Merge pull request #5581 from nesquena/release/stage-5543-r3

release #5543: fire-and-forget memory commit + drain-safe managed stop

CHANGELOG.md

@@ -14,6 +14,7 @@
 
 ### Fixed
 
+- **Starting a new chat is instant again — memory extraction no longer blocks it, and a managed restart no longer drops that memory.** Creating a new conversation could stall 1–5+ seconds because the previous session's memory-commit (the extraction call to the memory provider) ran synchronously on the request. That commit now runs in the background so "+ New Chat" responds immediately. To keep the deferred work safe across a restart, a `SIGTERM` handler drives an orderly stop so the server drains in-flight memory commits before exiting (previously a managed stop terminated instantly and lost them), the background-commit registry is bounded and self-cleaning, and the drain is wall-clock–bounded so a hung memory provider can't stall the stop (each commit runs under a shared time budget). Thanks @luperrypf. (#5543)
 - **Auto-scroll now resumes when you return to the bottom of a streaming reply.** If you scrolled up even once while a response was streaming, the chat permanently stopped following new output — scrolling back to the bottom didn't help; only starting a new turn re-engaged auto-follow. Returning to the very bottom of the transcript (within ~80px) now re-pins auto-scroll, so the view follows the stream again. It only re-pins at the true bottom and never while you're actively scrolling (wheel, keyboard, touch, or scrollbar drag), so reading back through history mid-stream still never yanks you down — preserving the #4295 behavior. Thanks @luperrypf. (#5544)
 - **Forking a long or compacted conversation now keeps a clean turn boundary.** "Fork from here" on a large session (one big enough to have been context-compacted) could cut the copied context in the middle of a turn — or right after a tool call, before its result — leaving the new conversation with a malformed context that the model choked on. The fork now aligns its cut to a real turn boundary in the same display coordinate space `/api/session` uses (matching both the compacted and non-compacted cases), erring toward keeping slightly less rather than slicing a turn, with send-time sanitization as a backstop. Small, non-compacted forks are unchanged. Thanks @b3nw. (#5563)
 - **Clearing a conversation now actually clears it — history no longer comes back after a refresh (P0 data-loss).** `Clear conversation` wiped the on-screen transcript but never recorded a truncation watermark, so the messages resurrected from the agent's state database on the next load: the cleared history reappeared after a refresh, and continuing the conversation still fed the full pre-clear context to the model. Clear now routes through the same truncate-to-empty path as the truncate action (setting the watermark that blocks state-db replay), verifies the empty state actually persisted, removes the stale backup, and — for a compressed-continuation session — detaches the compression-snapshot parent link so the archived transcript can't be stitched back in either (ordinary fork links are preserved). Thanks @rodboev and the maintainer fix. (#5556, #5553, #5532)

api/routes.py

@@ -13067,21 +13067,49 @@ def handle_post(handler, parsed) -> bool:
                 # the new session (#5420).
                 prev_session_id = None
             if prev_session_id:
-                try:
-                    from api.session_lifecycle import commit_session_memory
-                    from api.config import SESSION_AGENT_CACHE, SESSION_AGENT_CACHE_LOCK
-                    prev_agent = None
-                    with SESSION_AGENT_CACHE_LOCK:
-                        _cached = SESSION_AGENT_CACHE.get(prev_session_id)
-                        if _cached:
-                            prev_agent = _cached[0]
-                    commit_session_memory(prev_session_id, agent=prev_agent)
-                except Exception:
-                    logger.debug(
-                        "Lifecycle commit for prev_session %s failed",
-                        prev_session_id,
-                        exc_info=True,
-                    )
+                # Fire-and-forget: commit_memory_session() can take 1-5+ seconds
+                # (extraction call to the memory provider), and blocking the
+                # response here made "+ New Chat" feel slow/unresponsive.
+                # commit_session_memory() already serialises overlapping commits
+                # for a session via its own in-flight guard, so running it off
+                # the request thread is safe.
+                def _commit_prev_session_memory(_sid=prev_session_id):
+                    try:
+                        from api.session_lifecycle import commit_session_memory
+                        from api.config import SESSION_AGENT_CACHE, SESSION_AGENT_CACHE_LOCK
+                        prev_agent = None
+                        with SESSION_AGENT_CACHE_LOCK:
+                            _cached = SESSION_AGENT_CACHE.get(_sid)
+                            if _cached:
+                                prev_agent = _cached[0]
+                        commit_session_memory(_sid, agent=prev_agent)
+                    except Exception:
+                        logger.warning(
+                            "Lifecycle commit for prev_session %s failed",
+                            _sid,
+                            exc_info=True,
+                        )
+                    finally:
+                        # Self-unregister so the background-commit registry does
+                        # not leak completed threads; drain only tracks live ones.
+                        try:
+                            from api.session_lifecycle import _unregister_background_commit_thread
+                            _unregister_background_commit_thread(threading.current_thread())
+                        except Exception:
+                            pass
+
+                t = threading.Thread(
+                    target=_commit_prev_session_memory,
+                    daemon=True,
+                    name=f"commit-memory-{prev_session_id}",
+                )
+                from api.session_lifecycle import _register_background_commit_thread
+                # Refused only if shutdown draining has already begun; in that
+                # window the inline drain commits the pending generation instead,
+                # so skipping the worker start is safe (avoids a late daemon
+                # thread the drain snapshot already missed).
+                if _register_background_commit_thread(t):
+                    t.start()
         s = new_session(
             workspace=workspace,
             model=model,

api/session_lifecycle.py

@@ -45,6 +45,56 @@ _condition = threading.Condition(_lock)
 
 _sessions: dict[str, dict] = {}
 
+# Background commit threads spawned by fire-and-forget paths (e.g. POST /api/session/new).
+# Tracked so drain_all_on_shutdown() can join them before interpreter teardown,
+# preventing silent data loss when daemon threads are abandoned mid-commit.
+_background_commit_threads: set[threading.Thread] = set()
+_background_commit_threads_lock = threading.Lock()
+# Set once drain_all_on_shutdown() begins so no NEW background worker is started
+# after the drain has snapshotted the registry. A late arrival is not lost: its
+# uncommitted generation is still committed inline by the generation-drain loop
+# below (which re-snapshots _sessions each pass and commits with wait=True).
+_draining = False
+
+
+def _register_background_commit_thread(t: threading.Thread) -> bool:
+    """Register a fire-and-forget commit thread. Returns True if the caller
+    should start it. Returns False when shutdown draining has already begun —
+    the caller must NOT start a new worker in that window; the inline
+    generation-drain will commit the pending work instead."""
+    with _background_commit_threads_lock:
+        if _draining:
+            return False
+        _background_commit_threads.add(t)
+        return True
+
+
+def _unregister_background_commit_thread(t: threading.Thread) -> None:
+    # A completed worker removes itself so the registry does not grow without
+    # bound over the process lifetime; drain only needs threads still running.
+    with _background_commit_threads_lock:
+        _background_commit_threads.discard(t)
+
+
+def _drain_background_commit_threads(timeout: float = 5.0, deadline_fn=None) -> None:
+    """Join tracked fire-and-forget commit threads. When ``deadline_fn`` is given
+    (returns the remaining overall shutdown budget in seconds), each join is
+    clamped to that budget and the phase stops early once it is exhausted, so the
+    join phase cannot overrun ``drain_all_on_shutdown``'s deadline."""
+    with _background_commit_threads_lock:
+        threads = list(_background_commit_threads)
+        _background_commit_threads.clear()
+    for t in threads:
+        if not t.is_alive():
+            continue
+        if deadline_fn is None:
+            t.join(timeout)
+            continue
+        remaining = deadline_fn()
+        if remaining <= 0:
+            return
+        t.join(min(timeout, remaining))
+
 
 def _new_entry() -> dict:
     return {
@@ -225,16 +275,72 @@ def commit_session_memory(session_id: str, agent=None, *, wait: bool = False, ti
     return True
 
 
-def drain_all_on_shutdown() -> None:
+def drain_all_on_shutdown(deadline: float = 30.0) -> None:
+    # Signal that draining has begun so no NEW background commit worker is
+    # started after this point (see _register_background_commit_thread). Any
+    # already-registered worker is joined below; a caller that loses the race
+    # (registered after our snapshot, or refused a start) still has its pending
+    # generation committed by the inline loop.
+    global _draining
+    with _background_commit_threads_lock:
+        _draining = True
+
+    # Overall wall-clock budget so a stuck worker holding in_flight cannot hang
+    # the shutdown path forever. Start the clock BEFORE the join phase so the
+    # join and the inline drain share one budget. The inline commit runs in a
+    # daemon worker joined only for the remaining budget — otherwise a provider
+    # wedged inside commit_memory_session() (whether we are waiting on an
+    # existing in_flight commit, OR the inline drain owns the segment and calls
+    # the provider directly) would block indefinitely and the deadline check
+    # would never run again.
+    started = time.monotonic()
+
+    def _remaining() -> float:
+        return deadline - (time.monotonic() - started)
+
+    def _commit_within_budget(session_id: str, budget: float) -> bool:
+        """Run one commit in a daemon worker, join at most `budget` seconds.
+        Returns True only if the commit completed within budget AND made
+        progress. On timeout the daemon worker is left running (it will finish
+        or die at interpreter exit; either way it mutates generation state only
+        under _condition, so nothing is corrupted) and we report no progress so
+        the drain can honor its deadline instead of hanging on a wedged
+        provider."""
+        holder = {"ok": False}
+
+        def _run():
+            holder["ok"] = commit_session_memory(session_id, wait=True)
+
+        worker = threading.Thread(target=_run, daemon=True, name=f"drain-commit-{session_id}")
+        worker.start()
+        worker.join(budget)
+        return holder["ok"] if not worker.is_alive() else False
+
+    # Drain in-flight background commit threads first (fire-and-forget from
+    # POST /api/session/new) so their commit work completes before we drain any
+    # remaining uncommitted generations inline. Each join consumes the shared
+    # budget so it cannot overrun the overall deadline.
+    _drain_background_commit_threads(deadline_fn=_remaining)
+
     while True:
         with _lock:
             snapshot = [sid for sid, entry in _sessions.items() if entry["generation"] > entry["committed_generation"]]
         if not snapshot:
             return
+        remaining = _remaining()
+        if remaining <= 0:
+            logger.warning("drain_all_on_shutdown: deadline hit with uncommitted sessions: %s", sorted(snapshot))
+            return
 
         made_progress = False
         for sid in snapshot:
-            if commit_session_memory(sid, wait=True):
+            budget = _remaining()
+            if budget <= 0:
+                logger.warning("drain_all_on_shutdown: deadline hit with uncommitted sessions: %s", sorted(snapshot))
+                return
+            # Budget-bounded (covers both the wait-for-in_flight path AND the
+            # unbounded provider call when the inline drain owns the segment).
+            if _commit_within_budget(sid, budget):
                 made_progress = True
         if not made_progress:
             logger.debug("drain_all_on_shutdown: stopped with uncommitted sessions: %s", sorted(snapshot))

server.py

@@ -688,6 +688,36 @@ def main() -> None:
         print(f'  Remote access: ssh -N -L {PORT}:127.0.0.1:{PORT} <user>@<your-server>', flush=True)
     print(f'  Then open:     {scheme}://localhost:{PORT}', flush=True)
     print('', flush=True)
+
+    # ctl.sh stops the WebUI with SIGTERM. Python's default SIGTERM handler
+    # terminates the process WITHOUT unwinding the try/finally around
+    # serve_forever(), so drain_all_on_shutdown() (which flushes in-flight
+    # fire-and-forget memory commits) would never run on the normal managed
+    # stop. Install a handler that requests an orderly shutdown so
+    # serve_forever() returns and the existing `finally` block drains cleanly.
+    #
+    # httpd.shutdown() blocks until serve_forever() has exited and MUST NOT be
+    # called from the thread running serve_forever() (it would deadlock), so we
+    # dispatch it from a short-lived helper thread. The handler is idempotent
+    # and guards against double-shutdown (e.g. repeated SIGTERM/SIGINT).
+    _shutdown_requested = threading.Event()
+
+    def _request_shutdown(signum, _frame):
+        if _shutdown_requested.is_set():
+            return
+        _shutdown_requested.set()
+        threading.Thread(
+            target=httpd.shutdown,
+            name="webui-sigterm-shutdown",
+            daemon=True,
+        ).start()
+
+    try:
+        signal.signal(signal.SIGTERM, _request_shutdown)
+    except (ValueError, OSError):
+        # Not on the main thread (e.g. embedded/test harness); skip handler.
+        logger.debug("Could not install SIGTERM handler", exc_info=True)
+
     try:
         httpd.serve_forever()
     finally:

tests/test_issue5543_background_memory_commit.py

@@ -0,0 +1,358 @@
+"""Tests for issue #5543: fire-and-forget commit_session_memory + shutdown drain.
+
+PR #5543 (@luperrypf) moves commit_session_memory() off the POST
+/api/session/new request thread into a daemon "background commit" thread so
+"+ New Chat" returns immediately instead of blocking 1-5s on memory-provider
+extraction. The gate-certifier required three guarantees this file locks in:
+
+1. Fire-and-forget: spawning the background commit does not block the caller,
+   even while the underlying commit is still running.
+2. Registry hygiene: a completed background-commit worker unregisters itself so
+   ``_background_commit_threads`` does not grow without bound (Finding 2).
+3. Shutdown drain: drain_all_on_shutdown() joins in-flight background commit
+   threads AND flushes the pending generation, and the server's SIGTERM handler
+   routes the managed ``ctl.sh stop`` through the serve_forever() ``finally`` so
+   the drain actually runs (Finding 1).
+"""
+
+from __future__ import annotations
+
+import ast
+import importlib
+import socket
+import threading
+import time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+
+
+REPO = Path(__file__).resolve().parents[1]
+
+
+def _fresh_lifecycle():
+    """Import/reload lifecycle module and clear process-global test state."""
+    lifecycle = importlib.import_module("api.session_lifecycle")
+    lifecycle = importlib.reload(lifecycle)
+    reset = getattr(lifecycle, "_reset_for_tests", None)
+    if callable(reset):
+        reset()
+    # Ensure the background-commit registry starts empty for each test.
+    with lifecycle._background_commit_threads_lock:
+        lifecycle._background_commit_threads.clear()
+    return lifecycle
+
+
+class RecordingAgent:
+    """Agent whose commit blocks on an event so we can hold it "in flight"."""
+
+    def __init__(self):
+        self.calls = 0
+        self.entered = threading.Event()
+        self.release = threading.Event()
+
+    def commit_memory_session(self):
+        self.calls += 1
+        self.entered.set()
+        # Block until the test releases us (bounded so a buggy test can't hang).
+        self.release.wait(timeout=5)
+
+
+def _make_background_worker(lifecycle, sid, agent):
+    """Mirror the routes.py fire-and-forget worker: commit then self-unregister."""
+
+    def _worker():
+        try:
+            lifecycle.commit_session_memory(sid, agent=agent)
+        finally:
+            lifecycle._unregister_background_commit_thread(threading.current_thread())
+
+    t = threading.Thread(target=_worker, daemon=True, name=f"commit-memory-{sid}")
+    lifecycle._register_background_commit_thread(t)
+    return t
+
+
+# --------------------------------------------------------------------------- #
+# Finding 2: registry hygiene                                                   #
+# --------------------------------------------------------------------------- #
+
+
+def test_register_and_unregister_roundtrip():
+    lifecycle = _fresh_lifecycle()
+    t = threading.Thread(target=lambda: None)
+
+    # Registration returns True (caller should start) when not draining.
+    assert lifecycle._register_background_commit_thread(t) is True
+    with lifecycle._background_commit_threads_lock:
+        assert t in lifecycle._background_commit_threads
+
+    lifecycle._unregister_background_commit_thread(t)
+    with lifecycle._background_commit_threads_lock:
+        assert t not in lifecycle._background_commit_threads
+
+    # Idempotent: unregistering an absent thread is a harmless no-op.
+    lifecycle._unregister_background_commit_thread(t)
+
+
+def test_register_refused_once_draining():
+    """After drain begins, registration is refused (returns False) and the
+    thread is NOT added — the caller must skip start(); the inline generation
+    drain commits the pending work instead. This closes Codex's
+    register-before-start shutdown window."""
+    lifecycle = _fresh_lifecycle()
+    # Nothing pending, so drain_all_on_shutdown returns quickly but still sets
+    # the _draining flag.
+    lifecycle.drain_all_on_shutdown()
+
+    t = threading.Thread(target=lambda: None)
+    assert lifecycle._register_background_commit_thread(t) is False
+    with lifecycle._background_commit_threads_lock:
+        assert t not in lifecycle._background_commit_threads
+
+
+def test_completed_worker_self_unregisters():
+    """A finished background-commit worker must drop itself from the registry."""
+    lifecycle = _fresh_lifecycle()
+    agent = RecordingAgent()
+    agent.release.set()  # commit completes immediately
+    sid = "self-unregister"
+    lifecycle.register_agent(sid, agent)
+    lifecycle.mark_turn_completed(sid, agent=agent)
+
+    t = _make_background_worker(lifecycle, sid, agent)
+    t.start()
+    t.join(timeout=5)
+    assert not t.is_alive()
+
+    with lifecycle._background_commit_threads_lock:
+        assert t not in lifecycle._background_commit_threads
+        assert len(lifecycle._background_commit_threads) == 0
+    assert agent.calls == 1
+
+
+# --------------------------------------------------------------------------- #
+# Fire-and-forget: the caller does not block                                    #
+# --------------------------------------------------------------------------- #
+
+
+def test_fire_and_forget_does_not_block_caller():
+    """Starting the background commit returns immediately, even mid-commit."""
+    lifecycle = _fresh_lifecycle()
+    agent = RecordingAgent()  # blocks in commit until released
+    sid = "non-blocking"
+    lifecycle.register_agent(sid, agent)
+    lifecycle.mark_turn_completed(sid, agent=agent)
+
+    t = _make_background_worker(lifecycle, sid, agent)
+
+    start = time.monotonic()
+    t.start()
+    # Caller "returns" as soon as the thread is started; measure that dispatch.
+    elapsed = time.monotonic() - start
+
+    # The commit is genuinely in flight (proves we didn't run it inline).
+    assert agent.entered.wait(timeout=2)
+    assert elapsed < 1.0, f"caller blocked for {elapsed:.3f}s"
+    assert lifecycle.has_uncommitted_work(sid) is True
+
+    # Let the in-flight commit finish and clean up.
+    agent.release.set()
+    t.join(timeout=5)
+
+
+# --------------------------------------------------------------------------- #
+# Finding 1: shutdown drain flushes the pending commit                          #
+# --------------------------------------------------------------------------- #
+
+
+def test_drain_joins_inflight_background_commit_and_flushes():
+    """drain_all_on_shutdown() waits for the in-flight worker, then commits."""
+    lifecycle = _fresh_lifecycle()
+    agent = RecordingAgent()  # held in flight until we release it
+    sid = "drain-inflight"
+    lifecycle.register_agent(sid, agent)
+    lifecycle.mark_turn_completed(sid, agent=agent)
+
+    t = _make_background_worker(lifecycle, sid, agent)
+    t.start()
+    assert agent.entered.wait(timeout=2)  # commit is in flight
+
+    drained = threading.Event()
+
+    def _do_drain():
+        lifecycle.drain_all_on_shutdown()
+        drained.set()
+
+    dt = threading.Thread(target=_do_drain, daemon=True)
+    dt.start()
+
+    # While the background commit is still in flight, drain must be blocked
+    # joining it — it cannot have finished yet.
+    time.sleep(0.1)
+    assert not drained.is_set()
+
+    # Release the commit; drain should now join the worker and finish.
+    agent.release.set()
+    assert drained.wait(timeout=5)
+    dt.join(timeout=2)
+
+    assert agent.calls == 1
+    assert lifecycle.has_uncommitted_work(sid) is False
+    with lifecycle._background_commit_threads_lock:
+        assert t not in lifecycle._background_commit_threads
+
+
+def test_sigterm_style_shutdown_runs_finally_drain():
+    """The SIGTERM handler mechanism drives serve_forever() through its finally.
+
+    Replicates server.py's approach: a signal handler requests shutdown by
+    dispatching httpd.shutdown() from a short-lived helper thread (you cannot
+    call it from serve_forever()'s own thread). serve_forever() returns, the
+    ``finally`` runs, and drain_all_on_shutdown() flushes the pending commit.
+    """
+    lifecycle = _fresh_lifecycle()
+    agent = RecordingAgent()
+    agent.release.set()  # commit completes immediately during drain
+    sid = "sigterm-drain"
+    lifecycle.register_agent(sid, agent)
+    lifecycle.mark_turn_completed(sid, agent=agent)
+
+    class _Handler(BaseHTTPRequestHandler):
+        def log_message(self, *args):  # silence test noise
+            pass
+
+    httpd = ThreadingHTTPServer(("127.0.0.1", 0), _Handler)
+    finally_ran = threading.Event()
+
+    def _serve():
+        try:
+            httpd.serve_forever()
+        finally:
+            httpd.server_close()
+            lifecycle.drain_all_on_shutdown()
+            finally_ran.set()
+
+    serve_thread = threading.Thread(target=_serve, daemon=True)
+    serve_thread.start()
+
+    # Wait until the server is actually accepting connections before shutdown.
+    addr = httpd.server_address
+    for _ in range(100):
+        try:
+            with socket.create_connection(addr, timeout=0.1):
+                break
+        except OSError:
+            time.sleep(0.02)
+
+    # --- exact handler mechanism from server.py ---
+    shutdown_requested = threading.Event()
+
+    def _request_shutdown():
+        if shutdown_requested.is_set():
+            return
+        shutdown_requested.set()
+        threading.Thread(target=httpd.shutdown, daemon=True).start()
+
+    _request_shutdown()
+    _request_shutdown()  # idempotent: a repeated signal must not double-shutdown
+
+    serve_thread.join(timeout=5)
+    assert not serve_thread.is_alive(), "serve_forever() did not exit on shutdown"
+    assert finally_ran.wait(timeout=5), "finally block / drain did not run"
+    assert agent.calls == 1
+    assert lifecycle.has_uncommitted_work(sid) is False
+
+
+# --------------------------------------------------------------------------- #
+# Static wiring guards (the inline handler in main() cannot be imported)        #
+# --------------------------------------------------------------------------- #
+
+
+def test_server_installs_sigterm_handler_and_drains_in_finally():
+    """server.py must wire a SIGTERM handler and drain in serve_forever's finally."""
+    src = (REPO / "server.py").read_text(encoding="utf-8")
+    tree = ast.parse(src)
+
+    # A SIGTERM handler is installed via signal.signal(signal.SIGTERM, ...).
+    installs_sigterm = any(
+        isinstance(node, ast.Call)
+        and isinstance(node.func, ast.Attribute)
+        and node.func.attr == "signal"
+        and node.args
+        and ast.dump(node.args[0]).find("SIGTERM") != -1
+        for node in ast.walk(tree)
+    )
+    assert installs_sigterm, "server.py does not install a SIGTERM handler"
+
+    # httpd.shutdown() is dispatched (from a helper thread) rather than the
+    # default SIGTERM behavior that would skip the finally.
+    assert "httpd.shutdown" in src
+    # The existing shutdown drain is still wired.
+    assert "drain_all_on_shutdown" in src
+    # SIGPIPE handling must be preserved.
+    assert "SIGPIPE" in src
+
+
+# --------------------------------------------------------------------------- #
+# Deadline enforcement: a wedged worker must not hang the managed-stop drain     #
+# --------------------------------------------------------------------------- #
+
+
+def test_drain_deadline_returns_despite_stuck_worker():
+    """A background commit wedged in commit_memory_session() (e.g. a hung memory
+    provider) must NOT hang drain_all_on_shutdown() past its deadline. The inline
+    drain runs each commit in a budget-bounded daemon worker, so a stuck in_flight
+    session cannot block the drain. Regression for the dual-gate finding that the
+    top-of-loop deadline check never fired while blocked inside a stuck commit."""
+    lifecycle = _fresh_lifecycle()
+    agent = RecordingAgent()  # NEVER released → stays wedged in commit
+    sid = "stuck-drain"
+    lifecycle.register_agent(sid, agent)
+    lifecycle.mark_turn_completed(sid, agent=agent)
+
+    # Put the worker in flight, but keep it OUT of the registry so the 5s join
+    # phase does not dominate the timing — we are isolating the inline-drain
+    # deadline path.
+    worker = threading.Thread(
+        target=lambda: lifecycle.commit_session_memory(sid, agent=agent),
+        daemon=True,
+        name="stuck-commit",
+    )
+    worker.start()
+    assert agent.entered.wait(timeout=2)  # worker is now wedged holding in_flight
+
+    started = time.monotonic()
+    lifecycle.drain_all_on_shutdown(deadline=0.5)
+    elapsed = time.monotonic() - started
+
+    # Without the timeout fix this blocks until the worker is released (never);
+    # with it, the drain returns within roughly the deadline budget.
+    assert elapsed < 3.0, f"drain hung {elapsed:.1f}s past its 0.5s deadline"
+
+    # Let the wedged worker unwind so the test process exits cleanly.
+    agent.release.set()
+    worker.join(timeout=5)
+
+
+def test_drain_deadline_returns_when_inline_provider_hangs_no_prior_worker():
+    """Deeper case (2nd-round gate finding): a DIRTY session with NO pre-existing
+    in_flight worker, whose provider call hangs when the inline drain owns the
+    segment and calls commit_memory_session() DIRECTLY. The timeout= param only
+    bounds waiting on an already-in_flight commit; the direct provider call must
+    also be budget-bounded (run in a joined daemon worker), or the drain hangs."""
+    lifecycle = _fresh_lifecycle()
+    agent = RecordingAgent()  # NEVER released → the DIRECT provider call wedges
+    sid = "inline-hang"
+    lifecycle.register_agent(sid, agent)
+    lifecycle.mark_turn_completed(sid, agent=agent)
+    # No background worker started: the session is simply dirty
+    # (generation > committed_generation) with in_flight=False, so the inline
+    # drain itself acquires the segment and calls the hanging provider directly.
+
+    started = time.monotonic()
+    lifecycle.drain_all_on_shutdown(deadline=0.5)
+    elapsed = time.monotonic() - started
+
+    assert elapsed < 3.0, f"inline-drain hung {elapsed:.1f}s past its 0.5s deadline"
+
+    # Let the wedged inline commit worker unwind cleanly.
+    agent.release.set()