auth: spurious REFRESHGENERATIONS bump on early-cancel in coalescedrunrefreshcmd Summary…
- User
auth: spurious _REFRESH_GENERATIONS bump on early-cancel in _coalesced_run_refresh_cmd
Summary
_coalesced_run_refresh_cmdinsrc/notebooklm/_auth/refresh.py(moved fromsrc/notebooklm/auth.pyin tier-10 PR-B-high #814, but the code is byte-identical to origin/main since tier-9 E #803) can bump_REFRESH_GENERATIONS[refresh_key]even when no real refresh happened, in a narrow cancel-before-await race.Reproducer (race condition)
- Caller invokes
await _coalesced_run_refresh_cmd(refresh_key, path, profile)from the cancel-recovery loop in_fetch_tokens_with_refresh. - Task is cancelled in the window between
try:and the firstwith _REFRESH_STATE_LOCK:block inside_coalesced_run_refresh_cmd— so NO future is registered in_REFRESH_INFLIGHT_BY_LOOP[loop][refresh_key]. - Control jumps to
except asyncio.CancelledError:in the caller. registry.get(refresh_key)returnsNone.- The inner
if inflight is None or inflight.done():is taken; theif inflight is not None:block is skipped;subprocess_excstaysNone;break. - Outer
if subprocess_exc is not None:is False, so no raise. - Generation is bumped at the
with _REFRESH_STATE_LOCK: ... max(existing, refresh_generation + 1)line, even though no subprocess ran. - Then
if caller_cancelled: raise asyncio.CancelledError().
Impact
A concurrent waiter on a different event loop sharing the same
refresh_keywould observe the phantom generation bump on its lock-re-acquire path, treat the storage as freshly-refreshed, and skip its own_run_refresh_cmd— proceeding tobuild_httpx_cookies_from_storageagainst STALE on-disk state.Bounded blast radius: the stale-storage caller will fail again on the next token fetch and trigger its own real refresh. No data corruption. But it's a wasted round-trip.
Suggested fix (from CodeRabbit on PR #814)
Track whether
inflightwas ever non-None during the cancel-recovery loop. Only bump generation ifsubprocess_exc is None AND observed_inflight is True. Or per the alternative suggestion: raise immediately wheninflight is Noneat the top of the cancel handler (the subprocess never registered, so we can't be the leader-cancel-before-await case for an actually-running refresh).Why not fixed in PR #814
That PR's scope was "extract verbatim from auth.py into _auth/refresh.py + facade-forward". Changing async semantics was explicitly out of scope (
must_not_do). Filing here for a dedicated follow-up.References
- CodeRabbit comment: https://github.com/teng-lin/notebooklm-py/pull/814#discussion_r3256142201
- Original code:
src/notebooklm/_auth/refresh.py:487-537(tier-10),src/notebooklm/auth.py:961-1026(tier-9 E origin)
Interface
No new interfaces are introduced.
- Caller invokes
- Assistant
I'll inspect the refresh coalescing code and its cancel-recovery path to implement the fix.
- Assistant
Implementing the fix: re-raise
CancelledErrorwhen no inflight future was registered, and adding a regression test. - Assistant
Verifying the new test fails without the fix:
- Assistant
The test reproduces the bug. Restoring the fix.
- Assistant
Fixed the early-cancel phantom generation bump in
_fetch_tokens_with_refresh's cancel-recovery loop.Fix (
src/notebooklm/_auth/refresh.py)Split the
inflight is Noneandinflight.done()branches in theCancelledErrorhandler:inflight is None— cancellation happened before_coalesced_run_refresh_cmdregistered an in-flight future, so no subprocess ran. Re-raiseCancelledErrorimmediately instead of falling through to the success path.inflight.done()— unchanged: inspect the settled future's terminal state as before.
This matches the CodeRabbit alternative from PR #814: if no future was ever registered, we can't be in the leader-cancel-during-subprocess case.
Regression test (
tests/integration/concurrency/test_refresh_cmd_race.py)Added
test_early_cancel_before_inflight_register_does_not_bump, which gates_coalesced_run_refresh_cmdbefore it can register a future, cancels the caller, and asserts_REFRESH_GENERATIONSstays at0. Confirmed it fails on the old code (generation bumps to1) and passes with the fix.All 5 tests in
test_refresh_cmd_race.pypass.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.
Identical first prompt
- composer-2.5Compare side by side →
- composer-2.5Compare side by side →
- composer-2.5Compare side by side →
- composer-2.5Compare side by side →
- gpt-5.6-solCompare side by side →
- gpt-5.6-solCompare side by side →
- gpt-5.6-solCompare side by side →
- gpt-5.6-solCompare side by side →
- gpt-5.6-solCompare side by side →
- fable-5Compare side by side →
- fable-5Compare side by side →