Safe Device Control

Core Rule

Never execute risky control actions on physical devices without explicit go-ahead from Hobo. Risky = anything that could lock a device offline, change persistent settings, require physical intervention to reverse, or trigger firmware/account-level locks.

Pre-Action Checklist (Before Any Risky Control Command)

1. Read current state first — get full device state via the sensor endpoint
2. Classify the action into one of the risk tiers below
3. Surface the risk — tell Hobo what could go wrong
4. Get explicit go-ahead for write-risky and write-irreversible actions
5. Verify after — read state again, confirm intended change actually happened

Risk Tiers

Read-Only — no gate needed

• Get current state, list devices, read sensors, query API

Write-Low-Risk — execute, but verify state after

• Toggle Tapo plug on/off (state-verifiable via /plug/)
• Set fan_speed (queued, reversible, doesn't lock device)
• Set mode to a known preset (Auto/Strong/Sleep — reversible)
• Send learned IR code (reversible)

Write-Risky — require explicit go-ahead from Hobo

• Power toggle action on any cloud-controlled device (can trigger deep standby — see FP10 incident below)
• Factory reset, firmware update, account linking/unlinking
• Wipe learned IR codes, change device passwords
• First-time control of a device that hasn't been end-to-end-tested before
• Any action where the API returns code: 0 but device state didn't actually change (treat as suspicious)

Write-Irreversible — requires typed approval from Hobo

• Anything in AGENTS.md red lines (deleting files, configs, cron jobs)
• Anything destructive flagged in HEARTBEAT.md or LISTS.md
• Network/firewall changes, auth key rotations

Mistakes to Avoid (Lessons from FP10 Incident, 2026-07-17)

1. Don't blindly toggle power on a cloud-controlled device. The Dreame FP10 toggle action (siid=2, aiid=3) put the device into deep standby — went from online: True (with stale reads) to online: False (no reads at all). Required physical button press on the unit to recover.
2. code: 0 from cloud APIs is not confirmation. The Dreame cloud returned success for set_properties on power (siid=2, piid=1) when the device silently rejected it. The action toggle returned success and then bricked the connection. Cloud acceptance ≠ device execution.
3. Don't run untested control code on a real device. Test with dry_run=true, against a known-good baseline, or against a non-critical device first.
4. Don't assume state. Always read current state before controlling. If state is ambiguous (e.g., device shows online but power state is stuck, or returns stale values like fan_speed=8/10 while reporting power=2), stop and ask.
5. Don't chain control commands without per-step verification. Set one thing, verify it actually took effect, then proceed.
6. The "soft off" model matters. Some devices (Dreame FP10 included) have a soft-off state (e.g., Sleep mode + fan_speed=1) that keeps cloud alive, vs a deep standby (physical button press) that disconnects. Confirming which state we're in should be the first read before any control action.

Concrete API Guards (Already Implemented)

• POST /dreame/<name>/on requires {"force": true} in body — refuses without it (HTTP 403 with reason). Hobo must explicitly type the confirmation.
• POST /dreame/<name>/off — soft-off only (Sleep + fan_speed=1). No deep standby path.
• POST /dreame/<name>/fan_speed — value is clamped 1-10 (FP10 supports 10 manual speeds). No gate, since it's reversible.
• Future risky endpoints should follow the same force=true pattern.

Recovery Procedures

• Dreame FP10 deep standby: physical power button press on the unit (~3s hold). After wake, fan_speed queued value should be honored.
• Tapo plug offline: unplug 10s, replug, wait for DHCP lease + WiFi reconnect.
• Broadlink RM4 Mini timeout: power cycle the unit (unplug/replug).
• Anything else: check the device's official app first to confirm cloud-side state before re-trying from our API.

When to Use This Skill

Use this skill BEFORE any device control action that is:

• A new control path (first time we're sending this command to this device)
• Anything in the Write-Risky tier
• An action that has previously failed or behaved unexpectedly
• Part of a chained sequence where one link failed

Routine actions already validated end-to-end (e.g., turning the plant-light on/off, sending a learned IR button) don't need to re-trigger this skill every time — they're pre-approved by their existing log of successful runs.

Sources

• Original incident: memory/2026-07-17.md (21:11–21:49 SGT)
• MiOT spec verified from: https://github.com/CodyJon/dreame-ap10-integration
• Force-guard implementation: scripts/smart-home/app.py (dreame_on endpoint)
• HEARTBEAT.md: "Device Control — Safety Gate" section (loaded at every wake)