@agentsignalnet/claude-plugin 0.2.104 → 0.2.105
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.claude-plugin/plugin.json +1 -1
- package/package.json +1 -1
- package/scripts/inbox-daemon.mjs +22 -10
package/package.json
CHANGED
package/scripts/inbox-daemon.mjs
CHANGED
|
@@ -1475,16 +1475,6 @@ for (const identity of identities()) {
|
|
|
1475
1475
|
*/
|
|
1476
1476
|
async function loop() {
|
|
1477
1477
|
for (;;) {
|
|
1478
|
-
// Reconcile on every cycle, not only at startup. Registering and looking
|
|
1479
|
-
// are two steps, so a one-shot check at startup loses a race it is
|
|
1480
|
-
// supposed to settle: if the higher pid registers first and looks before
|
|
1481
|
-
// the lower one has registered, it sees an empty registry and stays -- and
|
|
1482
|
-
// the lower pid, looking afterwards, is the one entitled to stay. Both
|
|
1483
|
-
// keep watching, which is the exact duplicate this check exists to
|
|
1484
|
-
// prevent. Looking again each cycle converges whatever the order was,
|
|
1485
|
-
// because the late registration is visible by the next one.
|
|
1486
|
-
if (yieldToLowerPid(standDown)) return;
|
|
1487
|
-
|
|
1488
1478
|
const started = Date.now();
|
|
1489
1479
|
try {
|
|
1490
1480
|
await tick();
|
|
@@ -1500,4 +1490,26 @@ async function loop() {
|
|
|
1500
1490
|
}
|
|
1501
1491
|
}
|
|
1502
1492
|
|
|
1493
|
+
/*
|
|
1494
|
+
THE BACKSTOP GETS ITS OWN CLOCK, because the poll no longer has one.
|
|
1495
|
+
|
|
1496
|
+
Reconciling used to happen at the top of every cycle, and a cycle was two
|
|
1497
|
+
seconds, so a duplicate watcher stood down about that fast. Then the poll
|
|
1498
|
+
started holding its request open for twenty-five seconds -- and the check
|
|
1499
|
+
went with it, because it was riding the cycle rather than a timer. A
|
|
1500
|
+
duplicate that appeared just after one check now survived until the next,
|
|
1501
|
+
twelve times longer than before, delivering every message twice the whole
|
|
1502
|
+
while.
|
|
1503
|
+
|
|
1504
|
+
Neither change was wrong and neither test could see it: the long-poll tests
|
|
1505
|
+
assert request counts, the two-watcher tests start daemons and wait. The
|
|
1506
|
+
regression lived in the gap between two commits that were correct apart.
|
|
1507
|
+
|
|
1508
|
+
So it runs on `intervalMs` regardless of what the poll is doing. It is cheap
|
|
1509
|
+
-- a directory read and a pid check -- and it is the only thing standing
|
|
1510
|
+
between a failed lock and two watchers, which is not a thing to make twelve
|
|
1511
|
+
times slower as a side effect.
|
|
1512
|
+
*/
|
|
1513
|
+
setInterval(() => yieldToLowerPid(standDown), intervalMs);
|
|
1514
|
+
|
|
1503
1515
|
await loop();
|