cloudagent-daemon 0.5.3__tar.gz → 0.5.4__tar.gz

{cloudagent_daemon-0.5.3 → cloudagent_daemon-0.5.4}/CHANGELOG.md

@@ -4,6 +4,35 @@ 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`.
 
+## [0.5.4] — 2026-08-24
+
+### Changed
+- A model-side error no longer counts toward the crash quota, and can no longer
+  quarantine an agent. Three consecutive `crashed` reports quarantine an agent
+  server-side, after which every wake is dropped and nobody is told; a model
+  refusing one turn is not the same as the agent being broken. This is how a
+  whole delivery line went dark twice — six hours on 2026-08-21, four on 08-23
+  (the latter from a content-safety flag raised while reviewing security
+  material). The wake now reports `idle`, which keeps the counter still.
+
+### Added
+- A failed wake posts a notice into the originating channel carrying the
+  engine's own error text. Until now this path only wrote a log line and a
+  metric — "the silence must be VISIBLE" (U0.4) was true only for whoever reads
+  logs, and nobody does; the user found out by noticing an agent had stopped
+  answering, hours later. The notice is explicitly labelled as daemon-emitted so
+  it can never be mistaken for the agent's own answer, which is what the
+  no-fallback rule was protecting.
+
+### Notes
+- Status code and notification are now separate concerns: the code drives
+  counting, the notice does the telling. Model errors report `idle` AND still
+  trigger the a2a `wake_aborted` system reply — dropping the latter would leave
+  a dispatching chief waiting forever.
+- Process-level failures (binary missing, auth file unreadable) still count.
+  Retrying those forever fixes nothing, so the protection is kept — but they now
+  post a channel notice too, so they are no longer silent either.
+
 ## [0.5.3] — 2026-08-22
 
 ### Fixed

{cloudagent_daemon-0.5.3 → cloudagent_daemon-0.5.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: cloudagent-daemon
-Version: 0.5.3
+Version: 0.5.4
 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-0.5.3 → cloudagent_daemon-0.5.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "cloudagent-daemon"
-version = "0.5.3"
+version = "0.5.4"
 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-0.5.4/src/cloudagent_daemon/__init__.py

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

{cloudagent_daemon-0.5.3 → cloudagent_daemon-0.5.4}/src/cloudagent_daemon/claude_runner.py

@@ -984,21 +984,79 @@ The daemon discards stdout — only the mcp tool call reaches the channel.
         )
         ws.write_last_spawn(rc, spawn_record_session, stderr_tail)
 
+        # 主人拍板（2026-08-24）：**大模型端的报错一律直接回报给人，不再计入崩溃
+        # 配额、不再据此隔离 agent。** 模型拒绝一次不等于这个 agent 坏了；把它锁死
+        # 的代价是整条交付线静默停摆——8-21 六小时、8-23 四小时，两次都是这么来的
+        # （连续 3 次模型侧失败 → quarantined → 此后所有唤醒静默丢弃，没人被告知）。
+        #
+        # 报 idle 而非 crashed：idle 不进 record_agent_crash 的计数，且会顺带解除
+        # 历史遗留的隔离（record_agent_recovery）。失败本身**不靠这个状态码传达**，
+        # 由下面的频道通知负责——状态码只有运维看得到，人看的是频道。
+        model_side_error = attempt.had_api_error
         if timeout_hit:
             final_status = "timeout"
-        elif attempt.had_api_error:
-            # M8 — Claude Code swallows upstream Anthropic/Bedrock API
-            # errors: it prints a synthetic ``isApiErrorMessage:true``
-            # event and exits rc=0 *without* a terminating ``type=
-            # result``. Without this branch the wake would be reported
-            # as ``idle`` (cleanly completed) — leaving the dispatching
-            # chief silently waiting for a reply that will never come.
-            final_status = "crashed"
+        elif model_side_error:
+            final_status = "idle"
         elif rc == 0:
             final_status = "idle"
         else:
             final_status = "crashed"
 
+        # 主人拍板（2026-08-24）：**模型侧报错必须直接回到人眼前。**
+        #
+        # 此前这里什么都不发：U0.4 定的口径是"频道唤醒不做 fallback，沉默必须可见
+        # （log + metric）"。出发点对——别拿 daemon 的兜底伪造 agent 的回答——但
+        # "可见"只可见于日志和指标，**没有人在看**。实际后果：agent 崩了，频道里
+        # 一个字都没有，发指令的人只能靠"叫它没反应"发现，中间隔了 4 个小时。
+        #
+        # 所以这里补的**不是 agent 的回答，是一条明确署名的系统通知**：诚实口径不破
+        # （不冒充 agent），可见性拿回来。已有的 platform_send 兜底只在 rc==0 且有
+        # 输出时触发（"说了话但没调工具"），失败路径正是它漏掉的那半边。
+        failure_note: str | None = None
+        if model_side_error:
+            failure_note = (attempt.api_error_text or "").strip() or "模型侧返回错误（无详细信息）"
+        elif timeout_hit:
+            failure_note = f"本轮超时（{int(self._single_wake_timeout)} 秒内未结束）"
+        elif rc != 0:
+            failure_note = (attempt.stderr_tail or "").strip()[-400:] or f"进程非零退出（rc={rc}）"
+        notice_conv = payload.get("conversation_id")
+        notice_session = payload.get("session_id")
+        if (
+            failure_note
+            and isinstance(notice_conv, str) and notice_conv
+            and isinstance(notice_session, str) and notice_session
+        ):
+            kind_label = "模型侧报错" if model_side_error else ("超时" if timeout_hit else "进程异常")
+            body = (
+                f"⚠️ 我这一轮没能完成：**{kind_label}**\n\n"
+                f"```\n{failure_note[:600]}\n```\n"
+                + (
+                    "这是**模型端**返回的错误，不是这个 agent 坏了——它**没有被隔离**，"
+                    "下次被唤醒会照常重试。"
+                    if model_side_error
+                    else "这是运行时/进程层面的失败，重试未必能解决，可能需要看一下机器。"
+                )
+                + "\n（本条由 daemon 代发的系统通知，不是 agent 的回答。）"
+            )
+            try:
+                await self.outbound.send_tool_call(
+                    agent_id=agent_id,
+                    session_id=notice_session,
+                    tool_name="platform_send",
+                    args={
+                        "content": body,
+                        "conversation_id": notice_conv,
+                        "acknowledge_new_messages": True,
+                        **({"delivery_id": _effective_delivery_id} if _effective_delivery_id else {}),
+                    },
+                    timeout=15.0,
+                )
+            except Exception:  # noqa: BLE001 — 通知失败绝不能再把这一轮拖垮
+                logger.warning(
+                    "claude_runner: failed to post failure notice agent=%s conv=%s",
+                    agent_id, notice_conv, exc_info=True,
+                )
+
         # Wave C #3 — surface terminal status counts for /metrics +
         # doctor. Label-keyed so a single counter renders three
         # series (idle / crashed / timeout). Failure to publish must
@@ -1034,7 +1092,10 @@ The daemon discards stdout — only the mcp tool call reaches the channel.
             payload.get("parent_msg_id") or payload.get("message_id")
         )
         if (
-            final_status in ("crashed", "timeout")
+            # model_side_error 现在报 idle（不进崩溃配额），但**这条链仍要通知**：
+            # a2a 的 chief 是在等一条回复，不给它就是永远等下去。状态码用于计数，
+            # 这里用于告知——两件事，不能因为改了前者而丢掉后者。
+            (final_status in ("crashed", "timeout") or model_side_error)
             # This block only fires for task/peer_request/question kinds
             # (WAKE_REPLY_EXPECTED_KINDS, checked below), for which
             # fallback_tool is always send_message — so called_fallback_tool

{cloudagent_daemon-0.5.3 → cloudagent_daemon-0.5.4}/tests/unit/test_claude_runner.py

@@ -2461,11 +2461,22 @@ def _make_dispatch_payload(
     }
 
 
-async def test_isApiErrorMessage_promotes_wake_to_crashed(
+async def test_isApiErrorMessage_does_not_burn_the_crash_quota(
     fake_claude, tmp_path
 ):
-    """M8 — when claude emits ``isApiErrorMessage:true`` and exits rc=0,
-    the runner must report ``agent_status="crashed"`` (not ``idle``).
+    """2026-08-24 主人拍板改写：模型侧报错**不再**记 crashed。
+
+    这条测试原本断言的是相反的行为（M8：``isApiErrorMessage`` + rc=0 → 报
+    ``crashed``）。改写而不是删除，因为它当初的关切依然成立、只是解法换了：
+
+      当初的关切 —— "别报 idle，否则派活的 chief 会以为干完了、永远等下去"。
+      现在的解法 —— 状态码只负责**计数**，告知由 ``wake_aborted`` 的系统回复
+      和频道通知负责（见 ``test_wake_aborted_emits_system_reply_when_api_error``
+      与 ``test_model_side_error_is_reported_into_the_channel``）。两件事拆开了。
+
+    为什么必须拆：报 crashed 会累计进隔离配额，连续 3 次就把 agent 锁死，此后
+    **所有**唤醒被静默丢弃、无人被告知。模型拒绝一次不等于这个 agent 坏了。
+    2026-08-21 与 08-23 两次整条交付线停摆（6 小时 / 4 小时）都源于此。
     """
     mcp_cfg = tmp_path / "mcp.json"
     mcp_cfg.write_text("{}")
@@ -2493,8 +2504,8 @@ async def test_isApiErrorMessage_promotes_wake_to_crashed(
     statuses = [c for c in outbound.calls if c[0] == "agent_status"]
     final = [s for s in statuses if s[2] != "running"]
     assert len(final) == 1
-    assert final[0][2] == "crashed", (
-        f"isApiErrorMessage with rc=0 must report crashed, got {final[0][2]}"
+    assert final[0][2] == "idle", (
+        f"模型侧报错记成了 {final[0][2]} —— 会累计进隔离配额并最终锁死 agent"
     )
 
 
@@ -3590,3 +3601,130 @@ def test_context_markers_avoid_rewritable_punctuation():
             f"marker {marker!r} is not lowercase; _is_context_exhausted "
             f"lowercases the text, so this marker can never match"
         )
+
+
+# ---------------------------------------------------------------------------
+# 2026-08-24 主人拍板：模型侧报错直接回报给人，不计入崩溃配额、不据此隔离
+# ---------------------------------------------------------------------------
+
+
+def _platform_payload(**extra):
+    """一次频道唤醒的 payload —— 带 conversation_id 才谈得上把失败发回频道。"""
+    return {
+        **_make_payload(),
+        "kind": "platform_mention",
+        "conversation_id": "conv_test_1",
+        "session_id": "sess_test_1",
+        "runtime_engine": "codex",
+        **extra,
+    }
+
+
+async def test_model_side_error_is_not_counted_as_a_crash(fake_claude, tmp_path):
+    """模型报错必须报 idle —— 报 crashed 就会累计进隔离配额。
+
+    连续 3 次 crashed 会让服务端把 agent 隔离，此后**所有**唤醒被静默丢弃、无人
+    被告知。Dex 2026-08-23 就是这样消失了 4 小时：模型侧的内容安全拦截被当成
+    "这个 agent 坏了"。模型拒绝一次不等于 agent 坏了。
+    """
+    mcp_cfg = tmp_path / "mcp.json"
+    mcp_cfg.write_text("{}")
+    ws = FakeWorkspace(FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None))
+    outbound = FakeOutbound()
+    fake = fake_claude(
+        stdout_lines=[_codex_turn_failed(
+            "This content was flagged for possible cybersecurity risk"
+        )],
+        exit_code=1,
+    )
+    runner = _codex_runner(fake, ws, outbound)
+
+    await runner.run_one("agent-A", _platform_payload())
+
+    final = [c for c in outbound.calls if c[0] == "agent_status"][-1]
+    assert final[2] == "idle", (
+        f"模型侧报错被记成了 {final[2]} —— 会累计进隔离配额"
+    )
+
+
+async def test_model_side_error_is_reported_into_the_channel(fake_claude, tmp_path):
+    """失败必须出现在人看得见的地方，而且要带原始错误文本。
+
+    此前这条路径只写日志和指标（U0.4 的"沉默必须可见"），而日志没有人看——
+    使用者只能靠"叫它没反应"发现，中间隔了几个小时。
+    """
+    mcp_cfg = tmp_path / "mcp.json"
+    mcp_cfg.write_text("{}")
+    ws = FakeWorkspace(FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None))
+    outbound = FakeOutbound()
+    flagged = "This content was flagged for possible cybersecurity risk"
+    fake = fake_claude(stdout_lines=[_codex_turn_failed(flagged)], exit_code=1)
+    runner = _codex_runner(fake, ws, outbound)
+
+    await runner.run_one("agent-A", _platform_payload())
+
+    sends = [
+        inv for inv in outbound.tool_call_invocations
+        if inv.get("tool_name") == "platform_send"
+    ]
+    assert sends, "失败没有发回频道 —— 使用者无从知道它崩了"
+    content = sends[-1]["args"]["content"]
+    assert flagged in content, f"通知里没有原始错误文本：{content}"
+    assert sends[-1]["args"]["conversation_id"] == "conv_test_1"
+    # 诚实口径：这是 daemon 的系统通知，不能冒充成 agent 的回答。
+    assert "系统通知" in content and "不是 agent 的回答" in content
+    # 并且要说清楚它没被隔离 —— 否则人会以为又要去捞它。
+    assert "没有被隔离" in content
+
+
+async def test_process_level_failure_still_counts_as_a_crash(fake_claude, tmp_path):
+    """回归护栏：放宽的只是**模型侧**报错，进程级故障的保护不能一并拆掉。
+
+    二进制起不来、auth 文件权限坏这类，重试多少次都一样，仍应计入崩溃配额。
+    """
+    mcp_cfg = tmp_path / "mcp.json"
+    mcp_cfg.write_text("{}")
+    ws = FakeWorkspace(FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None))
+    outbound = FakeOutbound()
+    # 没有任何模型事件，只有非零退出 + stderr —— 典型的进程级失败。
+    fake = fake_claude(stdout_lines=[], stderr_lines=["exec: not found"], exit_code=127)
+    runner = _codex_runner(fake, ws, outbound)
+
+    await runner.run_one("agent-A", _platform_payload())
+
+    final = [c for c in outbound.calls if c[0] == "agent_status"][-1]
+    assert final[2] == "crashed", (
+        f"进程级失败被记成了 {final[2]} —— 保护被误拆"
+    )
+    # 但它同样要让人看见。
+    sends = [
+        inv for inv in outbound.tool_call_invocations
+        if inv.get("tool_name") == "platform_send"
+    ]
+    assert sends, "进程级失败也必须发回频道"
+    assert "进程异常" in sends[-1]["args"]["content"]
+
+
+async def test_no_conversation_id_means_no_notice_but_never_a_crash_of_the_wake(
+    fake_claude, tmp_path
+):
+    """拿不到频道就不发通知 —— 但绝不能因此把这一轮弄挂。"""
+    mcp_cfg = tmp_path / "mcp.json"
+    mcp_cfg.write_text("{}")
+    ws = FakeWorkspace(FakeAgentWorkspace(mcp_path=mcp_cfg, last_session=None))
+    outbound = FakeOutbound()
+    fake = fake_claude(
+        stdout_lines=[_codex_turn_failed("boom")], exit_code=1,
+    )
+    runner = _codex_runner(fake, ws, outbound)
+
+    payload = _platform_payload()
+    payload.pop("conversation_id")
+    result = await runner.run_one("agent-A", payload)  # 不抛异常
+
+    assert result is not None
+    sends = [
+        inv for inv in outbound.tool_call_invocations
+        if inv.get("tool_name") == "platform_send"
+    ]
+    assert not sends

{cloudagent_daemon-0.5.3 → cloudagent_daemon-0.5.4}/uv.lock

@@ -122,7 +122,7 @@ wheels = [
 
 [[package]]
 name = "cloudagent-daemon"
-version = "0.5.3"
+version = "0.5.4"
 source = { editable = "." }
 dependencies = [
     { name = "h2" },

cloudagent_daemon-0.5.3/src/cloudagent_daemon/__init__.py

@@ -1 +0,0 @@
-__version__ = "0.5.3"