From c9469a2f5f0aec2553cd3086ab9b9cc4f577a84b Mon Sep 17 00:00:00 2001
From: HCL <chenglunhu@gmail.com>
Date: Tue, 21 Apr 2026 11:02:32 +0800
Subject: [PATCH] fix(gateway): exclude CLI invoker ancestor chain in
 _append_unique_pid

Reporter @yes999zc in #13242: `hermes gateway run` from an interactive
CLI session got false-positive 'already running' and shut down
immediately. The `ps aux` scan in `_scan_gateway_pids` matches
generic substrings like 'hermes_cli.main gateway', which picks up the
invoking CLI process AND its ancestor chain (parent shell, tmux session,
launcher).

`_append_unique_pid` already filtered `os.getpid()` but not the
ancestor chain. The repo already has an ancestor-detection helper
(`_is_pid_ancestor_of_current_process`) used by
`_request_gateway_self_restart`; extend `_append_unique_pid` to
reuse it, so CLI-invoker self-matches are filtered at the same point
that covers `find_gateway_pids`, service probes, and scan results.

Fixes #13242
---
 hermes_cli/gateway.py            |  8 ++++++++
 tests/hermes_cli/test_gateway.py | 32 ++++++++++++++++++++++++++++++++
 2 files changed, 40 insertions(+)

diff --git a/hermes_cli/gateway.py b/hermes_cli/gateway.py
index bc809cadf93a..ef628a2c99c3 100644
--- a/hermes_cli/gateway.py
+++ b/hermes_cli/gateway.py
@@ -180,6 +180,14 @@ def _append_unique_pid(pids: list[int], pid: int | None, exclude_pids: set[int])
         return
     if pid == os.getpid() or pid in exclude_pids or pid in pids:
         return
+    # Exclude the CLI invoker's ancestor chain (parent shell, launcher, tmux
+    # session, etc.). Otherwise `hermes gateway run` from an interactive CLI
+    # gets a self false-positive: `ps aux` matches the current hermes CLI
+    # and/or its ancestor, and the gateway decides it's already running and
+    # exits immediately. The same ancestor-detection helper is already used
+    # by `_request_gateway_self_restart`. (#13242)
+    if _is_pid_ancestor_of_current_process(pid):
+        return
     pids.append(pid)
 
 
diff --git a/tests/hermes_cli/test_gateway.py b/tests/hermes_cli/test_gateway.py
index 07265b2c3ac4..91f3a44b594a 100644
--- a/tests/hermes_cli/test_gateway.py
+++ b/tests/hermes_cli/test_gateway.py
@@ -253,6 +253,10 @@ def test_find_gateway_pids_falls_back_to_pid_file_when_process_scan_fails(monkey
     monkeypatch.setattr(gateway, "_get_service_pids", lambda: set())
     monkeypatch.setattr(gateway, "is_windows", lambda: False)
     monkeypatch.setattr("gateway.status.get_running_pid", lambda: 321)
+    # _append_unique_pid now also excludes ancestor PIDs (#13242). Stub the
+    # ancestor check so this fallback test doesn't shell out to `ps -o ppid=`
+    # for every candidate pid (321 is a fake service pid, not a real ancestor).
+    monkeypatch.setattr(gateway, "_is_pid_ancestor_of_current_process", lambda pid: False)
 
     def fake_run(cmd, **kwargs):
         if cmd[:4] == ["ps", "-A", "eww", "-o"]:
@@ -264,6 +268,34 @@ def fake_run(cmd, **kwargs):
     assert gateway.find_gateway_pids() == [321]
 
 
+def test_append_unique_pid_excludes_current_process_ancestors(monkeypatch):
+    """#13242: `hermes gateway run` from an interactive CLI saw the CLI
+    process itself in `ps aux` and concluded 'already running'. The fix
+    extends `_append_unique_pid` to exclude not just `os.getpid()` but the
+    entire ancestor chain (parent shell, launcher, tmux/screen, etc.).
+    """
+    # Simulate: current pid = 5000, its parent = 4000 (a "hermes" CLI process
+    # that ps picked up and would have been returned). `_append_unique_pid`
+    # must reject BOTH 5000 (current) and 4000 (ancestor).
+    monkeypatch.setattr("os.getpid", lambda: 5000)
+
+    parent_chain = {5000: 4000, 4000: 1}  # pid 4000 is the direct parent
+
+    def fake_parent(pid):
+        return parent_chain.get(pid, 0)
+
+    monkeypatch.setattr(gateway, "_get_parent_pid", fake_parent)
+
+    pids: list[int] = []
+    gateway._append_unique_pid(pids, 5000, set())  # current
+    gateway._append_unique_pid(pids, 4000, set())  # ancestor
+    gateway._append_unique_pid(pids, 7777, set())  # unrelated real gateway
+    assert pids == [7777], (
+        "_append_unique_pid must exclude current pid AND its ancestor chain "
+        "so the CLI invoker doesn't self-detect as a running gateway (#13242)"
+    )
+
+
 # ---------------------------------------------------------------------------
 # _wait_for_gateway_exit
 # ---------------------------------------------------------------------------