Skip to content

feat(debug): live device screen mirror with remote control (protobufs #1054) - #6987

Draft
jamesarich wants to merge 14 commits into
mainfrom
screen-mirror-poc
Draft

feat(debug): live device screen mirror with remote control (protobufs #1054)#6987
jamesarich wants to merge 14 commits into
mainfrom
screen-mirror-poc

Conversation

@jamesarich

@jamesarich jamesarich commented Aug 31, 2026

Copy link
Copy Markdown
Collaborator

⚠️ Preview / draft — do not merge yet. This tracks org.meshtastic:protobufs:2.99.0-screen-mirror-poc-SNAPSHOT resolved from mavenLocal (useMavenLocal=true is committed on this branch) to pick up meshtastic/protobufs#1054, which isn't published anywhere yet — CI dependency resolution will fail by design. Re-pin to a released protobufs (and revert the two build-file lines) once meshtastic/protobufs#1054 ships in a tag. Firmware side: meshtastic/firmware#11681.

Adds a live Mirror tab to the Debug panel: the connected device's screen rendered in real time — in its actual theme colors on color-TFT devices — with full remote control.

Verified on hardware against RAK4631 (mono OLED) and LILYGO T-Deck (color, touch). Read the T-Deck result with care: the colour path is MUI/LVGL dirty rects, which need a device-ui carrying a display flush observer. That is not in any released device-ui, so on this branch the whole MUI path compiles out (HAS_MUI_MIRROR is 0 unless MESHTASTIC_MUI_MIRROR is defined) and the T-Deck run used a locally built device-ui. A reviewer building from these branches gets the mono path only.

🌟 New Features

  • Live screen mirror: FromRadio.display_frame chunks reassembled into frames and rendered once per frame into a 1:1 ImageBitmap, upscaled with nearest-neighbor filtering (crisp device pixels). Color devices stream a signature-keyed DisplayPalette; the renderer applies per-region RGB565 colors with the firmware's last-region-wins precedence, and a frame arriving before its palette self-heals to color when the palette lands.
  • Remote control, three ways: a circular 5-way D-pad ring (TV-remote layout: wedge hit-testing, hold-to-repeat at 10 Hz after 500 ms, long-press OK → SELECT_LONG, haptics, per-wedge semantics for screen readers); desktop keyboard capture on the mirror (chip- or click-toggled, visible focus ring; arrows/Enter/Esc, KeyUp consumed so keys never leak to the scroll container); swipe-to-navigate on the image, and on has_touch devices tap/long-press forwards real panel coordinates via send_input_event.
  • Adaptive layout: controls sit beside the mirror above 760 dp content width, below it otherwise.
  • Device awareness: DeviceMetadata.display gates behavior — e-ink devices get a hint steering toward one-shot Refresh; touch capability enables tap-to-touch.

Changes

  • core/repository: DisplayMirrorManager (+ MirrorFrame/MirrorPalette), AdminController.setDisplayMirror/requestDisplayFrame/sendInputEvent — non-suspend immediate sends so input never queues behind bulk admin traffic and lifecycle teardown can always stop the stream.
  • core/data: reassembly state machine — contiguous-offset chunks only, geometry pinned at the first chunk, format/size/region-cap validation before any allocation, reset() so a previous device's last screen is never shown as live. Unit-tested (frames, palettes, drops, restarts, caps).
  • core/service: admin senders via sendAdminImmediate.
  • feature/settings: the Mirror tab (DisplayMirror.kt, DisplayMirrorViewModel.kt, DpadRing.kt, MirrorFrameRenderer.kt) — stream stopped on tab dispose, ViewModel clear, and disconnect; toggle tracks connection state. Renderer resolver extracted pure and unit-tested (region precedence, RGB565 bit-replication).
  • All user-facing strings localized (core/resources), two Material Symbols chevrons added.

🤖 Generated with Claude Code


Status

Rebased onto main (2026-09-08). Baseline green: spotlessApply spotlessCheck detekt assembleDebug test allTests kmpSmokeCompile, 43 mirror tests passing.

Blocking, in release order:

  1. device-ui — the flush observer has no PR yet (branch screen-mirror-poc on a fork). Until it lands and a pin bumps, the colour/MUI path is compiled out everywhere.
  2. protobufs#1054 — must merge and tag. This branch then drops useMavenLocal=true and the 2.99.0-screen-mirror-poc-SNAPSHOT pin. Note main already tracks a protobufs develop-SNAPSHOT, so this may need no re-pin at all rather than a version bump.
  3. firmware#11681 — carries the producer plus the submodule bump.

Fixed since the last review pass:

  • Managed mode was bypassed. The Debug Panel is the one Settings route deliberately exempt from the connected && !isManaged gate, on the grounds that it "reads app-local logs only" — and the Mirror tab put an admin write and arbitrary input injection on it. Now refused in the ViewModel (disarm still allowed, so a live stream tears down when managed arrives mid-session) and gated in the UI.
  • Followed FromRadio.display_frame 20 → 21 and display_palette 21 → 22, which moved off tags open PRs Show full connectivity info in first line #980 and chore(deps): update coroutines_version to v1.8.1 #1030 already claimed.

Known gaps, not yet addressed: a cross-repo audit raised further findings on this branch — frame-render cost on the composition thread, and rect-reassembly panel-geometry stability across continuation chunks. Both are recorded and unfixed here.

Screenshots still to attach.


Feature stack

Five repos, and the release order runs top to bottom. Nothing below can ship until the piece above it lands.

Repo PR Role
meshtastic/design #142 Cross-platform feature spec
meshtastic/device-ui #394 MUI flush observer + opt-in input injection
meshtastic/protobufs #1054 DisplayFrame / DisplayPalette / DisplayInfo + the two admin verbs
meshtastic/firmware #11681 Producer: streams the framebuffer, bridges input
meshtastic/Meshtastic-Android #6987 Mirror tab with remote control
meshtastic/meshtastic-mcp #78 capture_display tooling

device-ui#394 is the current blocker: without the flush observer the MUI/colour path compiles out of every committed firmware configuration, leaving only the 1bpp mono path.

@coderabbitai

coderabbitai Bot commented Aug 31, 2026

Copy link
Copy Markdown
Contributor

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

  • Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

Comment @coderabbitai help to get the list of available commands.

@jamesarich

Copy link
Copy Markdown
Collaborator Author

Cross-platform feature spec (client-agnostic wire contract, reassembly rules, input mapping, UX and device requirements): meshtastic/design#142

@codecov

codecov Bot commented Aug 31, 2026

Copy link
Copy Markdown

⚠️ JUnit XML file not found

The CLI was unable to find any JUnit XML files to upload.
For more help, visit our troubleshooting guide.

jamesarich and others added 13 commits September 8, 2026 15:57
Consumes the DisplayFrame PoC protobufs (2.99.0-screen-mirror-poc-SNAPSHOT
from mavenLocal, useMavenLocal committed on this branch): DisplayMirrorManager
reassembles FromRadio.display_frame chunks into MirrorFrame snapshots,
AdminController gains setDisplayMirror + sendInputEvent, and the Debug panel
gains a Mirror tab rendering the device's 1bpp framebuffer live with a D-pad
driving the firmware InputBroker. PoC: tab strings deliberately unlocalized.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Cap the mirror canvas at 512.dp (4x for a 128px OLED) instead of filling
the window width, fold the frame info into the toggle row, and pair
Down/Back so the controls fit without scrolling on desktop.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Manager: validate format (MONO_VLSB only) and geometry (width/height must
describe exactly total_size bytes) so malformed device input can't reach
aspectRatio and crash; unit tests for reassembly, drops, restarts.
AdminController: setDisplayMirror/sendInputEvent become non-suspend
immediate sends (teardown-safe, input never queues behind bulk admin) and
gain requestDisplayFrame. Mirror tab: render each frame once into a 1:1
ImageBitmap scaled with FilterQuality.None (kills the fractional-scale
seams), gate controls on connection state, reset the toggle on disconnect,
stop the stream on tab dispose and ViewModel clear, add a one-shot Refresh,
inject AdminController/ConnectionStateProvider instead of RadioController.
Dispatch: display_frame test coverage incl. revoked-session.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Populates the new DisplayInfo during the handshake — dimensions from the
live display device, MONO_VLSB frame format, PanelClass from the panel
defines (EINK/HUB75/TFT/LCD/OLED), has_touch from HAS_TOUCHSCREEN — so
clients can gate mirroring UI and adapt to the panel. Absent on
display-less builds. Regenerates protos (DisplayInfo, DisplayFrame rect
fields reserved for dirty-rect streaming).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Replace the flat button stack with the researched control surface:
cross-shaped 5-way cluster (56dp circular keys, 8dp gaps, Material
Symbols chevrons with contentDescriptions), directions auto-repeat on
hold (500ms then 10Hz — slow enough for the radio to render), OK
long-press sends SELECT_LONG, Back sits below per TV-remote convention.
The mirror itself becomes an input surface: click to capture the
keyboard (arrows / Enter / Esc, with a visible focus ring and hint) and
swipe in a cardinal direction to navigate. Consumes the new
DeviceMetadata.display: shows an e-ink hint steering toward one-shot
Refresh. Adds keyboard_arrow_left/right icons (Material Symbols rounded,
transposed from the existing up glyph).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
DisplayMirrorManager reassembles FromRadio.display_palette chunks
(signature-keyed, defaults captured on the offset-0 chunk) into a
MirrorPalette; frames carry palette_signature and the renderer resolves
each pixel through the region table (last overlapping region wins,
matching firmware precedence) with RGB565 conversion, falling back to
monochrome when the referenced palette hasn't arrived. Renderer split
into MirrorFrameRenderer.kt.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Controls sit beside the mirror when the content width allows (>=760dp,
BoxWithConstraints) and stack below it otherwise; tapping the mirror now
toggles keyboard capture instead of only acquiring it, with the hint text
naming the way out.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Circular D-pad ring replaces the cross cluster (wedge hit-testing with a
radial dead band, hold-to-repeat, per-wedge semantics, haptics), keyboard
capture becomes an explicit chip with tap-to-toggle, and taps on touch
devices forward real panel coordinates (long-press = SELECT, matching the
physical touch driver). Manager gains reset() (wired to ViewModel creation
and disconnect), a palette region cap, and first-chunk geometry pinning;
renderer uses bit-replicated RGB565 and run-length row drawing with the
pixel resolver extracted pure and tested; KeyUp of captured keys is
consumed; ViewModel moved to its own file; all strings localized; the
stray .bak test file is gone.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
DisplayMirrorManager reassembles RGB565 rect chunks (sequence, geometry,
and panel-bound validation) and composites them into a persistent
little-endian RGB565 canvas emitted per completed rect; the renderer gains
a true-color branch using the same run-length row drawing. Mono frames are
untouched.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
RGB565 frames identify LVGL/MUI devices, whose UIs read their own input
drivers rather than the InputBroker remote events inject into — so the
mirror hides the D-pad, keyboard, swipe, and touch affordances and says
so, until an input-injection path exists. Utility helpers and the input
codes move to DisplayMirrorUtils.kt.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
device-ui now accepts injected input (virtual LVGL pointer + group-attached
keypad bridged from the InputBroker), so the view-only treatment goes away:
the D-pad navigates MUI widgets, taps forward panel coordinates, and the
keyboard types. The mirror_view_only string stays for the future
capability-flag path.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The mirror's keyboard capture only mapped navigation keys and silently
dropped everything else, so text entry did nothing: printable characters
never left the client even though the firmware already injects kb_char.
Characters now go out as ANYKEY with the code point, Backspace goes as a
character (so text fields delete rather than navigating back, which stays
Esc's job), and the handling moves into a Modifier.remoteKeyInput
extension.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Third adversarial review pass over the client side.

Rect reassembly moves out of DisplayMirrorManagerImpl into
MirrorRectCompositor: the manager was carrying two unrelated state
machines and 14 methods. Every rect field is now validated before
anything is allocated (total_size is device-controlled), geometry is
latched from the chunk at offset 0 so a later chunk cannot transpose
the blit, continuations must match the buffer that was actually
allocated rather than their own claim, and bounds use the subtraction
form because rect_x + rect_width overflows for hostile values. A rect
that omits width or height covers the full panel in that axis.

remoteKeyInput no longer swallows modified keys, so Cmd+Q, Ctrl+C and
Alt+Tab stay with the desktop instead of being forwarded to the device.

MirrorFrameRenderer tracks colorized runs as a raw Int with a sentinel
rather than a boxed Color?, and MirrorFrame.hashCode includes the
format so a mono and an RGB565 frame of the same size cannot collide.

Drops mirror_view_only, orphaned since MUI gained remote input.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Debug Panel is the one Settings route deliberately exempt from the
`connected && !isManaged` gate every sibling carries, justified by a comment
saying it "reads app-local logs only". The Mirror tab put an admin write
(set_display_mirror) and arbitrary input injection on that route, so managed
mode - which exists to lock a device out of configuration - was bypassed
entirely.

Enforced in the ViewModel, because the route will not do it: arming,
one-shot frame requests, key, character and touch sends are all refused while
managed. Disarming still goes through, so a stream already running is torn
down when managed mode arrives mid-session.

The UI gates on `active = connected && !managed` rather than offering controls
that silently do nothing, and the RadioConfig comment now says what the route
actually guarantees, since its old wording is what let this through.

MirrorHeaderRow is split out of DisplayMirrorContent, which was at detekt's
LongMethod limit and could not absorb another line.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

enhancement New feature or request

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant