# Next Session: Router Alerts — Live Validation + Deferred Items

## What Was Done (Session 388)

### All 3 Plan Items Implemented + Deployed
1. **model_tier default "nano" → "super"**: 2 code defaults + 5 YAMLs updated. 2 agents explicit `claude` (flight-search, orchestration-research), 10 on `super` (default). Nano fallback via `_get_client_for_tier()`.
2. **Telegram network query bypass (~7s savings)**: `classify_network_query()` in `router_monitor.py` + bypass in `text_llm.py` stream_chat round 0. 20 new tests. Device references fall through to LLM.
3. **E2E alert test**: cpu_temp alert fires at 77.4°C with 70°C threshold. Verified.

### Test Results
2436 passed, 0 failed. 2 commits pushed (`69a78f3`, `297eb3c`). Deployed to Titan, Annie restarted.

## What Needs Doing

### Priority 1: Live Telegram Timing Test
The bypass saves ~7s in theory (skips LLM call #1). Validate with real Telegram messages:
- Send "how's my internet?" — should respond in ~8-10s (was 15-17s)
- Send "who's on my wifi?" — should also be ~8-10s
- Send "show bandwidth for Titan" — should fall through to LLM (device reference), ~15s
- Compare timestamps to confirm speedup

### Priority 2: Monitor Scheduled Agents
- Check if anomaly detection ran (every 6h) — look at agent logs
- Check if daily report fired (10 PM IST) — look at Telegram
- Verify agents are actually hitting Beast (Super) now, not Nano
  ```bash
  ssh titan "grep 'model_used' ~/.her-os/annie/agent_runs/*.json 2>/dev/null | tail -10"
  ```

### Priority 3: Deferred Items
- **SMS sanitization**: `send_sms_alert()` passes unsanitized text to ADB shell command. Potential injection vector.
- **Session-per-query refactor**: Router monitoring creates a new `aiohttp` session per `router_status()` call. Should reuse sessions.
- **Reconnect downstairs devices**: New WiFi password `mango-tiger-92-pixel` for devices on the 3rd router.

## Key Files
| File | What |
|------|------|
| `services/annie-voice/router_monitor.py` | `classify_network_query()` at line ~862 |
| `services/annie-voice/text_llm.py` | Bypass logic at line ~1289 |
| `services/annie-voice/agent_discovery.py:72` | Default `model_tier: "super"` |
| `services/annie-voice/agent_context.py:924` | Scheduler default `model_tier: "super"` |

## First Commands
```bash
# Check Annie is running
ssh titan "pgrep -af server.py && tail -3 /tmp/annie-voice.log"

# Check agent model usage
ssh titan "grep -r model_used ~/.her-os/annie/agent_runs/ 2>/dev/null | tail -5"

# Check collector health
ssh titan "tail -3 ~/.her-os/annie/router/collector.log"
```

## Prompt
Continue router alerts from `docs/NEXT-SESSION-ROUTER-ALERTS-3.md`. Session 388 implemented the 3-item plan (model_tier upgrade, Telegram bypass, E2E alert test). Validate the bypass live via Telegram, check scheduled agents are using Super on Beast, and tackle deferred items if time allows.