Fixes #11205

The gateway session-expiry code path shut sessions down without invoking MemoryProvider.on_session_end(), so memory providers never got a chance to flush state for expired sessions. This patch invokes the hook on the expiry path, mirroring the graceful-end behavior.

What

_flush_memories_for_session (gateway/run.py:743) is the convergence point for all three gateway session-end paths:

1. Idle-timeout expiry from _session_expiry_watcher (:2046)
2. Daily scheduled reset (same watcher)
3. Explicit /reset from a platform handler (:4343, :6422)

Until now it only spawned a separate flush AIAgent to nudge the builtin memory tool — it never called memory_manager.on_session_end(history) on the live cached agent's memory manager. As a result, plugin MemoryProviders that implement on_session_end as their documented final-pass extraction hook fired exactly zero times on gateway platforms (per the bug report log: many periodic / pre_compress triggers, zero session_end).

This patch dispatches the hook before the bespoke flush runs, looking up the live agent in _agent_cache first (where idle agents live) and falling back to _running_agents (in case the agent is still mid-turn when expiry fires). The dispatch is wrapped in try/except so a misbehaving provider can't block the rest of the flush sequence — matching the per-provider error tolerance that already exists inside MemoryManager.on_session_end.

This is the same structural class of bug as #7193 / #7192: the hook contract exists, the dispatch layer works, but one caller was missing.

How

# gateway/run.py — _flush_memories_for_session, after the cron-skip guard
try:
    cached_agent = ...  # _agent_cache[session_key] -> _running_agents[session_key]
    if cached_agent is not None:
        _mm = getattr(cached_agent, "_memory_manager", None)
        if _mm is not None:
            _msgs = [...]  # cleaned transcript, role+content only
            _mm.on_session_end(_msgs)
except Exception as exc:
    logger.warning("MemoryProvider.on_session_end dispatch failed for session %s: %s", ...)

The bespoke flush-agent path is left fully intact (it addresses an orthogonal concern tracked in #6157).

Testing

New tests/gateway/test_flush_memory_session_end.py (6 tests, all passing) covers:

• Hook fires when the agent is in _agent_cache (idle path)
• Hook fires when the agent is in _running_agents (mid-turn fallback)
• No-op when no agent is cached for the session_key
• Provider failure does not break the bespoke flush downstream
• Cron sessions still bypass the entire flush (including the new hook)
• No session_key passed -> no cached-agent lookup, hook not called

Existing related tests (test_async_memory_flush.py, test_flush_memory_stale_guard.py, test_memory_provider.py) continue to pass — 82/82 green.

Test plan

• New unit tests for the dispatch + edge cases
• Existing memory-flush + memory-provider tests still pass
• Manual verification on a live gateway: install a plugin that logs from on_session_end, let an idle session expire, confirm the log line now appears

🤖 Generated with Claude Code