Measure: end-to-end latency idle vs busy receiver, native vs hub, with LLMMSG_POLL_MS=1000 as a third arm

Sub-tickets

Activity

• wi-cli-venus created · 7w ago
  parent=#2850
• wi-cli-venus completed · 7w ago
  MEASURED. Busy receiver (coder in 90s blocking tool): hub probe enqueued 05:48:03.319, native 05:48:03.855 (both <1s after send); BOTH held in queue and delivered at the same boundary 05:49:28.300 = first model turn after the tool returned. Busy case: no transport difference. POLL_MS arm (standalone shim sm-polltest, 10 probes each, send->notification): 3000ms med 2.30s (min 2.26 max 3.16, n=8); 1000ms med 0.30s (min 0.27, n=9, one 14s outlier = cursor warmup). Native idle was 0.09s (SENDMESS-2). POLL_MS=1000 closes the idle gap to ~0.2s.
• wi-cli-venus commented · 7w ago
  CORRECTION (Elazar challenge, code+docs re-read): the POLL_MS arm measured shim stdout emission only, not model-visible delivery. (a) -cc: delivery = shim poll (0..POLL_MS) + CC channel injection (idle: new turn; busy: queued to next turn per channels-reference; silent-drop class exists: CC issues #61797/#38736). SENDMESS-2's 1.71s already included injection, so idle-cc is poll-bound, but '~0.2s at 1000ms' was extrapolated, not measured end-to-end. (b) -ca (Codex): POLL_MS is irrelevant - Codex has no channels, the shim notification goes nowhere; delivery = hub tmux nudge 2s after landing (venus-local, scheduleCcNudge gated endsWith('-ca')) -> codex turn -> read_unread; off-venus -ca = context-monitor --drain poke, 2-min timer + >=2-tick persistence + 180s cooldown = minutes. (c) native SendMessage is CC-only (UDS socket per CC pid) - cannot reach -ca at all. (d) POLL_MS=1000 fleet-wide = 3x hub GET /unread load for a cc-idle-only gain.