Diagnostics

GPS (T-Deck Plus only) 

CMD> gps on    # start GPS background task with live status display
CMD> gps off   # stop GPS background task
CMD> gps test  # one-shot GPS coordinate read + display

GPS status bar icon: grey = off, yellow = searching, green = fixed.

Cold start takes ~4 minutes outdoors. Run gps on before trackme to pre-warm the fix.

test / tst — Hardware Self-Tests

The hardware tests are sub-commands of one test command (formerly the standalone spktest / mictest / loratest):

CMD> test spk       # I2S speaker
CMD> test mic       # ES7210 microphone
CMD> test lora      # LoRa SX1262
CMD> test touch     # GT911 touch (T-Deck / T-Deck-Plus)
CMD> test keydump   # raw byte from the I2C keyboard MCU (T-Deck / T-Deck-Plus)

test spk — Speaker Test 

CMD> test spk

I2S speaker hardware verification and notification level test.

Key Action
1–6 Raw tones (220–4000 Hz) — bypass NotificationManager
s Play C major scale
a / w / c / i / p Trigger ALERT / WARNING / SUCCESS / INFO / PING notification levels
q Quit

Keys a/w/c/i/p go through NotificationManager and play the configured WAV if one is set (respecting each level’s on/off toggle). Keys 1–6 bypass volume settings and play raw tones directly — use for hardware verification. To force-play a notification sound regardless of toggle, use nf test.

test mic — Microphone Test 

CMD> test mic

ES7210 microphone hardware verification (the mic is on both T-Deck and T-Deck Plus — only the GPS is Plus-exclusive).

  • Live level meter — peak-hold bar + scrolling bar-graph that react to sound.
  • Voice-activity detection — *** VOICE DETECTED *** flashes when speech is present (debounced).
  • Record + replay — capture 3 s to PSRAM and play it back through the speaker to confirm the full capture→playback path.

The capture de-duplicates the ES7210’s ALL_LEFT stream (which delivers two identical samples per audio sample) so playback is true 16 kHz mono — correct pitch and duration.

Key Action
r Record 3 s
p Replay the recording
+ / - Mic gain
q Quit

test lora — LoRa Test 

CMD> test lora

Initializes the LoRa SX1262, runs a TX test, then enters RX monitor mode. Press q to stop.

test keydump — Keyboard Raw-Byte Dump (T-Deck / T-Deck-Plus only) 

CMD> test keydump

Reads one byte at a time from the LilyGo I2C keyboard MCU (0x55) and prints it as 0xHH 'c'. Press q to exit.

Use it to discover what an unmapped Sym / Alt combo actually sends. The keyboard firmware resolves modifier combos internally and hands the ESP32 one already-resolved byte — Al-Anqa can only intercept bytes it can see. If you find a key or combo that types nothing but returns a specific non-zero byte in the dump, that byte can be remapped in boardReadKey() (see Keyboard Reference — SYM+0 → = was discovered exactly this way).

Not available on T-Pager (it has its own test keymap for the same purpose, tailored to the TCA8418 driver which does key resolution in our own code).

csidetect / csi — WiFi CSI Motion Detector [EXPERIMENTAL]

Not an actual radar. The circular sweep display is a style choice only — the device cannot locate, count, or aim at people. It detects motion energy from WiFi disturbance, nothing spatial. See “Honest limits” below.

CMD> cw <ssid> <password>   # connected mode: join a network first
CMD> csi                    # ...then open the detector (cleaner signal)

CMD> csi auto               # passive mode: NO router join needed

Two modes:

  • csi — uses the network you’re connected to (run cw first). A busy link = more frames = snappier, cleaner detection.
  • csi auto — scans, locks onto the strongest nearby AP’s beacons, and senses without associating to anything. Works against any router. Trade-off: beacon-rate is lower than a connected link, so it’s a touch laggier. It locks to a single AP on purpose (mixing multiple transmitters would read as fake motion).

A single-chip WiFi motion detector — believed to be the first WiFi-CSI sensing feature on a T-Deck. Every WiFi frame the radio receives carries Channel State Information (amplitude/phase of the OFDM subcarriers). A moving body changes the room’s multipath, so the CSI variance rises. csi tracks that, self-calibrates a floor/ceiling, and turns it into a live presence + motion readout on a double-buffered sweep-style display. No camera, no PIR, no extra hardware — it rides the WiFi your environment is already transmitting.

Reading the screen

Element Meaning
CLEAR (green centre) No movement detected
CONTACT (red centre, pulses) Movement near the device
MOTION % How much movement right now
Activity (STILL/FIDGET/WALK/RUN) Rough magnitude class
Blips / sectors Decorative. Blip intensity reflects real signal (how much a subcarrier band reacted), but the position/angle on the dial is arbitrary — the 8 sectors are mapped to fixed angles, NOT directions. A blip at “top” does not mean motion is in front of you.
Sweep cone Decorative only — a time-based animation, not scanning or aiming at anything.
AUTO c<ch> <ssid> / LINK c<ch> <ssid> Source AP: mode, channel + network name (truncated; full BSSID/SSID in the CSV log)
NBVI on/off Smart subcarrier weighting state (toggle with n)
fr:<n> + CSI live Bring-up health: CSI frames received + status

Keys

Key Action
a / l / trackball Sensitivity (− / +). In adaptive mode these nudge the noise-floor margin.
c Recalibrate — re-baseline the room (also resets the noise floor)
t Adaptive threshold on/off — auto-tracks the quiet-room noise floor
n NBVI smart-subcarrier weighting on/off (A/B test)
s SD logging on/off → /apps/csidetect/NNN.csv
h In-app help overlay
q Quit

Signal processing (what makes it more precise)

  • NBVI subcarrier weighting — instead of averaging all ~56 subcarriers equally, each subcarrier is weighted by its normalized variance, so the ones actually reacting to movement dominate and static ones contribute ~nothing. Auto-learned (no manual calibration); toggle n to compare on/off.
  • Hampel filter — a 7-sample median/MAD outlier reject on the motion signal, so a single RF glitch can’t false-trip CONTACT.
  • Adaptive threshold (t) — instead of a fixed sensitivity, the trip level sits a small margin above the continuously-learned quiet-room noise floor. When the source gets noisier (e.g. csi auto’s sparse beacons), the bar auto-raises to match → fewer false CONTACTs. a/l adjust the margin; the panel shows THR auto.
  • Single-source lock (auto mode) — only one AP’s frames feed the detector, avoiding false motion from channel traffic switching between transmitters.

SD logging (s)

Press s to record presence transitions to /apps/csidetect/NNN.csv (sequential, never overwritten — like wguard/bmon). A row is written each time the state flips CLEAR↔CONTACT, so the file stays small and answers “when was there motion.” Columns:

time,event,motion_pct,thresh_pct,zones,mode,channel,bssid,ssid

time is a real date-time when GPS/NTP is available, otherwise an @<uptime>ms counter (same convention as the other apps). mode is AUTO/LINK. channel/bssid/ssid identify the AP the motion was sensed off — the full BSSID (e.g. A1:B2:C3:D4:7F:78) and its network name (ssid, commas stripped for CSV; (hidden) if none). The panel shows L<n> next to fr: while recording. Writes are GDMA-guarded (promiscuous is paused for each write), so logging is safe even though CSI keeps the radio busy.

How to use it

  1. Either connect with cw then run csi, or just run csi auto (no join — it picks the strongest AP itself). A busier source gives more frames and snappier detection.
  2. Leave it still for ~10–15 s so it learns the empty room → CLEAR.
  3. Move or walk near it → CONTACT + the MOTION bar rises.
  4. Best confidence check: leave the room → CLEAR; walk back → CONTACT, repeated.

Honest limits

  • The radar layout is decorative. The circular dial, the sweeping cone, and the angle/position of every blip are a visual style only — not a map of where motion is. Only two things carry real meaning: the centre (CLEAR/CONTACT) and the MOTION % bar. A blip’s brightness ≈ how much a subcarrier band moved; its place on the dial is arbitrary.
  • Single antenna = one motion-energy signal. There is no true direction, no person count, and no localization — in either csi or csi auto. The 8 sectors are independent signal measurements spread around the dial for readability, not a compass bearing. (Reference projects either randomize the angle or use a multi-node mesh + ML to get real position; one T-Deck can’t.)
  • csi auto looks noisier than csi — passive beacon-rate (~10 frames/s) is far sparser than a connected link, so the signal jitters more. Press c to recalibrate while still, lower sensitivity with a, or use connected csi for the cleanest read.
  • Motion only — a perfectly still person can read CLEAR.
  • Environment-dependent — needs WiFi traffic, benefits from c recalibration when you change rooms. Best treated as a covert motion/occupancy indicator, not a precision sensor.

Adapted from the single-device CSI path of skizzophrenic/Cardputer-CSI-Human-Detector (MIT); subcarrier-band idea from ruvnet/ruview; the NBVI subcarrier weighting, Hampel filter, and passive operation are methodology from ESPectre (GPL — no code copied). See NOTICES.