An app can stop consuming input while it keeps drawing: Android delivers the tap, the app never acts on it. Phase 3 hit this three times against Spotify and the recipe failed on the label it was waiting for, which names the wrong problem. The new `expect_change <duration>` step fingerprints the screen (foreground package plus every visible label and its bounds) and fails when the screen does not change after the previous step, with the distinct reason "the app stopped consuming input". It supports `retry N`, which re-presses the step above a bounded number of times, so a dropped tap is still recovered and a wedged app still ends the run (FR-11). Co-Authored-By: Paperclip <noreply@paperclip.ing>
Artemis Control
Artemis is the phone's hands. It drives the Pixel 8a screen from .13 over
Tailscale using ADB wireless, so the Hub can do things no deep link allows.
The build spec is AGENTS.md. Follow the phases in section 11 in
order.
Status
| Phase | State |
|---|---|
| 1 — Prove the ADB link | ✅ proven — see docs/phase-1-adb-link.md |
| 2 — Act primitives | ✅ done — tap by text, type, scroll, back, home, launch |
| 3 — Spotify recipe | ✅ done — spotify_play plays the right song end to end (13 steps, 35s), see docs/phase-3-spotify-recipe.md |
| 4 — Song radio | ✅ done — spotify_radio starts a real radio (19 steps, 57s); the run log is now bounded at ~1 MB × 5, see docs/phase-4-song-radio.md |
| 5 — Maps recipe | not started |
| 6 — Action API | not started |
| 7 — Resilience | not started |
| 8 — Documentation | not started |
Layout
cmd/artemis/ command line entry point
internal/adb/ timeout-bounded wrapper around the adb binary
internal/ui/ read the screen (uiautomator dump) and act on it (input)
internal/recipe/ the recipes file format and its runner
recipes/ one hand-editable file per task
logs/ run log (JSON lines, not committed)
docs/ phase evidence and notes
Build and test
Go is not installed system-wide on .13; use a one-off nix shell. Nothing is
added to nixos-rebuild and no system module changes.
nix shell nixpkgs#go -c go build -o bin/artemis ./cmd/artemis
nix shell nixpkgs#go -c go vet ./...
nix shell nixpkgs#go -c go test ./...
Use
The phone address defaults to 100.101.49.17. Override with ARTEMIS_PHONE,
pin a serial with ARTEMIS_ADB_SERIAL, point at a different adb with
ARTEMIS_ADB, point at another recipes directory with ARTEMIS_RECIPES, and
move the run log with ARTEMIS_LOG.
./bin/artemis connect # scan for the wireless-debugging port, then connect
./bin/artemis devices
./bin/artemis dump # text nodes and bounds
./bin/artemis keyguard # lock-screen state; exits 1 when locked
./bin/artemis launch com.spotify.music
./bin/artemis tap "Search" # find by text, tap its centre (FR-4)
./bin/artemis type "Rolling Stones Honky Tonk"
./bin/artemis scroll down
./bin/artemis back
./bin/artemis home
./bin/artemis demo # Phase 2 DoD: launch an app, tap a named element
Recipes
A recipe is a plain-text file in recipes/ that Sam edits by hand. Steps name
what to find on screen; there are no coordinates anywhere.
./bin/artemis recipes # list them
./bin/artemis run spotify_play query="Rolling Stones Honky Tonk"
./bin/artemis run spotify_radio query="Rolling Stones Honky Tonk"
Read docs/phase-3-spotify-recipe.md for the
file format, the verbs, and why list rows need their own step, and
docs/phase-4-song-radio.md for the context menu,
menu_row, and the two ways a radio run fails on purpose.
Rules that the code enforces
- Find by text, never fixed coordinates.
tapresolves a text match to bounds and taps the centre;tap_rowscores result rows against the query and taps the winning row's centre;menu_rowopens the action menu of that same row, from the button inside it. - Never tap blindly. A missing or disabled element is a loud failure, not a
guessed coordinate. A named bad screen (
guard) fails with its own reason, and a step may name its own failure withelse "...". - Detect a wedged app. An app can stop consuming input while it keeps
drawing: Android delivers the tap, the app never acts on it.
expect_changefails withthe app stopped consuming inputwhen the screen does not change after the previous step, instead of blaming the label the next step waited for. - Never touch a locked phone. A secure keyguard stops Artemis with
phone is locked; unlock it and retry. No PIN is stored, read or typed (AGENTS.md section 13 Q2). - Confirm playback from the media session, not the screen. Spotify's
transport button keeps the label "Play" while music plays; the
playingstep asks Android what is actually playing and checks the track title. - Wake first. Every acting command wakes the screen before it acts (FR-7), and it checks the keyguard only after the screen is on.
- Bound every command. No adb call runs without a timeout, and a recipe has a per-step and a whole-run timeout (FR-11).
- Log every run. Each run appends a JSON line to
logs/runs.jsonl(FR-10). The file rotates at about 1 MB, keeping the recent runs (FR-10 is an audit trail, so history is bounded, never deleted wholesale). Tune it withARTEMIS_LOG_MAX_BYTESandARTEMIS_LOG_KEEP.