ares-do: send key events and take screenshots

[mariotaku]

Adds ares-do, a tenth tool: it presses buttons and captures the screen, so a UI flow can be scripted and replayed instead of walked by hand. The work that motivated it (webosbrew/samples#1) meant running luna-send once per keypress and diffing PNGs by hand.

ares-do -d tv key TAB TAB ENTER      # a web view starts with nothing focused
ares-do -d tv text demo
ares-do -d tv screenshot > shot.png
ares-do -d tv luna luna://com.webos.applicationManager/dev/running | jq
ares-do -d tv --json run - --out ./frames <<'EOF'
launch com.webos.app.livetv
wait 2s
shot 01-livetv
EOF

Everything is host-side over SSH. Nothing is installed on the device and nothing is written to its filesystem beyond one staging file per screenshot, which is deleted again.

Two undocumented services

test/sendKeyCode on com.webos.service.networkinput, and executeOneShot on com.webos.service.capture falling back to com.webos.service.tv.capture — a 49LK5900 only has the latter, so the fallback is not theoretical. Both are on the private bus, so a uid probe runs before anything is sent rather than after three keys have gone nowhere.

sendKeyCode reaches UInputWriter::sendKeyPress, which writes to /dev/uinput. That settles two things: the codes are ordinary evdev and stop at KEY_MAX, and there is no held modifier, so capitals cannot be typed at all. text refuses them up front naming every character, rather than a --shift flag that would quietly do nothing.

Which code each button sends

LG's choice, and often not the mainline name. /usr/share/X11/xkb/keycodes/lg in the firmware is the authority, and the surprises are aliases with notes in ares-do keys. The remote's Back is 412, not KEY_BACK (158) — confirmed on hardware, where 158 moved 970 pixels and 412 dismissed the launcher and moved 1.4M. Likewise the remote's Home is HOMEPAGE (172), not KEY_HOME.

The flow language is the CLI

A line is tokenized and handed to the same clap Command the shell uses, so the two cannot drift and the flags are identical in both. The whole file parses before anything runs, and every error is reported at once with line numbers.

Ergonomics

Follows adb and xdotool rather than inventing conventions: stdout is data and stderr is for people, so screenshot > shot.png works; --json emits NDJSON ending with an error object whose code is the exit code; exit codes are a documented table; and run - replays a flow from stdin over one SSH handshake instead of one per key. keys and --dry-run need no device at all.

Also fixes the shared Luna client

Luna::call read luna-send's stdout and stderr and then threw both away, collapsing "no such service", "not allowed on this bus" and "the channel died" into one value — and it read to EOF with no bound, so a method that subscribes instead of replying hung forever.

ares-connection-lib gains a public Exec::exec_timeout that drains both streams against a deadline, and Luna gains call_with/subscribe_with taking a LunaOptions carrying the bus and the timeout. call(uri, payload, public) stays as a defaulted delegate, so ares-launch and ares-install did not change a line and are bounded anyway — verified against a device.

Worth review: the 30s default is a behaviour change for those two, which previously waited forever. subscribe_with is deliberately left unbounded, since a subscription is meant to stay open. LunaError gained variants and is now #[non_exhaustive], which protects future ones but not this change — a downstream exhaustive match (dev-manager-desktop) would need a wildcard.

Not included

Pointer/tap is parked on backup/ares-do-tap. It works and lands on the pixel, but the TV runs dropbear, which has no direct-streamlocal, so a unix socket cannot be forwarded and it needs a helper on the far end. That is a bigger departure than this tool should make.

Verified

On a 49LK5900 (webOS 4.4.3): key injection, both capture services, PNG to stdout, a multi-step flow from stdin, remote temp cleanup, and every documented exit code. 67 unit tests, clippy clean at workspace pedantic, fmt --check clean, and the committed keycode table matches its generator.

The second commit is unrelated rustfmt churn from a newer nightly, split out so the ares-do diff stays about ares-do.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Gde2tcn5SbUHKhesuTNB7a