Bug: DefaultDictWithTimeout TTL refresh guard is inverted — expired keys never cleaned up (memory leak)

[karesansui-u]

Bug description

DefaultDictWithTimeout.__refresh() in pr_agent/servers/utils.py has an inverted guard condition on line 67. The comparison uses > when it should use <:

def __refresh(self):
    if self.__ttl is None:
        return
    request_time = self.__time()
    if request_time - self.__last_refresh > self.__refresh_interval:  # ← inverted
        return
    # cleanup logic below — only reached when NOT enough time has passed

When enough time has passed since the last refresh (i.e., cleanup is needed), the method returns early. The cleanup logic only runs when the interval has NOT elapsed — when there's nothing to clean up.

Impact

_duplicate_push_triggers and _pending_task_duplicate_push_conditions in the GitHub App server accumulate every processed PR URL and never release them. On long-running server instances, this is a memory leak that grows linearly with the number of PRs processed.

Steps to reproduce

from pr_agent.servers.utils import DefaultDictWithTimeout
import time

d = DefaultDictWithTimeout(list, ttl=1, refresh_interval=0.5)
d["key1"].append("val")
time.sleep(2)  # well past TTL
d["key2"]  # triggers __refresh via __getitem__
print("key1" in d)  # True — should be False, key1's TTL expired

Fix

PR #2255 — change > to < on line 67.

[JiwaniZakir]

The inverted > condition in DefaultDictWithTimeout.__refresh() (pr_agent/servers/utils.py, line 67) means the method returns early precisely when cleanup should run — when request_time - self.__last_refresh > self.__refresh_interval is true — and only executes the eviction logic during the window when the interval hasn't elapsed yet, making it a no-op in practice. This is a classic off-by-one logic inversion that's easy to miss in TTL/cache code because the "happy path" (keys being accessed frequently) masks the bug entirely until the server runs long enough to accumulate stale entries. The downstream effect on _duplicate_push_triggers and _pending_task_duplicate_push_conditions is a straightforward unbounded growth, since those dicts rely solely on this refresh mechanism for cleanup and never shrink. Flipping > to < on that line restores the intended semantics.