diff --git a/cron/lifecycle_guard.py b/cron/lifecycle_guard.py
index 6c70c1af8ae87..cf6c3e0739fc4 100644
--- a/cron/lifecycle_guard.py
+++ b/cron/lifecycle_guard.py
@@ -51,18 +51,48 @@ class GatewayLifecycleBlocked(ValueError):
     # `start` is intentionally excluded: starting a gateway from inside a
     # gateway is benign (a no-op or "already running" error), and a
     # legitimate cron job might start a sibling profile's gateway.
-    r"(?:hermes\s+gateway\s+(?:restart|stop))"
+    # The separator class is ["',\s]+ rather than \s+ throughout: the same
+    # command reaches this guard both as a shell string ("launchctl bootout X")
+    # and as a Python argv list from execute_code
+    # (subprocess.run(["launchctl", "bootout", "X"])), where the tokens are
+    # separated by quotes and commas instead of spaces. Verified 2026-07-19:
+    # with \s+ the list form was the one variant that slipped past every other
+    # branch.
+    r"(?:hermes[\"',\s]+gateway[\"',\s]+(?:restart|stop))"
     # Branch B: launchctl ops on a hermes-gateway label. macOS launchd
     # labels look like `ai.hermes.gateway` / `hermes-gateway`. Requiring the
     # gateway identifier prevents blocking unrelated hermes services (e.g.
     # `launchctl unload ai.hermes.update-checker.plist`).
-    r"|(?:launchctl\s+(?:kickstart|unload|load|stop|restart)\b[^\n]*\bhermes[.\-]?gateway)"
+    # `bootout`/`bootstrap` (the modern launchctl verbs) are the worst of the
+    # set and were missing: `bootout` *unloads* the job, so KeepAlive can no
+    # longer revive it, and the `bootstrap` that would re-register it dies with
+    # the gateway it just killed. Real incident 2026-07-13: the agent fell
+    # through this gap, booted the job out, and the Mac gateway stayed dead for
+    # 26h until a manual restart. `disable`/`remove` are blocked for the same
+    # reason (they leave the label un-startable).
+    r"|(?:launchctl[\"',\s]+(?:kickstart|unload|load|stop|restart|bootout|bootstrap|disable|remove)\b[^\n]*\bhermes[.\-]?gateway)"
     # Branch C: systemctl ops on a hermes-gateway unit.
     r"|(?:systemctl\s+(?:-\S+\s+)*(?:restart|stop|start)\b[^\n]*\bhermes[.\-]?gateway)"
     # Branch D: pkill / kill targeting the hermes gateway process. Both
     # token orders because real reproductions show both.
-    r"|(?:p?kill\b[^\n]*\bhermes\b[^\n]*\bgateway)"
+    # No right-hand \b after `hermes`: the process as it actually appears in
+    # `ps` is `hermes_cli.main gateway run`, and `\bhermes\b` does NOT match
+    # `hermes_cli` because `_` is a word character — so the single most likely
+    # kill command (`pkill -f "hermes_cli.main gateway"`) sailed straight
+    # through this guard until 2026-07-19. Matching `hermes` as a prefix also
+    # covers `hermes-gateway`, `hermesd`, etc. `gateway` must still appear.
+    r"|(?:p?kill\b[^\n]*\bhermes[^\n]*\bgateway)"
     r"|(?:p?kill\b[^\n]*\bgateway\b[^\n]*\bhermes)"
+    # Branch E: starting a SECOND gateway with `--replace`. Branch A excludes
+    # bare `start`/`run` on the theory that starting a gateway from inside one
+    # is benign ("already running"), and that holds — but only WITHOUT this
+    # flag. `--replace` means "SIGTERM whoever holds the slot, then take it",
+    # i.e. it is a restart wearing a start's clothes. Real incident 2026-07-19:
+    # a second `gateway run --replace` SIGTERM'd the live gateway mid-task; the
+    # victim logged `parent_pid=1` (its own launchd parent), which reads like
+    # launchd killed it and hides the actual caller. Covers both the CLI shape
+    # and the module shape (`python -m hermes_cli.main gateway run --replace`).
+    r"|(?:gateway\s+(?:run|start)\b[^\n]*--replace)"
 )
 
 
diff --git a/tools/code_execution_tool.py b/tools/code_execution_tool.py
index 650ee5b31d341..707503c91379e 100644
--- a/tools/code_execution_tool.py
+++ b/tools/code_execution_tool.py
@@ -1195,6 +1195,34 @@ def execute_code(
     if not code or not code.strip():
         return tool_error("No code provided.")
 
+    # Hard-block gateway-lifecycle commands, mirroring the guard in
+    # terminal_tool. Without this, execute_code is a straight bypass: the
+    # terminal() path refuses `launchctl bootout ai.hermes.gateway`, but the
+    # identical command inside `os.system(...)` / `subprocess.run([...])` here
+    # sailed through and SIGTERM'd the gateway mid-task — verified 2026-07-19,
+    # every variant (bootout / kickstart -k / hermes gateway restart) returned
+    # approved=True. That matters because several bundled skills still document
+    # a bootout recipe in their troubleshooting sections (grsai-provider, which
+    # the agent reads while generating images, among them), so the model has a
+    # standing invitation to run exactly this. Killing the gateway from inside
+    # it also kills this very subprocess, so the restart may never complete.
+    if os.environ.get("_HERMES_GATEWAY") == "1":
+        from hermes_cli.cron import _contains_gateway_lifecycle_command
+        if _contains_gateway_lifecycle_command(code):
+            return json.dumps({
+                "status": "error",
+                "error": (
+                    "Blocked: cannot restart or stop the gateway from inside the "
+                    "gateway process. The gateway would kill this script before it "
+                    "could complete (SIGTERM propagates to child processes). Run "
+                    "`hermes gateway restart` from a shell outside the gateway, or "
+                    "let the watchdog (~/.hermes/scripts/gateway-watchdog.sh) "
+                    "handle recovery."
+                ),
+                "tool_calls_made": 0,
+                "duration_seconds": 0,
+            }, ensure_ascii=False)
+
     # Dispatch: remote backends use file-based RPC, local uses UDS
     from tools.terminal_tool import _get_env_config, _docker_has_host_access
     _env_config = _get_env_config()