fix(sandbox): keep output and exit code when an e2b command fails (#6441)

* fix(sandbox): keep output and exit code when an e2b command fails

* docs(changelog): link #6441

* fix(sandbox): keep list_dir output when an e2b listing exits nonzero

---------

Co-authored-by: Willem Jiang <willem.jiang@gmail.com>

CHANGELOG.md

@@ -804,6 +804,18 @@ This release closes that milestone with **439 merged pull requests**.
 
 ### Fixed
 
+- **sandbox:** A failed command in the e2b sandbox now keeps its output and exit
+  code. The e2b SDK raises `CommandExitException` on a nonzero exit instead of
+  returning a result, so `E2BSandbox.execute_command` returned
+  `Error: Command exited with code N and error: ...`: stdout was dropped, the
+  `Exit Code: N` marker was never added, and `_bash_evidence_status` fell back to
+  `deerflow_tool_meta` (`success`), so a failed `pytest` could satisfy a
+  `tests_passed` acceptance criterion. The exception carries the command's
+  stdout, stderr and exit code, and is now formatted like a returned result. A
+  failed command whose stderr mentions "sandbox not found" no longer marks the
+  sandbox as reaped. `list_dir` on e2b handles the same exception, so a missing
+  directory raises `FileNotFoundError` and a listing truncated at 500 entries
+  (SIGPIPE 141) is returned instead of failing with `OSError`. ([#6441])
 - **uploads:** The Gateway's startup sweep of orphaned `.upload-*.part` staging
   files now skips files younger than 24 hours. The sweep removed every staging
   file it found, which was right for one Gateway but not for several replicas
@@ -9164,6 +9176,7 @@ with **180 merged pull requests** since the first 2.0 milestone tag.
 [#6407]: https://github.com/bytedance/deer-flow/pull/6407
 [#6412]: https://github.com/bytedance/deer-flow/pull/6412
 [#6426]: https://github.com/bytedance/deer-flow/pull/6426
+[#6441]: https://github.com/bytedance/deer-flow/pull/6441
 [#6445]: https://github.com/bytedance/deer-flow/pull/6445
 [#6447]: https://github.com/bytedance/deer-flow/pull/6447
 [#6448]: https://github.com/bytedance/deer-flow/pull/6448

CHANGELOG_zh.md

@@ -694,6 +694,14 @@
 
 ### 修复
 
+- **沙箱：** e2b 沙箱中失败的命令现在会保留输出和退出码。e2b SDK 在退出码非零时抛出
+  `CommandExitException`，而不是返回结果，因此 `E2BSandbox.execute_command` 返回的是
+  `Error: Command exited with code N and error: ...`：stdout 被丢弃，`Exit Code: N`
+  标记从未追加，`_bash_evidence_status` 退回到报告 `success` 的 `deerflow_tool_meta`，
+  失败的 `pytest` 可能满足 `tests_passed` 验收条件。该异常本身带有命令的 stdout、stderr
+  与退出码，现在会按返回结果的方式格式化。stderr 中恰好包含 "sandbox not found" 的失败命令
+  也不再把沙箱标记为已回收。e2b 的 `list_dir` 也会处理同一异常：目录不存在时抛出
+  `FileNotFoundError`，超过 500 条被截断（SIGPIPE 141）的列表会正常返回，不再以 `OSError` 失败。([#6441])
 - **前端：** 附件上传失败后重试发送时，现在会保留原先附带的上下文。输入框此前在发送开始时（附件上传之前）就清除了
   引用、对话引用、已暂存的项目文件和已保存的草稿，因此上传失败后文字和文件虽仍在，重试发送却缺少这些上下文。
   现在这些一次性状态只在发送真正派发（上传完成）后才清除；若上传完成时用户已切换对话或离开页面，只清除该次发送
@@ -7506,4 +7514,5 @@ DeerFlow 2.0 是围绕"超级智能体"框架的彻底重写，核心包含子
 [#6407]: https://github.com/bytedance/deer-flow/pull/6407
 [#6412]: https://github.com/bytedance/deer-flow/pull/6412
 [#6426]: https://github.com/bytedance/deer-flow/pull/6426
+[#6441]: https://github.com/bytedance/deer-flow/pull/6441
 [#6448]: https://github.com/bytedance/deer-flow/pull/6448

backend/packages/harness/deerflow/community/e2b_sandbox/e2b_sandbox.py

@@ -7,7 +7,7 @@ import shlex
 import threading
 from typing import TYPE_CHECKING
 
-from e2b import FileNotFoundException
+from e2b import CommandExitException, FileNotFoundException
 from e2b_code_interpreter import Sandbox as E2BClientSandbox
 
 from deerflow.config.paths import VIRTUAL_PATH_PREFIX
@@ -165,7 +165,13 @@ class E2BSandbox(Sandbox):
                     kwargs["envs"] = env
                 if timeout is not None:
                     kwargs["timeout"] = timeout
-                result = client.commands.run(command, **kwargs)
+                try:
+                    result = client.commands.run(command, **kwargs)
+                except CommandExitException as exc:
+                    # The SDK raises on a nonzero exit instead of returning a
+                    # result. The exception is itself a ``CommandResult``, so
+                    # format it like one to keep stdout and the exit marker.
+                    result = exc
                 stdout = getattr(result, "stdout", "") or ""
                 stderr = getattr(result, "stderr", "") or ""
                 exit_code = getattr(result, "exit_code", 0)
@@ -347,6 +353,12 @@ class E2BSandbox(Sandbox):
                 raise RuntimeError("sandbox client has been closed")
             try:
                 result = client.commands.run(remote_list_dir_command(resolved, max_depth))
+            except CommandExitException as exc:
+                # The listing script exits nonzero on reachable outcomes (missing
+                # root: 1; head truncating a large listing: SIGPIPE 141). The SDK
+                # raises for those, but the exception carries stdout, whose
+                # status marker the parser trusts over the exit code.
+                result = exc
             except Exception as e:
                 logger.error("Failed to list_dir %s in e2b sandbox: %s", resolved, e)
                 raise OSError(f"Failed to list_dir {resolved} in e2b sandbox: {e}") from e

backend/tests/test_e2b_sandbox_provider.py

@@ -5,11 +5,13 @@ from __future__ import annotations
 import asyncio
 import hashlib
 import importlib
+import importlib.util
 import json
 import os
 import shutil
 import signal
 import subprocess
+import sys
 import threading
 import time
 from collections import OrderedDict
@@ -22,7 +24,8 @@ from typing import Any
 from unittest.mock import MagicMock
 
 import pytest
-from e2b import FileNotFoundException, TimeoutException
+from e2b import CommandExitException, FileNotFoundException, TimeoutException
+from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
 from pydantic import ValidationError
 
 from deerflow.community.e2b_sandbox.capacity import (
@@ -1751,15 +1754,79 @@ def test_execute_command_returns_stdout_on_success():
     assert sb.is_dead is False
 
 
+def _raise_exit(stdout: str, stderr: str, exit_code: int):
+    """The real SDK raises ``CommandExitException`` on a nonzero exit; it
+    never returns a result with ``exit_code != 0``."""
+
+    def run(_cmd: str) -> Any:
+        raise CommandExitException(stderr=stderr, stdout=stdout, exit_code=exit_code, error=f"exit status {exit_code}")
+
+    return run
+
+
 def test_execute_command_appends_exit_marker_when_failure_has_output():
     """LocalSandbox parity: a nonzero exit must survive in the output text
     even when the command produced output, so evidence consumers (acceptance
     checklist) recover the actual shell status."""
-    client = FakeClient(commands=FakeCommandsAPI([SimpleNamespace(stdout="5 passed, 1 error\n", stderr="", exit_code=1)]))
+    client = FakeClient(commands=FakeCommandsAPI([_raise_exit("5 passed, 1 error\n", "", 1)]))
     sb = _make_sandbox(client)
     assert sb.execute_command("make test") == "5 passed, 1 error\n\nExit Code: 1"
 
 
+def test_execute_command_keeps_stdout_and_stderr_on_nonzero_exit():
+    client = FakeClient(commands=FakeCommandsAPI([_raise_exit("12 passed\n", "coverage below 80%\n", 2)]))
+    sb = _make_sandbox(client)
+    assert sb.execute_command("pytest -q") == "12 passed\n\ncoverage below 80%\n\nExit Code: 2"
+    assert sb.is_dead is False
+
+
+def test_execute_command_reports_exit_code_when_failure_has_no_output():
+    client = FakeClient(commands=FakeCommandsAPI([_raise_exit("", "", 1)]))
+    sb = _make_sandbox(client)
+    assert sb.execute_command("false") == "Command exited with code 1"
+
+
+def test_execute_command_nonzero_exit_mentioning_not_found_does_not_mark_dead():
+    """A command's own stderr can contain "sandbox not found"; that is a failed
+    command, not a reaped VM."""
+    client = FakeClient(commands=FakeCommandsAPI([_raise_exit("", "error: sandbox not found\n", 1)]))
+    sb = _make_sandbox(client)
+    assert sb.execute_command("mytool status") == "error: sandbox not found\n\nExit Code: 1"
+    assert sb.is_dead is False
+
+
+def test_execute_command_nonzero_exit_is_harvested_as_error():
+    """End-to-end with the real executor harvest. ``tests/conftest.py`` mocks
+    ``deerflow.subagents.executor``, so it is loaded under a unique name."""
+    path = Path(__file__).parents[1] / "packages/harness/deerflow/subagents/executor.py"
+    spec = importlib.util.spec_from_file_location("_e2b_exit_marker_executor", path)
+    assert spec is not None and spec.loader is not None
+    module = importlib.util.module_from_spec(spec)
+    try:
+        sys.modules[spec.name] = module
+        spec.loader.exec_module(module)
+        from deerflow.subagents import acceptance_checks
+
+        client = FakeClient(commands=FakeCommandsAPI([_raise_exit("12 passed\n", "coverage below 80%\n", 1)]))
+        out = _make_sandbox(client).execute_command("pytest -q")
+
+        cmd = "pytest -q"
+        ai = AIMessage(content="", tool_calls=[{"name": "bash", "args": {"command": cmd}, "id": "tc-1", "type": "tool_call"}])
+        tm = ToolMessage(content=out, tool_call_id="tc-1", name="bash")
+        executions = module._harvest_bash_executions({"messages": [HumanMessage(content="task"), ai, tm]})
+        assert executions, "harvest returned no bash executions"
+        for entry in executions:
+            entry["shell_persistent"] = False
+        assert executions[-1]["status"] == "error"
+        assert executions[-1]["status_marker"] == "Exit Code: 1"
+        assert acceptance_checks._check_tests_passed_leaf(cmd, executions)["holds"] is False
+    finally:
+        sys.modules.pop(spec.name, None)
+        shutdown = getattr(module, "_shutdown_isolated_subagent_loop", None)
+        if shutdown is not None:
+            shutdown()
+
+
 def test_execute_command_does_not_mark_dead_on_unrelated_error():
 
     def boom(_cmd: str, **kwargs) -> Any:
@@ -5586,8 +5653,9 @@ def test_list_dir_raises_when_client_closed():
 
 @pytest.mark.parametrize("marker, error", [("missing", FileNotFoundError), ("1", OSError)])
 def test_list_dir_classifies_empty_failure(marker, error):
-    listing = SimpleNamespace(stdout=f"\n__DF_FIND_STATUS__:{marker}\n", stderr="", exit_code=1)
-    client = FakeClient(commands=FakeCommandsAPI([listing]))
+    # The listing script exits 1 for both; the SDK raises CommandExitException,
+    # whose stdout still carries the status marker the parser classifies by.
+    client = FakeClient(commands=FakeCommandsAPI([_raise_exit(f"\n__DF_FIND_STATUS__:{marker}\n", "", 1)]))
     sb = _make_sandbox(client)
 
     with pytest.raises(error) as exc:
@@ -5595,9 +5663,19 @@ def test_list_dir_classifies_empty_failure(marker, error):
     assert type(exc.value) is error
 
 
+def test_list_dir_returns_entries_when_head_truncation_exits_141():
+    """``head`` closing the pipe on a large listing kills ``find`` with SIGPIPE
+    (141), a successful truncation. The SDK raises for the nonzero exit; the
+    listing must still be returned, not surface as ``OSError``."""
+    stdout = "/home/user\n/home/user/a\n/home/user/b\n\n__DF_FIND_STATUS__:141\n"
+    client = FakeClient(commands=FakeCommandsAPI([_raise_exit(stdout, "", 141)]))
+    sb = _make_sandbox(client)
+
+    assert sb.list_dir("/home/user") == ["/home/user", "/home/user/a", "/home/user/b"]
+
+
 def test_list_dir_raises_oserror_when_find_exit_is_not_missing_path():
-    listing = SimpleNamespace(stdout="", stderr="", exit_code=127)
-    client = FakeClient(commands=FakeCommandsAPI([listing]))
+    client = FakeClient(commands=FakeCommandsAPI([_raise_exit("", "", 127)]))
     sb = _make_sandbox(client)
 
     with pytest.raises(OSError, match="exited with code 127"):