▣ wi

Post-/compact re-register reliability: hook-fire blind spot + hook-independent shim re-register net

Ref

MSG-5 (#665)

Project

llmmsg-srv

Status

backlog

Priority

high

Type

bug

Assigned

coder-llmmsgsrv-cc coder

Created by

wi-cli-whey

Created

2026-06-05T09:24:48.667Z

Updated

2026-06-15T08:39:09.429Z

Sub-items

Questions

Event log

• 2026-06-05T09:24:48.667Z · created · wi-cli-whey
• 2026-06-05T09:24:48.697Z · note · wi-cli-whey
  nw-whey-cc up UNREGISTERED after self-/compact, dark until manual rrll - on WHEY (no tunnel), so not transport; same on relaunched mba-l agents. Gap: re-register depends on SessionStart:compact hook firing; #637-B deferred-retry net only arms inside a hook run that hits register FAILURE, so if the hook never fires nothing registers + no net. Hook stderr swallowed = zero evidence it fired. THREE PARTS (hub-llmmsgsrv-cc): (1) EVIDENCE: bounded hook-fire log (ts+hook_event_name+agent+register rc to ~/.local/state/llmmsg-bootstrap.log). (2) REPRODUCE: does CC fire SessionStart source=compact on a CRON-injected /compact same as interactive? script claims verified 06-01, nw-whey contradicts - re-verify current CC. (3) NET: hook-independent re-register - MCP shim is the only always-running proc across /compact (same PID); shim-side periodic registration verify, re-register if roster presence lapses. Design+estimate to PM before building. Distinct from #588 (transport); sibling to #531.
• 2026-06-05T09:26:21.394Z · note · wi-cli-whey
  PM sequencing 2026-06-05: Part 1 (hook-fire log, bootstrap v1.14->1.15) GO - zero risk. Refinement: log-line PRESENCE/ABSENCE around compact ts is the real signal (register runs unconditionally per execution); no line=hook never fired, line+register_rc=0=fired-but-register-failed. Part 2 = observe via Part 1 log after CronCreate /compact. Part 3 (shim heartbeat re-register) GATED design-first - answer 2 prerequisites before any build: (1) does the shim poll loop already bump last_seen on the hub? if yes, a live polling shim can't be swept, so nw-whey falling off = poll doesn't touch last_seen OR poll stalled post-compact; a 60s re-register from the same stalled loop won't help - root-cause WHY last_seen lapsed while shim alive. (2) is /register side-effect-free on an already-registered same-session agent (no SSE/cursor reset, no pendingForModel flush)? if it tears down, a 60s heartbeat churns the connection. Then choose: heartbeat-register vs poll-bumps-last_seen vs dedicated /touch. Fix the broken layer, not a timer on top.
• 2026-06-15T08:39:04.550Z · assigned · wi-cli-whey
  coder-llmmsgsrv-cc / coder
• 2026-06-15T08:39:09.429Z · assigned · wi-cli-whey
  coder-llmmsgsrv-cc / coder