watch: a stale cursor asks for twice as much; work: probes do not ask for a metadata service; msgpack 1.2.4

deploy/docker-compose.work.yml

@@ -43,6 +43,14 @@ x-probe-proxy: &probe-proxy
   HTTP_PROXY: ${HOOKPROBE_EGRESS_PROXY:-http://egress-proxy:8888}
   HTTPS_PROXY: ${HOOKPROBE_EGRESS_PROXY:-http://egress-proxy:8888}
   NO_PROXY: ${HOOKPROBE_NO_PROXY:-127.0.0.1,localhost,host.docker.internal,hookrelay,sink,egress-proxy,mcp-gate,watch-signer}
+  # No instance metadata service exists on the probes' network, and their AWS
+  # credentials are files. Without this, an AWS CLI call whose profile names no
+  # region asks 169.254.169.254 for one through the proxy above, and each ask
+  # lands in the egress ledger as a refusal that is nobody's attempt to leave —
+  # a second deployment running copies of these files saw six in one session.
+  # The CLI now says "specify a region" instead; the posture check skips the
+  # metadata question for the same reason.
+  AWS_EC2_METADATA_DISABLED: "true"
 
 services:
   hookrelay:

deploy/watch/tests/test_scan.py

@@ -232,3 +232,38 @@ def test_a_bot_that_says_so_in_its_name_never_reaches_the_digest(monkeypatch):
     )
     found = watch_scan.scan_chat(chat, {}, watch_scan.when("2026-10-27T02:10:00Z"), [])
     assert [(name, [who for _, who, _ in msgs]) for name, msgs in found] == [("ops room", ["Talbot"])]
+
+
+class _Feeds:
+    """A chat service with the given feeds; records every search it is asked."""
+
+    def __init__(self, feeds):
+        self.feeds, self.asked = feeds, []
+
+    def call(self, tool, **args):
+        if tool == "list_folder_feeds":
+            return {"feeds": self.feeds}
+        assert tool == "search_chat_records", tool
+        self.asked.append(args)
+        return {"messages": []}
+
+
+def test_a_stale_cursor_asks_for_twice_as_much_up_to_the_servers_ceiling(monkeypatch):
+    """A Monday's first round: the weekend group's cursor is Friday's evening."""
+    now = watch_scan.when("2026-11-02T01:40:00Z")
+    feeds = [
+        {"name": "weekend group", "chat_type_label": "group chat", "last_message_send_at": "2026-11-01T07:00:00Z"},
+        {"name": "weekday group", "chat_type_label": "group chat", "last_message_send_at": "2026-11-02T01:30:00Z"},
+    ]
+    cursors = {
+        "weekend group": watch_scan.when("2026-10-30T11:20:00Z"),
+        "weekday group": watch_scan.when("2026-11-02T01:00:00Z"),
+    }
+    monkeypatch.setattr(watch_scan, "PER_CHAT", 60)
+    chat = _Feeds(feeds)
+    watch_scan.scan_chat(chat, dict(cursors), now, [])
+    assert {a["contact_name"]: a["limit"] for a in chat.asked} == {"weekend group": 120, "weekday group": 60}
+    monkeypatch.setattr(watch_scan, "PER_CHAT", 150)
+    chat = _Feeds(feeds)
+    watch_scan.scan_chat(chat, dict(cursors), now, [])
+    assert {a["contact_name"]: a["limit"] for a in chat.asked} == {"weekend group": 200, "weekday group": 150}

deploy/watch/watch_scan.py

@@ -93,6 +93,12 @@
 # 60 而不是 12：截断要罕见到算异常。上限是 200，一个 20 分钟的窗口里 60 条已经
 # 是刷屏了。真截断了下面会明说，因为结果是新的在前，被丢掉的是更早的那些。
 PER_CHAT = int(os.environ.get("WATCH_MAX_PER_CHAT") or 60)
+# 游标比这还旧的会话（周末之后、长假之后、刚出现的会话）积压的不是一轮的量：
+# 按常量取，取满了就把更早的丢掉。另一套跑着这些文件副本的部署，在两个周一的首轮
+# 各有两个会话「一次取满 60 条」。这种会话单次多取一倍，仍不超过服务端对 limit 的
+# 上限；工作日的常规轮次不变。
+STALE_CURSOR_SECONDS = 36 * 3600
+CHAT_LIMIT_MAX = 200  # search_chat_records 的 limit 上限，超过会被整个调用拒掉
 # 服务端对这个参数的上限就是 100，传 200 会被整个调用拒掉——而一次被拒的
 # 列表调用意味着这一轮完全没看聊天。
 FEED_LIMIT = int(os.environ.get("WATCH_FEED_LIMIT") or 100)
@@ -396,13 +402,14 @@ def scan_chat(
                 continue
             # 到期放行。取完（或取失败走 unreachable）都在下面清 pending。
 
+        limit = PER_CHAT if now - base < STALE_CURSOR_SECONDS else min(PER_CHAT * 2, CHAT_LIMIT_MAX)
         args = {
             "contact_name": name,
             # from_time 是闭区间。原样传游标会把「设定这个游标的那条消息」再取
             # 一遍——2026-09-04 就是这么把一条五小时前的 @All 当成新消息发出去的。
             "from_time": str(int(base) + 1),
             "to_time": str(int(now)),
-            "limit": PER_CHAT,
+            "limit": limit,
         }
         if feed.get("chat_type_label") == "group chat":
             args["search_group"] = True
@@ -467,9 +474,9 @@ def scan_chat(
             unreachable.pop(name, None)  # 又够得着了；下次再断会重新说一次
         if pending is not None:
             pending.pop(name, None)  # 攒的一批已经端出去了；下一批重新起算
-        if len(got) >= PER_CHAT:
+        if len(got) >= limit:
             notes.append(
-                f"{name}: 一次取满 {PER_CHAT} 条，更早的可能没取到（结果是新的在前，游标仍推进到最新，否则会一直重报）"
+                f"{name}: 一次取满 {limit} 条，更早的可能没取到（结果是新的在前，游标仍推进到最新，否则会一直重报）"
             )
         if picked:
             findings.append((name, picked))

hookjudge/requirements.lock

@@ -24,7 +24,7 @@ librt==0.16.0
 license-expression==30.4.4
 markdown-it-py==4.2.0
 mdurl==0.1.2
-msgpack==1.2.1
+msgpack==1.2.4
 mypy==2.4.0
 mypy_extensions==1.1.0
 opentelemetry-api==1.45.1

hookprobe/requirements.lock

@@ -33,7 +33,7 @@ markdown-it-py==4.2.0
 mcp==2.3.0
 mcp-types==2.3.0
 mdurl==0.1.2
-msgpack==1.2.1
+msgpack==1.2.4
 mypy==2.4.0
 mypy_extensions==1.1.0
 opentelemetry-api==1.45.1

hookrelay/requirements.lock

@@ -24,7 +24,7 @@ librt==0.16.0
 license-expression==30.4.4
 markdown-it-py==4.2.0
 mdurl==0.1.2
-msgpack==1.2.1
+msgpack==1.2.4
 mypy==2.4.0
 mypy_extensions==1.1.0
 opentelemetry-api==1.45.1