1 Commits

Include renames

Author	SHA1	Message	Date
3 people	2af41c561b	fix(plugins): make update scheduling atomic so update() cannot run twice at once (#437) ... * fix(plugins): make update scheduling atomic so update() cannot run twice at once Closes #401. `run_scheduled_updates()` decided whether to update a plugin with a check-then-act sequence: `can_execute()` and `set_state(RUNNING)` were separate calls with nothing between them, so two scheduler threads could both observe ENABLED and both go on to call the same plugin's `update()`. `update_all_plugins()` had the identical pattern. Two schedulers really do run at once. The render loop calls `_tick_plugin_updates()`, and Vegas mode fires its own `vegas-plugin-tick` daemon thread that is never joined when `VegasModeCoordinator.play()` returns — a slow `update()` still in flight overlaps the next tick from the main loop. A plugin running `update()` twice concurrently is unsafe unless it happens to be reentrant; shared mutable state, a non-thread-safe HTTP session or cache all break. The async path was already covered by the `_pending_lock` dedup in `_enqueue_update`, so the live exposure was the synchronous kill-switch path and `update_all_plugins()`. Both now claim the plugin through `_reserve_for_update()`, which holds one lock across the eligibility check, the due-time check and the RUNNING transition — and nothing more. Holding it across `execute_update()` would serialize slow plugins behind each other and reintroduce the render stall the async worker exists to avoid. The due-time check moved inside the lock deliberately. Left outside, a thread that had already decided "due" could claim the plugin the instant the winner finished, running `update()` twice within one interval. Two supporting changes fall out of it: - `_enqueue_update()` no longer sets RUNNING (the reservation did), and hands the reservation back if the pending-dedup ever fires. Otherwise a reserved-but-unqueued plugin would sit in RUNNING with nothing left to release it, and `can_execute()` would refuse it forever. - `_finish()` now clears the pending entry *before* flipping the state back to ENABLED. The old order left a window where a scheduler saw ENABLED, reserved the plugin, then had its enqueue silently dropped by the dedup — harmless as a missed tick before, a stuck plugin once a reservation is involved. Regression suite added and enrolled in CI, along with test_async_plugin_updates.py which was not previously run there. The overlap tests delay `can_execute()` to hold every thread inside the check-then-act gap: the real window is a couple of bytecodes wide, so a plain hammering test passes against the unfixed scheduler and proves nothing. With that delay the suite reports `update() ran 8x concurrently` on both affected paths before the fix, and passes after. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Udr6MfaFLUPhX5Fgo67Jf5 * test: fix a race in the reservation suite's own wait loops `test_async_path_never_overlaps` failed in CI with "update() ran 0x concurrently" — the test's bug, not the scheduler's. It polled `plugin._active` to wait for the update to finish, but before the worker picks the item up nothing is active yet, so the loop fell straight through and asserted on a plugin that had never run. Both async waits now key on `update_calls >= 1` as well, so they wait for an update to have started *and* finished. The stranded-state test gets the same guard for a second reason: ENABLED is also the starting state, so without it that assertion passes vacuously on a plugin that was never scheduled. Verified over 12 consecutive local runs, 12 passed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Udr6MfaFLUPhX5Fgo67Jf5 * fix(plugins): roll back the claim when dispatch fails _enqueue_update() reserved the plugin and added it to the pending set, then started the worker and queued the item. Thread.start() raises RuntimeError when the OS refuses a new thread — not hypothetical on a Pi under memory or thread pressure — and nothing is queued at that point to release the plugin. It stayed RUNNING with a stale pending entry, so can_execute() refused it for the rest of the process, and the exception escaped run_scheduled_updates() and skipped every remaining plugin in that tick. That is the same stranded-RUNNING failure the reservation was introduced to prevent, just reached through the dispatch rather than the dedup, so it is handled the same way: discard the pending entry, hand the reservation back, log the cause. Swallowed rather than raised so one plugin failing to queue cannot abort the others' turn. Both new tests fail against the un-rolled-back version — the second on the escaping RuntimeError itself — and pass with it. 74 tests across the reservation, async-update, plugin-system, health, Vegas-adapter and controller-toggle suites still pass. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01WexvwNDtWLVymGVqKD7BGk --------- Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>	2026-08-05 19:38:11 -04:00