personal memory agent
cortex: add pgid teardown for talent wrappers
Start talent wrapper subprocesses in their own process groups and terminate the group when cortex stops a run, so the 2026-04-26 21:00-22:05 cortex meltdown failure mode cannot leave Gemini CLI descendants behind after wrapper shutdown. Talent main now installs asyncio-integrated SIGTERM/SIGINT cancellation handlers and removes them during cleanup; the live dev-host Gemini smoke remains deferred until API credentials are available.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
+35
-2
+22
-2
think/cortex.py
+22
-2
think/cortex.py
···
20
20
import json
21
21
import logging
22
22
import os
23
+
import signal
23
24
import subprocess
24
25
import sys
25
26
import threading
···
58
59
59
60
if self.process.poll() is None:
60
61
# First try SIGTERM for graceful shutdown
61
-
self.process.terminate()
62
+
try:
63
+
self.process.terminate()
64
+
except ProcessLookupError:
65
+
pass
66
+
self._signal_process_group(signal.SIGTERM)
62
67
try:
63
68
self.process.wait(timeout=10) # Give more time for graceful shutdown
64
69
except subprocess.TimeoutExpired:
65
70
logging.getLogger(__name__).warning(
66
71
f"Talent {self.use_id} didn't stop gracefully, killing"
67
72
)
68
-
self.process.kill()
73
+
self._signal_process_group(signal.SIGKILL)
74
+
try:
75
+
self.process.kill()
76
+
except ProcessLookupError:
77
+
pass
69
78
self.process.wait() # Ensure zombie is reaped
79
+
80
+
def _signal_process_group(self, sig: int) -> None:
81
+
try:
82
+
pgid = os.getpgid(self.process.pid)
83
+
except ProcessLookupError:
84
+
return
85
+
try:
86
+
os.killpg(pgid, sig)
87
+
except ProcessLookupError:
88
+
return
70
89
71
90
72
91
class CortexService:
···
321
340
env=env,
322
341
bufsize=1,
323
342
cwd=subprocess_cwd,
343
+
start_new_session=True,
324
344
)
325
345
326
346
# Send input and close stdin
+13
think/talents.py
+13
think/talents.py
···
18
18
import json
19
19
import logging
20
20
import os
21
+
import signal
21
22
import sys
22
23
import traceback
23
24
from datetime import datetime
···
1239
1240
1240
1241
app_logger = setup_logging(args.verbose)
1241
1242
event_writer = JSONEventWriter(None)
1243
+
loop = asyncio.get_running_loop()
1244
+
main_task = asyncio.current_task()
1245
+
registered_signals: list[signal.Signals] = []
1246
+
if main_task:
1247
+
for sig in (signal.SIGTERM, signal.SIGINT):
1248
+
try:
1249
+
loop.add_signal_handler(sig, main_task.cancel)
1250
+
registered_signals.append(sig)
1251
+
except (NotImplementedError, RuntimeError):
1252
+
LOG.debug("Signal handler registration unavailable for %s", sig)
1242
1253
1243
1254
def emit_event(data: Event) -> None:
1244
1255
if "ts" not in data:
···
1301
1312
emit_event(err)
1302
1313
raise
1303
1314
finally:
1315
+
for sig in registered_signals:
1316
+
loop.remove_signal_handler(sig)
1304
1317
event_writer.close()
1305
1318
1306
1319