Files
hermes-agent/tests/tools/test_process_checkpoint_worker_handoff.py
alt-glitch 0cbf579e63 refactor(environments-terminal): resolve the process checkpoint path per call
Delete CHECKPOINT_PATH and _CHECKPOINT_PATH_AT_IMPORT from tools/process_registry.py.
_checkpoint_path() returned the constant only when a test had patched it; production
never patches it, so every real call already took get_hermes_home() / "processes.json".
The constant was also an import-time read of HERMES_HOME. No other production reader,
not in plugin_api.json or the frozen updater surface.

Tests and evals/completion_backlog_probe.py now patch _checkpoint_path itself.
2026-11-16 21:56:33 +05:30

53 lines
2.9 KiB
Python

"""processes.json keeps the worker handoffs its writer does not track (ahrazzle B2-04).
A managed worker's background process reaches the next interpreter only through the profile
checkpoint. The owner (or a second worker) rewrote that file from its own table alone, so its
next process start/exit dropped the worker's entry: nobody tailed the log, notify_on_complete
never fired and poll/kill could not see the process. Real processes, real checkpoint file.
"""
import json
import subprocess
import sys
import time
import psutil
import pytest
@pytest.mark.platforms("linux", "macos")
def test_owner_checkpoint_write_keeps_a_live_worker_handoff(tmp_path, monkeypatch):
monkeypatch.setenv('HERMES_HOME', str(tmp_path))
from tools import process_registry as pr
from tools.process_registry_output_log import output_dir
checkpoint = tmp_path / 'processes.json'
monkeypatch.setattr(pr, '_checkpoint_path', lambda: checkpoint)
logs = output_dir()
logs.mkdir(parents=True)
worker_proc = subprocess.Popen([sys.executable, '-c', 'import time; time.sleep(60)'], stdin=subprocess.DEVNULL)
owner_proc = subprocess.Popen([sys.executable, '-c', 'import time; time.sleep(60)'], stdin=subprocess.DEVNULL)
try:
handoff = {'session_id': 'proc_worker', 'command': 'dev-server', 'pid': worker_proc.pid, 'pid_scope': 'host',
'host_start_time': pr.ProcessRegistry._safe_host_start_time(worker_proc.pid),
'started_at': time.time(), 'output_log': str(logs / 'proc_worker.log'), 'notify_on_complete': True}
(logs / 'proc_worker.log').write_text('serving\n')
gone = {**handoff, 'session_id': 'proc_gone', 'pid': 2 ** 22 + 7, 'output_log': str(logs / 'proc_gone.log')}
(logs / 'proc_gone.log').write_text('')
checkpoint.write_text(json.dumps([handoff, gone])) # what the exiting worker left behind
owner = pr.ProcessRegistry()
owner._running['proc_owner'] = pr.ProcessSession(
id='proc_owner', command='owner-job', task_id='t', started_at=time.time(), pid=owner_proc.pid,
pid_scope='host', host_start_time=pr.ProcessRegistry._safe_host_start_time(owner_proc.pid))
owner._write_checkpoint() # the owner's own process start
ids = {entry['session_id'] for entry in json.loads(checkpoint.read_text())}
assert ids == {'proc_owner', 'proc_worker'}, ids
# The next interpreter still adopts the worker's process from the file.
successor = pr.ProcessRegistry()
successor._running['proc_owner'] = owner._running['proc_owner']
assert successor.recover_from_checkpoint() == 1
assert successor.get('proc_worker') is not None
finally:
for proc in (worker_proc, owner_proc):
proc.kill()
proc.wait(timeout=10)
assert not psutil.pid_exists(worker_proc.pid) or psutil.Process(worker_proc.pid).status() == psutil.STATUS_ZOMBIE