fix(persistence): discard the staged temp when the write itself fails (#7504)

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

backend/packages/harness/deerflow/persistence/agents/file.py

@@ -246,12 +246,10 @@ class FileAgentStore(AgentStore):
         try:
             if config is not None:
                 config_text = yaml.dump(config, default_flow_style=False, allow_unicode=True, sort_keys=False)
-                config_tmp = _stage_temp(agent_dir / "config.yaml", config_text)
-                staged.append(config_tmp)
+                config_tmp = _stage_temp(agent_dir / "config.yaml", config_text, staged)
                 pending.append((config_tmp, agent_dir / "config.yaml"))
             if soul is not None:
-                soul_tmp = _stage_temp(agent_dir / SOUL_FILENAME, soul)
-                staged.append(soul_tmp)
+                soul_tmp = _stage_temp(agent_dir / SOUL_FILENAME, soul, staged)
                 pending.append((soul_tmp, agent_dir / SOUL_FILENAME))
             for tmp, target in pending:
                 tmp.replace(target)
@@ -263,10 +261,14 @@ class FileAgentStore(AgentStore):
                 cleanup_failed_atomic_write(tmp)
 
 
-def _stage_temp(target: Path, text: str) -> Path:
+def _stage_temp(target: Path, text: str, staged: list[Path]) -> Path:
     # newline="" keeps the operator's/agent's bytes: the default text write
     # translates LF into os.linesep, so SOUL.md and config.yaml landed with
     # CRLF on Windows while every reader saw LF.
     with tempfile.NamedTemporaryFile("w", encoding="utf-8", newline="", dir=target.parent, suffix=".tmp", delete=False) as f:
+        # Register the path before writing: a failing ``write`` must still reach
+        # the caller's cleanup, or the staged file leaks with no attempt and no log.
+        path = Path(f.name)
+        staged.append(path)
         f.write(text)
-        return Path(f.name)
+        return path

backend/packages/harness/deerflow/persistence/managed_subagents/file.py

@@ -90,10 +90,12 @@ class FileManagedSubagentStore(ManagedSubagentStore):
         tmp_path: Path | None = None
         try:
             with tempfile.NamedTemporaryFile("w", encoding="utf-8", newline="", dir=path.parent, suffix=".tmp", delete=False) as handle:
+                # Record the staged path before writing: a failing ``write``/``fsync``
+                # must still reach the cleanup below, or the temp file leaks silently.
+                tmp_path = Path(handle.name)
                 handle.write(payload)
                 handle.flush()
                 os.fsync(handle.fileno())
-                tmp_path = Path(handle.name)
             os.replace(tmp_path, path)
             tmp_path = None
         finally:

backend/tests/test_atomic_cleanup_census.py

@@ -69,3 +69,59 @@ def test_agent_store_write_keeps_the_publish_error_when_cleanup_fails(tmp_path,
 
     with pytest.raises(OSError, match="replace failed"):
         FileAgentStore._write(agent_dir, {"name": "planner"}, "# SOUL")
+
+
+def test_managed_subagent_write_removes_the_temp_when_the_write_fails(tmp_path, monkeypatch):
+    """A failure before the publish still has to discard the staged file."""
+    from deerflow.persistence.managed_subagents import file as store_module
+    from deerflow.persistence.managed_subagents.base import ManagedSubagentDefinition
+
+    definition = ManagedSubagentDefinition(
+        name="planner",
+        description="plans",
+        system_prompt="plan things",
+    )
+
+    def _failing_fsync(*_args, **_kwargs):
+        raise OSError("fsync failed")
+
+    monkeypatch.setattr(store_module.os, "fsync", _failing_fsync)
+
+    with pytest.raises(OSError, match="fsync failed"):
+        store_module.FileManagedSubagentStore._atomic_write(tmp_path / "planner.json", definition)
+
+    assert [p.name for p in tmp_path.iterdir()] == []
+
+
+def test_agent_store_write_removes_the_temp_when_the_write_fails(tmp_path, monkeypatch):
+    """The staging helper must register its path before writing the payload."""
+    from deerflow.persistence.agents import file as store_module
+
+    agent_dir = tmp_path / "agents"
+    agent_dir.mkdir()
+
+    real_named_tempfile = store_module.tempfile.NamedTemporaryFile
+
+    class _FailingWrite:
+        """A real staged file whose payload write fails."""
+
+        def __init__(self, *args, **kwargs):
+            self._inner = real_named_tempfile(*args, **kwargs)
+            self.name = self._inner.name
+
+        def __enter__(self):
+            self._inner.__enter__()
+            return self
+
+        def __exit__(self, *exc_info):
+            return self._inner.__exit__(*exc_info)
+
+        def write(self, *_args, **_kwargs):
+            raise OSError("write failed")
+
+    monkeypatch.setattr(store_module.tempfile, "NamedTemporaryFile", _FailingWrite)
+
+    with pytest.raises(OSError, match="write failed"):
+        store_module.FileAgentStore._write(agent_dir, {"name": "planner"}, "# SOUL")
+
+    assert [p.name for p in agent_dir.iterdir()] == []