fix(gateway): tz-aware datetime compare in suspend_recent_sessions

[shagarwal]

Problem

SessionStore.suspend_recent_sessions() (the startup session-suspension safety net from #7536) compared entry.updated_at (a datetime) to a float cutoff derived from time.time():

cutoff = _time.time() - max_age_seconds
...
if not entry.suspended and entry.updated_at >= cutoff:

This raised TypeError: >= not supported between instances of datetime.datetime and float on gateway startup, silently disabling the whole loop-recovery mechanism.

Impact

When a gateway restart happened mid-conversation (e.g. to stop an ignored-chat loop), in-flight sessions were blindly resumed instead of reset, resulting in apparent multi-hour hangs where the agent appeared stuck and never responded to the user.

Observed in the wild: gateway restart at 06:25, user session sat suspended-but-resumed until ~08:02, never processing queued messages.

Fix

• Compute cutoff as a tz-aware datetime using _now() - timedelta(...)
• Normalize entry.updated_at tzinfo to match the cutoff before comparing
• Guard entry.updated_at is None

No behavior change when timestamps are well-formed; just stops the routine from crashing.

[alt-glitch]

Likely duplicate of #8026 and #8129 — all fix the same TypeError comparing float vs datetime in suspend_recent_sessions(). See also #7995 and #7973.

[teknium1]

Thanks for the detailed write-up and the real-world reproduction timeline, @shagarwal! This is a valid bug report.

After reviewing main, the fix is already in place — automated hermes-sweeper review found that suspend_recently_active() on current main already uses a datetime-based cutoff:

• gateway/session.py line 1074: cutoff = _now() - timedelta(seconds=max_age_seconds) — the function imports timedelta from datetime and calls _now() (which returns datetime.now()), so the comparison is datetime-to-datetime. No time.time() float anywhere in the file.
• The original float-based bug was introduced in commit 2d328d5c7 (fix(gateway): break stuck session resume loops on restart #7536) and was corrected as part of the broader session-resume overhaul in commit cb4addaca (fix(gateway): auto-resume sessions after drain-timeout restart #11852).
• @alt-glitch also noted this is likely a duplicate of fix(gateway): use datetime cutoff in suspend_recently_active #8026, fix: compare session suspend cutoff as datetime #8129, fix(gateway): TypeError in suspend_recently_active on startup #7995, and [codex] fix gateway startup session suspension cutoff #7973 — all addressing the same TypeError.

Closing as implemented on main. The resume_pending skip guard added in cb4addaca also covers the broader stuck-resume scenario described in the impact section.