cloudagent-daemon 0.5.4__tar.gz → 0.5.6__tar.gz

{cloudagent_daemon-0.5.4 → cloudagent_daemon-0.5.6}/CHANGELOG.md

@@ -4,6 +4,55 @@ All notable changes to the local daemon. Format follows
 [Keep a Changelog](https://keepachangelog.com/); versions are
 [PEP 440](https://peps.python.org/pep-0440/). Release with `scripts/release.sh`.
 
+## [0.5.6] — 2026-08-30
+
+### Changed
+- Every claude_code wake now passes `--autocompact 200k`. A lane is re-read in
+  full on every wake, so an unbounded one charges its whole history each time —
+  one agent's lane climbed to 1.07M cached tokens and was replaying all of it 96
+  times a day to add a few thousand tokens of new content. An explicit size is
+  used rather than `auto` because `auto` tracks the model's own window (1M on
+  opus), which is precisely the runaway being bounded. This prevents what
+  `_is_context_exhausted` (0.5.x, #114) only cures after a wake has already
+  failed and its session been discarded.
+- A timer-fired wake (`message_id="remind:<id>:<n>"`) now runs in a throwaway
+  session: it does not resume a lane, does not record one, and its failure may
+  not clear one. A reminder tick is a stateless patrol — what it must decide
+  comes from the room's current state, which it reads live, not from what it
+  said on the previous tick.
+
+### Notes
+- Recording a patrol's session as the room's lane was the same bleed as #95,
+  arriving by another door: the next human to speak in that room inherited the
+  timer's accumulated history. Ordinary wakes are unaffected — a human message
+  in that same room still resumes the room's lane and still records its session.
+- Hosts on 0.5.5 will see supervision-style agents start each tick fresh. That
+  is the change, not a regression: such an agent should be reading the room, not
+  replaying itself.
+
+## [0.5.5] — 2026-08-27
+
+### Changed
+- Every engine now gets one runtime session per conversation, not one per agent.
+  An agent resumed the same CLI session no matter which channel woke it, so
+  #debug, the DM and #SSS all shared a single context: the per-channel prompt
+  leaked between rooms, and each wake replayed the union of every room's
+  history — muddled context and a token bill that grew without bound. Lanes
+  already fixed this for qoder; the bleed was never qoder-specific, so
+  claude_code and codex now resume from the same per-conversation lane, keyed
+  on the waking conversation (channel, DM and thread each carry their own).
+  Wakes with no conversation — timers, spawn chains — keep the global lane.
+
+### Notes
+- The lane file is now `.meta/sessions.json`. `qoder-sessions.json` is still
+  read so an upgraded host keeps its lanes; the first write folds it into the
+  new file and removes it, or a cleared lane would come back to life from the
+  stale value left behind.
+- Agents carrying a global session into this version start fresh on their first
+  wake in each conversation. That is the fix rather than a side effect —
+  inheriting it is precisely the bleed being removed — but expect an agent to
+  look like it "forgot" an in-flight thread once, right after the upgrade.
+
 ## [0.5.4] — 2026-08-24
 
 ### Changed

{cloudagent_daemon-0.5.4 → cloudagent_daemon-0.5.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: cloudagent-daemon
-Version: 0.5.4
+Version: 0.5.6
 Summary: Local daemon for cloudagent — bridges Claude Code, Codex, and Qoder CLI agents to the cloudagent backend.
 Project-URL: Homepage, https://cloudagent.chat
 Project-URL: Repository, https://codeup.aliyun.com/69bd0a4d29ad98af4065ce71/cloudagent

{cloudagent_daemon-0.5.4 → cloudagent_daemon-0.5.6}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "cloudagent-daemon"
-version = "0.5.4"
+version = "0.5.6"
 description = "Local daemon for cloudagent — bridges Claude Code, Codex, and Qoder CLI agents to the cloudagent backend."
 authors = [{name = "chaoyin", email = "tudoucyx@gmail.com"}]
 readme = "README.md"

cloudagent_daemon-0.5.6/src/cloudagent_daemon/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.5.6"

{cloudagent_daemon-0.5.4 → cloudagent_daemon-0.5.6}/src/cloudagent_daemon/claude_runner.py

@@ -474,6 +474,16 @@ class ClaudeRunner:
         "no rollout found for thread",
     )
 
+    #: #114 followup — ``--autocompact`` window for claude_code wakes.
+    #:
+    #: An explicit size rather than ``auto``: ``auto`` tracks the model's own
+    #: window (1M on opus), which is the runaway this exists to bound — a lane
+    #: is replayed in full on every wake, so "compacts just before overflow"
+    #: still means every wake pays for a near-window-sized history. 200k keeps
+    #: days of one room's working context while cutting per-wake replay ~5x.
+    #: The CLI accepts ``auto`` or 100k–1M (``claude --help``).
+    AUTOCOMPACT_WINDOW: str = "200k"
+
     #: #114 — 上下文窗口撑满。这是**永久性**前提失败,不是瞬时错误:退避几秒
     #: 再把同一份历史送一次,必然再次溢出。codex 自己给出的处置就是 "Start a
     #: new thread or clear earlier history",与 stale-session 的补救动作完全
@@ -680,11 +690,10 @@ The daemon discards stdout — only the mcp tool call reaches the channel.
             stale_conv = (
                 str(payload["conversation_id"]) if payload.get("conversation_id") else None
             )
-            stale_is_qoder = self._runtime_engine_for(ws, payload) == "qoder"
-            qoder_lane_sid = (
-                self._qoder_lane_session(ws, stale_conv)
-                if stale_is_qoder and stale_conv else None
-            )
+            # #95 — the lane IS the session this wake resumed from (the
+            # global last_session only when there is no conversation), so
+            # it is also the one to drop once it proves unusable.
+            stale_lane_sid = self._lane_session(ws, stale_conv)
             # #114 — 上下文撑满与 stale-session 是同一类问题：**当前会话已经
             # 不能再用了**，区别只在于一个是引擎不认识它、另一个是它太大。补救
             # 动作完全相同（清掉会话、免 resume 重跑一次），所以走同一条路径，
@@ -693,6 +702,10 @@ The daemon discards stdout — only the mcp tool call reaches the channel.
             context_exhausted = self._is_context_exhausted(attempt)
             if (
                 self._run_mode_for(payload) != "read_analyze"
+                # A periodic wake resumed nothing, so a stale/exhausted session
+                # is not its to drop — clearing here would evict the lane the
+                # room's humans are still using.
+                and not self._is_periodic_wake(payload)
                 and attempt.exit_code != 0
                 and (
                     any(
@@ -701,12 +714,12 @@ The daemon discards stdout — only the mcp tool call reaches the channel.
                     )
                     or context_exhausted
                 )
-                and (ws.last_session_id() is not None or qoder_lane_sid is not None)
+                and stale_lane_sid is not None
             ):
-                stale_sid = qoder_lane_sid or ws.last_session_id()
-                # #95 — qoder 的过期会话在 lane 里，清 lane 才能真正脱离它
-                if stale_is_qoder and stale_conv:
-                    self._qoder_set_lane(ws, stale_conv, None)
+                stale_sid = stale_lane_sid
+                # #95 — 过期会话就在 lane 里，清 lane 才能真正脱离它
+                if stale_conv:
+                    self._set_lane(ws, stale_conv, None)
                 logger.warning(
                     "claude_runner: dropping unusable session_id=%s "
                     "(cause=%s, detail=%r). Clearing and retrying "
@@ -782,21 +795,24 @@ The daemon discards stdout — only the mcp tool call reaches the channel.
         # 防外人注入的上下文渗进 owner 后续 full 会话。仍上报 session_start
         # 供实时展示,但不落 last-spawn(下次 full 唤醒仍 resume 干净的 owner 轨)。
         is_read_analyze = self._run_mode_for(payload) == "read_analyze"
-        # #95 — qoder 按 conversation_id 分轨存会话；全局 last_session 仍写
-        # （last-spawn.json 可观测性 + 非平台唤醒兜底），但 qoder 平台唤醒的
-        # resume 只读 lane（见 _build_cmd），不会被别的频道的会话覆盖。
+        # #114 followup — a periodic wake ran in a throwaway session (it never
+        # resumed one). Recording it would hand the room's humans a lane whose
+        # history is a patrol tick, which is the bleed we just removed.
+        is_periodic = self._is_periodic_wake(payload)
+        # #95 — 所有引擎都按 conversation_id 分轨存会话；全局 last_session 仍写
+        # （last-spawn.json 可观测性 + 非平台唤醒兜底），但平台唤醒的 resume
+        # 只读 lane（见 _build_cmd），不会被别的频道的会话覆盖。
         conv_for_lane = (
             str(payload["conversation_id"]) if payload.get("conversation_id") else None
         )
-        if (
-            captured_session
-            and not is_read_analyze
-            and conv_for_lane
-            and self._runtime_engine_for(ws, payload) == "qoder"
-        ):
-            self._qoder_set_lane(ws, conv_for_lane, captured_session)
-        if captured_session and captured_session != ws.last_session_id():
-            if not is_read_analyze:
+        # 比较基准取**本轨**上一个会话，不是全局 last_session：分轨后全局会被
+        # 最后说话的频道盖掉，拿它比会让 A/B 频道交替唤醒每轮都像是新会话，
+        # 把 session_start 刷成噪音。必须在 _set_lane 之前读。
+        prev_lane_session = self._lane_session(ws, conv_for_lane)
+        if captured_session and captured_session != prev_lane_session:
+            if not is_read_analyze and not is_periodic:
+                if conv_for_lane:
+                    self._set_lane(ws, conv_for_lane, captured_session)
                 ws.set_last_session_id(captured_session)
             await self.outbound.send_agent_session_start(
                 agent_id, captured_session
@@ -1382,6 +1398,7 @@ The daemon discards stdout — only the mcp tool call reaches the channel.
         model, reasoning_effort = self._model_and_effort_for(ws, payload)
         runtime_env = self._runtime_env_for(ws, payload)
         run_mode = self._run_mode_for(payload)
+        periodic = self._is_periodic_wake(payload)
         scratch_dir = (
             self._readonly_scratch_dir(ws, payload)
             if run_mode == "read_analyze"
@@ -1397,6 +1414,7 @@ The daemon discards stdout — only the mcp tool call reaches the channel.
             reasoning_effort=reasoning_effort,
             run_mode=run_mode,
             scratch_dir=scratch_dir,
+            periodic=periodic,
             conversation_id=(
                 str(payload["conversation_id"])
                 if payload.get("conversation_id") else None
@@ -1920,6 +1938,27 @@ TASKS (频道任务板) — 频道里的分工协作走任务认领原语，纪
         raw = payload.get("run_mode")
         return "read_analyze" if raw == "read_analyze" else "full"
 
+    @staticmethod
+    def _is_periodic_wake(payload: dict) -> bool:
+        """True for a timer-fired wake — one occurrence of a reminder.
+
+        Backend stamps these with the synthetic message id
+        ``remind:<reminder_id>:<occurrence>``.
+
+        A periodic wake is a *stateless patrol*: what it has to decide comes
+        from the room's CURRENT state, which it reads live (platform_read),
+        not from what it said on the previous tick. So it neither resumes a
+        lane nor writes one back — it runs in its own throwaway session.
+
+        Before this, a 15-minute supervision timer resumed its channel's lane
+        96 times a day, replaying an ever-growing history to contribute a few
+        thousand tokens of new content — the replay reached ~1M tokens per
+        wake — and left that history behind for the humans in the room to
+        inherit on their next message.
+        """
+        mid = payload.get("message_id")
+        return isinstance(mid, str) and mid.startswith("remind:")
+
     def _readonly_scratch_dir(
         self, ws: AgentWorkspaceLike, payload: dict
     ) -> "Path":
@@ -2098,53 +2137,79 @@ TASKS (频道任务板) — 频道里的分工协作走任务认领原语，纪
         )
 
     # ------------------------------------------------------------------
-    # #95 — qoder runtime-session lanes, one per conversation.
+    # #95 — runtime-session lanes, one per conversation.
     #
-    # qodercli ``--resume`` does not reliably adopt a new
-    # ``--append-system-prompt`` on a resumed session, and a single global
-    # last_session shared by DM/#debug/#SSS made the per-channel prompt
-    # (#89) bleed across rooms. Each conversation therefore gets its own
-    # lane: ``.meta/qoder-sessions.json`` = {conversation_id: session_id}.
+    # A single global last_session shared by DM/#debug/#SSS made every room
+    # bleed into every other one: the agent resumed the same CLI session no
+    # matter which channel woke it, so #89's per-channel prompt leaked
+    # across rooms and each wake replayed the union of all rooms' history
+    # (context noise + token burn). Each conversation therefore gets its own
+    # lane: ``.meta/sessions.json`` = {conversation_id: session_id}.
     # Non-platform wakes (no conversation_id) keep the legacy global lane.
+    #
+    # Lanes started out qoder-only (qodercli ``--resume`` additionally
+    # refuses to adopt a new ``--append-system-prompt``), but the bleed is
+    # engine-independent — claude_code and codex resumed one global session
+    # too. All engines now share this path; the superseded qoder file is
+    # still read so an upgraded host keeps its existing lanes.
     # ------------------------------------------------------------------
 
+    #: Superseded lane file — read-only, for hosts upgraded from the
+    #: qoder-only era. Writes always land in ``sessions.json``.
+    LEGACY_QODER_LANES_FILENAME = "qoder-sessions.json"
+
     @staticmethod
-    def _qoder_lanes_path(ws: AgentWorkspaceLike) -> "Path":
-        return ws.path / ".meta" / "qoder-sessions.json"
+    def _lanes_path(ws: AgentWorkspaceLike) -> "Path":
+        return ws.path / ".meta" / "sessions.json"
+
+    @classmethod
+    def _read_lanes(cls, ws: AgentWorkspaceLike) -> dict:
+        """Current lane file layered over the superseded qoder one."""
+        merged: dict = {}
+        for path in (
+            ws.path / ".meta" / cls.LEGACY_QODER_LANES_FILENAME,
+            cls._lanes_path(ws),
+        ):
+            try:
+                data = json.loads(path.read_text(encoding="utf-8"))
+            except (FileNotFoundError, json.JSONDecodeError, OSError):
+                continue
+            if isinstance(data, dict):
+                merged.update(data)
+        return merged
 
-    def _qoder_lane_session(
+    def _lane_session(
         self, ws: AgentWorkspaceLike, conversation_id: str | None
     ) -> str | None:
         if not conversation_id:
             return ws.last_session_id()
-        try:
-            data = json.loads(self._qoder_lanes_path(ws).read_text(encoding="utf-8"))
-        except (FileNotFoundError, json.JSONDecodeError, OSError):
-            return None
-        sid = data.get(conversation_id) if isinstance(data, dict) else None
+        # Deliberately NO fallback to the global last_session: the first
+        # wake in a conversation must start fresh, or it would inherit
+        # whatever room spoke last — the exact bleed lanes exist to remove.
+        sid = self._read_lanes(ws).get(conversation_id)
         return sid if isinstance(sid, str) and sid else None
 
-    def _qoder_set_lane(
+    def _set_lane(
         self, ws: AgentWorkspaceLike, conversation_id: str, session_id: str | None
     ) -> None:
         """Set (or clear with None) the lane. Best-effort; never raises."""
-        p = self._qoder_lanes_path(ws)
+        p = self._lanes_path(ws)
+        legacy = ws.path / ".meta" / self.LEGACY_QODER_LANES_FILENAME
         try:
-            try:
-                data = json.loads(p.read_text(encoding="utf-8"))
-                if not isinstance(data, dict):
-                    data = {}
-            except (FileNotFoundError, json.JSONDecodeError, OSError):
-                data = {}
+            data = self._read_lanes(ws)
             if session_id:
                 data[conversation_id] = session_id
             else:
                 data.pop(conversation_id, None)
             p.parent.mkdir(parents=True, exist_ok=True)
             p.write_text(json.dumps(data, indent=2), encoding="utf-8")
+            # The merge above already carried every legacy entry across, so
+            # drop the old file — otherwise a *cleared* lane would come back
+            # to life on the next read from the stale value left behind.
+            legacy.unlink(missing_ok=True)
         except OSError:
             logger.warning(
-                "claude_runner: qoder lane write failed conv=%s", conversation_id,
+                "claude_runner: lane write failed conv=%s", conversation_id,
                 exc_info=True,
             )
 
@@ -2193,6 +2258,7 @@ TASKS (频道任务板) — 频道里的分工协作走任务认领原语，纪
         run_mode: str = "full",
         scratch_dir: "Path | None" = None,
         conversation_id: str | None = None,
+        periodic: bool = False,
     ) -> list[str]:
         """Build the ``claude`` argv list for one wake.
 
@@ -2214,8 +2280,12 @@ TASKS (频道任务板) — 频道里的分工协作走任务认领原语，纪
         故此模式下 ``last_session`` 恒为 None。
         """
         read_analyze = run_mode == "read_analyze"
+        # #95 — 每个 conversation 一条轨（无 conversation_id 的非平台唤醒落回
+        # 全局轨）。原先这里直接读 ws.last_session_id()，于是同一个 agent 无论
+        # 在哪个频道被唤醒都 --resume 同一条会话。
         last_session = (
-            ws.last_session_id() if (with_resume and not read_analyze) else None
+            self._lane_session(ws, conversation_id)
+            if (with_resume and not read_analyze and not periodic) else None
         )
         if runtime_engine == "codex":
             binary = self.runtime_binaries.get("codex") or "codex"
@@ -2290,14 +2360,10 @@ TASKS (频道任务板) — 频道里的分工协作走任务认领原语，纪
             # 输出走既有 stream_json 消费器（联调时如发现事件形状偏差再适配）。
             # 认证：QODER_PERSONAL_ACCESS_TOKEN 环境变量或 qodercli login。
             binary = self.runtime_binaries.get("qoder") or "qodercli"
-            # #95 — 会话按 conversation_id 分轨（DM/各频道各自续聊），不再共用
-            # 全局 last_session；无 conversation_id 的非平台唤醒沿用全局轨。
+            # #95 — 分轨已上提到 last_session（所有引擎共用），这里直接复用。
             # #107 — read_analyze 与 claude/codex 一样绝不 resume owner 轨，
             # 防外部指挥上下文渗进后续 full 会话。
-            lane_session = (
-                self._qoder_lane_session(ws, conversation_id)
-                if with_resume and not read_analyze else None
-            )
+            lane_session = last_session
             cmd = [binary]
             # r6 同款：模型 + 推理强度由 chairman 在编辑页选（qoder 档位
             # low|medium|high|xhigh|max，与 claude_code 同集）。
@@ -2354,6 +2420,16 @@ TASKS (频道任务板) — 频道里的分工协作走任务认领原语，纪
             cmd += ["--model", model]
         if reasoning_effort:
             cmd += ["--effort", reasoning_effort]
+        # #114 followup — cap the context BEFORE it overflows.
+        #
+        # ``_is_context_exhausted`` is the cure and it is expensive: the wake
+        # has already failed, and the remedy throws the whole session away.
+        # ``--autocompact`` is the prevention — the CLI folds older turns into
+        # a summary at the threshold, so a long-lived lane keeps its thread
+        # instead of being discarded whole. It also bounds what every wake
+        # REPLAYS, which is the larger bill: a lane is re-read in full on each
+        # wake, so an unbounded one charges its entire history every time.
+        cmd += ["--autocompact", self.AUTOCOMPACT_WINDOW]
         if last_session:
             cmd += ["--resume", last_session]
         # 1i Bug F — inject operational contract via --append-system-prompt.

{cloudagent_daemon-0.5.4 → cloudagent_daemon-0.5.6}/tests/unit/test_claude_runner.py

@@ -3293,14 +3293,18 @@ async def test_qoder_sessions_are_isolated_per_conversation(
     a3 = _read_argv_n(tmp_path, 3)
     assert a3[a3.index("--resume") + 1] == "sess_dm"
 
-    lanes = json.loads((tmp_path / ".meta" / "qoder-sessions.json").read_text())
+    lanes = json.loads((tmp_path / ".meta" / "sessions.json").read_text())
     assert lanes == {"conv_debug": "sess_debug", "conv_dm": "sess_dm"}
 
 
 async def test_qoder_stale_resume_clears_only_that_lane(
     stateful_fake_claude, tmp_path
 ):
-    """lane 里的过期会话被 CLI 拒绝 → 清该 lane 并免 resume 重试；别的 lane 不动。"""
+    """lane 里的过期会话被 CLI 拒绝 → 清该 lane 并免 resume 重试；别的 lane 不动。
+
+    seed 故意写在**旧** ``qoder-sessions.json`` 上，于是这条同时验收升级路径：
+    旧 lane 仍可 resume，首次写入把两条都搬进 ``sessions.json`` 并删掉旧文件。
+    """
     mcp_cfg = tmp_path / "mcp.json"; mcp_cfg.write_text("{}")
     ws = FakeWorkspace(FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None))
     (tmp_path / ".meta").mkdir(parents=True, exist_ok=True)
@@ -3319,8 +3323,122 @@ async def test_qoder_stale_resume_clears_only_that_lane(
     a0 = _read_argv_n(tmp_path, 0); a1 = _read_argv_n(tmp_path, 1)
     assert a0[a0.index("--resume") + 1] == "old_A"
     assert "--resume" not in a1
-    lanes = json.loads((tmp_path / ".meta" / "qoder-sessions.json").read_text())
+    lanes = json.loads((tmp_path / ".meta" / "sessions.json").read_text())
     assert lanes == {"conv_A": "new_A", "conv_B": "keep_B"}
+    # 迁移后旧文件必须消失，否则被清掉的 lane 会从残留值里复活
+    assert not (tmp_path / ".meta" / "qoder-sessions.json").exists()
+
+
+# ---------------------------------------------------------------------------
+# #95 — lanes are engine-independent: claude_code / codex get them too
+# ---------------------------------------------------------------------------
+
+
+def _claude_ok(sid: str) -> dict:
+    return {
+        "stdout": [
+            json.dumps({"type": "system", "subtype": "init", "session_id": sid}),
+            json.dumps({"type": "result", "subtype": "success", "result": "ok"}),
+        ],
+        "exit_code": 0,
+    }
+
+
+async def test_claude_code_sessions_are_isolated_per_conversation(
+    stateful_fake_claude, tmp_path
+):
+    """claude_code 同样按 conversation 分轨：频道/DM 各自续聊，互不串味。
+
+    这条是本次改动的本体 —— 分轨此前只有 qoder 走，claude_code 不管在哪个
+    频道被唤醒都 --resume 同一条全局会话。
+    """
+    mcp_cfg = tmp_path / "mcp.json"; mcp_cfg.write_text("{}")
+    ws = FakeWorkspace(FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None))
+    fake = stateful_fake_claude([
+        _claude_ok("sess_debug"),   # 0: #debug 首轮
+        _claude_ok("sess_dm"),      # 1: DM 首轮 —— 不得继承 #debug 的会话
+        _claude_ok("sess_debug"),   # 2: 回 #debug —— resume 自己的轨
+        _claude_ok("sess_dm"),      # 3: 回 DM —— resume DM 的轨
+    ])
+    runner = _build_runner(fake, ws, FakeOutbound())
+    dbg = {**_make_payload(), "conversation_id": "conv_debug"}
+    dm = {**_make_payload(), "conversation_id": "conv_dm"}
+
+    await runner.run_one("agent-A", dbg)
+    assert "--resume" not in _read_argv_n(tmp_path, 0)
+    await runner.run_one("agent-A", dm)
+    assert "--resume" not in _read_argv_n(tmp_path, 1)
+    await runner.run_one("agent-A", dbg)
+    a2 = _read_argv_n(tmp_path, 2)
+    assert a2[a2.index("--resume") + 1] == "sess_debug"
+    await runner.run_one("agent-A", dm)
+    a3 = _read_argv_n(tmp_path, 3)
+    assert a3[a3.index("--resume") + 1] == "sess_dm"
+
+    lanes = json.loads((tmp_path / ".meta" / "sessions.json").read_text())
+    assert lanes == {"conv_debug": "sess_debug", "conv_dm": "sess_dm"}
+
+
+async def test_new_conversation_never_inherits_the_global_session(
+    stateful_fake_claude, tmp_path
+):
+    """升级现场的回归闸：老 agent 带着全局 last_session 进入本版本，新频道
+    首轮必须从零开始 —— 继承它正是分轨要消灭的那条串味路径。"""
+    mcp_cfg = tmp_path / "mcp.json"; mcp_cfg.write_text("{}")
+    ws = FakeWorkspace(
+        FakeAgentWorkspace(mcp_path=mcp_cfg, last_session="global_old")
+    )
+    fake = stateful_fake_claude([_claude_ok("sess_fresh")])
+    runner = _build_runner(fake, ws, FakeOutbound())
+
+    await runner.run_one(
+        "agent-A", {**_make_payload(), "conversation_id": "conv_new"}
+    )
+    assert "--resume" not in _read_argv_n(tmp_path, 0)
+
+
+async def test_non_platform_wake_still_uses_the_global_lane(
+    stateful_fake_claude, tmp_path
+):
+    """无 conversation_id 的唤醒（定时器 / spawn 链）沿用全局轨，不受分轨影响。"""
+    mcp_cfg = tmp_path / "mcp.json"; mcp_cfg.write_text("{}")
+    ws = FakeWorkspace(
+        FakeAgentWorkspace(mcp_path=mcp_cfg, last_session="global_old")
+    )
+    fake = stateful_fake_claude([_claude_ok("global_old")])
+    runner = _build_runner(fake, ws, FakeOutbound())
+
+    await runner.run_one("agent-A", _make_payload())
+    a0 = _read_argv_n(tmp_path, 0)
+    assert a0[a0.index("--resume") + 1] == "global_old"
+
+
+async def test_codex_sessions_are_isolated_per_conversation(
+    stateful_fake_claude, tmp_path
+):
+    """codex 分轨；resume 形态是 ``exec resume … <thread_id> -``。"""
+    mcp_cfg = tmp_path / "mcp.json"; mcp_cfg.write_text("{}")
+    ws = FakeWorkspace(FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None))
+    fake = stateful_fake_claude([
+        {"stdout": _codex_ok("thread_a"), "exit_code": 0},   # 0: conv_A 首轮
+        {"stdout": _codex_ok("thread_b"), "exit_code": 0},   # 1: conv_B 首轮
+        {"stdout": _codex_ok("thread_a"), "exit_code": 0},   # 2: 回 conv_A
+    ])
+    runner = _codex_runner(fake, ws, FakeOutbound())
+    a = {**_make_payload(), "runtime_engine": "codex", "conversation_id": "conv_A"}
+    b = {**_make_payload(), "runtime_engine": "codex", "conversation_id": "conv_B"}
+
+    await runner.run_one("agent-A", a)
+    assert "resume" not in _read_argv_n(tmp_path, 0)
+    await runner.run_one("agent-A", b)
+    assert "resume" not in _read_argv_n(tmp_path, 1)
+    await runner.run_one("agent-A", a)
+    a2 = _read_argv_n(tmp_path, 2)
+    assert "resume" in a2
+    assert a2[-2] == "thread_a"          # …<thread_id> "-"
+
+    lanes = json.loads((tmp_path / ".meta" / "sessions.json").read_text())
+    assert lanes == {"conv_A": "thread_a", "conv_B": "thread_b"}
 
 
 # ---------------------------------------------------------------------------
@@ -3728,3 +3846,113 @@ async def test_no_conversation_id_means_no_notice_but_never_a_crash_of_the_wake(
         if inv.get("tool_name") == "platform_send"
     ]
     assert not sends
+
+
+# ---------------------------------------------------------------------------
+# #114 followup — bound what a wake replays (A: --autocompact), and keep
+# timers off the room's thread (B: a periodic wake is a throwaway session).
+#
+# Both target the same bill from opposite ends. A lane is re-read IN FULL on
+# every wake, so an unbounded one charges its whole history each time; a
+# 15-minute timer made that 96 times a day while contributing a few thousand
+# tokens of new content, and left the accumulated patrol history behind for
+# whoever spoke in the room next.
+# ---------------------------------------------------------------------------
+
+
+def _reminder_payload(reminder_id: str = "rem_1", occurrence: int = 0, **extra) -> dict:
+    """A timer wake as the backend stamps it: ``remind:<reminder_id>:<n>``."""
+    return {
+        **_make_payload(message_id=f"remind:{reminder_id}:{occurrence}"),
+        **extra,
+    }
+
+
+async def test_every_claude_wake_caps_its_context_window(
+    stateful_fake_claude, tmp_path
+):
+    """A — the cap is on every wake, not only the ones that already grew."""
+    mcp_cfg = tmp_path / "mcp.json"; mcp_cfg.write_text("{}")
+    ws = FakeWorkspace(FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None))
+    runner = _build_runner(
+        stateful_fake_claude([_claude_ok("s1")]), ws, FakeOutbound()
+    )
+
+    await runner.run_one("agent-A", _make_payload())
+
+    argv = _read_argv_n(tmp_path, 0)
+    assert argv[argv.index("--autocompact") + 1] == ClaudeRunner.AUTOCOMPACT_WINDOW
+
+
+async def test_periodic_wake_does_not_resume_the_rooms_lane(
+    stateful_fake_claude, tmp_path
+):
+    """B — a patrol reads the room's CURRENT state; it must not replay its past."""
+    mcp_cfg = tmp_path / "mcp.json"; mcp_cfg.write_text("{}")
+    (tmp_path / ".meta").mkdir(exist_ok=True)
+    (tmp_path / ".meta" / "sessions.json").write_text(
+        json.dumps({"conv_room": "human_thread"})
+    )
+    ws = FakeWorkspace(FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None))
+    runner = _build_runner(
+        stateful_fake_claude([_claude_ok("patrol_thread")]), ws, FakeOutbound()
+    )
+
+    await runner.run_one("agent-A", _reminder_payload(conversation_id="conv_room"))
+
+    assert "--resume" not in _read_argv_n(tmp_path, 0)
+
+
+async def test_periodic_wake_leaves_the_rooms_lane_untouched(
+    stateful_fake_claude, tmp_path
+):
+    """B — and it writes nothing back: neither the room's lane nor the global.
+
+    This is the regression itself. A timer that records its throwaway session
+    hands its patrol history to the next human who speaks in that room.
+    """
+    mcp_cfg = tmp_path / "mcp.json"; mcp_cfg.write_text("{}")
+    (tmp_path / ".meta").mkdir(exist_ok=True)
+    (tmp_path / ".meta" / "sessions.json").write_text(
+        json.dumps({"conv_room": "human_thread"})
+    )
+    agent_ws = FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None)
+    runner = _build_runner(
+        stateful_fake_claude([_claude_ok("patrol_thread")]),
+        FakeWorkspace(agent_ws),
+        FakeOutbound(),
+    )
+
+    await runner.run_one("agent-A", _reminder_payload(conversation_id="conv_room"))
+
+    lanes = json.loads((tmp_path / ".meta" / "sessions.json").read_text())
+    assert lanes == {"conv_room": "human_thread"}   # patrol_thread did NOT land
+    assert agent_ws.set_calls == []                 # nor did the global lane move
+
+
+async def test_ordinary_wake_in_that_room_still_resumes_and_records(
+    stateful_fake_claude, tmp_path
+):
+    """B control — only TIMER wakes are throwaway. A human message in the same
+    room still resumes that room's lane and still records its session."""
+    mcp_cfg = tmp_path / "mcp.json"; mcp_cfg.write_text("{}")
+    (tmp_path / ".meta").mkdir(exist_ok=True)
+    (tmp_path / ".meta" / "sessions.json").write_text(
+        json.dumps({"conv_room": "human_thread"})
+    )
+    agent_ws = FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None)
+    runner = _build_runner(
+        stateful_fake_claude([_claude_ok("human_thread_2")]),
+        FakeWorkspace(agent_ws),
+        FakeOutbound(),
+    )
+
+    await runner.run_one(
+        "agent-A", {**_make_payload(), "conversation_id": "conv_room"}
+    )
+
+    argv = _read_argv_n(tmp_path, 0)
+    assert argv[argv.index("--resume") + 1] == "human_thread"
+    lanes = json.loads((tmp_path / ".meta" / "sessions.json").read_text())
+    assert lanes == {"conv_room": "human_thread_2"}
+    assert agent_ws.set_calls == ["human_thread_2"]

{cloudagent_daemon-0.5.4 → cloudagent_daemon-0.5.6}/uv.lock

@@ -122,7 +122,7 @@ wheels = [
 
 [[package]]
 name = "cloudagent-daemon"
-version = "0.5.4"
+version = "0.5.6"
 source = { editable = "." }
 dependencies = [
     { name = "h2" },

cloudagent_daemon-0.5.4/src/cloudagent_daemon/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.5.4"