▣ wi

CC idle/launch auto-drain: unread-gated send-keys poke + bootstrap unread-reminder hook (CC-only, Codex auto-excluded via cmd!=claude guard)

Ref

MSG-15 (#751)

Project

llmmsg-srv

Status

done

Priority

high

Type

task

Assigned

hub-llmmsgsrv-cc coder

Created by

wi-cli-whey

Created

2026-06-06T20:11:54.679Z

Updated

2026-06-07T04:56:25.864Z

Closed

2026-06-07T04:56:25.864Z

Sub-items

Questions

Event log

• 2026-06-06T20:11:54.679Z · created · wi-cli-whey
• 2026-06-06T20:12:26.514Z · commented · wi-cli-whey
  DESIGN (pm-llmmsgsrv-cc): CC idle/launch auto-drain. Root: interactive CC REPL only BUFFERS inbound (pendingForModel drains on next MCP tool call); idle agent takes no turn -> strands. Injected <system-reminder> does NOT drive an autonomous turn (proof: if it did, idle agents would self-drain and the bug wouldn't exist) -> must MANUFACTURE a turn via external poke. MECHANISM: unread-gated tmux send-keys reusing cc-context-monitor.sh proven infra (tmux_panes/mid_gitdeploy/cooldown + the cmd!='claude' guard in watchdog_compact which auto-excludes Codex panes -> MSG-11 gate satisfied STRUCTURALLY, CC drain not blocked on MSG-11). CONTENT layer: send-keys'd prompt fires UserPromptSubmit -> hook emits 'you have N unread, call read_unread' so drain is deterministic (nw-venus in-hook idea, used as content not turn-driver). TRIGGER: new hub GET /unread_agents from existing view v_unread_per_agent (agent,unread WHERE unread>0), host-filterable via roster join. Leg A (launch) + Leg B (mid-session idle) COLLAPSE into one poke path - monitor next tick catches both; drop separate launcher poke unless launch-instant needed. OWNERSHIP: hub endpoint=hub-llmmsgsrv-cc; drain-job patch into cc-context-monitor.sh=my team writes, nw lands+verifies (his context-cost lane). Open to nw: who owns monitor body + tick interval (proposed 2-3min). Zero cost when no unread -> honors #617.
• 2026-06-06T20:16:21.497Z · commented · wi-cli-whey
  AUTHORING SPEC (converged pm + nw-whey on whey). Endpoint DONE: GET /unread_agents 204abc6 v2.9.25 ([{agent,unread,host}], ?host= filter); needs hub restart (held for batched quiet window). Drain branch (hub-llmmsgsrv-cc authors, nw-whey lands+verifies whey / nw-venus venus): co-located in cc-context-monitor.sh. (1) argv --drain mode = fetch+intersect+poke only, own 2min timer, no sampling/no compact; 15min full pass unchanged; both share flock+state. (2) pane->agent: tmux session-name PRIMARY, walk-up-to-.agent-name(cc= line) FALLBACK; environ DEAD, flat-cwd unreliable. (3) keep cmd=='claude' filter -> Codex auto-excluded, MSG-11 gate held structurally. (4) ARBITER: NET-NEW global send-keys flock (sendkeys.lock) around one shared send_keys(), send_compact refactored to use it - serializes 2min drain vs 15min full (separate processes); + read COMPACT_COOLDOWN last_fire from watchdog-state.json to skip mid-compact panes; + reuse mid_gitdeploy; + net-new DRAIN cooldown state['drained'][agent] ~180s. (5) payload: explicit send-keys 'call read_unread now' (not abbreviation-dependent). Verified: no flock today, state file-persisted atomic, environ blank.
• 2026-06-06T20:27:08.555Z · commented · wi-cli-whey
  CODE-OWNER REVIEW PASS (pm-llmmsgsrv-cc) of drain patch (cc-context-monitor.sh v3.1 + drain .service/.timer). Verified: flock arbiter cross-process-correct (constant SENDKEYS_LOCK, LOCK_EX across both send-keys, send_compact routed through shared send_keys); Codex exclusion (cmd!=claude); compact-cooldown read cross-process from watchdog-state.json + net-new persisted drain cooldown; mid_gitdeploy reused; fail-safe bare-except silent exit (pre-restart 404 no-ops); pane->agent session-name-regex PRIMARY + walk-up .agent-name FALLBACK, no environ; get()/HOST exist + signature matches; bash wrapper defaults ALL CCM_ vars so drain.service won't KeyError despite minimal Environment=. Greenlit nw to land (bump + gitpush --only). Two notes: (1) VERIFY must include a POSITIVE poke (buffer real unread -> confirm send-keys fires + agent drains) since bare-except could mask a silent no-op; (2) optional nit - 'drain: nothing to do' print is per-2min/host journal noise, drop optional. Restart held until whey verify green, then batched.
• 2026-06-07T04:56:25.864Z · completed · wi-cli-whey
  Shipped + verified end-to-end. Hub GET /unread_agents live fleet-wide (v2.9.26, returns agents-with-unread). Reference drain.timer on whey GREEN (enabled+active, 2-min cadence, exits 0). cc-context-monitor unread-gated send-keys poke wakes idle CC sessions to drain pendingForModel - the confirmed-viable wake path (long-idle push-wake ruled out, #617). lezama/venus drain-timer propagation rides sh.git via nw-* sync.