Add Xingzhi Cube 1.83" TFT WiFi (2mic) board support

[nguyenduchoai]

Summary

Adds a third reference port: Xingzhi Cube 1.83" TFT WiFi (2-mic variant) — ESP32-S3-N16R8 + 284×240 landscape display + 3 physical buttons.

Despite "ST7789" on every forum/silkscreen reference for this board, the panel is actually NV3023. Both Arduino_GFX and TFT_eSPI ST7789 init sequences silently fail — the panel needs a specific ~30-command init that lives in the xiaozhi-esphome project. This port transcribes that sequence into a small hand-rolled SPI driver (~190 lines, no GFX library dependency for this board).

Two commits

1. hal: add display_hal_lv_color_format() — new HAL function so boards whose MIPI-SPI controller wants pixels MSB-first (like NV3023) can ask LVGL to write swapped bytes directly, avoiding a per-board CPU byte-swap in draw_bitmap. Existing ports return LV_COLOR_FORMAT_RGB565 and behave identically.
2. Add Xingzhi Cube 1.83" TFT WiFi (2mic) board support — the actual board folder + platformio env + CLAUDE.md entry.

Board hardware summary

• MCU: ESP32-S3-N16R8 (16 MB flash + 8 MB OPI PSRAM)
• Display: NV3023 at 284×240 landscape, CASET column offset 36, MADCTL 0xA0 (no BGR bit), no INVON
• Backlight: GPIO13 via LEDC PWM (plain digitalWrite does NOT light it)
• Buttons: BOOT/GPIO0 → primary (Space), VOL_DOWN/GPIO40 → secondary (Shift+Tab), VOL_UP/GPIO39 → cycle screens (routed through power_hal, no PMU on this board)
• Battery: ADC2 voltage divider on GPIO17 + charging-status on GPIO38
• Power latch: GPIO21 must be HIGH or the board self-shuts on USB unplug
• No touch, no IMU, no PMU IC, no IO expander

What was tested

• ✅ ESP32-S3 boots, USB-CDC enumerates, BLE advertises with correct MAC
• ✅ BoardCaps reports Xingzhi Cube 1.83, 284x240
• ✅ NV3023 init sequence accepted; backlight lights via LEDC PWM
• ✅ LVGL splash renders Clawd at correct geometry and orientation
• ✅ Primary R/G/B fillScreen colors render correctly
• ✅ Builds clean: pio run -e xingzhi_cube_183 → Flash 19.8%, RAM 31.3%

Known limitations

• Buttons (GPIO 0 / 39 / 40) and battery ADC verified by code path only — not on real BLE HID receiver yet.
• LVGL anti-aliased mid-tones render slightly desaturated on this panel (Clawd's salmon body shows as warm brown). Acceptable for a status display; can be tuned later via NV3023 gamma registers (0xE0 / 0xE3 / 0xE5 in the init sequence).

Test plan for reviewers

• pio run -e waveshare_amoled_216 still builds (no regressions from HAL change)
• pio run -e waveshare_amoled_18 still builds
• pio run -e xingzhi_cube_183 builds and produces a valid factory.bin
• Flash a real Xingzhi Cube 1.83" device → splash renders, BLE advertises
• (Optional) Verify the 3 buttons emit Space / Shift+Tab / cycle-screens via BLE HID

🤖 Generated with Claude Code

[nguyenduchoai]

Pushed a follow-up commit (bffb238) that switches this board to a high-contrast monochrome render in display_hal_draw_bitmap.

The NV3023 on this specific hardware never settles into clean colors with the standard INVOFF / MADCTL / LV_COLOR_FORMAT_RGB565* combinations — anti-aliased mid-tones drift through several hues across init permutations, and pure black/white pixels come out with inverted polarity in some configs. After exhausting the obvious knobs, the firmware now thresholds each pixel's max channel against MONO_THRESHOLD (150/255), pre-inverts the result, and pushes pure black or pure white. Clawd reads as a crisp white silhouette on solid black; text + battery icon stay sharp.

#define MONOCHROME_RENDER in display.cpp can be commented out for anyone who wants to keep exploring color tuning later — no other boards see this change.

[nguyenduchoai]

Discovered + fixed one more issue during real-hardware verification (5e97d41):

On the splash screen, PWR-press was hard-coded to call splash_next() (cycle animations), with the implicit assumption that a touch tap is the way to leave splash. On the Xingzhi Cube — which has no touch panel — that left the user stuck on splash forever with no way to reach Usage or Bluetooth.

Added BoardCaps.has_touch (true for both Waveshare ports, false for Xingzhi). When has_touch == false, the splash-screen PWR-press now routes to ui_toggle_splash() instead, taking the user straight to Usage. Cycle behavior on Usage ↔ Bluetooth is unchanged.

Verified on hardware: BOOT (GPIO 0), VOL_UP (GPIO 39), VOL_DOWN (GPIO 40) all toggle correctly, and PWR now successfully transitions Splash → Usage → Bluetooth on the Xingzhi.

[nguyenduchoai]

Followed up with 4efb07d — adds a third UI breakpoint ("tiny", height < 280) so the Usage screen also re-scales for the 284×240 Xingzhi panel.

Previously the Usage-screen widgets hardcoded font_tiempos_56 / font_styrene_48 / _28, sized for the 368-px and 480-px panels. On the Xingzhi the title overlapped the 80×80 logo, the reset text overlapped the bar, and the percent number didn't fit inside its panel.

Layout struct now carries Usage-screen fonts, bar height, pill padding, and a show_chrome flag. The new tiny breakpoint hides logo + battery icons (no room) and picks tighter fonts; the existing large/compact branches keep their previous values verbatim, so Waveshare boards see no behavioral change.

Verified on real Xingzhi hardware with screenshot.sh: title, two usage panels, and the "Divining…" footer all fit within 284×240 without overlap. Reset text bumped to styrene_20 for legibility — easy to read on the panel.

[nguyenduchoai]

Added one more commit (880c3a0) — BoardCaps.auto_cycle_ms for opt-in screen auto-cycling. Touchless boards can set this to > 0 to have main.cpp call ui_cycle_screen() on a global timer; existing ports leave it at 0 and behave unchanged.

The Xingzhi opts in at 5000 ms: after VOL_UP toggles out of splash into Usage, the panel auto-flips Usage ↔ Bluetooth every 5 seconds. The timer pauses on splash and while idle_is_asleep(), and is reset by every manual PWR-press so taps never fight the cadence.

Verified on hardware.

(The user also asked about adding Codex usage tracking on the same screens — that's a meaningful scope expansion since OpenAI doesn't expose Claude-style *-utilization headers, so it'd need a separate daemon path, BLE payload extension, and UI rework. Keeping it out of this PR; happy to open a separate issue if there's interest in pursuing it.)

[nguyenduchoai]

One more commit (c6acb4f) — full Codex usage tracking, end-to-end.

Turns out the Codex CLI (https://github.com/openai/codex) hits chatgpt.com/backend-api/codex/responses and the response carries the same shape of rate-limit signal Claude does:

```
x-codex-primary-used-percent → cs (5-hour window)
x-codex-primary-reset-after-seconds → csr
x-codex-secondary-used-percent → cw (weekly / 7d window)
x-codex-secondary-reset-after-seconds → cwr
```

Daemon: load_codex_auth() reads ~/.codex/auth.json for the ChatGPT access token + account id; poll_codex() makes a minimal SSE call to /codex/responses and parses the headers above. Codex polling is best-effort — if the token is missing/expired the poll returns None and Claude data still ships on its own. Verified against a real ChatGPT Pro account: cs=3, cw=14.

Firmware: new SCREEN_USAGE_CODEX (title "Codex", "5h"/"7d" pills, same panel chrome as the Claude Usage screen). ui_cycle_screen() weaves it in only after the daemon has sent Codex data — have_codex_data flag — so a Codex-less setup never sees an empty "Codex" screen. Auto-cycle then rotates: Usage(Claude) → Usage(Codex) → Bluetooth → Usage…

Screenshot from the Xingzhi at 284×240 (test data injected for QA):

• "3% / 5h" with "Resets in 3h 38m"
• "14% / 7d" with "Resets in 4d 17h"
• "* Elucidating…" spinner footer

Happy to drop this commit if you'd prefer to keep this PR scoped strictly to the board port — it's clean enough to lift to a separate PR.

[nguyenduchoai]

Final commit (61898ee) — Antigravity (Google's agent IDE) tracking.

Turns out Antigravity ships a `language_server` binary that already talks to Google's cloud-code API on the user's behalf and exposes a `GetUserStatus` RPC on a random localhost port. Polling that locally means no OAuth refresh dance is needed and we get monthly prompt + flow credit balances, plan name, and per-model quotas directly:

```
{
"userStatus": {
"planStatus": {
"planInfo": {
"monthlyPromptCredits": 50000,
"monthlyFlowCredits": 150000,
...
},
"availablePromptCredits": 500,
"availableFlowCredits": 100
},
"userTier": { "name": "Google AI Ultra", ... }
}
}
```

The daemon auto-discovers the endpoint (pgrep `language_server` → scrape `--csrf_token` from `ps -ww` → `lsof` for the listening port → POST to `/exa.language_server_pb.LanguageServerService/GetUserStatus`) and adds `as`/`asr`/`aw`/`awr`/`apn`/`apf`/`aff` to the BLE payload.

Firmware: new `SCREEN_USAGE_ANTIG` with title "Antig <plan>", panels "Prompt" + "Flow", credit counts ("500/50K") rendered in place of the reset clock. The cycle order with all three services is now:

```
Splash → Usage(Claude) → Usage(Codex) → Usage(Antigravity) → Bluetooth → …
```

Each "extra" screen is only inserted once we've seen its data, so a Claude-only user still just sees Usage ↔ Bluetooth.

Verified live on Xingzhi against my actual Google AI Ultra subscription — device shows `Antig Ultra` / `99% / 500/50K` / `100% / 100/150K`.

Maintainer: if this scope is too wide for a single board-port PR, all four Codex/Antigravity commits (`c6acb4f`, `61898ee`, plus the screen + auto-cycle infra) sit on the back of the branch and can be cherry-picked into a follow-up PR. The base port commits (`ec0395f` through `5e97d41`) are independent.