_at, expires_at)"
+        " VALUES(?, ?, ?, ?, ?, ?, 'pending', ?,"
+        " NULL, NULL, NULL, ?, ?, ?)",
+        (item_id, task_id, project_id, kind, summary, detail_str,
+         requested_by, now, now, expires_at),
+    )
+    conn.commit()
+    return approval_to_dict(conn.execute(
+        "SELECT * FROM task_approvals WHERE id = ?",
+        (item_id,)).fetchone())
+
+
+def get_approval(conn, approval_id):
+    """取一条，缺失回 None(不抛，让调用方定 404 还是 409)。"""
+    ensure_approvals_table(conn)
+    return approval_to_dict(conn.execute(
+        "SELECT * FROM task_approvals WHERE id = ?",
+        (approval_id,)).fetchone())
+
+
+def list_approvals(conn, task_id=None, project_id=None, status=None):
+    """列审批项：纯读，不扫荡过期(扫荡是写操作，调用方在
+    需要时先调 expire_overdue)。过滤全是可选的精确匹配。"""
+    ensure_approvals_table(conn)
+    query = "SELECT * FROM task_approvals WHERE 1 = 1"
+    args = []
+    if task_id is not None:
+        query += " AND task_id = ?"
+        args.append(task_id)
+    if project_id is not None:
+        query += " AND project_id = ?"
+        args.append(project_id)
+    if status is not None:
+        query += " AND status = ?"
+        args.append(status)
+    query += " ORDER BY created_at, id"
+    rows = conn.execute(query, args).fetchall()
+    return [approval_to_dict(r) for r in rows]
+
+
+def decide_approval(conn, approval_id, decision, decided_by):
+    """写决定。返回 (status, item)：decided 表示本次写赢；
+    already_decided 表示该项已有终态(批准、拒绝或过期)，
+    原决定绝不被覆盖；not_found 表示无此项。
+
+    决定是 CAS：UPDATE 只改 status='pending' 的行，rowcount
+    为 0 说明并发下已被别人先决定，调用方按 already_decided
+    处理，不要按成功回。成功与过期扫荡都 commit。"""
+    if decision not in DECISIONS:
+        raise ValueError("decision must be approved or denied")
+    ensure_approvals_table(conn)
+    item = get_approval(conn, approval_id)
+    if item is None:
+        return "not_found", None
+    if item["status"] != "pending":
+        return "already_decided", item
+    now = time.time()
+    cur = conn.execute(
+        "UPDATE task_approvals SET status = ?, decision = ?,"
+        " decided_by = ?, decided_at = ?, updated_at = ?"
+        " WHERE id = ? AND status = 'pending'",
+        (decision, decision, decided_by, now, now, approval_id))
+    conn.commit()
+    if cur.rowcount == 1:
+        return "decided", get_approval(conn, approval_id)
+    return "already_decided", get_approval(conn, approval_id)
+
+
+def expire_overdue(conn, now=None):
+    """过期扫荡：把已到 expires_at 的 pending 项置为 expired，
+    同时把 effective decision 记为 denied(超时默认拒绝，绝不
+    放行)，decided_by 记为 system-timeout。返回过期条数。"""
+    ensure_approvals_table(conn)
+    now = time.time() if now is None else now
+    cur = conn.execute(
+        "UPDATE task_approvals SET status = 'expired',"
+        " decision = 'denied', decided_by = 'system-timeout',"
+        " decided_at = ?, updated_at = ?"
+        " WHERE status = 'pending' AND expires_at IS NOT NULL"
+        " AND expires_at <= ?",
+        (now, now, now))
+    conn.commit()
+    return cur.rowcount
+
+
+def elicitation_response(decision):
+    """未来 host hook 的回包映射(本片未接实时流，只定映射)：
+    批准 = accept 空 content，与今天 approval=auto 的行为一致；
+    拒绝或过期 = decline。decision 取 approved/denied/expired
+    之外的任何值都按拒绝处理，绝不默认放行。"""
+    if decision == "approved":
+        return {"action": "accept", "content": {}}
+    return {"action": "decline"}
+
+
+def command_response(decision):
+    """未来命令审批 hook 的回包映射，按新协议词表写(accept
+    / decline)，不是 host.py 现通用回包的 approve/deny——后者
+    与新协议不符，见模块 docstring 的遗留清单。非批准一律
+    decline，绝不默认放行。"""
+    if decision == "approved":
+        return {"decision": "accept"}
+    return {"decision": "decline"}
diff --git a/cloud/automations.py b/cloud/automations.py
new file mode 100644
index 0000000..4c21cf1
--- /dev/null
+++ b/cloud/automations.py
@@ -0,0 +1,525 @@
+"""自动化与分诊收件箱(P1 第 2 项)。
+
+与 host.py 现有 automations / run-due 的关系(先说清楚，避免
+两套调度打架)：
+
+- host.py 的自动化存在 host state JSON 文件里(字段 name /
+  task / repo / interval / last_run)，run-due 按
+  now - last_run >= interval 判断到期，到期派发一个 host 任务
+  跑一轮，跑完才回写 last_run。daemon tick 与 crontab 都只
+  调这一个 run-due，它是 host 侧唯一的执行器。
+- 本模块是云端控制面的另一套登记口径：自动化定义存在控制面
+  数据库里(automations 表，带 project_id)，不是 host state
+  文件。两者存储不相交，同一个自动化绝不会同时出现在两边，
+  所以本批不存在双重执行。
+- 本批不给云端自动化加执行器、不动 host.py 的 run-due 区：
+  host 的 run-due 跑的是 host 任务模型(codex 协议线程)，云
+  端任务模型(POST /tasks 建、worker 认领)是另一条链路，把
+  云端定义硬塞进 host run-due 会造出第二个调度器。云端执行
+  器是后续一步，它必须走 claim_due_automation 先抢单(条件
+  UPDATE，rowcount 为 1 才算抢到)，抢到后再建云端任务、再
+  记 automation_runs 一行；抢不到的执行器绝不能建任务。这
+  个 claim 门就是防重复执行的唯一关口，比 host run-due 的
+  “跑完才写 last_run”更严(中途被杀也不会在当间隔内重跑，
+  代价是崩掉的那次要等到下个间隔，见 claim 文档)。
+- 分诊收件箱是只读视图：到期自动化、失败或中断的任务、待
+  处理的后续消息、最近的自动化产出记录。它不跑模型、不改
+  任务状态，只把待处理项聚到一处。
+
+通知接入(本批结论：遗留，不新写通知代码)：
+
+- server 没有接 webhooks.NotificationService(它只有可注入
+  sender 的占位实现，server docstring 也写明未接成)；真正
+  的终态派发是 worker 走 cloud/push.py dispatch_event(订阅
+  制，completed / failed / goal-unmet 等终态才发)。
+- 自动化本身没有终态事件可派：定义创建不是完成，取消不是
+  失败；硬在创建/取消时发通知是伪造成功。本批只记审计
+  (automation_created / automation_cancelled)。等执行器
+  落地后，它建的云端任务自然走 worker 现有终态派发(任务
+  id 记在 automation_runs.task_id 里，派发与任务绑定，不用
+  为自动化另起一套通知系统)。
+
+本模块只依赖标准库，不 import cloud 其它模块，避免循环
+import；表迁移写法仿 peer_messages(CREATE TABLE IF NOT
+EXISTS + PRAGMA 查列补齐，SQLite 与 Postgres 适配层同款)。
+"""
+
+import time
+import uuid
+
+MIN_INTERVAL_SECONDS = 60
+MAX_INTERVAL_SECONDS = 30 * 86400
+MAX_NAME_LEN = 100
+MAX_PROMPT_LEN = 20000
+MAX_REPO_URL_LEN = 1000
+TRIAGE_LIMIT = 50
+
+ACTIVE = "active"
+CANCELLED = "cancelled"
+
+
+def row_get(row, name):
+    """行取值容错：sqlite3.Row 缺列报 IndexError、dict 缺键报
+    KeyError、空行报 TypeError，一律回 None(老库行缺列时
+    读自动化列表不能因此炸掉整页)。"""
+    try:
+        return row[name]
+    except (KeyError, IndexError, TypeError):
+        return None
+
+
+AUTOMATION_COLUMNS = (
+    "id", "project_id", "name", "prompt", "repo_url",
+    "interval_seconds", "status", "last_run_at", "last_claimed_at",
+    "created_by", "created_at", "cancelled_at")
+RUN_COLUMNS = (
+    "id", "automation_id", "project_id", "task_id", "status",
+    "detail", "created_at")
+
+
+def _row_dict(row, columns):
+    """把行统一成可按名取值的形态：dict 与 sqlite3.Row 原样
+    用；裸 sqlite3 连接给的是 tuple(按 SELECT 列顺序)，
+    按建表列顺序 zip 成 dict——直接调模块的调用方不一定设
+    row_factory，模块不能因此在取 row[id] 处炸掉。"""
+    if row is None or isinstance(row, dict):
+        return row
+    try:
+        row["id"]
+        return row
+    except TypeError:
+        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_automations_tables(conn):
+    """幂等建自动化表与产出记录表，并补齐新增列。
+
+    automations：一条定义；status 只取 active / cancelled，
+    取消是软删除(行保留、审计可查，名字可重用)。
+    automation_runs：一次执行产出；task_id 为空表示执行器
+    抢单后尚未建任务(只应短暂存在，执行器建任务后补记)。
+    调用方在写成功后 commit(server 的 handler 与 peer 模式
+    同款：建表后由业务写操作一并 commit)。"""
+    conn.execute(
+        "CREATE TABLE IF NOT EXISTS automations("
+        "id TEXT PRIMARY KEY, "
+        "project_id TEXT, "
+        "name TEXT, "
+        "prompt TEXT, "
+        "repo_url TEXT, "
+        "interval_seconds INTEGER, "
+        "status TEXT, "
+        "last_run_at DOUBLE PRECISION, "
+        "last_claimed_at DOUBLE PRECISION, "
+        "created_by TEXT, "
+        "created_at DOUBLE PRECISION, "
+        "cancelled_at DOUBLE PRECISION)"
+    )
+    conn.execute(
+        "CREATE TABLE IF NOT EXISTS automation_runs("
+        "id TEXT PRIMARY KEY, "
+        "automation_id TEXT, "
+        "project_id TEXT, "
+        "task_id TEXT, "
+        "status TEXT, "
+        "detail TEXT, "
+        "created_at DOUBLE PRECISION)"
+    )
+    existing = _existing_columns(conn, "automations")
+    for col, ddl in (
+            ("repo_url", "TEXT"),
+            ("last_claimed_at", "DOUBLE PRECISION"),
+            ("cancelled_at", "DOUBLE PRECISION")):
+        if col not in existing:
+            try:
+                conn.execute(
+                    "ALTER TABLE automations ADD COLUMN %s %s"
+                    % (col, ddl))
+            except Exception:
+                pass
+    existing_runs = _existing_columns(conn, "automation_runs")
+    for col, ddl in (
+            ("task_id", "TEXT"),
+            ("detail", "TEXT")):
+        if col not in existing_runs:
+            try:
+                conn.execute(
+                    "ALTER TABLE automation_runs ADD COLUMN %s %s"
+                    % (col, ddl))
+            except Exception:
+                pass
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_automations_project"
+        " ON automations(project_id, status)")
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_automation_runs"
+        " ON automation_runs(automation_id, created_at)")
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_automation_runs_project"
+        " ON automation_runs(project_id, created_at)")
+
+
+def next_run_at(auto, now=None):
+    """下次到期时刻：从未跑过时就是创建时间(建完即到期，
+    执行器下一轮即可认领)；跑过时为上次执行 + 间隔。"""
+    last = row_get(auto, "last_run_at")
+    interval = row_get(auto, "interval_seconds") or 0
+    if last is None:
+        return row_get(auto, "created_at")
+    return last + interval
+
+
+def is_due(auto, now=None):
+    """是否到期：仅 active 的定义可到期；now 缺省取当前时间。
+    到期边界用 >=，与 host run-due 的 now - last_run >=
+    interval 同款(含边界那一秒)。"""
+    if row_get(auto, "status") != ACTIVE:
+        return False
+    if now is None:
+        now = time.time()
+    nxt = next_run_at(auto)
+    if nxt is None:
+        return True
+    return now >= nxt
+
+
+def automation_to_dict(row, now=None):
+    row = _row_dict(row, AUTOMATION_COLUMNS)
+    nxt = next_run_at(row)
+    return {
+        "id": row["id"],
+        "project_id": row["project_id"],
+        "name": row["name"],
+        "prompt": row["prompt"],
+        "repo_url": row_get(row, "repo_url"),
+        "interval_seconds": row["interval_seconds"],
+        "status": row["status"],
+        "last_run_at": row_get(row, "last_run_at"),
+        "last_claimed_at": row_get(row, "last_claimed_at"),
+        "created_by": row["created_by"],
+        "created_at": row["created_at"],
+        "cancelled_at": row_get(row, "cancelled_at"),
+        "next_run_at": nxt,
+        "due": is_due(row, now),
+    }
+
+
+def _validate_fields(name, prompt, interval_seconds, repo_url):
+    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")
+    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 repo_url is not None:
+        if not isinstance(repo_url, str):
+            raise ValueError("repo_url must be a string")
+        repo_url = repo_url.strip() or None
+        if repo_url is not None and len(repo_url) > MAX_REPO_URL_LEN:
+            raise ValueError("repo_url too long")
+    return name, prompt, interval_seconds, repo_url
+
+
+def create_automation(conn, project_id, name, prompt, created_by,
+                      interval_seconds=86400, repo_url=None, now=None):
+    """创建自动化定义：同项目内 active 的名字唯一(取消过的
+    名字可重用，与环境名唯一规则同款语义)。成功时 commit
+    并返回 dict；校验失败抛 ValueError(调用方转 400)。"""
+    name, prompt, interval_seconds, repo_url = _validate_fields(
+        name, prompt, interval_seconds, repo_url)
+    if now is None:
+        now = time.time()
+    ensure_automations_tables(conn)
+    existing = conn.execute(
+        "SELECT id FROM automations WHERE project_id = ?"
+        " AND name = ? AND status = ?",
+        (project_id, name, ACTIVE)).fetchone()
+    if existing is not None:
+        raise ValueError("automation name already exists in this project")
+    auto_id = uuid.uuid4().hex
+    conn.execute(
+        "INSERT INTO automations(id, project_id, name, prompt,"
+        " repo_url, interval_seconds, status, last_run_at,"
+        " last_claimed_at, created_by, created_at, cancelled_at)"
+        " VALUES(?, ?, ?, ?, ?, ?, ?, NULL, NULL, ?, ?, NULL)",
+        (auto_id, project_id, name, prompt, repo_url,
+         interval_seconds, ACTIVE, created_by, now))
+    conn.commit()
+    row = conn.execute(
+        "SELECT * FROM automations WHERE id = ?", (auto_id,)).fetchone()
+    return automation_to_dict(row, now)
+
+
+def list_automations(conn, project_id, now=None):
+    """列出某项目全部自动化定义(含已取消，取消的 due 为
+    False)，按创建时间排序。调用方须先 ensure 表。"""
+    if now is None:
+        now = time.time()
+    ensure_automations_tables(conn)
+    rows = conn.execute(
+        "SELECT * FROM automations WHERE project_id = ?"
+        " ORDER BY created_at, id",
+        (project_id,)).fetchall()
+    return [automation_to_dict(r, now) for r in rows]
+
+
+def get_automation(conn, automation_id):
+    """取单条定义；不存在回 None(调用方转 404)。"""
+    ensure_automations_tables(conn)
+    row = conn.execute(
+        "SELECT * FROM automations WHERE id = ?", (automation_id,)
+    ).fetchone()
+    if row is None:
+        return None
+    return automation_to_dict(row)
+
+
+def cancel_automation(conn, automation_id, now=None):
+    """取消自动化(软删除)：仅 active 的行能取消。
+
+    返回 (ok, auto)：ok 为 True 时 auto 是取消后的 dict；
+    行不存在时 (False, None)；已取消时 (False, 旧 dict)，
+    调用方据此区分 404 与 409。成功时 commit。"""
+    if now is None:
+        now = time.time()
+    ensure_automations_tables(conn)
+    row = conn.execute(
+        "SELECT * FROM automations WHERE id = ?", (automation_id,)
+    ).fetchone()
+    if row is None:
+        return False, None
+    row = _row_dict(row, AUTOMATION_COLUMNS)
+    if row["status"] != ACTIVE:
+        return False, automation_to_dict(row, now)
+    cur = conn.execute(
+        "UPDATE automations SET status = ?, cancelled_at = ?"
+        " WHERE id = ? AND status = ?",
+        (CANCELLED, now, automation_id, ACTIVE))
+    conn.commit()
+    if cur.rowcount != 1:
+        return False, automation_to_dict(row, now)
+    row = conn.execute(
+        "SELECT * FROM automations WHERE id = ?", (automation_id,)
+    ).fetchone()
+    return True, automation_to_dict(row, now)
+
+
+def due_automations(conn, project_id, now=None):
+    """列出某项目当前到期的 active 定义(纯读，不抢单)。
+    SQL 层算到期：从未跑过(last_run_at 为空)或上次执行距
+    今已满间隔；与 is_due 的 Python 算法同款。"""
+    if now is None:
+        now = time.time()
+    ensure_automations_tables(conn)
+    rows = conn.execute(
+        "SELECT * FROM automations WHERE project_id = ?"
+        " AND status = ?"
+        " AND (last_run_at IS NULL"
+        " OR last_run_at + interval_seconds <= ?)"
+        " ORDER BY created_at, id",
+        (project_id, ACTIVE, now)).fetchall()
+    return [automation_to_dict(r, now) for r in rows]
+
+
+def claim_due_automation(conn, automation_id, now=None):
+    """抢单：未来执行器的唯一入口。
+
+    条件 UPDATE 只有一行能成功：同一自动化在同一间隔内，
+    两个并发执行器只有一个拿到 rowcount 1，另一个拿到 0
+    就必须放弃建任务——这是防重复执行的硬门。抢单即把
+    last_run_at 写成 now(不是等执行完才写，与 host run-due
+    相反)：好处是崩掉的执行不会在当间隔内被另一执行器重
+    跑；代价是抢单后崩掉的那次产出会缺失，直到下个间隔才
+    补跑。执行器抢单成功后必须调 record_automation_run 记
+    产出，否则分诊收件箱里这次执行不可见。成功时 commit。
+    返回 True 表示抢到，False 表示不存在、已取消或未到期。
+    """
+    if now is None:
+        now = time.time()
+    ensure_automations_tables(conn)
+    cur = conn.execute(
+        "UPDATE automations SET last_run_at = ?,"
+        " last_claimed_at = ?"
+        " WHERE id = ? AND status = ?"
+        " AND (last_run_at IS NULL"
+        " OR last_run_at + interval_seconds <= ?)",
+        (now, now, automation_id, ACTIVE, now))
+    conn.commit()
+    return cur.rowcount == 1
+
+
+def record_automation_run(conn, automation_id, project_id, task_id=None,
+                          status="created", detail="", now=None):
+    """记一次自动化产出：执行器建任务后调它，task_id 为建好
+    的云端任务 id；status 取 created / completed / failed 等
+    执行器自定的终态词(本模块不校验任务状态机，只原样存)。
+    自动化不存在时抛 ValueError(执行器绝不能给不存在的
+    定义记产出)。成功时 commit 并返回 run dict。"""
+    if now is None:
+        now = time.time()
+    ensure_automations_tables(conn)
+    auto = conn.execute(
+        "SELECT id FROM automations WHERE id = ?",
+        (automation_id,)).fetchone()
+    if auto is None:
+        raise ValueError("automation not found")
+    run_id = uuid.uuid4().hex
+    conn.execute(
+        "INSERT INTO automation_runs(id, automation_id, project_id,"
+        " task_id, status, detail, created_at)"
+        " VALUES(?, ?, ?, ?, ?, ?, ?)",
+        (run_id, automation_id, project_id, task_id, status,
+         detail or "", now))
+    conn.commit()
+    return {
+        "id": run_id,
+        "automation_id": automation_id,
+        "project_id": project_id,
+        "task_id": task_id,
+        "status": status,
+        "detail": detail or "",
+        "created_at": now,
+    }
+
+
+def list_automation_runs(conn, automation_id, limit=TRIAGE_LIMIT):
+    """列某自动化的最近产出记录，按时间倒序。"""
+    ensure_automations_tables(conn)
+    rows = conn.execute(
+        "SELECT * FROM automation_runs WHERE automation_id = ?"
+        " ORDER BY created_at DESC, id LIMIT ?",
+        (automation_id, limit)).fetchall()
+    out = []
+    for raw in rows:
+        r = _row_dict(raw, RUN_COLUMNS)
+        out.append({
+            "id": r["id"],
+            "automation_id": r["automation_id"],
+            "project_id": r["project_id"],
+            "task_id": row_get(r, "task_id"),
+            "status": r["status"],
+            "detail": row_get(r, "detail"),
+            "created_at": r["created_at"],
+        })
+    return out
+
+
+def triage_inbox(conn, project_id, now=None):
+    """分诊收件箱(只读，不写业务行、不改任务状态)：
+
+    - automations_due：当前到期的自动化定义。
+    - failed_tasks：失败或中断的任务(最需要人工看的项)，
+      按更新时间倒序。
+    - pending_followups：仍 pending 的后续消息(含任务间
+      消息排队的那部分，带 sender_task_id)，按创建时间排序；
+      followups 表不存在时该段为空(不因此炸掉整页)。
+    - recent_runs：最近的自动化产出记录，带对应任务的当前
+      状态(任务删了则 task_status 为 None，记录本身保留)。
+
+    调用方(server handler)须先 ensure 自动化表与 followups
+    表；本函数内 followups 缺表时只吞缺表错，其它错照常抛
+    出(调用方转 502，不伪造一页空收件箱)。"""
+    if now is None:
+        now = time.time()
+    ensure_automations_tables(conn)
+    due = due_automations(conn, project_id, now)
+    task_rows = conn.execute(
+        "SELECT id, status, prompt, created_at, updated_at,"
+        " error_text FROM tasks WHERE project_id = ?"
+        " AND status IN ('failed', 'interrupted')"
+        " ORDER BY updated_at DESC, id LIMIT ?",
+        (project_id, TRIAGE_LIMIT)).fetchall()
+    failed = []
+    for raw in task_rows:
+        r = _row_dict(raw, ("id", "status", "prompt", "created_at",
+                            "updated_at", "error_text"))
+        failed.append({
+            "id": r["id"],
+            "status": r["status"],
+            "prompt": (r["prompt"] or "")[:300],
+            "created_at": r["created_at"],
+            "updated_at": r["updated_at"],
+            "error_text": (row_get(r, "error_text") or "")[:300],
+        })
+    try:
+        fu_rows = conn.execute(
+            "SELECT f.id, f.task_id, f.body, f.mode, f.status,"
+            " f.sender_task_id, f.created_at FROM task_followups f"
+            " JOIN tasks t ON t.id = f.task_id"
+            " WHERE f.project_id = ? AND f.status = 'pending'"
+            " ORDER BY f.created_at, f.id LIMIT ?",
+            (project_id, TRIAGE_LIMIT)).fetchall()
+        pending = []
+        for raw in fu_rows:
+            r = _row_dict(raw, ("id", "task_id", "body", "mode",
+                                "status", "sender_task_id",
+                                "created_at"))
+            pending.append({
+                "id": r["id"],
+                "task_id": r["task_id"],
+                "body": (r["body"] or "")[:300],
+                "mode": r["mode"],
+                "status": r["status"],
+                "sender_task_id": row_get(r, "sender_task_id"),
+                "created_at": r["created_at"],
+            })
+    except Exception as exc:
+        msg = str(exc)
+        if "no such table" in msg or "does not exist" in msg:
+            pending = []
+        else:
+            raise
+    run_rows = conn.execute(
+        "SELECT r.id, r.automation_id, r.task_id, r.status,"
+        " r.detail, r.created_at, t.status AS task_status"
+        " FROM automation_runs r"
+        " LEFT JOIN tasks t ON t.id = r.task_id"
+        " WHERE r.project_id = ?"
+        " ORDER BY r.created_at DESC, r.id LIMIT ?",
+        (project_id, TRIAGE_LIMIT)).fetchall()
+    runs = []
+    for raw in run_rows:
+        r = _row_dict(raw, ("id", "automation_id", "task_id", "status",
+                            "detail", "created_at", "task_status"))
+        runs.append({
+            "id": r["id"],
+            "automation_id": r["automation_id"],
+            "task_id": row_get(r, "task_id"),
+            "status": r["status"],
+            "detail": row_get(r, "detail"),
+            "created_at": r["created_at"],
+            "task_status": row_get(r, "task_status"),
+        })
+    return {
+        "project_id": project_id,
+        "automations_due": due,
+        "failed_tasks": failed,
+        "pending_followups": pending,
+        "recent_runs": runs,
+        "counts": {
+            "automations_due": len(due),
+            "failed_tasks": len(failed),
+            "pending_followups": len(pending),
+            "recent_runs": len(runs),
+        },
+    }
diff --git a/cloud/caps.py b/cloud/caps.py
new file mode 100644
index 0000000..01364cd
--- /dev/null
+++ b/cloud/caps.py
@@ -0,0 +1,76 @@
+"""Worker 并发上限：配置解析与认领判断（纯函数，无 DB、无 IO）。
+
+配置来源与优先级（高到低）：
+1. 显式值：Worker 构造参数 capacity / CLI --worker-capacity。
+2. 环境变量 CLOUD_WORKER_CAPACITY（CLI 未给时生效）。
+3. 默认值 1：默认行为不得弱于现状，即旧串行行为。
+非法值（非整数、<= 0）一律归一化为 1，绝不出现 0 上限导致
+worker 永远认领不到任务，也绝不出现负数上限。
+
+workers 表的 capacity 是注册表展示值，不是控制面下推：
+本地 worker 只执行自己启动时的配置，表里被改大的值不会
+悄悄抬高本 worker 的上限（否则 stale 行或误写会放大并发）。
+
+上限口径（单 worker 本地计数）：
+- running 数 = Worker._running 的条目数（_running_lock 下读）。
+- 启动中的任务也算：_claim 占位时就先在 _running 里建空条目，
+  执行器起 proc / 容器后 _running_set 只往该条目里补字段，
+  不新增条目；所以“已认领未起进程”与“已起进程”都占一个槽。
+- 同一任务重复占位不重复计数（dict 按 task_id 去重）。
+- 本模块只管单 worker 上限；多 worker 全局上限不在此列，
+  见汇报里的诚实边界（项目级上限同样是遗留项）。
+"""
+
+import os
+
+DEFAULT_CAPACITY = 1
+ENV_VAR = "CLOUD_WORKER_CAPACITY"
+
+
+def parse_capacity(value):
+    """把任意输入归一化为 >= 1 的整数上限；非法值返回 1。"""
+    try:
+        capacity = int(value)
+    except (TypeError, ValueError):
+        return DEFAULT_CAPACITY
+    if capacity < 1:
+        return DEFAULT_CAPACITY
+    return capacity
+
+
+def resolve_capacity(explicit=None, env=None):
+    """按“显式 > 环境变量 > 默认 1”解析上限。
+
+    explicit 为 None 表示调用方未给（CLI 未传 --worker-capacity
+    时 argparse 默认即 None）；env 默认读进程环境，单测可注入
+    普通 dict。显式值即使非法也按非法处理（归一为 1），不回落
+    到环境变量：显式给错了就该按 1 跑，而不是悄悄用环境值。
+    """
+    if explicit is not None:
+        return parse_capacity(explicit)
+    source = env if env is not None else os.environ
+    try:
+        raw = source.get(ENV_VAR)
+    except Exception:
+        raw = None
+    if raw is None or (isinstance(raw, str) and not raw.strip()):
+        return DEFAULT_CAPACITY
+    return parse_capacity(raw)
+
+
+def effective_capacity(capacity):
+    """运行期上限：构造后 self.capacity 可能来自旧调用方，统一归一。"""
+    return parse_capacity(capacity)
+
+
+def running_count(running_snapshot):
+    """_running 快照的条目数：空条目（启动中）同样计数。"""
+    try:
+        return len(running_snapshot)
+    except TypeError:
+        return 0
+
+
+def can_claim(running_count_value, capacity):
+    """未满返回 True；running 数达到上限时不再认领。"""
+    return running_count_value < effective_capacity(capacity)
diff --git a/cloud/console.html b/cloud/console.html
index f2bdd7c..4674f2c 100644
--- a/cloud/console.html
+++ b/cloud/console.html
@@ -250,6 +250,24 @@
 <pre id="detailBody" style="display:none"></pre>
 <h3>子代理链路</h3>
 <div id="agentBox" class="muted">点任务行查看详情后，这里显示子代理链路</div>
+<h3>评审队列</h3>
+<div id="reviewForm">
+  <select id="reviewTargetType">
+    <option value="uncommittedChanges">未提交改动（工作区）</option>
+    <option value="baseBranch">与分支的差异</option>
+    <option value="commit">指定提交</option>
+    <option value="custom">自定义评审指令</option>
+  </select>
+  <input id="reviewBranch" placeholder="与分支的差异用：分支名（如 main）" style="width:100%">
+  <input id="reviewSha" placeholder="指定提交用：提交 sha" style="width:100%">
+  <textarea id="reviewInstructions" placeholder="自定义评审指令用：评审指令正文"></textarea>
+  <div class="bar">
+    <button id="btnCreateReview" class="primary" onclick="postReview()">发起评审</button>
+    <button onclick="if (currentTask) loadReviews(currentTask.id)">刷新队列</button>
+    <span id="reviewMsg" class="muted"></span>
+  </div>
+</div>
+<div id="reviewList" class="muted">点任务行查看详情后，点“刷新队列”查看评审记录</div>
 <h3>实时链路</h3>
 <div id="traceBox" class="muted">点任务行查看详情后，这里显示过程链路</div>
 <div id="costBox"></div>
@@ -263,6 +281,10 @@
   </div>
 </div>
 <div id="followupList" class="muted"></div>
+<h3>运行时审批队列</h3>
+<div class="bar"><button id="btnRefreshApprovals" onclick="if(currentTask)loadApprovals(currentTask.id)">刷新审批队列</button><span id="approvalMsg" class="muted"></span></div>
+<div class="muted">待审批项需人工点允许或拒绝；超时默认拒绝。本队列的决定会留回执；实时请求的事前拦截待 host 接入后生效，在此之前实时行为不变。</div>
+<div id="approvalList" class="muted"></div>
 <h3>给其他任务发消息</h3>
 <div id="peerMessageForm">
   <input id="peerTarget" placeholder="目标任务 id（与本任务同项目；对方运行中排队执行、已终态则续跑复活，绝不打断当前轮）" style="width:100%">
@@ -274,6 +296,22 @@
   </div>
 </div>
 <div id="peerMessageList" class="muted"></div>
+<h3>自动化</h3>
+<div id="automationForm">
+  <input id="autoName" placeholder="自动化名称(同项目内启用中的名称唯一)" style="width:100%">
+  <textarea id="autoPrompt" placeholder="定时任务内容(创建只登记定义与到期状态；执行由后续执行器经 claim 抢单后建任务，不另起调度器)"></textarea>
+  <div class="bar">
+    <input id="autoInterval" placeholder="间隔秒数(默认 86400，最小 60)" style="width:40%">
+    <input id="autoRepo" placeholder="仓库(可选)" style="width:40%">
+    <button id="btnCreateAutomation" class="primary">创建自动化</button>
+    <button id="btnLoadAutomations">刷新自动化</button>
+    <span id="autoMsg" class="muted"></span>
+  </div>
+</div>
+<div id="automationList" class="muted"></div>
+<h3>分诊收件箱</h3>
+<div class="bar"><button id="btnLoadTriage">刷新分诊</button><span id="triageMsg" class="muted"></span></div>
+<div id="triageBox" class="muted"></div>
 </section>
 
 <script>
@@ -399,14 +437,46 @@ async function loadProjects() {
   } catch (e) { $("#projectMsg").textContent = "加载失败：" + e.message; }
 }
 
+async function loadApprovals(taskId) {
+  var box = $("#approvalList");
+  try {
+    var d = await api("/tasks/" + encodeURIComponent(taskId) + "/approvals", "GET");
+    var list = d.approvals || [];
+    if (!list.length) { box.innerHTML = "（暂无审批项）"; return; }
+    box.innerHTML = list.map(function (a) {
+      var receipt = "";
+      if (a.status === "approved" || a.status === "denied" || a.status === "expired") {
+        receipt = "｜回执：" + esc(a.decision || a.status) + " by " + esc(String(a.decided_by || "").slice(0, 8)) + " " + fmt(a.decided_at);
+      }
+      var btns = "";
+      if (a.status === "pending") {
+        btns = " <button onclick=\"decideApproval('" + esc(a.id) + "', 'approved')\">允许</button>" +
+          " <button onclick=\"decideApproval('" + esc(a.id) + "', 'denied')\">拒绝</button>";
+      }
+      return "[" + esc(a.kind) + "｜" + esc(a.status) + "] " + esc(a.summary) + receipt + btns;
+    }).join("<br>");
+  } catch (e) { box.innerHTML = "<span class='err'>审批队列加载失败：" + esc(e.message) + "</span>"; }
+}
+async function decideApproval(approvalId, decision) {
+  if (!currentTask) return;
+  var verb = decision === "approved" ? "approve" : "deny";
+  try {
+    await api("/tasks/" + encodeURIComponent(currentTask.id) + "/approvals/" + encodeURIComponent(approvalId) + "/" + verb, "POST", {});
+    $("#approvalMsg").innerHTML = "<span class='ok'>已" + (decision === "approved" ? "允许" : "拒绝") + "</span>";
+    loadApprovals(currentTask.id);
+  } catch (e) { $("#approvalMsg").innerHTML = "<span class='err'>决定失败：" + esc(e.message) + "</span>"; }
+}
 async function loadWorkers() {
   try {
     var d = await api("/workers", "GET");
     var list = d.workers || d || [];
     if (!list.length) { $("#workers").textContent = "无 worker 注册"; return; }
     $("#workers").innerHTML = list.map(function (w) {
+      var run = (w.running === undefined || w.running === null) ? "?" : w.running;
+      var cap = (w.capacity === undefined || w.capacity === null) ? "?" : w.capacity;
+      var live = (w.alive === false) ? " / 离线（占用为上次上报值）" : "";
       return esc(w.worker_id || w.id || "?") + " / " + esc(w.executor || "") +
-        " / capacity " + esc(w.capacity || "") +
+        " / 占用 " + esc(run) + "/" + esc(cap) + live +
         " / last_seen " + esc(w.last_seen || w.last_seen_seconds_ago || "");
     }).join("<br>");
   } catch (e) { $("#workers").textContent = "加载失败：" + e.message; }
@@ -639,6 +709,52 @@ async function startTrace(taskId, status) {
   if (done) { loadTasks(); showDetail(taskId); }
 }
 
+async function loadReviews(taskId) {
+  var box = $("#reviewList");
+  try {
+    var d = await api("/tasks/" + encodeURIComponent(taskId) + "/reviews", "GET");
+    var list = d.reviews || [];
+    if (!list.length) { box.innerHTML = "（该任务暂无评审记录）"; return; }
+    box.innerHTML = list.map(function (r) {
+      var target = esc(r.target_type);
+      if (r.branch) target += "：" + esc(r.branch);
+      if (r.sha) target += "：" + esc(String(r.sha).slice(0, 12));
+      if (r.instructions) target += "：" + esc(String(r.instructions).slice(0, 80));
+      var html = "<div style='margin:6px 0;padding:8px 10px;border:1px solid var(--border);border-radius:var(--radius);background:#fff'>";
+      html += "<b>评审</b> " + esc(String(r.id).slice(0, 8)) + " ｜ " + target +
+        " ｜ <span class='st-" + esc(r.status) + "'>" + esc(r.status) + "</span>" +
+        (r.created_at ? " ｜ " + fmt(r.created_at) : "") + "<br>";
+      html += r.findings ? esc(String(r.findings).slice(0, 1000)) : "<span class='muted'>（暂无 findings：评审尚未执行或尚未写回）</span>";
+      html += "<br><button onclick=\"submitReviewFindings('" + esc(r.id) + "')\">写 findings</button>";
+      html += "</div>";
+      return html;
+    }).join("");
+  } catch (e) { box.innerHTML = "<span class='err'>评审队列加载失败：" + esc(e.message) + "</span>"; }
+}
+async function postReview() {
+  if (!currentTask) return;
+  var body = { target_type: $("#reviewTargetType").value };
+  var branch = $("#reviewBranch").value.trim();
+  var sha = $("#reviewSha").value.trim();
+  var instructions = $("#reviewInstructions").value.trim();
+  if (branch) body.branch = branch;
+  if (sha) body.sha = sha;
+  if (instructions) body.instructions = instructions;
+  try {
+    await api("/tasks/" + encodeURIComponent(currentTask.id) + "/reviews", "POST", body);
+    $("#reviewMsg").innerHTML = "<span class='ok'>评审已登记（requested）</span>";
+    loadReviews(currentTask.id);
+  } catch (e) { $("#reviewMsg").innerHTML = "<span class='err'>发起评审失败：" + esc(e.message) + "</span>"; }
+}
+async function submitReviewFindings(reviewId) {
+  if (!currentTask) return;
+  var text = prompt("输入该条评审的 findings（将写回记录并标为 completed）：");
+  if (text === null) return;
+  try {
+    await api("/tasks/" + encodeURIComponent(currentTask.id) + "/reviews/" + encodeURIComponent(reviewId) + "/findings", "POST", { findings: text });
+    loadReviews(currentTask.id);
+  } catch (e) { $("#reviewMsg").innerHTML = "<span class='err'>写 findings 失败：" + esc(e.message) + "</span>"; }
+}
 async function showDetail(taskId) {
   stopTrace();
   try {
@@ -696,6 +812,75 @@ async function doAction(kind) {
   } catch (e) { $("#actionMsg").innerHTML = "<span class='err'>" + kind + " 失败：" + esc(e.message) + "</span>"; }
 }
 
+async function loadAutomations() {
+  var pid = $("#projectSelect").value;
+  var box = $("#automationList");
+  if (!pid) { box.innerHTML = "(先选项目)"; return; }
+  try {
+    var d = await api("/projects/" + encodeURIComponent(pid) + "/automations", "GET");
+    var list = d.automations || [];
+    if (!list.length) { box.innerHTML = "(暂无自动化)"; return; }
+    box.innerHTML = list.map(function (a) {
+      var html = "[" + esc(a.status) + (a.due ? "｜到期" : "") + "] " + esc(a.name) +
+        "｜每 " + esc(a.interval_seconds) + " 秒｜下次 " + esc(fmt(a.next_run_at));
+      if (a.status === "active") html += " <button data-auto-cancel='" + esc(a.id) + "'>取消</button>";
+      return html;
+    }).join("<br>");
+    box.querySelectorAll("[data-auto-cancel]").forEach(function (btn) {
+      btn.onclick = function () { cancelAutomation(btn.getAttribute("data-auto-cancel")); };
+    });
+  } catch (e) { box.innerHTML = "<span class='err'>自动化加载失败：" + esc(e.message) + "</span>"; }
+}
+async function createAutomation() {
+  var pid = $("#projectSelect").value;
+  if (!pid) { $("#autoMsg").textContent = "先选项目"; return; }
+  var name = $("#autoName").value.trim();
+  var prompt = $("#autoPrompt").value.trim();
+  if (!name) { $("#autoMsg").textContent = "自动化名称必填"; return; }
+  if (!prompt) { $("#autoMsg").textContent = "定时任务内容必填"; return; }
+  var body = { name: name, prompt: prompt };
+  var iv = $("#autoInterval").value.trim();
+  if (iv) body.interval_seconds = parseInt(iv, 10);
+  var repo = $("#autoRepo").value.trim();
+  if (repo) body.repo_url = repo;
+  try {
+    await api("/projects/" + encodeURIComponent(pid) + "/automations", "POST", body);
+    $("#autoMsg").innerHTML = "<span class='ok'>自动化已创建(登记定义，到期后由执行器认领)</span>";
+    $("#autoName").value = ""; $("#autoPrompt").value = ""; $("#autoInterval").value = ""; $("#autoRepo").value = "";
+    loadAutomations();
+  } catch (e) { $("#autoMsg").innerHTML = "<span class='err'>创建失败：" + esc(e.message) + "</span>"; }
+}
+async function cancelAutomation(autoId) {
+  var pid = $("#projectSelect").value;
+  if (!pid) return;
+  try {
+    await api("/projects/" + encodeURIComponent(pid) + "/automations/" + encodeURIComponent(autoId) + "/cancel", "POST", {});
+    $("#autoMsg").innerHTML = "<span class='ok'>自动化已取消</span>";
+    loadAutomations();
+  } catch (e) { $("#autoMsg").innerHTML = "<span class='err'>取消失败：" + esc(e.message) + "</span>"; }
+}
+async function loadTriage() {
+  var pid = $("#projectSelect").value;
+  var box = $("#triageBox");
+  if (!pid) { box.innerHTML = "(先选项目)"; return; }
+  try {
+    var d = await api("/projects/" + encodeURIComponent(pid) + "/triage", "GET");
+    var html = "";
+    html += "<b>到期自动化(" + (d.counts.automations_due || 0) + ")</b><br>";
+    html += (d.automations_due || []).length ? (d.automations_due || []).map(function (a) { return esc(a.name) + "｜每 " + esc(a.interval_seconds) + " 秒"; }).join("<br>") : "(无)";
+    html += "<br><b>失败/中断任务(" + (d.counts.failed_tasks || 0) + ")</b><br>";
+    html += (d.failed_tasks || []).length ? (d.failed_tasks || []).map(function (t) { return esc(String(t.id).slice(0, 8)) + "｜" + esc(t.status) + "｜" + esc(String(t.prompt || "").slice(0, 80)); }).join("<br>") : "(无)";
+    html += "<br><b>待处理后续消息(" + (d.counts.pending_followups || 0) + ")</b><br>";
+    html += (d.pending_followups || []).length ? (d.pending_followups || []).map(function (f) { return "任务 " + esc(String(f.task_id).slice(0, 8)) + "｜" + esc(String(f.body || "").slice(0, 80)); }).join("<br>") : "(无)";
+    html += "<br><b>最近自动化产出(" + (d.counts.recent_runs || 0) + ")</b><br>";
+    html += (d.recent_runs || []).length ? (d.recent_runs || []).map(function (r) { return esc(String(r.automation_id).slice(0, 8)) + "｜" + esc(r.status) + (r.task_id ? "｜任务 " + esc(String(r.task_id).slice(0, 8)) : "") + (r.task_status ? "｜" + esc(r.task_status) : ""); }).join("<br>") : "(无)";
+    box.innerHTML = html;
+  } catch (e) { box.innerHTML = "<span class='err'>分诊加载失败：" + esc(e.message) + "</span>"; }
+}
+$("#btnCreateAutomation").onclick = createAutomation;
+$("#btnLoadAutomations").onclick = loadAutomations;
+$("#btnLoadTriage").onclick = loadTriage;
+$("#projectSelect").addEventListener("change", function () { loadAutomations(); loadTriage(); });
 async function loadFollowups(taskId) {
   var box = $("#followupList");
   try {
diff --git a/cloud/review.py b/cloud/review.py
new file mode 100644
index 0000000..f8e60b7
--- /dev/null
+++ b/cloud/review.py
@@ -0,0 +1,189 @@
+"""自动代码评审：评审队列记录。
+
+评审队列是服务端的一张登记表，不跑模型：调用方对某个任务
+发起评审时记一条 requested 记录，评审执行方（目前是
+host.py code-review，走协议 review/start）跑完后把 findings
+写回同一条记录，控制台评审区据此展示队列与结果。
+
+与 peer_messages 同款分工：本模块只含纯表操作与协议目标
+构造，不 import cloud 其它模块，避免循环 import；服务端
+的表迁移委托到 ensure_reviews_table，失败不伪造成功。
+
+协议锚点（protocol/codex_app_server_protocol.v2.schemas.json，
+逐个对过枚举，不编造）：
+- 方法名 review/start，参数 ReviewStartParams：threadId 与
+  target 必填，delivery 可空（inline 为默认，detached 已
+  废弃，故本模块不产生 detached）。
+- ReviewTarget 四种：uncommittedChanges（工作区未提交改动）、
+  baseBranch（须带 branch）、commit（须带 sha，title 可空）、
+  custom（须带 instructions）。
+
+状态机：requested（已登记，未执行）-> running（执行中，
+目前仅 host CLI 执行时可由调用方先置位，服务端不自动
+流转）-> completed / failed（写 findings 时落终态）。
+服务端不替执行方编造 findings：findings 为空时记录仍是
+requested，控制台如实显示“暂无 findings”。
+"""
+
+import time
+import uuid
+
+REVIEW_TARGET_TYPES = (
+    "uncommittedChanges", "baseBranch", "commit", "custom")
+REVIEW_STATUSES = ("requested", "running", "completed", "failed")
+
+_REVIEW_COLUMNS = (
+    "id", "task_id", "project_id", "target_type", "branch", "sha",
+    "instructions", "status", "findings", "created_by",
+    "created_at", "updated_at")
+
+
+def build_review_target(target_type, branch=None, sha=None,
+                        instructions=None):
+    """按协议构造 ReviewTarget dict；组合非法时 ValueError，
+    调用方转 400，绝不把非法目标发给 app-server。"""
+    if target_type == "uncommittedChanges":
+        return {"type": "uncommittedChanges"}
+    if target_type == "baseBranch":
+        if not branch or not isinstance(branch, str) or not branch.strip():
+            raise ValueError("baseBranch target requires branch")
+        return {"type": "baseBranch", "branch": branch.strip()}
+    if target_type == "commit":
+        if not sha or not isinstance(sha, str) or not sha.strip():
+            raise ValueError("commit target requires sha")
+        return {"type": "commit", "sha": sha.strip()}
+    if target_type == "custom":
+        if not instructions or not isinstance(instructions, str) \
+                or not instructions.strip():
+            raise ValueError("custom target requires instructions")
+        return {"type": "custom", "instructions": instructions.strip()}
+    raise ValueError("unknown target_type: %r" % (target_type,))
+
+
+def ensure_reviews_table(conn):
+    """幂等建评审队列表。调用方（服务端接口）在每次读写前
+    调它，老库第一次跑到这里时自动补表，不丢数据。"""
+    conn.execute(
+        "CREATE TABLE IF NOT EXISTS task_reviews("
+        "id TEXT PRIMARY KEY, "
+        "task_id TEXT, "
+        "project_id TEXT, "
+        "target_type TEXT, "
+        "branch TEXT, "
+        "sha TEXT, "
+        "instructions TEXT, "
+        "status TEXT, "
+        "findings TEXT, "
+        "created_by TEXT, "
+        "created_at DOUBLE PRECISION, "
+        "updated_at DOUBLE PRECISION)"
+    )
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_reviews_task"
+        " ON task_reviews(task_id, created_at)")
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_reviews_project"
+        " ON task_reviews(project_id, created_at)")
+
+
+def _row_to_dict(row):
+    """行转 dict：兼容 dict 行、sqlite3.Row 与裸元组（按
+    _REVIEW_COLUMNS 的 SELECT 顺序）。"""
+    if row is None:
+        return None
+    if isinstance(row, dict):
+        return {c: row.get(c) for c in _REVIEW_COLUMNS}
+    try:
+        keys = row.keys()
+    except AttributeError:
+        keys = None
+    if keys is not None:
+        d = dict(row)
+        return {c: d.get(c) for c in _REVIEW_COLUMNS}
+    return dict(zip(_REVIEW_COLUMNS, row))
+
+
+def _select_cols():
+    return ", ".join(_REVIEW_COLUMNS)
+
+
+def create_review(conn, task_id, project_id, target_type, created_by,
+                  branch=None, sha=None, instructions=None):
+    """登记一条评审，返回记录 dict。先校验目标组合（非法
+    时 ValueError，调用方转 400，表里不留脏行），成功才
+    commit。"""
+    build_review_target(target_type, branch, sha, instructions)
+    ensure_reviews_table(conn)
+    now = time.time()
+    review_id = uuid.uuid4().hex
+    conn.execute(
+        "INSERT INTO task_reviews(id, task_id, project_id,"
+        " target_type, branch, sha, instructions, status,"
+        " findings, created_by, created_at, updated_at)"
+        " VALUES(?, ?, ?, ?, ?, ?, ?, 'requested', NULL, ?, ?, ?)",
+        (review_id, task_id, project_id, target_type,
+         branch.strip() if isinstance(branch, str) else None,
+         sha.strip() if isinstance(sha, str) else None,
+         instructions.strip() if isinstance(instructions, str) else None,
+         created_by, now, now),
+    )
+    conn.commit()
+    return get_review(conn, review_id)
+
+
+def list_reviews(conn, task_id):
+    """某任务的评审队列，按创建时间升序（先登记的先审）。"""
+    ensure_reviews_table(conn)
+    rows = conn.execute(
+        "SELECT %s FROM task_reviews WHERE task_id = ?"
+        " ORDER BY created_at, id" % _select_cols(),
+        (task_id,)).fetchall()
+    return [_row_to_dict(r) for r in rows]
+
+
+def list_project_reviews(conn, project_id):
+    """某项目的评审队列（供未来项目级队列端点复用；本批
+    服务端只暴露任务级队列，函数先备好不接线）。"""
+    ensure_reviews_table(conn)
+    rows = conn.execute(
+        "SELECT %s FROM task_reviews WHERE project_id = ?"
+        " ORDER BY created_at, id" % _select_cols(),
+        (project_id,)).fetchall()
+    return [_row_to_dict(r) for r in rows]
+
+
+def get_review(conn, review_id):
+    """取单条评审记录，不存在时返回 None（调用方转 404）。"""
+    ensure_reviews_table(conn)
+    row = conn.execute(
+        "SELECT %s FROM task_reviews WHERE id = ?" % _select_cols(),
+        (review_id,)).fetchone()
+    return _row_to_dict(row)
+
+
+def update_review(conn, review_id, status=None, findings=None):
+    """更新评审状态与 findings，返回更新后记录，不存在时
+    返回 None。status 非法时 ValueError（调用方转 400）；
+    findings 为字符串时照存（空串也存，区别于未写回的
+    NULL）；不传 status 时 findings 写回默认落 completed，
+    即“有结果即完成”，调用方可显式传 failed 覆盖。"""
+    if status is not None and status not in REVIEW_STATUSES:
+        raise ValueError("unknown review status: %r" % (status,))
+    if findings is not None and not isinstance(findings, str):
+        raise ValueError("findings must be a string")
+    ensure_reviews_table(conn)
+    record = get_review(conn, review_id)
+    if record is None:
+        return None
+    new_status = status
+    if new_status is None:
+        new_status = "completed" if findings is not None \
+            else record["status"]
+    new_findings = findings if findings is not None \
+        else record["findings"]
+    conn.execute(
+        "UPDATE task_reviews SET status = ?, findings = ?,"
+        " updated_at = ? WHERE id = ?",
+        (new_status, new_findings, time.time(), review_id))
+    conn.commit()
+    return get_review(conn, review_id)
diff --git a/cloud/server.py b/cloud/server.py
index 7c59495..65e37d7 100644
--- a/cloud/server.py
+++ b/cloud/server.py
@@ -66,6 +66,17 @@ from cloud.peer_messages import deliver_peer_message as _deliver_peer_message
 from cloud.peer_messages import ensure_followups_table as _peer_ensure_followups
 from cloud.peer_messages import peer_message_prefix as _peer_prefix
 from cloud.peer_messages import row_get as _row_get
+from cloud.automations import cancel_automation as _auto_cancel
+from cloud.automations import create_automation as _auto_create
+from cloud.automations import ensure_automations_tables as _auto_ensure
+from cloud.automations import list_automations as _auto_list
+from cloud.automations import triage_inbox as _auto_triage
+from cloud.approvals import create_approval as _appr_create
+from cloud.approvals import decide_approval as _appr_decide
+from cloud.approvals import ensure_approvals_table as _appr_ensure
+from cloud.approvals import expire_overdue as _appr_expire
+from cloud.approvals import get_approval as _appr_get
+from cloud.approvals import list_approvals as _appr_list
 from cloud.kb import KBStore
 from cloud.memory import MemoryStore
 from cloud.pr import PrSkipped, PrStepError
@@ -1480,8 +1491,27 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
                 self._create_peer_message(user_id, rest[: -len("/messages")])
             elif rest.endswith("/messages") and method == "GET":
                 self._list_peer_messages(user_id, rest[: -len("/messages")])
+            elif rest.endswith("/reviews") and method == "POST":
+                self._create_review(user_id, rest[: -len("/reviews")])
+            elif rest.endswith("/reviews") and method == "GET":
+                self._list_reviews(user_id, rest[: -len("/reviews")])
+            elif "/reviews/" in rest and rest.endswith("/findings") \
+                    and method == "POST":
+                _task_part, _rev_part = rest.split("/reviews/", 1)
+                self._update_review_findings(
+                    user_id, _task_part, _rev_part.split("/")[0])
             elif rest.endswith("/fork") and method == "POST":
                 self._fork_task(user_id, rest[: -len("/fork")])
+            elif rest.endswith("/approvals") and method == "GET":
+                self._list_approvals(user_id, rest[: -len("/approvals")])
+            elif rest.endswith("/approvals") and method == "POST":
+                self._create_approval(user_id, rest[: -len("/approvals")])
+            elif rest.endswith("/approve") and method == "POST":
+                self._decide_approval(
+                    user_id, rest[: -len("/approve")], "approved")
+            elif rest.endswith("/deny") and method == "POST":
+                self._decide_approval(
+                    user_id, rest[: -len("/deny")], "denied")
             elif rest.endswith("/continue") and method == "POST":
                 self._continue_task(user_id, rest[: -len("/continue")])
             else:
@@ -1974,6 +2004,46 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
                                    subscription_id=parts[3])
             self._send_json(200, {"deleted": bool(deleted)})
             return
+        # /projects/{id}/automations
+        if len(parts) == 2 and parts[1] == "automations":
+            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._automation_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._automation_list(project_id)
+            else:
+                self._send_json(404, {"error": "not found"})
+            return
+        # /projects/{id}/automations/{aid}/cancel
+        if len(parts) == 4 and parts[1] == "automations" \
+                and parts[3] == "cancel":
+            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._automation_cancel(user_id, project_id, parts[2])
+            return
+        # /projects/{id}/triage
+        if len(parts) == 2 and parts[1] == "triage":
+            if method != "GET":
+                self._send_json(404, {"error": "not found"})
+                return
+            err = self._check_project(user_id, project_id, "read")
+            if err:
+                self._send_json(err[0], {"error": err[1]})
+                return
+            self._triage_inbox(project_id)
+            return
         self._send_json(404, {"error": "not found"})
 
     def _memory_create(self, user_id, project_id):
@@ -2998,6 +3068,127 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
         self._send_json(200, {
             "messages": [_followup_to_dict(r) for r in rows]})
 
+    def _create_review(self, user_id, task_id):
+        """发起评审：对某任务登记一条评审队列记录（status
+        requested），不跑模型、不伪造 findings——执行方跑完
+        后经 /findings 写回。权限：项目 create_task；目标组合
+        非法时 400，表里不留脏行；记审计 review_requested。"""
+        from cloud.review import build_review_target
+        from cloud.review import create_review as _review_create
+        body, err = self._read_json()
+        if err:
+            self._send_json(400, {"error": err})
+            return
+        row = self._fetch_task(task_id)
+        if row is None:
+            self._send_json(404, {"error": "task not found"})
+            return
+        task = _task_to_dict(row)
+        err = self._check_project(user_id, task["project_id"],
+                                  "create_task")
+        if err:
+            self._send_json(err[0], {"error": err[1]})
+            return
+        target_type = body.get("target_type", "uncommittedChanges")
+        branch = body.get("branch")
+        sha = body.get("sha")
+        instructions = body.get("instructions")
+        for name, value in (("branch", branch), ("sha", sha),
+                            ("instructions", instructions)):
+            if value is not None and not isinstance(value, str):
+                self._send_json(400, {"error": "%s must be a string" % name})
+                return
+        try:
+            build_review_target(target_type, branch, sha, instructions)
+        except ValueError as exc:
+            self._send_json(400, {"error": str(exc)})
+            return
+        conn = self._db()
+        try:
+            record = _review_create(
+                conn, task_id, task["project_id"], target_type,
+                user_id, branch=branch, sha=sha,
+                instructions=instructions)
+        finally:
+            conn.close()
+        with _auth_store(self.server.db_path) as store:
+            store.audit(user_id, task["project_id"],
+                        "review_requested",
+                        "%s -> %s" % (task_id, record["id"]))
+        self._send_json(201, record)
+
+    def _list_reviews(self, user_id, task_id):
+        """评审队列查看：列出某任务的全部评审记录（含 findings，
+        未写回的 findings 为 null，如实展示不编造）。权限：
+        项目 read。"""
+        from cloud.review import list_reviews as _review_list
+        row = self._fetch_task(task_id)
+        if row is None:
+            self._send_json(404, {"error": "task not found"})
+            return
+        task = _task_to_dict(row)
+        err = self._check_project(user_id, task["project_id"], "read")
+        if err:
+            self._send_json(err[0], {"error": err[1]})
+            return
+        conn = self._db()
+        try:
+            reviews = _review_list(conn, task_id)
+        finally:
+            conn.close()
+        self._send_json(200, {"reviews": reviews})
+
+    def _update_review_findings(self, user_id, task_id, review_id):
+        """评审结果写回：执行方把 findings 记到指定记录上。
+        不传 status 时 findings 写回默认落 completed；显式传
+        failed 可标失败。权限：项目 create_task；记录不属于
+        该任务时 404；记审计 review_findings_updated。"""
+        from cloud.review import get_review as _review_get
+        from cloud.review import update_review as _review_update
+        body, err = self._read_json()
+        if err:
+            self._send_json(400, {"error": err})
+            return
+        row = self._fetch_task(task_id)
+        if row is None:
+            self._send_json(404, {"error": "task not found"})
+            return
+        task = _task_to_dict(row)
+        err = self._check_project(user_id, task["project_id"],
+                                  "create_task")
+        if err:
+            self._send_json(err[0], {"error": err[1]})
+            return
+        findings = body.get("findings")
+        status = body.get("status")
+        if findings is None and status is None:
+            self._send_json(400, {
+                "error": "findings or status required"})
+            return
+        if findings is not None and not isinstance(findings, str):
+            self._send_json(400, {"error": "findings must be a string"})
+            return
+        conn = self._db()
+        try:
+            record = _review_get(conn, review_id)
+            if record is None or record["task_id"] != task_id:
+                self._send_json(404, {"error": "review not found"})
+                return
+            try:
+                updated = _review_update(conn, review_id,
+                                         status=status,
+                                         findings=findings)
+            except ValueError as exc:
+                self._send_json(400, {"error": str(exc)})
+                return
+        finally:
+            conn.close()
+        with _auth_store(self.server.db_path) as store:
+            store.audit(user_id, task["project_id"],
+                        "review_findings_updated",
+                        "%s -> %s" % (task_id, review_id))
+        self._send_json(200, updated)
+
     def _fork_task(self, user_id, task_id):
         """会话分支：从已结束任务的 thread 分叉出新任务。
 
@@ -3180,6 +3371,136 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
             store.audit(user_id, task["project_id"], "merge_task", task_id)
         self._send_json(200, task)
 
+    # -- 运行时审批队列 ------------------------------------------------
+    def _list_approvals(self, user_id, task_id):
+        """列出某任务的审批项（含已决定的，回执一并带出）。
+
+        权限：项目 read。列之前先扫荡过期：已到期的 pending
+        在这里变成 expired（effective decision 为 denied），
+        列表里看到的就是最终态，不会把已过期项继续显示为
+        待审批。下游异常不拦：抛给 _dispatch_safe 回 500，
+        绝不回一条伪造的空列表或成功。"""
+        row = self._fetch_task(task_id)
+        if row is None:
+            self._send_json(404, {"error": "task not found"})
+            return
+        task = _task_to_dict(row)
+        with _auth_store(self.server.db_path) as store:
+            if not store.check_permission(user_id, task["project_id"],
+                                          "read"):
+                self._send_json(403, {"error": "forbidden"})
+                return
+        conn = self._db()
+        try:
+            _appr_ensure(conn)
+            _appr_expire(conn)
+            items = _appr_list(conn, task_id=task_id)
+        finally:
+            conn.close()
+        self._send_json(200, {"approvals": items})
+
+    def _create_approval(self, user_id, task_id):
+        """给某任务建一条待审批项（系统入口：未来 host hook 与
+        手动补录都走这里，实时 elicitation 本片未接，见
+        approvals 模块 docstring 的诚实边界）。
+
+        权限：项目 create_task（建审批项是写动作，与后续消息
+        同级）。kind 缺省 elicitation；summary 必填；
+        timeout_seconds 只接受 60 到 86400 秒整数，缺省 3600，
+        超出范围回 400。成功记审计 approval_created。"""
+        row = self._fetch_task(task_id)
+        if row is None:
+            self._send_json(404, {"error": "task not found"})
+            return
+        task = _task_to_dict(row)
+        with _auth_store(self.server.db_path) as store:
+            if not store.check_permission(user_id, task["project_id"],
+                                          "create_task"):
+                self._send_json(403, {"error": "forbidden"})
+                return
+        body, err = self._read_json()
+        if err:
+            self._send_json(400, {"error": err})
+            return
+        kind = body.get("kind", "elicitation")
+        summary = body.get("summary")
+        detail = body.get("detail")
+        if detail is not None and not isinstance(
+                detail, (dict, list, str)):
+            self._send_json(400, {"error": "detail must be an object"})
+            return
+        timeout = body.get("timeout_seconds", 3600)
+        if not isinstance(timeout, int) or isinstance(timeout, bool) \
+                or timeout < 60 or timeout > 86400:
+            self._send_json(
+                400, {"error": "timeout_seconds must be 60..86400"})
+            return
+        conn = self._db()
+        try:
+            try:
+                item = _appr_create(
+                    conn, task_id, task["project_id"], kind=kind,
+                    summary=summary, detail=detail,
+                    requested_by=body.get("requested_by"),
+                    timeout_seconds=timeout)
+            except ValueError as exc:
+                self._send_json(400, {"error": str(exc)})
+                return
+        finally:
+            conn.close()
+        with _auth_store(self.server.db_path) as store:
+            store.audit(user_id, task["project_id"],
+                        "approval_created",
+                        "%s %s" % (item["id"], item["kind"]))
+        self._send_json(201, {"approval": item})
+
+    def _decide_approval(self, user_id, prefix, decision):
+        """批准或拒绝一条审批项：prefix 形如
+        <task_id>/approvals/<approval_id>（路由从 /approve
+        或 /deny 后缀剥出来的剩余部分）。
+
+        权限：项目 create_task（决定是写动作）。审批项不属于
+        该任务时回 404，不跨任务决定。已终态（批准、拒绝或
+        过期）时回 409 并带回执，原决定绝不被覆盖。成功记
+        审计 approval_decided，回执里有 decided_by 与
+        decided_at。"""
+        parts = prefix.split("/approvals/")
+        if len(parts) != 2 or not parts[0] or not parts[1]:
+            self._send_json(404, {"error": "not found"})
+            return
+        task_id, approval_id = parts
+        row = self._fetch_task(task_id)
+        if row is None:
+            self._send_json(404, {"error": "task not found"})
+            return
+        task = _task_to_dict(row)
+        with _auth_store(self.server.db_path) as store:
+            if not store.check_permission(user_id, task["project_id"],
+                                          "create_task"):
+                self._send_json(403, {"error": "forbidden"})
+                return
+        conn = self._db()
+        try:
+            _appr_ensure(conn)
+            _appr_expire(conn)
+            existing = _appr_get(conn, approval_id)
+            if existing is None or existing["task_id"] != task_id:
+                self._send_json(404, {"error": "approval not found"})
+                return
+            status, item = _appr_decide(
+                conn, approval_id, decision, user_id)
+        finally:
+            conn.close()
+        if status == "already_decided":
+            self._send_json(409, {"error": "already decided",
+                                  "approval": item})
+            return
+        with _auth_store(self.server.db_path) as store:
+            store.audit(user_id, task["project_id"],
+                        "approval_decided",
+                        "%s %s by %s" % (approval_id, decision, user_id))
+        self._send_json(200, {"approval": item})
+
     def _continue_task(self, user_id, task_id):
         """旧任务接回来继续会话：终端任务（completed/failed/
         interrupted/merge-failed）带一条新消息 + 全部选项重排队，
@@ -3566,8 +3887,9 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
 
     def _workers_heartbeat(self, user_id):
         """worker 心跳：body 含 worker_id（必填，缺了 400）与可选
-        executor/capacity。心跳本身就是注册：worker_id 不存在时
-        自动注册，绝不 404。"""
+        executor/capacity/running。心跳本身就是注册：worker_id
+        不存在时自动注册，绝不 404。running 是 worker 上报的
+        当前占用槽数（启动中占位也算）；未给时保留旧值。"""
         body, err = self._read_json()
         if err:
             self._send_json(400, {"error": err})
@@ -3584,10 +3906,20 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
             except (TypeError, ValueError):
                 self._send_json(400, {"error": "capacity must be an integer"})
                 return
+        running = body.get("running")
+        if running is not None:
+            try:
+                running = int(running)
+            except (TypeError, ValueError):
+                self._send_json(400, {"error": "running must be an integer"})
+                return
+            if running < 0:
+                self._send_json(400, {"error": "running must be >= 0"})
+                return
         try:
             worker = _registry_heartbeat(
                 self.server.db_path, worker_id,
-                executor=executor, capacity=capacity)
+                executor=executor, capacity=capacity, running=running)
         except ValueError as exc:
             self._send_json(400, {"error": str(exc)})
             return
@@ -4534,6 +4866,106 @@ class CloudControlPlaneHandler(BaseHTTPRequestHandler):
         self._send_json(200, result)
 
     # -- webhooks（签名认证，不走 X-API-Key）---------------------------
+    # -- 自动化与分诊收件箱(P1 第 2 项)---------------------------
+    # 权限口径与其它项目级写操作同款：创建/取消用 create_task
+    # (不是 manage_project，因为自动化是项目成员的日常编排，
+    # 不是改项目归属/仓库这类 owner 专属操作；评审/审批等写
+    # 操作同样用 create_task)；读用 read。状态变更记审计。
+    # 失败口径：校验失败 400；模块/数据库层意外失败时回 502
+    # 透传错误原文(截断 500 字)，绝不回 201 假装建成功——
+    # 注意 _dispatch_safe 的兜底是 500，这里显式 502 优先于它。
+    def _automation_create(self, user_id, project_id):
+        """创建自动化定义：只登记定义与到期状态，不跑模型、
+        不另起调度器(执行器后续走 claim 门，见 automations
+        模块文档)。body：name、prompt 必填；interval_seconds
+        可选(默认 86400，最小 60)；repo_url 可选。"""
+        body, err = self._read_json()
+        if err:
+            self._send_json(400, {"error": err})
+            return
+        interval = body.get("interval_seconds", 86400)
+        if interval is None:
+            interval = 86400
+        conn = self._db()
+        try:
+            try:
+                auto = _auto_create(
+                    conn, project_id, body.get("name"),
+                    body.get("prompt"), user_id, interval,
+                    body.get("repo_url"))
+            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, "automation_created",
+                        auto["id"])
+        self._send_json(201, auto)
+
+    def _automation_list(self, project_id):
+        """列出某项目全部自动化定义(含已取消)，附到期计数。"""
+        conn = self._db()
+        try:
+            try:
+                autos = _auto_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,
+            "automations": autos,
+            "due_count": sum(1 for a in autos if a["due"]),
+        })
+
+    def _automation_cancel(self, user_id, project_id, automation_id):
+        """取消自动化(软删除)：行不存在或属于别的项目时一
+        律 404(不泄露跨项目存在性)；已取消时 409。"""
+        conn = self._db()
+        try:
+            try:
+                ok, auto = _auto_cancel(conn, automation_id)
+            except Exception as exc:
+                self._send_json(502, {"error": str(exc)[:500]})
+                return
+        finally:
+            conn.close()
+        if auto is None or auto["project_id"] != project_id:
+            self._send_json(404, {"error": "automation not found"})
+            return
+        if not ok:
+            self._send_json(409, {
+                "error": "automation already cancelled"})
+            return
+        with _auth_store(self.server.db_path) as store:
+            store.audit(user_id, project_id, "automation_cancelled",
+                        automation_id)
+        self._send_json(200, auto)
+
+    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
+        triage_inbox 的字段说明。"""
+        conn = self._db()
+        try:
+            try:
+                _auto_ensure(conn)
+                _ensure_followups_table(conn)
+                inbox = _auto_triage(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, inbox)
+
     def _handle_github_webhook(self):
         """GitHub webhook：验签 -> delivery 去重 -> 解析 -> 执行建任务。
 
diff --git a/cloud/tests/test_approvals.py b/cloud/tests/test_approvals.py
new file mode 100644
index 0000000..cad5619
--- /dev/null
+++ b/cloud/tests/test_approvals.py
@@ -0,0 +1,157 @@
+"""运行时审批队列模块单测：cloud/approvals.py。
+
+不跑真实模型、不联网：内存 SQLite。覆盖批准放行、拒绝拦截、
+回执留痕、二次决定不覆盖、超时默认拒绝、非法输入、回包映射。
+"""
+
+import sqlite3
+import sys
+import unittest
+from pathlib import Path
+
+try:
+    from cloud.approvals import (
+        command_response,
+        create_approval,
+        decide_approval,
+        elicitation_response,
+        ensure_approvals_table,
+        expire_overdue,
+        get_approval,
+        list_approvals,
+    )
+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.approvals import (
+        command_response,
+        create_approval,
+        decide_approval,
+        elicitation_response,
+        ensure_approvals_table,
+        expire_overdue,
+        get_approval,
+        list_approvals,
+    )
+
+
+class ApprovalsModuleTestCase(unittest.TestCase):
+    def setUp(self):
+        self.conn = sqlite3.connect(":memory:")
+        self.conn.row_factory = sqlite3.Row
+        ensure_approvals_table(self.conn)
+        self.conn.commit()
+
+    def tearDown(self):
+        self.conn.close()
+
+    def _mk(self, **kw):
+        kw.setdefault("task_id", "task-1")
+        kw.setdefault("project_id", "proj-1")
+        kw.setdefault("summary", "MCP 服务请求表单输入")
+        return create_approval(self.conn, **kw)
+
+    def test_create_defaults_to_pending_with_timeout(self):
+        item = self._mk()
+        self.assertEqual(item["status"], "pending")
+        self.assertEqual(item["kind"], "elicitation")
+        self.assertIsNone(item["decided_by"])
+        self.assertIsNone(item["decision"])
+        self.assertIsNotNone(item["expires_at"])
+        self.assertGreater(item["expires_at"], item["created_at"])
+        pending = list_approvals(self.conn, task_id="task-1",
+                                 status="pending")
+        self.assertEqual(len(pending), 1)
+
+    def test_approve_releases_with_receipt(self):
+        item = self._mk()
+        status, decided = decide_approval(
+            self.conn, item["id"], "approved", "user-9")
+        self.assertEqual(status, "decided")
+        self.assertEqual(decided["status"], "approved")
+        self.assertEqual(decided["decision"], "approved")
+        self.assertEqual(decided["decided_by"], "user-9")
+        self.assertIsNotNone(decided["decided_at"])
+        self.assertEqual(elicitation_response(decided["decision"]),
+                         {"action": "accept", "content": {}})
+        self.assertEqual(command_response(decided["decision"]),
+                         {"decision": "accept"})
+
+    def test_deny_blocks_with_receipt(self):
+        item = self._mk()
+        status, decided = decide_approval(
+            self.conn, item["id"], "denied", "user-9")
+        self.assertEqual(status, "decided")
+        self.assertEqual(decided["status"], "denied")
+        self.assertEqual(decided["decision"], "denied")
+        self.assertEqual(decided["decided_by"], "user-9")
+        self.assertEqual(elicitation_response(decided["decision"]),
+                         {"action": "decline"})
+        self.assertEqual(command_response(decided["decision"]),
+                         {"decision": "decline"})
+
+    def test_second_decision_never_overwrites_first(self):
+        item = self._mk()
+        decide_approval(self.conn, item["id"], "approved", "user-9")
+        status, kept = decide_approval(
+            self.conn, item["id"], "denied", "user-7")
+        self.assertEqual(status, "already_decided")
+        self.assertEqual(kept["status"], "approved")
+        self.assertEqual(kept["decided_by"], "user-9")
+
+    def test_expire_overdue_defaults_to_deny_not_allow(self):
+        item = self._mk(now=1000.0, timeout_seconds=60)
+        self.assertEqual(item["expires_at"], 1060.0)
+        count = expire_overdue(self.conn, now=1061.0)
+        self.assertEqual(count, 1)
+        done = get_approval(self.conn, item["id"])
+        self.assertEqual(done["status"], "expired")
+        self.assertEqual(done["decision"], "denied")
+        self.assertEqual(done["decided_by"], "system-timeout")
+        self.assertEqual(elicitation_response(done["decision"]),
+                         {"action": "decline"})
+        status, kept = decide_approval(
+            self.conn, item["id"], "approved", "user-9")
+        self.assertEqual(status, "already_decided")
+        self.assertEqual(kept["status"], "expired")
+
+    def test_not_yet_due_is_not_expired(self):
+        self._mk(now=1000.0, timeout_seconds=3600)
+        self.assertEqual(expire_overdue(self.conn, now=2000.0), 0)
+        self.assertEqual(len(list_approvals(self.conn, status="pending")),
+                         1)
+
+    def test_decide_missing_is_not_found(self):
+        status, item = decide_approval(self.conn, "nope", "approved",
+                                       "user-9")
+        self.assertEqual(status, "not_found")
+        self.assertIsNone(item)
+
+    def test_invalid_inputs_raise_value_error(self):
+        with self.assertRaises(ValueError):
+            self._mk(kind="bogus")
+        with self.assertRaises(ValueError):
+            self._mk(summary="   ")
+        with self.assertRaises(ValueError):
+            self._mk(summary="x" * 2001)
+        with self.assertRaises(ValueError):
+            self._mk(timeout_seconds=-1)
+        with self.assertRaises(ValueError):
+            decide_approval(self.conn, "x", "maybe", "user-9")
+
+    def test_ensure_is_idempotent_and_keeps_rows(self):
+        item = self._mk()
+        ensure_approvals_table(self.conn)
+        self.conn.commit()
+        self.assertEqual(get_approval(self.conn, item["id"])["id"],
+                         item["id"])
+
+    def test_response_mapping_never_defaults_to_allow(self):
+        for weird in ("expired", "bogus", None, ""):
+            self.assertEqual(elicitation_response(weird),
+                             {"action": "decline"})
+            self.assertEqual(command_response(weird),
+                             {"decision": "decline"})
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/cloud/tests/test_approvals_server.py b/cloud/tests/test_approvals_server.py
new file mode 100644
index 0000000..5850929
--- /dev/null
+++ b/cloud/tests/test_approvals_server.py
@@ -0,0 +1,222 @@
+"""运行时审批队列服务端单测：审批端点。
+
+不跑真实模型、不联网：本地临时端口 + 临时 SQLite，与
+test_peer_messages_server.py 同款骨架。覆盖创建、列表、批准、
+拒绝、回执、二次决定 409、过期扫荡、权限、校验、跨任务隔离。
+"""
+
+import http.client
+import json
+import os
+import sqlite3
+import sys
+import tempfile
+import threading
+import time
+import unittest
+from pathlib import Path
+
+try:
+    from cloud.approvals import ensure_approvals_table
+    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.approvals import ensure_approvals_table
+    from cloud.auth import AuthStore
+    from cloud.server import create_server
+
+
+class ApprovalsServerTestCase(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 _create_task(self, project_id):
+        status, task = self._req(
+            "POST", "/tasks",
+            {"project_id": project_id, "prompt": "做事"}, self.owner_key)
+        self.assertEqual(status, 201, task)
+        return task
+
+    def _create_approval(self, task_id, **extra):
+        body = {"summary": "MCP 服务请求表单输入"}
+        body.update(extra)
+        return self._req(
+            "POST", "/tasks/%s/approvals" % task_id, body, self.owner_key)
+
+    def _insert_expired_row(self, task_id, project_id):
+        conn = sqlite3.connect(self.db_path)
+        try:
+            ensure_approvals_table(conn)
+            conn.execute(
+                "INSERT INTO task_approvals(id, task_id, project_id,"
+                " kind, summary, detail, status, requested_by,"
+                " decided_by, decided_at, decision,"
+                " created_at, updated_at, expires_at)"
+                " VALUES(?, ?, ?, 'elicitation', '过期项', NULL,"
+                " 'pending', NULL, NULL, NULL, NULL, ?, ?, ?)",
+                ("expired-1", task_id, project_id,
+                 time.time() - 7200, time.time() - 7200,
+                 time.time() - 3600))
+            conn.commit()
+        finally:
+            conn.close()
+
+    def test_create_list_approve_receipt(self):
+        task = self._create_task(self.project["id"])
+        status, body = self._create_approval(task["id"])
+        self.assertEqual(status, 201, body)
+        item = body["approval"]
+        self.assertEqual(item["status"], "pending")
+        self.assertEqual(item["kind"], "elicitation")
+        status, body = self._req(
+            "GET", "/tasks/%s/approvals" % task["id"],
+            None, self.owner_key)
+        self.assertEqual(status, 200)
+        self.assertEqual(len(body["approvals"]), 1)
+        status, body = self._req(
+            "POST", "/tasks/%s/approvals/%s/approve" % (
+                task["id"], item["id"]), {}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        decided = body["approval"]
+        self.assertEqual(decided["status"], "approved")
+        self.assertEqual(decided["decision"], "approved")
+        self.assertEqual(decided["decided_by"], self.owner_id)
+        self.assertIsNotNone(decided["decided_at"])
+        status, body = self._req(
+            "GET", "/tasks/%s/approvals" % task["id"],
+            None, self.owner_key)
+        self.assertEqual(body["approvals"][0]["status"], "approved")
+
+    def test_deny_and_second_decision_conflicts(self):
+        task = self._create_task(self.project["id"])
+        _, body = self._create_approval(task["id"])
+        item = body["approval"]
+        status, body = self._req(
+            "POST", "/tasks/%s/approvals/%s/deny" % (
+                task["id"], item["id"]), {}, self.owner_key)
+        self.assertEqual(status, 200)
+        self.assertEqual(body["approval"]["status"], "denied")
+        self.assertEqual(body["approval"]["decided_by"], self.owner_id)
+        status, body = self._req(
+            "POST", "/tasks/%s/approvals/%s/approve" % (
+                task["id"], item["id"]), {}, self.owner_key)
+        self.assertEqual(status, 409, body)
+        self.assertEqual(body["approval"]["status"], "denied")
+
+    def test_list_sweeps_overdue_to_expired_deny(self):
+        task = self._create_task(self.project["id"])
+        self._insert_expired_row(task["id"], self.project["id"])
+        status, body = self._req(
+            "GET", "/tasks/%s/approvals" % task["id"],
+            None, self.owner_key)
+        self.assertEqual(status, 200)
+        self.assertEqual(len(body["approvals"]), 1)
+        swept = body["approvals"][0]
+        self.assertEqual(swept["status"], "expired")
+        self.assertEqual(swept["decision"], "denied")
+        self.assertEqual(swept["decided_by"], "system-timeout")
+
+    def test_validation_and_not_found(self):
+        task = self._create_task(self.project["id"])
+        base = "/tasks/%s/approvals" % task["id"]
+        status, _ = self._req("POST", base, {}, self.owner_key)
+        self.assertEqual(status, 400)
+        status, _ = self._req(
+            "POST", base, {"summary": "x", "kind": "bogus"},
+            self.owner_key)
+        self.assertEqual(status, 400)
+        status, _ = self._req(
+            "POST", base, {"summary": "x", "timeout_seconds": 10},
+            self.owner_key)
+        self.assertEqual(status, 400)
+        status, _ = self._req(
+            "POST", base, {"summary": "x", "timeout_seconds": 90000},
+            self.owner_key)
+        self.assertEqual(status, 400)
+        status, _ = self._req(
+            "GET", "/tasks/nope/approvals", None, self.owner_key)
+        self.assertEqual(status, 404)
+        status, _ = self._req(
+            "POST", base, {"summary": "x"}, None)
+        self.assertEqual(status, 401)
+
+    def test_forbidden_for_outsider(self):
+        task = self._create_task(self.project["id"])
+        status, body = self._req(
+            "POST", "/tasks/%s/approvals" % task["id"],
+            {"summary": "x"}, self.outsider_key)
+        self.assertEqual(status, 403)
+        status, _ = self._req(
+            "GET", "/tasks/%s/approvals" % task["id"],
+            None, self.outsider_key)
+        self.assertEqual(status, 403)
+        _, body = self._create_approval(task["id"])
+        item = body["approval"]
+        status, _ = self._req(
+            "POST", "/tasks/%s/approvals/%s/approve" % (
+                task["id"], item["id"]), {}, self.outsider_key)
+        self.assertEqual(status, 403)
+
+    def test_cross_task_approval_id_is_not_found(self):
+        task_a = self._create_task(self.project["id"])
+        task_b = self._create_task(self.project["id"])
+        _, body = self._create_approval(task_a["id"])
+        item = body["approval"]
+        status, _ = self._req(
+            "POST", "/tasks/%s/approvals/%s/approve" % (
+                task_b["id"], item["id"]), {}, self.owner_key)
+        self.assertEqual(status, 404)
+        status, body = self._req(
+            "GET", "/tasks/%s/approvals" % task_b["id"],
+            None, self.owner_key)
+        self.assertEqual(body["approvals"], [])
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/cloud/tests/test_automations.py b/cloud/tests/test_automations.py
new file mode 100644
index 0000000..fb1cb23
--- /dev/null
+++ b/cloud/tests/test_automations.py
@@ -0,0 +1,232 @@
+"""自动化与分诊收件箱模块单测(cloud/automations.py)。
+
+不跑真实模型、不联网：临时 SQLite 文件 + 模块函数直调。
+覆盖：建表幂等、创建校验与同名规则、到期边界、claim 抢单
+防重、取消语义、产出记录、分诊四段聚合。
+"""
+
+import os
+import sqlite3
+import sys
+import tempfile
+import time
+import unittest
+from pathlib import Path
+
+try:
+    from cloud.automations import (
+        cancel_automation,
+        claim_due_automation,
+        create_automation,
+        due_automations,
+        ensure_automations_tables,
+        get_automation,
+        is_due,
+        list_automation_runs,
+        list_automations,
+        record_automation_run,
+        triage_inbox,
+    )
+    from cloud.peer_messages import ensure_followups_table
+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.automations import (
+        cancel_automation,
+        claim_due_automation,
+        create_automation,
+        due_automations,
+        ensure_automations_tables,
+        get_automation,
+        is_due,
+        list_automation_runs,
+        list_automations,
+        record_automation_run,
+        triage_inbox,
+    )
+    from cloud.peer_messages import ensure_followups_table
+
+
+class AutomationsTestCase(unittest.TestCase):
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.db_path = os.path.join(self.tmp.name, "control.db")
+        self.conn = sqlite3.connect(self.db_path)
+        self.conn.row_factory = sqlite3.Row
+        self.conn.execute(
+            "CREATE TABLE tasks(id TEXT PRIMARY KEY, project_id TEXT,"
+            " prompt TEXT, status TEXT, created_at DOUBLE PRECISION,"
+            " updated_at DOUBLE PRECISION, error_text TEXT)")
+        ensure_automations_tables(self.conn)
+        ensure_followups_table(self.conn)
+        self.conn.commit()
+        self.now = 1700000000.0
+        self.project = "proj-1"
+
+    def tearDown(self):
+        self.conn.close()
+        self.tmp.cleanup()
+
+    def _make(self, name="daily", interval=3600, **kw):
+        kw.setdefault("project_id", self.project)
+        return create_automation(
+            self.conn, kw.pop("project_id"), name, "做每日检查",
+            "user-1", interval, now=self.now, **kw)
+
+    def test_ensure_is_idempotent(self):
+        ensure_automations_tables(self.conn)
+        ensure_automations_tables(self.conn)
+        self.conn.commit()
+        rows = self.conn.execute(
+            "SELECT name FROM sqlite_master WHERE type = 'table'"
+            " AND name IN ('automations', 'automation_runs')").fetchall()
+        self.assertEqual(len(rows), 2)
+
+    def test_create_fields_and_due_immediately(self):
+        auto = self._make()
+        self.assertEqual(auto["status"], "active")
+        self.assertEqual(auto["interval_seconds"], 3600)
+        self.assertIsNone(auto["last_run_at"])
+        self.assertEqual(auto["next_run_at"], self.now)
+        self.assertTrue(auto["due"])
+        self.assertTrue(is_due(auto, self.now))
+        got = get_automation(self.conn, auto["id"])
+        self.assertEqual(got["name"], "daily")
+        self.assertIsNone(get_automation(self.conn, "missing"))
+
+    def test_create_validation(self):
+        for name, prompt, interval in (
+                ("", "x", 3600),
+                ("   ", "x", 3600),
+                ("n", "", 3600),
+                ("n", "x", 59),
+                ("n", "x", True),
+                ("n", "x", 3600.5),
+                ("x" * 101, "x", 3600),
+                ("n", "x" * 20001, 3600)):
+            with self.assertRaises(ValueError, msg=(name, interval)):
+                create_automation(
+                    self.conn, self.project, name, prompt, "user-1",
+                    interval, now=self.now)
+
+    def test_duplicate_active_name_rejected_reusable_after_cancel(self):
+        self._make(name="dup")
+        with self.assertRaises(ValueError):
+            self._make(name="dup")
+        autos = list_automations(self.conn, self.project, self.now)
+        self.assertEqual(len(autos), 1)
+        ok, cancelled = cancel_automation(
+            self.conn, autos[0]["id"], self.now)
+        self.assertTrue(ok)
+        self.assertEqual(cancelled["status"], "cancelled")
+        again = self._make(name="dup")
+        self.assertEqual(again["status"], "active")
+
+    def test_due_boundary_matches_host_rundue_rule(self):
+        auto = self._make(interval=3600)
+        self.conn.execute(
+            "UPDATE automations SET last_run_at = ? WHERE id = ?",
+            (self.now - 3599, auto["id"]))
+        self.conn.commit()
+        self.assertEqual(due_automations(
+            self.conn, self.project, self.now), [])
+        self.conn.execute(
+            "UPDATE automations SET last_run_at = ? WHERE id = ?",
+            (self.now - 3600, auto["id"]))
+        self.conn.commit()
+        due = due_automations(self.conn, self.project, self.now)
+        self.assertEqual([d["id"] for d in due], [auto["id"]])
+
+    def test_claim_is_the_duplicate_gate(self):
+        auto = self._make(interval=3600)
+        self.assertTrue(claim_due_automation(
+            self.conn, auto["id"], self.now))
+        # 同一间隔内第二次抢单必须失败：这就是防重复执行。
+        self.assertFalse(claim_due_automation(
+            self.conn, auto["id"], self.now))
+        self.assertFalse(claim_due_automation(
+            self.conn, auto["id"], self.now + 3599))
+        # 下个间隔到了才能再抢。
+        self.assertTrue(claim_due_automation(
+            self.conn, auto["id"], self.now + 3600))
+        self.assertFalse(claim_due_automation(
+            self.conn, "missing", self.now))
+        ok, _ = cancel_automation(self.conn, auto["id"], self.now)
+        self.assertTrue(ok)
+        self.assertFalse(claim_due_automation(
+            self.conn, auto["id"], self.now + 7200))
+
+    def test_cancel_semantics(self):
+        ok, auto = cancel_automation(self.conn, "missing", self.now)
+        self.assertFalse(ok)
+        self.assertIsNone(auto)
+        made = self._make()
+        ok, auto = cancel_automation(self.conn, made["id"], self.now)
+        self.assertTrue(ok)
+        self.assertEqual(auto["status"], "cancelled")
+        self.assertFalse(auto["due"])
+        ok, auto = cancel_automation(self.conn, made["id"], self.now)
+        self.assertFalse(ok)
+        self.assertEqual(auto["status"], "cancelled")
+
+    def test_record_run_and_list_runs(self):
+        auto = self._make()
+        run = record_automation_run(
+            self.conn, auto["id"], self.project, task_id="task-1",
+            status="created", detail="claimed", now=self.now)
+        self.assertEqual(run["task_id"], "task-1")
+        self.assertEqual(run["status"], "created")
+        runs = list_automation_runs(self.conn, auto["id"])
+        self.assertEqual(len(runs), 1)
+        with self.assertRaises(ValueError):
+            record_automation_run(
+                self.conn, "missing", self.project, now=self.now)
+
+    def test_triage_inbox_four_sections(self):
+        auto = self._make()
+        other = self._make(name="weekly", interval=86400)
+        cancel_automation(self.conn, other["id"], self.now)
+        self.conn.execute(
+            "INSERT INTO tasks(id, project_id, prompt, status,"
+            " created_at, updated_at, error_text)"
+            " VALUES('t-fail', ?, '坏任务', 'failed', ?, ?, 'boom')",
+            (self.project, self.now, self.now))
+        self.conn.execute(
+            "INSERT INTO tasks(id, project_id, prompt, status,"
+            " created_at, updated_at, error_text)"
+            " VALUES('t-run', ?, '跑中任务', 'running', ?, ?, NULL)",
+            (self.project, self.now, self.now))
+        self.conn.execute(
+            "INSERT INTO task_followups(id, task_id, project_id, body,"
+            " mode, status, created_by, sender_task_id, created_at,"
+            " updated_at) VALUES('f-1', 't-run', ?, '请看一下',"
+            " 'queued', 'pending', 'user-1', NULL, ?, ?)",
+            (self.project, self.now, self.now))
+        self.conn.execute(
+            "INSERT INTO tasks(id, project_id, prompt, status,"
+            " created_at, updated_at, error_text)"
+            " VALUES('t-done', ?, '已完成', 'completed', ?, ?, NULL)",
+            (self.project, self.now, self.now))
+        record_automation_run(
+            self.conn, auto["id"], self.project, task_id="t-done",
+            status="created", detail="", now=self.now)
+        self.conn.commit()
+        inbox = triage_inbox(self.conn, self.project, self.now)
+        self.assertEqual(
+            [d["id"] for d in inbox["automations_due"]], [auto["id"]])
+        self.assertEqual(
+            [t["id"] for t in inbox["failed_tasks"]], ["t-fail"])
+        self.assertEqual(
+            [f["id"] for f in inbox["pending_followups"]], ["f-1"])
+        self.assertEqual(len(inbox["recent_runs"]), 1)
+        self.assertEqual(inbox["recent_runs"][0]["task_status"],
+                         "completed")
+        self.assertEqual(inbox["counts"], {
+            "automations_due": 1,
+            "failed_tasks": 1,
+            "pending_followups": 1,
+            "recent_runs": 1,
+        })
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/cloud/tests/test_automations_server.py b/cloud/tests/test_automations_server.py
new file mode 100644
index 0000000..96ea4a5
--- /dev/null
+++ b/cloud/tests/test_automations_server.py
@@ -0,0 +1,238 @@
+"""自动化与分诊收件箱服务端单测：POST/GET /projects/{id}/
+automations、POST /projects/{id}/automations/{aid}/cancel、
+GET /projects/{id}/triage。
+
+不跑真实模型、不联网：本地临时端口 + 临时 SQLite，与
+test_peer_messages_server.py 同款骨架。覆盖鉴权、权限、
+校验、取消语义(404/409)、分诊聚合与审计。
+"""
+
+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.automations import record_automation_run
+    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.automations import record_automation_run
+    from cloud.server import create_server
+
+
+class AutomationServerTestCase(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)
+        status, self.project2 = self._req(
+            "POST", "/projects", {"name": "other"}, 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 _autos_url(self, project_id=None):
+        return "/projects/%s/automations" % (
+            project_id or self.project["id"])
+
+    def _create_auto(self, project_id=None, **extra):
+        body = {"name": "daily", "prompt": "做每日检查"}
+        body.update(extra)
+        return self._req(
+            "POST", self._autos_url(project_id), body, self.owner_key)
+
+    def _audit_actions(self, project_id):
+        conn = sqlite3.connect(self.db_path)
+        try:
+            rows = conn.execute(
+                "SELECT action FROM audit_log WHERE project_id = ?",
+                (project_id,)).fetchall()
+        finally:
+            conn.close()
+        return [r[0] for r in rows]
+
+    def _create_task(self, project_id):
+        status, task = self._req(
+            "POST", "/tasks",
+            {"project_id": project_id, "prompt": "做事"}, self.owner_key)
+        self.assertEqual(status, 201, task)
+        return task
+
+    def _set_status(self, task_id, status):
+        conn = sqlite3.connect(self.db_path)
+        try:
+            conn.execute("UPDATE tasks SET status = ? WHERE id = ?",
+                         (status, task_id))
+            conn.commit()
+        finally:
+            conn.close()
+
+    def test_create_and_list(self):
+        status, auto = self._create_auto(interval_seconds=3600)
+        self.assertEqual(status, 201, auto)
+        self.assertEqual(auto["status"], "active")
+        self.assertEqual(auto["interval_seconds"], 3600)
+        self.assertTrue(auto["due"])
+        self.assertIn("automation_created",
+                      self._audit_actions(self.project["id"]))
+        status, body = self._req(
+            "GET", self._autos_url(), None, self.owner_key)
+        self.assertEqual(status, 200)
+        self.assertEqual(len(body["automations"]), 1)
+        self.assertEqual(body["due_count"], 1)
+        self.assertEqual(body["project_id"], self.project["id"])
+
+    def test_create_defaults_and_validation(self):
+        status, auto = self._create_auto()
+        self.assertEqual(status, 201, auto)
+        self.assertEqual(auto["interval_seconds"], 86400)
+        for body in (
+                {"prompt": "x"},
+                {"name": "n"},
+                {"name": "n", "prompt": "x", "interval_seconds": 30},
+                {"name": "n", "prompt": "x",
+                 "interval_seconds": "3600"}):
+            status, _ = self._req(
+                "POST", self._autos_url(), body, self.owner_key)
+            self.assertEqual(status, 400, body)
+        status, _ = self._create_auto()
+        self.assertEqual(status, 400)
+
+    def test_auth_and_permission(self):
+        status, _ = self._req(
+            "POST", self._autos_url(),
+            {"name": "n", "prompt": "x"}, None)
+        self.assertEqual(status, 401)
+        status, _ = self._req(
+            "GET", self._autos_url(), None, None)
+        self.assertEqual(status, 401)
+        status, _ = self._req(
+            "POST", self._autos_url(),
+            {"name": "n", "prompt": "x"}, self.outsider_key)
+        self.assertEqual(status, 403)
+        status, _ = self._req(
+            "GET", self._autos_url(), None, self.outsider_key)
+        self.assertEqual(status, 403)
+        status, _ = self._req(
+            "POST", "/projects/nope/automations",
+            {"name": "n", "prompt": "x"}, self.owner_key)
+        self.assertEqual(status, 404)
+
+    def test_cancel_semantics(self):
+        _, auto = self._create_auto()
+        url = "%s/%s/cancel" % (self._autos_url(), auto["id"])
+        status, body = self._req("POST", url, {}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["status"], "cancelled")
+        self.assertFalse(body["due"])
+        self.assertIn("automation_cancelled",
+                      self._audit_actions(self.project["id"]))
+        status, _ = self._req("POST", url, {}, self.owner_key)
+        self.assertEqual(status, 409)
+        status, _ = self._req(
+            "POST", "%s/missing/cancel" % self._autos_url(), {},
+            self.owner_key)
+        self.assertEqual(status, 404)
+        # 别的项目的自动化 id 在本项目下取消也一律 404。
+        _, other = self._create_auto(project_id=self.project2["id"])
+        status, _ = self._req(
+            "POST", "%s/%s/cancel" % (self._autos_url(), other["id"]),
+            {}, self.owner_key)
+        self.assertEqual(status, 404)
+        status, _ = self._req(
+            "POST", url, {}, self.outsider_key)
+        self.assertEqual(status, 403)
+
+    def test_triage_inbox(self):
+        _, auto = self._create_auto()
+        task = self._create_task(self.project["id"])
+        self._set_status(task["id"], "failed")
+        runner = self._create_task(self.project["id"])
+        self._set_status(runner["id"], "running")
+        status, _ = self._req(
+            "POST", "/tasks/%s/followups" % runner["id"],
+            {"body": "请看一下", "mode": "queued"}, self.owner_key)
+        self.assertEqual(status, 201)
+        conn = sqlite3.connect(self.db_path)
+        try:
+            record_automation_run(
+                conn, auto["id"], self.project["id"],
+                task_id=task["id"], status="created", detail="")
+        finally:
+            conn.close()
+        status, body = self._req(
+            "GET", "/projects/%s/triage" % self.project["id"], None,
+            self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(
+            [d["id"] for d in body["automations_due"]], [auto["id"]])
+        self.assertEqual(
+            [t["id"] for t in body["failed_tasks"]], [task["id"]])
+        self.assertEqual(len(body["pending_followups"]), 1)
+        self.assertEqual(
+            body["pending_followups"][0]["task_id"], runner["id"])
+        self.assertEqual(len(body["recent_runs"]), 1)
+        self.assertEqual(body["recent_runs"][0]["task_status"],
+                         "failed")
+        self.assertEqual(body["counts"]["automations_due"], 1)
+        status, _ = self._req(
+            "GET", "/projects/%s/triage" % self.project["id"], None,
+            self.outsider_key)
+        self.assertEqual(status, 403)
+        status, _ = self._req(
+            "GET", "/projects/%s/triage" % self.project["id"], None,
+            None)
+        self.assertEqual(status, 401)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/cloud/tests/test_caps.py b/cloud/tests/test_caps.py
new file mode 100644
index 0000000..b698c28
--- /dev/null
+++ b/cloud/tests/test_caps.py
@@ -0,0 +1,223 @@
+"""并发上限单测（caps）。
+
+覆盖：配置解析优先级与默认值、认领门的未满 / 已满 / 释放后
+恢复、启动中任务计入占用、CAS 抢输后占位不泄漏、缺 key 门
+提前返回后占位释放、双线程同时认领只赢一个。
+
+只用注入的假环境与本地 SQLite，不跑真实模型、不联网。
+"""
+
+import os
+import sqlite3
+import sys
+import tempfile
+import threading
+import time
+import unittest
+from pathlib import Path
+
+try:
+    from cloud.caps import (
+        can_claim,
+        effective_capacity,
+        resolve_capacity,
+        running_count,
+    )
+    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]))
+    from cloud.caps import (
+        can_claim,
+        effective_capacity,
+        resolve_capacity,
+        running_count,
+    )
+    from cloud.server import init_db
+    from cloud.worker import Worker
+
+
+def _insert_task(db_path, task_id, project_id="p1"):
+    now = time.time()
+    conn = sqlite3.connect(db_path)
+    try:
+        conn.execute(
+            "INSERT INTO tasks(id, project_id, prompt, status,"
+            " idempotency_key, depends_on, priority, created_by,"
+            " created_at, updated_at, last_heartbeat)"
+            " VALUES(?, ?, ?, 'queued', NULL, '[]', 0, 'tester', ?, ?, NULL)",
+            (task_id, project_id, "do work", now, now),
+        )
+        conn.commit()
+    finally:
+        conn.close()
+
+
+def _status(db_path, task_id):
+    conn = sqlite3.connect(db_path)
+    try:
+        row = conn.execute(
+            "SELECT status FROM tasks WHERE id = ?",
+            (task_id,)).fetchone()
+    finally:
+        conn.close()
+    return row[0] if row else None
+
+
+class CapsResolveTestCase(unittest.TestCase):
+    def test_explicit_beats_env(self):
+        env = {"CLOUD_WORKER_CAPACITY": "3"}
+        self.assertEqual(resolve_capacity(2, env=env), 2)
+
+    def test_env_used_when_no_explicit(self):
+        env = {"CLOUD_WORKER_CAPACITY": "3"}
+        self.assertEqual(resolve_capacity(None, env=env), 3)
+
+    def test_default_is_1(self):
+        self.assertEqual(resolve_capacity(None, env={}), 1)
+        self.assertEqual(resolve_capacity(), 1)
+
+    def test_invalid_values_fall_back_to_1(self):
+        self.assertEqual(resolve_capacity(0, env={}), 1)
+        self.assertEqual(resolve_capacity(-2, env={}), 1)
+        self.assertEqual(resolve_capacity("abc", env={}), 1)
+        self.assertEqual(resolve_capacity(None, env={
+            "CLOUD_WORKER_CAPACITY": "abc"}), 1)
+        self.assertEqual(resolve_capacity(None, env={
+            "CLOUD_WORKER_CAPACITY": "0"}), 1)
+        self.assertEqual(resolve_capacity(None, env={
+            "CLOUD_WORKER_CAPACITY": "  "}), 1)
+
+    def test_effective_capacity(self):
+        self.assertEqual(effective_capacity(2), 2)
+        self.assertEqual(effective_capacity(0), 1)
+        self.assertEqual(effective_capacity(None), 1)
+
+    def test_can_claim_boundaries(self):
+        self.assertTrue(can_claim(0, 1))
+        self.assertFalse(can_claim(1, 1))
+        self.assertTrue(can_claim(1, 2))
+        self.assertFalse(can_claim(2, 2))
+
+    def test_running_count_counts_starting_entries(self):
+        # 空条目 = 已认领未起进程的启动中任务，同样占一个槽。
+        self.assertEqual(running_count({"a": {}, "b": {"proc": 1}}), 2)
+        self.assertEqual(running_count({}), 0)
+
+
+class WorkerClaimGateTestCase(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.env = {"XCM_API_KEY": "test-placeholder"}
+
+    def tearDown(self):
+        self.tmp.cleanup()
+
+    def _worker(self, **kwargs):
+        kwargs.setdefault("env", self.env)
+        kwargs.setdefault("retry_backoff_seconds", 0)
+        kwargs.setdefault("max_retries", 0)
+        return Worker(self.db_path, "/fake/host.py", **kwargs)
+
+    def test_default_capacity_is_1(self):
+        worker = self._worker()
+        self.assertEqual(worker.capacity, 1)
+
+    def test_env_capacity_and_explicit_priority(self):
+        env = {"CLOUD_WORKER_CAPACITY": "3"}
+        self.assertEqual(self._worker(env=env).capacity, 3)
+        self.assertEqual(self._worker(env=env, capacity=2).capacity, 2)
+
+    def test_full_worker_refuses_claim_and_task_stays_queued(self):
+        _insert_task(self.db_path, "A")
+        _insert_task(self.db_path, "B")
+        worker = self._worker(capacity=1)
+        self.assertTrue(worker._claim("A"))
+        self.assertEqual(_status(self.db_path, "A"), "running")
+        self.assertFalse(worker._claim("B"))
+        self.assertEqual(_status(self.db_path, "B"), "queued")
+        self.assertEqual(set(worker._running_snapshot()), {"A"})
+
+    def test_release_restores_claim(self):
+        _insert_task(self.db_path, "A")
+        _insert_task(self.db_path, "B")
+        worker = self._worker(capacity=1)
+        self.assertTrue(worker._claim("A"))
+        self.assertFalse(worker._claim("B"))
+        worker._running_clear("A")
+        self.assertTrue(worker._claim("B"))
+        self.assertEqual(_status(self.db_path, "B"), "running")
+
+    def test_starting_task_counts_toward_capacity(self):
+        _insert_task(self.db_path, "B")
+        worker = self._worker(capacity=1)
+        # 空条目模拟“已认领、执行器还没起进程”的启动中任务。
+        worker._running_set("STARTING")
+        self.assertEqual(worker._running_count(), 1)
+        self.assertFalse(worker._claim("B"))
+        self.assertEqual(_status(self.db_path, "B"), "queued")
+        worker._running_clear("STARTING")
+        self.assertTrue(worker._claim("B"))
+
+    def test_capacity_2_allows_two_blocks_third(self):
+        for task_id in ("A", "B", "C"):
+            _insert_task(self.db_path, task_id)
+        worker = self._worker(capacity=2)
+        self.assertTrue(worker._claim("A"))
+        self.assertTrue(worker._claim("B"))
+        self.assertFalse(worker._claim("C"))
+        self.assertEqual(_status(self.db_path, "C"), "queued")
+        worker._running_clear("A")
+        self.assertTrue(worker._claim("C"))
+
+    def test_lost_cas_race_releases_slot(self):
+        _insert_task(self.db_path, "A")
+        _insert_task(self.db_path, "B")
+        first = self._worker(capacity=1)
+        second = self._worker(capacity=1)
+        self.assertTrue(first._claim("A"))
+        # second 的 CAS 抢输：返回 False，且它自己的占位已清，
+        # 不能留下一个空槽把自己堵死。
+        self.assertFalse(second._claim("A"))
+        self.assertEqual(second._running_snapshot(), {})
+        self.assertTrue(second._claim("B"))
+
+    def test_two_threads_claim_only_one_wins(self):
+        _insert_task(self.db_path, "A")
+        _insert_task(self.db_path, "B")
+        worker = self._worker(capacity=1)
+        barrier = threading.Barrier(2, timeout=5)
+        results = {}
+
+        def _try(task_id):
+            barrier.wait()
+            results[task_id] = worker._claim(task_id)
+
+        threads = [threading.Thread(target=_try, args=(t,))
+                   for t in ("A", "B")]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join(timeout=10)
+        self.assertEqual(sorted(results.values()), [False, True])
+        self.assertEqual(worker._running_count(), 1)
+
+    def test_missing_key_gate_releases_slot(self):
+        _insert_task(self.db_path, "A")
+        _insert_task(self.db_path, "B")
+        # env 里无 key：run_one 在缺 key 门提前返回 failed，
+        # 占位必须随之释放，否则 capacity=1 的 worker 会被堵死。
+        worker = self._worker(env={}, capacity=1)
+        result = worker.run_one()
+        self.assertTrue(result["dispatched"])
+        self.assertEqual(result["status"], "failed")
+        self.assertEqual(_status(self.db_path, "A"), "failed")
+        self.assertEqual(worker._running_snapshot(), {})
+        self.assertTrue(worker._claim("B"))
+        self.assertEqual(_status(self.db_path, "B"), "running")
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/cloud/tests/test_caps_server.py b/cloud/tests/test_caps_server.py
new file mode 100644
index 0000000..b1d7206
--- /dev/null
+++ b/cloud/tests/test_caps_server.py
@@ -0,0 +1,125 @@
+"""并发上限服务端面单测（/workers 占用展示）。
+
+覆盖：GET /workers 行里含 capacity 与 running；心跳上报 running
+后列表可见；心跳未给 running 时保留旧值；非法 running 400；
+未认证 401；/workers 读权限口径是任何已认证用户（与既有
+_workers_list 一致，不按项目成员过滤）。
+
+只用标准库 http.client，不跑真实模型，不联网。
+"""
+
+import http.client
+import json
+import os
+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
+
+
+class CapsServerTestCase(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.viewer_id, self.viewer_key = self.auth.create_user("viewer")
+        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()
+
+    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 _workers(self, key):
+        status, body = self._req("GET", "/workers", key=key)
+        self.assertEqual(status, 200, body)
+        return {w["worker_id"]: w for w in body["workers"]}
+
+    def test_workers_list_shows_capacity_and_running(self):
+        status, body = self._req(
+            "POST", "/workers/heartbeat",
+            {"worker_id": "w1", "executor": "local",
+             "capacity": 4, "running": 2},
+            self.owner_key)
+        self.assertEqual(status, 200, body)
+        # 读权限：任何已认证用户可读（viewer 也能看到基础设施行）。
+        workers = self._workers(self.viewer_key)
+        self.assertEqual(workers["w1"]["capacity"], 4)
+        self.assertEqual(workers["w1"]["running"], 2)
+
+    def test_heartbeat_without_running_preserves_old(self):
+        self._req("POST", "/workers/heartbeat",
+                  {"worker_id": "w1", "capacity": 2, "running": 1},
+                  self.owner_key)
+        status, body = self._req(
+            "POST", "/workers/heartbeat", {"worker_id": "w1"},
+            self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["running"], 1)
+        self.assertEqual(self._workers(self.owner_key)["w1"]["running"], 1)
+
+    def test_heartbeat_invalid_running_400(self):
+        status, _ = self._req(
+            "POST", "/workers/heartbeat",
+            {"worker_id": "w1", "running": "abc"}, self.owner_key)
+        self.assertEqual(status, 400)
+        status, _ = self._req(
+            "POST", "/workers/heartbeat",
+            {"worker_id": "w1", "running": -1}, self.owner_key)
+        self.assertEqual(status, 400)
+
+    def test_workers_requires_auth(self):
+        status, _ = self._req("GET", "/workers")
+        self.assertEqual(status, 401)
+
+    def test_heartbeat_registers_unknown_worker_with_defaults(self):
+        status, body = self._req(
+            "POST", "/workers/heartbeat", {"worker_id": "new1"},
+            self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["capacity"], 1)
+        self.assertEqual(body["running"], 0)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/cloud/tests/test_cloud_dispatch_pr.py b/cloud/tests/test_cloud_dispatch_pr.py
index 62e4ab7..8d27c62 100644
--- a/cloud/tests/test_cloud_dispatch_pr.py
+++ b/cloud/tests/test_cloud_dispatch_pr.py
@@ -240,12 +240,13 @@ class TaskExecutorFilterTestCase(unittest.TestCase):
     def tearDown(self):
         self.tmp.cleanup()
 
-    def _docker_worker(self):
+    def _docker_worker(self, capacity=None):
         return Worker(
             self.db_path, "host.py", config_path=None, repo=None,
             executor=DockerExecutor(runner=lambda cmd: _result()),
             env={"XCM_API_KEY": "test-placeholder"}, env_key=None,
-            task_timeout=1800, max_retries=0, retry_backoff_seconds=0)
+            task_timeout=1800, max_retries=0, retry_backoff_seconds=0,
+            capacity=capacity)
 
     def test_claim_rejects_other_pool_task(self):
         worker = self._docker_worker()
@@ -260,7 +261,9 @@ class TaskExecutorFilterTestCase(unittest.TestCase):
         self.assertEqual(status, "queued")
 
     def test_claim_accepts_unassigned_and_same_pool(self):
-        worker = self._docker_worker()
+        # 本测试连续认领两个任务, 测的是执行器过滤不是上限;
+        # caps 默认上限 1 会挡第二单, 故显式给 capacity=2。
+        worker = self._docker_worker(capacity=2)
         self.assertTrue(worker._claim("t-none", executor="docker"))
         self.assertTrue(worker._claim("t-docker", executor="docker"))
 
diff --git a/cloud/tests/test_review.py b/cloud/tests/test_review.py
new file mode 100644
index 0000000..738cfa2
--- /dev/null
+++ b/cloud/tests/test_review.py
@@ -0,0 +1,134 @@
+"""自动代码评审模块单测：cloud/review.py。
+
+不跑真实模型、不联网：纯 sqlite 内存库 + 协议目标构造。
+覆盖四种 ReviewTarget、非法组合、登记/队列/取单条/写回
+状态机（findings 写回默认 completed、显式 failed 可覆盖）。
+"""
+
+import sqlite3
+import sys
+import unittest
+from pathlib import Path
+
+try:
+    from cloud.review import build_review_target
+    from cloud.review import create_review
+    from cloud.review import ensure_reviews_table
+    from cloud.review import get_review
+    from cloud.review import list_project_reviews
+    from cloud.review import list_reviews
+    from cloud.review import update_review
+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.review import build_review_target
+    from cloud.review import create_review
+    from cloud.review import ensure_reviews_table
+    from cloud.review import get_review
+    from cloud.review import list_project_reviews
+    from cloud.review import list_reviews
+    from cloud.review import update_review
+
+
+class ReviewTargetTestCase(unittest.TestCase):
+    def test_four_targets(self):
+        self.assertEqual(
+            build_review_target("uncommittedChanges"),
+            {"type": "uncommittedChanges"})
+        self.assertEqual(
+            build_review_target("baseBranch", branch="main"),
+            {"type": "baseBranch", "branch": "main"})
+        self.assertEqual(
+            build_review_target("commit", sha="abc123"),
+            {"type": "commit", "sha": "abc123"})
+        self.assertEqual(
+            build_review_target("custom", instructions="look closely"),
+            {"type": "custom", "instructions": "look closely"})
+
+    def test_target_validation(self):
+        with self.assertRaises(ValueError):
+            build_review_target("baseBranch")
+        with self.assertRaises(ValueError):
+            build_review_target("baseBranch", branch="  ")
+        with self.assertRaises(ValueError):
+            build_review_target("commit")
+        with self.assertRaises(ValueError):
+            build_review_target("custom")
+        with self.assertRaises(ValueError):
+            build_review_target("nope")
+
+    def test_whitespace_trimmed(self):
+        self.assertEqual(
+            build_review_target("baseBranch", branch="  main  "),
+            {"type": "baseBranch", "branch": "main"})
+
+
+class ReviewStoreTestCase(unittest.TestCase):
+    def setUp(self):
+        self.conn = sqlite3.connect(":memory:")
+        ensure_reviews_table(self.conn)
+
+    def tearDown(self):
+        self.conn.close()
+
+    def test_create_and_list(self):
+        first = create_review(
+            self.conn, "task-a", "proj-a", "uncommittedChanges",
+            "owner-1")
+        self.assertEqual(first["status"], "requested")
+        self.assertIsNone(first["findings"])
+        self.assertEqual(first["task_id"], "task-a")
+        second = create_review(
+            self.conn, "task-a", "proj-a", "baseBranch", "owner-1",
+            branch="main")
+        self.assertEqual(second["target_type"], "baseBranch")
+        self.assertEqual(second["branch"], "main")
+        queue = list_reviews(self.conn, "task-a")
+        self.assertEqual(
+            [r["id"] for r in queue], [first["id"], second["id"]])
+        self.assertEqual(list_reviews(self.conn, "task-other"), [])
+        proj = list_project_reviews(self.conn, "proj-a")
+        self.assertEqual(len(proj), 2)
+
+    def test_create_invalid_target_writes_nothing(self):
+        with self.assertRaises(ValueError):
+            create_review(
+                self.conn, "task-a", "proj-a", "commit", "owner-1")
+        self.assertEqual(list_reviews(self.conn, "task-a"), [])
+
+    def test_get_missing_returns_none(self):
+        self.assertIsNone(get_review(self.conn, "no-such-id"))
+
+    def test_findings_writeback_defaults_completed(self):
+        record = create_review(
+            self.conn, "task-a", "proj-a", "custom", "owner-1",
+            instructions="check it")
+        updated = update_review(
+            self.conn, record["id"], findings="two issues found")
+        self.assertEqual(updated["status"], "completed")
+        self.assertEqual(updated["findings"], "two issues found")
+        self.assertEqual(
+            get_review(self.conn, record["id"])["status"], "completed")
+
+    def test_explicit_failed_status_kept(self):
+        record = create_review(
+            self.conn, "task-a", "proj-a", "uncommittedChanges",
+            "owner-1")
+        updated = update_review(
+            self.conn, record["id"], status="failed",
+            findings="reviewer crashed")
+        self.assertEqual(updated["status"], "failed")
+
+    def test_update_validation(self):
+        record = create_review(
+            self.conn, "task-a", "proj-a", "uncommittedChanges",
+            "owner-1")
+        with self.assertRaises(ValueError):
+            update_review(self.conn, record["id"], status="bogus")
+        with self.assertRaises(ValueError):
+            update_review(self.conn, record["id"], findings=123)
+        self.assertIsNone(
+            update_review(self.conn, "no-such-id", findings="x"))
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/cloud/tests/test_review_server.py b/cloud/tests/test_review_server.py
new file mode 100644
index 0000000..aa42fae
--- /dev/null
+++ b/cloud/tests/test_review_server.py
@@ -0,0 +1,240 @@
+"""自动代码评审服务端单测：POST /tasks/<id>/reviews、
+GET /tasks/<id>/reviews、POST /tasks/<id>/reviews/<rid>/findings。
+
+不跑真实模型、不联网：本地临时端口 + 临时 SQLite，与
+test_peer_messages_server.py 同款骨架。覆盖鉴权、权限、
+目标校验、队列查看、findings 写回与跨任务隔离。
+"""
+
+import http.client
+import json
+import os
+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
+
+
+class ReviewServerTestCase(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()
+        self.trace_dir = os.path.join(self.tmp.name, "host-traces")
+        os.makedirs(self.trace_dir, exist_ok=True)
+        self.server.trace_dir = self.trace_dir
+        status, self.project = self._req(
+            "POST", "/projects", {"name": "demo"}, self.owner_key)
+        self.assertEqual(status, 201)
+        status, self.project2 = self._req(
+            "POST", "/projects", {"name": "other"}, 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 _create_task(self, project_id):
+        status, task = self._req(
+            "POST", "/tasks",
+            {"project_id": project_id, "prompt": "做事"},
+            self.owner_key)
+        self.assertEqual(status, 201, task)
+        return task
+
+    def test_create_default_and_list(self):
+        task = self._create_task(self.project["id"])
+        status, created = self._req(
+            "POST", "/tasks/%s/reviews" % task["id"], {},
+            self.owner_key)
+        self.assertEqual(status, 201, created)
+        self.assertEqual(created["status"], "requested")
+        self.assertIsNone(created["findings"])
+        self.assertEqual(created["target_type"], "uncommittedChanges")
+        status, body = self._req(
+            "GET", "/tasks/%s/reviews" % task["id"], None,
+            self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(len(body["reviews"]), 1)
+        self.assertEqual(body["reviews"][0]["id"], created["id"])
+
+    def test_create_all_targets(self):
+        task = self._create_task(self.project["id"])
+        cases = [
+            ({"target_type": "baseBranch", "branch": "main"}, "main"),
+            ({"target_type": "commit", "sha": "abc123"}, "abc123"),
+            ({"target_type": "custom", "instructions": "look"}, None),
+        ]
+        for payload, _expected in cases:
+            status, body = self._req(
+                "POST", "/tasks/%s/reviews" % task["id"], payload,
+                self.owner_key)
+            self.assertEqual(status, 201, body)
+        status, body = self._req(
+            "GET", "/tasks/%s/reviews" % task["id"], None,
+            self.owner_key)
+        self.assertEqual(len(body["reviews"]), 3)
+
+    def test_create_target_validation(self):
+        task = self._create_task(self.project["id"])
+        bad = [
+            {"target_type": "baseBranch"},
+            {"target_type": "commit"},
+            {"target_type": "custom"},
+            {"target_type": "nope"},
+            {"target_type": "baseBranch", "branch": 123},
+        ]
+        for payload in bad:
+            status, body = self._req(
+                "POST", "/tasks/%s/reviews" % task["id"], payload,
+                self.owner_key)
+            self.assertEqual(status, 400, (payload, body))
+        status, body = self._req(
+            "GET", "/tasks/%s/reviews" % task["id"], None,
+            self.owner_key)
+        self.assertEqual(body["reviews"], [])
+
+    def test_findings_writeback(self):
+        task = self._create_task(self.project["id"])
+        _status, review = self._req(
+            "POST", "/tasks/%s/reviews" % task["id"], {},
+            self.owner_key)
+        status, body = self._req(
+            "POST",
+            "/tasks/%s/reviews/%s/findings" % (task["id"], review["id"]),
+            {"findings": "two issues"}, self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["status"], "completed")
+        self.assertEqual(body["findings"], "two issues")
+        status, body = self._req(
+            "GET", "/tasks/%s/reviews" % task["id"], None,
+            self.owner_key)
+        self.assertEqual(body["reviews"][0]["findings"], "two issues")
+
+    def test_findings_explicit_failed(self):
+        task = self._create_task(self.project["id"])
+        _status, review = self._req(
+            "POST", "/tasks/%s/reviews" % task["id"], {},
+            self.owner_key)
+        status, body = self._req(
+            "POST",
+            "/tasks/%s/reviews/%s/findings" % (task["id"], review["id"]),
+            {"status": "failed", "findings": "reviewer error"},
+            self.owner_key)
+        self.assertEqual(status, 200, body)
+        self.assertEqual(body["status"], "failed")
+
+    def test_findings_validation_and_isolation(self):
+        task = self._create_task(self.project["id"])
+        other = self._create_task(self.project["id"])
+        _status, review = self._req(
+            "POST", "/tasks/%s/reviews" % task["id"], {},
+            self.owner_key)
+        url = "/tasks/%s/reviews/%s/findings" % (
+            task["id"], review["id"])
+        status, _body = self._req("POST", url, {}, self.owner_key)
+        self.assertEqual(status, 400)
+        status, _body = self._req(
+            "POST", url, {"status": "bogus"}, self.owner_key)
+        self.assertEqual(status, 400)
+        status, _body = self._req(
+            "POST", url, {"findings": 123}, self.owner_key)
+        self.assertEqual(status, 400)
+        # 记录属于 task，挂在 other 任务下写回时 404，不泄露。
+        status, _body = self._req(
+            "POST",
+            "/tasks/%s/reviews/%s/findings" % (
+                other["id"], review["id"]),
+            {"findings": "x"}, self.owner_key)
+        self.assertEqual(status, 404)
+        status, _body = self._req(
+            "POST",
+            "/tasks/%s/reviews/%s/findings" % (
+                task["id"], "no-such-review"),
+            {"findings": "x"}, self.owner_key)
+        self.assertEqual(status, 404)
+
+    def test_auth_and_permissions(self):
+        task = self._create_task(self.project["id"])
+        status, _body = self._req(
+            "POST", "/tasks/%s/reviews" % task["id"], {})
+        self.assertEqual(status, 401)
+        status, _body = self._req(
+            "GET", "/tasks/%s/reviews" % task["id"], None)
+        self.assertEqual(status, 401)
+        status, _body = self._req(
+            "POST", "/tasks/%s/reviews" % task["id"], {},
+            self.outsider_key)
+        self.assertEqual(status, 403)
+        status, _body = self._req(
+            "GET", "/tasks/%s/reviews" % task["id"], None,
+            self.outsider_key)
+        self.assertEqual(status, 403)
+        _status, review = self._req(
+            "POST", "/tasks/%s/reviews" % task["id"], {},
+            self.owner_key)
+        status, _body = self._req(
+            "POST",
+            "/tasks/%s/reviews/%s/findings" % (
+                task["id"], review["id"]),
+            {"findings": "x"}, self.outsider_key)
+        self.assertEqual(status, 403)
+
+    def test_task_not_found(self):
+        status, _body = self._req(
+            "POST", "/tasks/no-such-task/reviews", {},
+            self.owner_key)
+        self.assertEqual(status, 404)
+        status, _body = self._req(
+            "GET", "/tasks/no-such-task/reviews", None,
+            self.owner_key)
+        self.assertEqual(status, 404)
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/cloud/worker.py b/cloud/worker.py
index 7b624c9..94a2861 100644
--- a/cloud/worker.py
+++ b/cloud/worker.py
@@ -147,6 +147,7 @@ from cloud.repos import RepoError, checkout_dir, resolve_repo
 from cloud.sandbox import CloudExecutor, DockerExecutor, LocalExecutor
 from cloud.uploads import upload_root as _upload_root
 from cloud.sandbox import _default_runner as _sandbox_default_runner
+from cloud.caps import effective_capacity, resolve_capacity
 from cloud.workers import heartbeat_worker, register_worker
 
 # 与 server.TERMINAL_STATUSES 对齐：这些状态的任务不再被认领或取消。
@@ -282,7 +283,7 @@ class Worker:
                  task_timeout=1800, max_retries=1,
                  retry_backoff_seconds=5.0, repos_dir="./repos",
                  repos_runner=None, heartbeat_file=None,
-                 worker_id=None, capacity=1, fallback_models=None,
+                 worker_id=None, capacity=None, fallback_models=None,
                  tenant_isolation=False, tenants_root=None,
                  tenant_template_home=None, tenant_template_repo=None,
                  egress_proxy_port=0, claim_executor_only=False,
@@ -412,10 +413,11 @@ class Worker:
         # 自然变 stale，不要用随机 uuid 让注册表无法对应同一台机器。
         self.worker_id = worker_id or "%s-%d" % (
             socket.gethostname(), os.getpid())
-        try:
-            self.capacity = int(capacity)
-        except (TypeError, ValueError):
-            self.capacity = 1
+        # 并发上限配置来源（见 cloud/caps.py）：显式构造参数 /
+        # CLI --worker-capacity > 环境变量 CLOUD_WORKER_CAPACITY >
+        # 默认 1；非法值归一为 1。self._env 在上面已定好，单测
+        # 注入的普通 dict 同样可读环境变量（无该键时默认 1）。
+        self.capacity = resolve_capacity(capacity, env=self._env)
         # 常驻启动时的代码签名：run_loop 每轮边界比对，变了就自退
         # （见 run_loop 的代码变化自退检查）。
         self._code_signature = _code_signature()
@@ -436,7 +438,8 @@ class Worker:
         try:
             register_worker(
                 self.db_path, self.worker_id,
-                executor=self._executor_name(), capacity=self.capacity)
+                executor=self._executor_name(), capacity=self.capacity,
+                running=self._running_count())
         except Exception as exc:
             log_event(self.logger, "worker_register_failed",
                       message="worker register failed: %s" % exc)
@@ -446,7 +449,8 @@ class Worker:
         try:
             heartbeat_worker(
                 self.db_path, self.worker_id,
-                executor=self._executor_name(), capacity=self.capacity)
+                executor=self._executor_name(), capacity=self.capacity,
+                running=self._running_count())
         except Exception as exc:
             log_event(self.logger, "worker_registry_heartbeat_failed",
                       message="worker registry heartbeat failed: %s" % exc)
@@ -531,7 +535,23 @@ class Worker:
         claim_executor_only 开启时（宿主级沙箱 worker 用）：WHERE
         只加 executor = ?，NULL 任务绝不认领（否则宿主 worker
         会抢走本该进 Docker 的任务）。
+
+        并发上限门（caps）：CAS 之前先在 _running_lock 下占位。
+        running 数达到 self.capacity 时直接返回 False，任务保持
+        queued，不碰数据库；占位成功才走 CAS。同一任务已在
+        _running 里时不重复计数（重复占位不增槽）。CAS 失败
+        （已被别的 worker 抢走）时，只清掉本次新建的占位，绝
+        不动已在跑任务的条目。占位条目在执行路径的 finally
+        里清（本地 _execute_via_process / Docker 同步路径），
+        认领后提前返回的三道门（缺 key / 仓库 / 配额）在
+        run_one 里显式清占位；CAS 语义本身一句未改。
         """
+        with self._running_lock:
+            already_running = task_id in self._running
+            if not already_running and \
+                    len(self._running) >= effective_capacity(self.capacity):
+                return False
+            self._running.setdefault(task_id, {})
         now = time.time()
         conn = _db_connect(self.db_path)
         try:
@@ -563,7 +583,10 @@ class Worker:
                     (now, now, self.host_name, task_id, executor),
                 )
             conn.commit()
-            return cur.rowcount == 1
+            claimed = cur.rowcount == 1
+            if not claimed and not already_running:
+                self._running_clear(task_id)
+            return claimed
         finally:
             conn.close()
 
@@ -908,6 +931,16 @@ class Worker:
             return {tid: dict(entry)
                     for tid, entry in self._running.items()}
 
+    def _running_count(self):
+        """当前占用槽位数（_running_lock 下计数）。
+
+        启动中的任务已在 _claim 里占位（空条目），所以这里数
+        到的既包括已起进程 / 容器的任务，也包括已认领未起
+        进程的任务；注册表心跳与认领门共用这个口径。
+        """
+        with self._running_lock:
+            return len(self._running)
+
     def _wait_for_running(self, timeout=None):
         """等 _running 变空；timeout 为 None 时无限等。
 
@@ -2082,6 +2115,9 @@ class Worker:
                       project_id=task["project_id"], status="failed")
             self._notify_task_finished(
                 task_id, task["project_id"], "failed", error=error)
+            # 认领占位在此释放：缺 key 门在执行器起进程之前返回，
+            # 不清占位会把槽位漏到 worker 退出（见 _claim 文档）。
+            self._running_clear(task_id)
             return {"dispatched": True, "task_id": task_id,
                     "host_name": self.executor._host_name(task)
                     if hasattr(self.executor, "_host_name") else task_id,
@@ -2106,6 +2142,8 @@ class Worker:
                       project_id=task["project_id"], status="failed")
             self._notify_task_finished(
                 task_id, task["project_id"], "failed", error=repo_error)
+            # 同缺 key 门：仓库解析失败在起进程之前返回，先清占位。
+            self._running_clear(task_id)
             return {"dispatched": True, "task_id": task_id,
                     "host_name": self.executor._host_name(task)
                     if hasattr(self.executor, "_host_name") else task_id,
@@ -2129,6 +2167,8 @@ class Worker:
             log_event(self.logger, "worker_quota_stop", task_id=task_id,
                       project_id=task["project_id"],
                       status="interrupted")
+            # 同缺 key 门：配额门在起进程之前返回，先清占位。
+            self._running_clear(task_id)
             return {"dispatched": True, "task_id": task_id,
                     "host_name": self.executor._host_name(task)
                     if hasattr(self.executor, "_host_name") else task_id,
@@ -2736,6 +2776,11 @@ class Worker:
                 if stop_event is not None and stop_event.is_set():
                     break
                 self._heartbeat_running_tasks()
+                # 注册表心跳也每 30 秒刷一次：长任务跑满一轮时
+                # run_all 久久不返回，只靠轮次结束的心跳会让
+                # /workers 的 running 数长期停在旧值；这里刷的
+                # running 含启动中占位，与认领门口径一致。
+                self._registry_heartbeat()
                 hb_stop.wait(30)
 
         hb_thread = threading.Thread(target=_hb_loop, daemon=True)
@@ -2795,6 +2840,14 @@ class Worker:
                     time.sleep(0.2)
         finally:
             hb_stop.set()
+            # 等心跳线程彻底退出再返回：否则它手头一次未写完的
+            # 注册表心跳可能在调用方清理 DB 目录时重建 control.db，
+            # 导致 tearDown rmtree 报 Directory not empty（全量套件
+            # 里 test_code_changed_exits_loop_before_claiming 偶发
+            # error 即此竞态）。daemon 线程 + set() 会唤醒 wait，
+            # 正常情况下 join 瞬间返回；5 秒超时兜底，避免写挂时
+            # 拖住退出。
+            hb_thread.join(timeout=5)
             # 优雅排空：先等正在跑的任务自然结束，超时后才杀。
             # 串行模式下 run_all 返回时任务已跑完，这里立刻通过；
             # 并行模式下停机信号退出时后台线程可能还在跑，先等
@@ -2900,10 +2953,12 @@ def main(argv=None):
     parser.add_argument("--worker-id", default=None,
                         help="多 worker 池身份（默认：主机名-进程号；"
                              "生产部署应配稳定 id，否则重启后旧行变 stale）")
-    parser.add_argument("--worker-capacity", type=int, default=1,
+    parser.add_argument("--worker-capacity", type=int, default=None,
                         help="worker 并发上限：同一 worker 内最多同时跑"
-                             "几个任务（默认 1，即旧串行行为；>1 时用"
-                             "线程池并行，认领仍是 CAS，成本写库持锁）")
+                             "几个任务（未给时读环境变量"
+                             " CLOUD_WORKER_CAPACITY，未设则默认 1，"
+                             "即旧串行行为；>1 时用线程池并行，认领"
+                             "仍是 CAS，成本写库持锁）")
     parser.add_argument("--host-name", default=None,
                         help="本 worker 所在主机名（HA 阶段 1）：认领任务"
                              "时写进 tasks.worktree_host，/pr 据此判断"
diff --git a/cloud/workers.py b/cloud/workers.py
index 73d16f9..a3df77c 100644
--- a/cloud/workers.py
+++ b/cloud/workers.py
@@ -62,10 +62,12 @@ def ensure_workers_table(db_path):
             "worker_id TEXT PRIMARY KEY, "
             "executor TEXT, "
             "capacity INTEGER, "
+            "running INTEGER, "
             "last_seen DOUBLE PRECISION, "
             "created_at DOUBLE PRECISION)"
         )
         conn.commit()
+        _ensure_running_column(conn)
         # 老库升精度只在确实是 Postgres REAL 列时做一次：每次心跳都
         # ALTER 会反复拿 ACCESS EXCLUSIVE 锁，没必要。先查
         # information_schema（SQLite 下这张表不存在，查询失败即跳过，
@@ -98,7 +100,59 @@ def ensure_workers_table(db_path):
         conn.close()
 
 
-def register_worker(db_path, worker_id, executor=None, capacity=1):
+def _ensure_running_column(conn):
+    """老库补 running 列（新表建表时已含该列）。
+
+    running 是 worker 上次心跳上报的当前占用槽数（启动中占位
+    也算，见 cloud/caps.py）；列缺失时 _worker_to_dict 默认 0，
+    所以补列失败只回滚、不抛，绝不炸注册与心跳。
+    SQLite 用 PRAGMA 查列，Postgres 用 information_schema。
+    """
+    try:
+        rows = conn.execute("PRAGMA table_info(workers)").fetchall()
+        names = set()
+        for row in rows:
+            try:
+                names.add(row[1])
+            except (IndexError, TypeError):
+                pass
+        if rows and "running" not in names:
+            conn.execute(
+                "ALTER TABLE workers ADD COLUMN running INTEGER DEFAULT 0")
+            conn.commit()
+        return
+    except Exception:
+        try:
+            conn.rollback()
+        except Exception:
+            pass
+    try:
+        row = conn.execute(
+            "SELECT COUNT(*) FROM information_schema.columns "
+            "WHERE table_name = 'workers' AND column_name = 'running'"
+            ).fetchone()
+        exists = bool(row and row[0])
+    except Exception:
+        try:
+            conn.rollback()
+        except Exception:
+            pass
+        return
+    if not exists:
+        try:
+            conn.execute(
+                "ALTER TABLE workers ADD COLUMN IF NOT EXISTS "
+                "running INTEGER DEFAULT 0")
+            conn.commit()
+        except Exception:
+            try:
+                conn.rollback()
+            except Exception:
+                pass
+
+
+def register_worker(db_path, worker_id, executor=None, capacity=1,
+                    running=None):
     """注册（或刷新）一个 worker：不存在 INSERT，存在 UPDATE。
 
     返回注册行 dict。worker_id 为空时抛 ValueError，调用方
@@ -112,14 +166,17 @@ def register_worker(db_path, worker_id, executor=None, capacity=1):
     try:
         cur = conn.execute(
             "UPDATE workers SET executor = ?, capacity = ?,"
-            " last_seen = ? WHERE worker_id = ?",
-            (executor, capacity, now, worker_id),
+            " running = ?, last_seen = ? WHERE worker_id = ?",
+            (executor, capacity, running if running is not None else 0,
+             now, worker_id),
         )
         if cur.rowcount == 0:
             conn.execute(
                 "INSERT INTO workers(worker_id, executor, capacity,"
-                " last_seen, created_at) VALUES(?, ?, ?, ?, ?)",
-                (worker_id, executor, capacity, now, now),
+                " running, last_seen, created_at)"
+                " VALUES(?, ?, ?, ?, ?, ?)",
+                (worker_id, executor, capacity,
+                 running if running is not None else 0, now, now),
             )
         conn.commit()
         row = conn.execute(
@@ -130,7 +187,8 @@ def register_worker(db_path, worker_id, executor=None, capacity=1):
     return _worker_to_dict(row)
 
 
-def heartbeat_worker(db_path, worker_id, executor=None, capacity=None):
+def heartbeat_worker(db_path, worker_id, executor=None, capacity=None,
+                     running=None):
     """心跳：就是带 last_seen 刷新的注册。
 
     executor/capacity 未给时保留旧值（已注册的 worker 只刷时间，
@@ -148,17 +206,19 @@ def heartbeat_worker(db_path, worker_id, executor=None, capacity=None):
         if row is None:
             conn.execute(
                 "INSERT INTO workers(worker_id, executor, capacity,"
-                " last_seen, created_at) VALUES(?, ?, ?, ?, ?)",
+                " running, last_seen, created_at)"
+                " VALUES(?, ?, ?, ?, ?, ?)",
                 (worker_id, executor, capacity if capacity is not None
-                 else 1, now, now),
+                 else 1, running if running is not None else 0, now, now),
             )
         else:
             old = _worker_to_dict(row)
             conn.execute(
                 "UPDATE workers SET executor = ?, capacity = ?,"
-                " last_seen = ? WHERE worker_id = ?",
+                " running = ?, last_seen = ? WHERE worker_id = ?",
                 (executor if executor is not None else old["executor"],
                  capacity if capacity is not None else old["capacity"],
+                 running if running is not None else old["running"],
                  now, worker_id),
             )
         conn.commit()
@@ -197,10 +257,12 @@ def _worker_to_dict(row):
             return row[name]
         except (KeyError, IndexError, TypeError):
             return None
+    running = _get("running")
     return {
         "worker_id": _get("worker_id"),
         "executor": _get("executor"),
         "capacity": _get("capacity"),
+        "running": running if running is not None else 0,
         "last_seen": _get("last_seen"),
         "created_at": _get("created_at"),
     }
diff --git a/host.py b/host.py
index a8d8924..db5d435 100644
--- a/host.py
+++ b/host.py
@@ -511,6 +511,18 @@ class Server:
             raise RuntimeError(f"thread/resume 失败: {resp}")
         return resp["result"]
 
+    def start_review(self, thread_id, target, timeout=600):
+        # 协议 review/start：params 必填 threadId 与 target，
+        # target 四种见 protocol v2 的 ReviewTarget（uncommitted
+        # Changes / baseBranch / commit / custom）；delivery 缺
+        # 省即 inline，不发 detached（已废弃）。失败直接抛错，
+        # 调用方不得把 error 响应当成功打印。
+        params = {"threadId": thread_id, "target": target}
+        resp = self.request("review/start", params, timeout=timeout)
+        if not resp or "error" in resp:
+            raise RuntimeError(f"review/start 失败: {resp}")
+        return resp["result"]
+
     def fork_thread(self, thread_id, cwd=None, sandbox=None):
         params = {"threadId": thread_id}
         if cwd:
@@ -2012,6 +2024,13 @@ def main():
     p = sub.add_parser("archive"); p.add_argument("--task", required=True)
     p = sub.add_parser("fork"); p.add_argument("--task", required=True); p.add_argument("--name", required=True); p.add_argument("--sandbox", default=None)
     p = sub.add_parser("compact"); p.add_argument("--task", required=True)
+    p = sub.add_parser("code-review"); p.add_argument("--task", required=True)
+    p.add_argument("--base-branch", default=None,
+                   help="评审目标：与该分支的差异（baseBranch）")
+    p.add_argument("--commit", default=None,
+                   help="评审目标：指定提交 sha（commit）")
+    p.add_argument("--instructions", default=None,
+                   help="评审目标：自定义评审指令（custom）")
     p = sub.add_parser("run"); p.add_argument("--task", required=True); p.add_argument("--prompt", required=True)
     p = sub.add_parser("interrupt"); p.add_argument("--task", required=True)
     p = sub.add_parser("delegate")
@@ -2460,6 +2479,34 @@ def main():
             resp = srv.request("thread/compact/start",
                                {"threadId": t["thread_id"]}, timeout=120)
             print(f"compact {args.task}: {resp}")
+        elif args.cmd == "code-review":
+            # 注意命名：host.py 已有另一条线的 review 子命令
+            # （PR 评审评论），本命令换名 code-review，不覆盖它。
+            # 目标优先级：commit > instructions > base-branch >
+            # 缺省 uncommittedChanges；与 cloud/review.py 的
+            # build_review_target 同款组合，非法组合直接报错。
+            # compact 同款原因先 resume 再 review（未加载的旧
+            # thread 直接 review 会报 thread not found）。
+            # app-server 返回 error 时只打印失败并以非零退出，
+            # 绝不把失败结果当成评审结论。
+            t = state["tasks"][args.task]
+            if args.commit:
+                target = {"type": "commit", "sha": args.commit}
+            elif args.instructions:
+                target = {"type": "custom",
+                          "instructions": args.instructions}
+            elif args.base_branch:
+                target = {"type": "baseBranch",
+                          "branch": args.base_branch}
+            else:
+                target = {"type": "uncommittedChanges"}
+            try:
+                srv.resume_thread(t["thread_id"])
+                result = srv.start_review(t["thread_id"], target)
+            except Exception as exc:
+                print(f"code-review failed: {exc}")
+                raise SystemExit(1)
+            print(f"code-review {args.task}: {result}")
         elif args.cmd == "archive":
             t = state["tasks"][args.task]
             srv.archive_thread(t["thread_id"])