decnet/agent/executor.py

@@ -152,6 +152,7 @@ async def self_destruct() -> None:
     install footprint. Returns immediately so the HTTP response can drain
     before the reaper starts deleting files out from under the agent."""
     import os
+    import shutil
     import subprocess  # nosec B404
     import tempfile
 
@@ -172,17 +173,39 @@ async def self_destruct() -> None:
         os.close(fd)
     os.chmod(path, 0o700)  # nosec B103 — root-owned reaper, needs exec
 
-    # start_new_session detaches from the agent process group so the
+    # The reaper MUST run outside decnet-agent.service's cgroup — otherwise
-    # reaper isn't killed when systemctl stop decnet-agent fires.
+    # `systemctl stop decnet-agent` SIGTERMs the whole cgroup (reaper included)
+    # before rm -rf completes. `start_new_session=True` gets us a fresh POSIX
+    # session but does NOT escape the systemd cgroup. So we prefer
+    # `systemd-run --scope` (launches the command in a transient scope
+    # detached from the caller's service), falling back to a bare Popen if
+    # systemd-run is unavailable (non-systemd host / container).
+    systemd_run = shutil.which("systemd-run")
+    if systemd_run:
+        argv = [
+            systemd_run,
+            "--collect",
+            "--unit", f"decnet-reaper-{os.getpid()}",
+            "--description", "DECNET agent self-destruct reaper",
+            "/bin/bash", path,
+        ]
+        spawn_kwargs = {"start_new_session": True}
+    else:
+        argv = ["/bin/bash", path]
+        spawn_kwargs = {"start_new_session": True}
+
     subprocess.Popen(  # nosec B603
-        ["/bin/bash", path],
+        argv,
         stdin=subprocess.DEVNULL,
         stdout=subprocess.DEVNULL,
         stderr=subprocess.DEVNULL,
-        start_new_session=True,
         close_fds=True,
+        **spawn_kwargs,
+    )
+    log.warning(
+        "self_destruct: reaper spawned path=%s via=%s — agent will die in ~3s",
+        path, "systemd-run" if systemd_run else "popen",
     )
-    log.warning("self_destruct: reaper spawned path=%s — agent will die in ~3s", path)
 
 
 async def status() -> dict[str, Any]:

tests/swarm/test_agent_app.py

@@ -71,8 +71,12 @@ def test_self_destruct_spawns_reaper_and_returns_fast(monkeypatch, tmp_path) ->
     assert resp.json()["status"] == "self_destruct_scheduled"
     assert len(spawned) == 1
     assert spawned[0]["kw"].get("start_new_session") is True
-    script_path = spawned[0]["args"][1]
+    script_candidates = [
-    assert script_path.startswith("/tmp/decnet-reaper-")
+        a for a in spawned[0]["args"]
+        if isinstance(a, str) and a.startswith("/tmp/decnet-reaper-")
+    ]
+    assert len(script_candidates) == 1, spawned[0]["args"]
+    script_path = script_candidates[0]
     # Reaper content sanity check — covers the paths the operator asked for.
     import pathlib
     body = pathlib.Path(script_path).read_text()