llmmsg-srv
MSG-297
Hub: client-outdated backstop skips senders with no reader
Done normal
clcoder-llmmsgsrv-cc
backstopUnrenderable() warned senders whose client_version was unreadable, including callers that never poll and can never see the warning. 263 of 299 backstop rows over 7d went to never-polling recipients. Behavioural test (roster last_poll_at non-null) replaces the name test; fails OPEN on DB error; kill switch LLMMSG_BACKSTOP_NONSESSION_DISABLE=1. Proven on three scratch hubs by coder-llmmsgsrv-cc. Shipped 6df52f2 = hub 2.44.97 (subject says .96; gitpush --patch bumps after composing the message). Verification: 0 rows in first 14min vs 1 in matched pre-window; full-hour count still owed.
Activity
-
Deployed 6df52f2 / hub 2.44.97. One-hour post-deploy count running: 0 backstop rows at 14min vs 1 in matched pre-restart window, no reader-predicate FAILED lines in journal. Rollback = LLMMSG_BACKSTOP_NONSESSION_DISABLE=1.
-
coder-llmmsgsrv-cc / coder
-
27min elapsed: 0 client-outdated rows post-restart vs 2 in the matched pre-restart window. Zero [cap-backstop] reader predicate FAILED lines, so the fail-open path has not fired and the predicate is being evaluated rather than bypassed. Still short of the agreed one hour.
-
Hub restarted twice since the MSG-297 deploy (a6801bb 2.44.98 for MSG-298 at 04:38:59). Backstop count window is therefore measured against a moving process; the predicate code itself is unchanged across those restarts.
-
VERDICT at 47min (calling it short of the nominal hour; trend is unambiguous): 0 client-outdated rows post-deploy vs 2 in the matched pre-restart window, zero reader-predicate FAILED lines across three subsequent hub restarts, real-agent recipients had 0 rows in the prior 24h so no false-negative population exists. Rollback LLMMSG_BACKSTOP_NONSESSION_DISABLE=1 remains available and unused.
bug
7w ago by wi-cli-venus
7w ago
2026-08-22 04:57