Can I take this video call from an off-grid homestead?
One Go binary that reads your solar/battery system (Solar Assistant over MQTT), your Starlink dish (local gRPC) and the weather (Open-Meteo), and answers three questions every few seconds:
- Can I take a video call now? — GO / MAYBE / NO, with the reasons.
- How many hours of workstation do I have at this SOC? — conservative (no sun) and forecast-integrated ("until 02:40"), with an honesty band.
- What should I turn off before the call? — ranked list with the hours each load buys you.
It publishes the answer over MQTT with Home Assistant discovery (zero-config entities), serves a phone-sized status page, and ships as a HA add-on, a systemd service, or a container. It never writes to your inverter.
$ deadzone check
{ "verdict": { "call_ok": "amber",
"call_reasons": ["uplink:drop_5m=0.031>0.02", "uplink:lat_p90=61ms<=150ms",
"power:usable=5222Wh>=call_budget=938Wh", "power:soc_trust=low"],
"hours_workstation": 7.0, "hours_workstation_with_pv": 11.8, "until": "2026-11-01T05:21:33Z",
"shed": [{"load":"water_heater","w":1200,"gain_h":3.1}] }, ... }
$ echo $?
1 # 0 = go, 1 = maybe, 2 = no
One click adds this repository to your Home Assistant; then install deadzone from the store. Or by hand:
- Settings → Apps → App store → ⋮ → Repositories (HA 2026.2 renamed
add-ons to apps; on older releases it is Settings → Add-ons → Add-on
store) → add
https://github.com/rafe-walker/deadzone→ close → find deadzone → Install. - Configuration tab: set MQTT host to your Solar Assistant Pi's IP (its broker must be started under SA → Configuration → MQTT), or leave it empty if SA bridges into your Mosquitto app. Set Latitude, Longitude, Battery capacity. Save.
- Start. Open the sidebar entry for the status page. Within a minute a
deadzone device appears under Settings → Devices (MQTT) with
sensor.deadzone_verdict,binary_sensor.deadzone_call_ok,sensor.deadzone_hours_at_current_loadand friends.
The app builds the Go binary on the Yellow the first time (a few minutes
on a CM4; a prebuilt image is coming, see Maintainer: publishing an
image below). Starlink needs 192.168.100.1 reachable from the HA box;
in bypass mode add a static route on your router.
On your phone: open http://<ha-ip>:8099/ (enable the port on the
app's Network section) and use "Add to Home Screen" — the page ships a web
manifest and icons, so it installs as a standalone app. The ingress URL
works too but its session expires.
# grab the release tarball for your arch (linux-arm64 for Pi 4/5, linux-amd64 for NUC)
tar xzf deadzone-v0.2.0-linux-arm64.tar.gz && cd deadzone-v0.2.0-linux-arm64
sudo install -m 755 deadzone /usr/local/bin/deadzone
sudo mkdir -p /etc/deadzone /var/lib/deadzone
sudo cp deadzone.example.yaml /etc/deadzone/deadzone.yaml # edit mqtt.host, forecast.latitude/longitude, battery.capacity_kwh
sudo useradd -r -s /usr/sbin/nologin deadzone; sudo chown deadzone /var/lib/deadzone
sudo cp deploy/deadzone.service /etc/systemd/system/ && sudo systemctl enable --now deadzone
curl -s localhost:8099/api/state | jq .verdictOr docker build -f deploy/Dockerfile -t deadzone . and run with
DEADZONE_* env vars (see below). From source: go build ./cmd/deadzone
(Go 1.24, no CGO).
deadzone run [--config FILE] [--interval 5s] [--log-level debug]
deadzone check [--config FILE] [--wait 15s] # one-shot JSON, exit 0/1/2
deadzone version
YAML file, every key overridable by a DEADZONE_* environment variable.
deadzone.example.yaml documents all of them. The ones
you must set or measure:
| Key | Env | What |
|---|---|---|
mqtt.host |
DEADZONE_MQTT_HOST |
Solar Assistant broker (its Pi's IP) or a broker SA bridges into |
mqtt.sa_prefix |
DEADZONE_SA_PREFIX |
solar_assistant unless you changed it |
starlink.addr |
DEADZONE_STARLINK_ADDR |
192.168.100.1:9200 |
forecast.latitude / longitude |
DEADZONE_LAT / DEADZONE_LON |
required when forecast is enabled |
forecast.tilt_deg / azimuth_deg / array_kwp |
DEADZONE_TILT_DEG / DEADZONE_AZIMUTH_DEG / DEADZONE_ARRAY_KWP |
array geometry; azimuth 0 = south, −90 = east, 90 = west |
battery.capacity_kwh |
DEADZONE_BATTERY_KWH |
packs × Ah × 51.2 V (two Eco-Worthy 100 Ah = 10.24) |
battery.reserve_soc_pct |
DEADZONE_RESERVE_SOC |
SOC floor, default 20 |
battery.parasitic_w |
DEADZONE_PARASITIC_W |
inverter idle draw not in SA's load number |
workstation.watts |
DEADZONE_WORKSTATION_W |
laptop + monitor + dish; used when the measured load is lower |
workstation.call_length_min |
DEADZONE_CALL_LENGTH_MIN |
power budget horizon (60) |
thresholds.* |
— | uplink cut-offs, stale window (30 s), call margin |
shed[] |
— | {name, watts, priority, control} loads you could turn off |
http.listen |
DEADZONE_HTTP_LISTEN |
:8099 |
Hours at current SOC (Q2). Usable energy is capacity × (SOC − reserve).
The load is a 10-minute moving average of Solar Assistant's total load, or
your workstation profile if the workstation is evidently off, plus the
inverters' idle draw. Divide energy by load and you get the conservative
number (no sun). For the optimistic number deadzone walks the Open-Meteo
15-minute irradiance forecast, converts it to watts with your array size
and a performance ratio it learns by comparing forecast to the PV you
actually produced, and keeps subtracting until the reserve is hit — that
gives a clock time. If your SOC comes from the inverter's voltage guess
(Solar Assistant "Use inverter values"), the estimate is shown as a band
and labelled low-trust; LiFePO4 voltage is flat between ~30 % and ~90 %
and a BMS read is the single biggest accuracy upgrade.
Can I take the call (Q1). Two gates, the verdict is the worse one.
Uplink: over the last five minutes of 1-second dish history, mean ping
drop ≤ 2 %, p90 latency ≤ 150 ms, jitter ≤ 30 ms and no outage ≥ 5 s is
green; ≤ 5 %, ≤ 250 ms, ≤ 50 ms and at most one such outage is amber;
worse is red. Dish not CONNECTED, thermal shutdown, or currently
obstructed is red. So are the account/policy states a pinging dish hides:
disablement_code other than OKAY (NO_ACTIVE_ACCOUNT,
DATA_OVERAGE_SANDBOX_POLICY, ROAM_RESTRICTED, ...), stowed,
overage_rate_limited (all from get_diagnostics, polled once a
minute) and an OVERAGE_LIMIT rate-limit reason in get_status; other
policy limits and a failed hardware self-test cap at amber. Dish heating,
"lower signal than predicted", or more than 2 mm of rain forecast in the
next two hours caps it at amber ("storm risk"). Power: red if SOC is
within 5 points of reserve or an inverter reports a fault; green if
usable energy covers (workstation + dish) × call length × 1.25;
otherwise amber. The dish watts are the 5-minute mean of
get_history.power_in (the dish reports no power in get_status), with
the newest sample and then the configured starlink.dish_watts as
fallbacks; the reason list says which (dish_w=62(dish_mean_5m)). If
either side has no data (SA silent for 30 s, dish unreachable) it reads
unknown and the published verdict is amber with the reason — deadzone
does not guess.
What to shed (Q3). Each configured load is ranked by watts ÷ priority. For each one, the hours gained are energy ÷ (load − watts) − energy ÷ load. The list grows until the runtime reaches your target
hours. v0.1 only recommends; your HA automation flips the switches.
Retained deadzone/state carries the whole JSON document (GET /api/state
serves the same plus history_15m, a 180-point 5-second downsample of the
dish's 1-s drop/p90 ring for the sparkline — HTTP only, to keep the
retained message small). Convenience topics: deadzone/verdict (green|amber|red),
deadzone/hours_at_soc, deadzone/shed_list, deadzone/verdict/call_reasons,
deadzone/verdict/hours_with_pv, deadzone/verdict/until,
deadzone/power/soc_trust, deadzone/uplink/starlink/score,
deadzone/forecast/pv_wh_next_6h, and deadzone/availability (LWT).
Discovery configs go to homeassistant/{sensor,binary_sensor}/deadzone/<key>/config
and all template the one state topic. sensor.deadzone_verdict carries
call_reasons, uplink_light, power_light, until and shed as
attributes; sensor.deadzone_starlink_power_w is the dish's own draw and
the diagnostic sensor.deadzone_starlink_disablement_code tells you why a
red verdict is an account problem rather than a link problem.
Or Settings → Automations & scenes → Blueprints → Import blueprint and
paste
https://github.com/rafe-walker/deadzone/blob/main/blueprints/automation/deadzone/call_ok_notify.yaml.
It triggers on binary_sensor.deadzone_call_ok after a settle time
(default 20 s, so a Starlink hiccup does not page you), restricts itself
to the hours you take calls, and sends the verdict, hours left and the
reasons through the notify action you pick (notify.mobile_app_<phone>).
deadzone only recommends (sensor.deadzone_shed_recommendation is the
ranked list, verdict.shed on the state topic has the watts and the hours
each load buys). The automation that flips the switch is yours, so the
"never writes to the inverter" promise stays with the binary:
alias: Shed water heater when the call verdict goes red
triggers:
- trigger: state
entity_id: sensor.deadzone_verdict
to: red
for: "00:00:30"
conditions:
- condition: template
value_template: "{{ 'water_heater' in states('sensor.deadzone_shed_recommendation') }}"
actions:
- action: switch.turn_off
target: { entity_id: switch.water_heater }
- action: notify.mobile_app_phone
data:
message: "Verdict NO: water heater off, {{ states('sensor.deadzone_hours_at_current_load') }} h at current load."Add a second automation that turns the load back on when the verdict has been green for 15 minutes, and give heaters/pumps a minimum off time so a flapping link cannot cycle them.
- It does not control anything. No
/setwrites to Solar Assistant, no switch flipping. Theactionsinterface exists; live actuation is v0.2 and opt-in. - It does not read your BMS. SOC is whatever Solar Assistant publishes. If that is the inverter's voltage estimate, deadzone says so and widens the band. For the real number read the pack directly — jbd-up does that for Eco-Worthy / JBD UP16S racks and feeds Solar Assistant or MQTT.
- It does not replace Solar Assistant. SA does the serial protocol work and owns the inverter port; deadzone is a consumer of its MQTT output. Direct PI30 serial is on the roadmap as a replacement source for people without SA, not an addition.
- It does not know about cellular failover, EcoFlow units or an active probe to the internet yet. See ROADMAP.md.
- It does not ask the dish for its location.
get_locationneeds "Allow access on local network" in the Starlink app and is off on most residential plans;get_diagnostics.location.enabledis published so you can see the setting, and the forecast uses the coordinates you configure. - It does not predict Starlink rain fade with any proven lead time. The storm cap is a prior from the forecast, not a measurement; the SIGSPATIAL '25 terminal study that motivated it found correlation, not a predictor.
- Thresholds are opinions, tuned for WebRTC-style calls. Change them in
thresholds.uplink.
go test ./... # ~150 tests incl. subtests, no network
go vet ./...
make cross # linux/amd64, linux/arm64, darwin/arm64 into dist/Fixtures under internal/source/starlink/testdata follow the dish's
reflection output with original field names (get_status,
get_history, get_diagnostics); the diagnostics shapes come from the
DishGetDiagnosticsResponse message (disablement_code = 6, stowed =
10, overage_rate_limited = 14) in SpaceX's published device proto.
The app builds on the device by default (addon/deadzone/build.yaml,
DEADZONE_REF pinned to the release tag). .github/workflows/addon-image.yml
builds the same Dockerfile with the composable home-assistant/builder
actions on every v* tag and pushes
ghcr.io/rafe-walker/aarch64-deadzone:<version>
ghcr.io/rafe-walker/amd64-deadzone:<version>
ghcr.io/rafe-walker/deadzone:<version> # multi-arch manifest
where <version> is the tag without the v — the string the Supervisor
pulls when config.yaml names an image. Switching users over to the
prebuilt image is a one-time, ordered procedure, because the Supervisor
pulls image:version the moment it sees a new version:
- Push the tag (
vX.Y.Z); wait for the addon-image workflow to go green (the per-arch jobs run natively onubuntu-24.04/ubuntu-24.04-arm). - Make the three packages public: GitHub → your profile → Packages →
deadzone,aarch64-deadzone,amd64-deadzone→ Package settings → Danger zone → Change visibility → Public. A private package fails the install with "pull access denied". Do this once; later pushes inherit it. - Verify from any machine:
docker pull ghcr.io/rafe-walker/deadzone:X.Y.Z. - In
addon/deadzone/config.yamluncommentimage: "ghcr.io/rafe-walker/deadzone"(keepbuild.yaml; the Supervisor ignores it onceimageis set anddocker buildstill works). Commit onmain. From then on, for every release: tag first, wait for the image, then bumpversioninconfig.yamlin a separate commit — never the other way round, and never re-push a different image under an existing tag. workflow_dispatchon the same workflow rebuilds any tag by hand (ref: vX.Y.Z,publish: true).
All fixtures are synthetic and labelled as such; a real
mosquitto_sub -v -t 'solar_assistant/#' dump from your bench makes a
better one — please contribute it (scrubbed). See
CONTRIBUTING.md.
Apache-2.0. Starlink is a trademark of SpaceX; Solar Assistant is a trademark of its owners; deadzone is affiliated with neither.
Built and maintained by IsoKron.