perf: stop three loops billing by the invocation

bin/iak-mcp-daemon.mjs

@@ -84,16 +84,32 @@ if (!apiKey) {
 } else if (!room) {
   console.warn('[iak-mcp-daemon] mcp.confirmations.room missing — chat-reply poller disabled');
 } else {
+  // Poll cadence is a COST decision, not a detail: at the old hard-wired 5000 ms
+  // this one poller made 17,280 requests a day per device, and the host bills per
+  // invocation (2026-09-12: the fleet's Vercel bill passed 200 USD/month, ~99% of
+  // it request volume). The adjacent poller.interval_sec key already existed in
+  // config and was silently ignored here. Default is unchanged so nobody's
+  // latency moves without them asking.
+  // Two different jobs share this one loop, and they want opposite things.
+  // Reading the room is paid for per request. Carrying petrus's /approve tap is
+  // the only interval in the fleet a HUMAN feels: 30 seconds after tapping
+  // Approve reads as broken, not thrifty (claudeMB on PR #101). So the gate gets
+  // its own key and falls back to the cheap one, then to the old default — which
+  // means nobody's latency moves unless they set something.
+  const pollerIntervalMs = Math.max(1000, Number(
+    config?.mcp?.confirmations?.interval_sec
+    ?? config?.poller?.interval_sec
+    ?? 5) * 1000);
   startChatReplyPoller({
-    apiKey, room, intervalMs: 5000,
+    apiKey, room, intervalMs: pollerIntervalMs,
     // Exact owner identities allowed to settle intents (config, with the
     // fleet's known surfaces as the default). Every surface the owner taps
     // from must be listed — an unlisted one gets a VISIBLE rejection reply,
     // never silence.
     owners: config?.mcp?.confirmations?.owners || ['petrus', 'petrus-boox'],
     log: (msg) => console.log(`[iak-mcp-daemon] ${msg}`),
   });
-  console.log(`[iak-mcp-daemon] chat-reply poller watching room "${room}" every 5s`);
+  console.log(`[iak-mcp-daemon] chat-reply poller watching room "${room}" every ${pollerIntervalMs / 1000}s`);
   // Mirror every intent/action transition to the central action_status store
   // (antfarm PR #43) so CodeWatch renders durable button state off-LAN.
   const pushBase = config?.groupmind?.base_url || config?.groupmind?.baseUrl || 'https://groupmind.one/api/v1';

scripts/claudemm-uik-daemon.sh

@@ -18,6 +18,11 @@ export INTENT_USER_ID="$(python3 -c "import json;print(json.load(open('$CONFIG')
 export INTENT_AGENT_HANDLE="@claudemm"
 export INTENT_DEVICE_ID="mac-mini"
 
+# 2026-09-12: uik-daemon defaults to POLL_INTERVAL_MS=30000, i.e. 2,880 requests/day
+# per device. With the fleet on the same cadence Vercel billing passed 200 USD/month.
+# 120s still detects a dead agent inside two minutes, at a quarter of the traffic.
+export POLL_INTERVAL_MS="${POLL_INTERVAL_MS:-120000}"
+
 # launchd starts this with a bare PATH (no Homebrew): resolve node explicitly
 # or the job dies with "exec: node: not found" - which is what kept the
 # publisher an orphan started by hand instead of a KeepAlive service.

scripts/mini-vitals.sh

@@ -33,5 +33,9 @@ print(f'{total:.1f} {avail:.1f}')")"
     mem_total_gb="$MEMT" mem_available_gb="$MEMA" \
     memory="${MEMT}GB total, ${MEMA}GB available" \
     --config "$CONFIG" >/dev/null 2>&1 || true
-  sleep 30
+  # 2026-09-12: was 30s = 2,880 API writes/day from this loop alone, and Vercel
+  # billing passed 200 USD/month on fleet request volume (petrus). Load and free
+  # memory in a dashboard row do not need 30-second freshness; 5 minutes keeps the
+  # row visibly live at a tenth of the requests.
+  sleep "${VITALS_INTERVAL_S:-300}"
 done

src/confirmations.mjs

@@ -1102,6 +1102,11 @@ export function composeAnnouncers(map) {
 // Ported from the Mini's field-hardened fork (branch mini-local-fork-rescue),
 // security-reviewed by codexmb 2026-08-27; `owner` kept as an alias so
 // existing call sites keep working.
+// NOTE on `intervalMs`: this default is a COST decision as much as a latency one.
+// The host bills per request, so 5000 ms is 17,280 requests/day/device for this
+// poller alone — the single largest source of our 2026-09 hosting bill. It is kept
+// at 5000 for backward compatibility, but callers should pass a value explicitly;
+// iak-mcp-daemon.mjs derives one from mcp.confirmations.interval_sec.
 export function startChatReplyPoller({ apiKey, room, intervalMs = 5000, log, owners, owner = 'petrus' }) {
   if (!apiKey || !room) {
     process.stderr.write('[iak-mcp] chat-reply poller: missing apiKey or room — disabled\n');