Two defects found by running spotify_play against the real phone. - Step durations were always zero in the run log: executeStep returned a copy, so the deferred assignment never reached the caller (FR-10). It now returns a named value, with a regression test. - The phone was sitting on the Android Settings app, and wait "Search" was satisfied by Settings' own "Search Settings" entry, so the recipe tapped the Settings search bar. New "foreground" step waits until the recipe's app really is in front; spotify_play uses it right after launch. docs/phase-3-spotify-recipe.md now logs all six live attempts and states plainly that the end-to-end playback run has not completed: the phone is showing a Settings "Verify that it's you" biometric prompt, which needs Sam and is left untouched.
9.5 KiB
Phase 3 — the recipes file and spotify_play
Phase 3 of AGENTS.md section 11: the first named recipe, plus the resolved
answer to section 13 Q2 (the locked screen).
Date: 2026-10-16. Phone: Pixel_8a (akita) over Tailscale at
100.101.49.17:46689 (the port moves on every reboot; nothing hardcodes it).
What was built
| File | What it holds |
|---|---|
recipes/spotify_play.recipe |
The recipe, written to be edited by hand. |
internal/recipe/recipe.go |
The file format: directives, guards, steps, parameters. |
internal/recipe/engine.go |
Runs a recipe step by step with timeouts and a run log. |
internal/ui/rows.go |
Picks a list row by text and taps its centre. |
internal/ui/keyguard.go |
Reads the lock state and refuses to act when locked. |
cmd/artemis/run.go |
artemis recipes, artemis run <task> k=v, JSON run log. |
The recipes file
One file per task in recipes/, plain text, no build step. It names the app,
the parameters the caller must supply, the screens Artemis must never tap on,
and the ordered steps.
task: spotify_play
app: com.spotify.music
param: query
row_id: com.spotify.music:id/row_root
guard "Log in" => Spotify is signed out; sign in on the phone and retry
steps:
- launch
- wait "Search" 30s
- tap "Search"
- tap_if "Clear search query"
- wait "What do you want to listen to?" 30s
- tap "What do you want to listen to?"
- type "{query}"
- wait "Clear search query" 30s
- wait_gone "Recents" 30s
- tap_row "{query}" 40s
- wait exact "Pause" 40s
Steps express intent. There is no coordinate anywhere in the file (FR-4).
Run it with:
./bin/artemis run spotify_play query="Rolling Stones Honky Tonk"
The verbs are launch, foreground, wait, wait_gone, tap, tap_if,
tap_row, type, key, back, home and sleep. tap exact "Pause"
matches the whole text, not a substring. tap_if taps only when the text is
already on screen, which lets one recipe cope with an app that remembers its
last screen. foreground waits until the recipe's app is really in front, so a
word like "Search" cannot be matched from a different app that happens to be
open — every recipe should use it straight after launch.
Why tap_row exists
A plain text tap picked the wrong element on the real Spotify screen. On the
results list, artemis tap "Honky Tonk Women" tapped
ImageButton[context_menu_button] text="More options for song Honky Tonk Women - Live"
bounds=[828,2388][954,2400]
That is the smallest clickable match, and it sits at the bottom edge of the screen. It was the wrong target: a "More options" button, not the song.
tap_row scores every result row against the query and taps the row itself:
- A candidate row is a clickable container that carries a title and a subtitle. A one-line search suggestion (Spotify's "rolling stones honky tonk women single") is therefore not a result row.
- Text inside a clickable child is skipped, because that is an action button ("More options", "Add to library", "Add suggestion"), not the row's own text.
- A row clipped by the bottom of the list is dropped. It is much shorter than a row that is fully drawn, and its centre sits under the navigation area.
- Rows are ranked by F1 score over the query words, so an exact title beats a noisy deluxe-edition title. Equal scores go to the highest row, because search results are ranked by relevance.
On the real screen for Rolling Stones Honky Tonk, the top hit wins:
Honky Tonk Women / Song • The Rolling Stones bounds=[0,794][1080,962] F1=0.727
and the tap lands at the row's centre, (540,878) — on the title, clear of the
action buttons. internal/ui/rows_test.go runs this against a trimmed copy of
the real dump.
Failing cleanly (FR-8)
Every failure names its cause, and they are distinguishable:
| Situation | Message |
|---|---|
| App missing | app is not installed: com.spotify.music |
| Signed out | Spotify is signed out; sign in on the phone and retry (phone is showing "Log in") |
| Update nag | Spotify wants an update; update it on the phone and retry |
| Element missing | text "Search" did not appear within 30s; foreground=com.spotify.music; screen shows: "Log in", "Sign up free" |
| Phone locked | phone is locked; unlock it and retry |
A guard is checked when a step cannot find its target, so a bad screen produces its own reason instead of a generic "element not found".
Timeouts (FR-11)
step_timeout: 20s is the default per step, total_timeout: 150s bounds the
whole run. A step may name its own timeout (wait "Search" 30s). The engine
checks the remaining budget before each step and stops when it runs out. Every
adb call is already bounded by adb.DefaultTimeout.
Logging (FR-10)
Every run appends one JSON line to logs/runs.jsonl (override with
ARTEMIS_LOG): task, params, each step with its duration, the result and the
total duration. Two real entries:
{"time":"2026-10-16T13:23:31+11:00","task":"spotify_play","params":{"query":"Rolling Stones Honky Tonk"},"ok":false,"error":"phone is locked; unlock it and retry","steps":null,"duration_ms":832}
The same steps print to the terminal as they run.
The locked screen (section 13 Q2)
Decision: Artemis fails loudly. No PIN is stored and none is ever typed.
ui.KeyguardState reads two things: whether the lock screen is showing
(dumpsys window, the KeyguardServiceDelegate block) and whether the device
has a secure lock at all (cmd lock_settings is-user-secure). Acting commands
wake the screen first and then refuse if the keyguard is showing and secure. A
non-secure keyguard is still dismissed, as before. An unreadable lock state is a
failure, not a guess.
Evidence
The prescribed test (put the phone to sleep, wake it, Artemis must refuse):
adb -s 100.101.49.17:46689 shell input keyevent 26 # POWER = sleep
adb -s 100.101.49.17:46689 shell input keyevent 224 # WAKEUP
$ adb shell dumpsys window policy | grep -E 'showing=|screenState'
showing=true
screenState=SCREEN_STATE_ON
$ adb shell cmd lock_settings is-user-secure
true
$ ./bin/artemis keyguard
keyguard showing=true secure=true
artemis: phone is locked; unlock it and retry
exit=1
$ ./bin/artemis tap "Search"
artemis: phone is locked; unlock it and retry
exit=1
$ ./bin/artemis run spotify_play query="Rolling Stones Honky Tonk"
task spotify_play (Search Spotify for a song and play it.)
FAILED spotify_play after 833ms: phone is locked; unlock it and retry
exit=1
No PIN was entered, nothing was read from the phone, and the lock screen was still on screen afterwards — no blind tap happened. The screen dump after the refusals still showed the lock screen clock and date.
A dozing phone reports showing=false, because the lock screen is not drawn
yet. That is why every acting command wakes the screen before it checks the
keyguard; reading the state first would say "unlocked" about a locked phone.
Verification status
| Item | State |
|---|---|
| Recipes file format, guards, parameters, timeouts | unit tested (internal/recipe) |
| Row selection against a real Spotify dump | unit tested (internal/ui) |
| Keyguard parsing and secure-lock detection | unit tested (internal/ui) |
| Step durations reach the run log | unit tested (internal/recipe) |
| All acting commands refuse on a locked phone | live, verified (above) |
artemis run spotify_play refuses on a locked phone |
live, verified (above) |
spotify_play plays the right song |
not yet run end to end — see below |
| Spotify search reaches the result list | live (Home → Search → type → results) |
Live run log
Six attempts, in order. Each failure was clean, named its screen and tapped nothing random — which is the point of the phase.
- Refused, locked phone.
phone is locked; unlock it and retry. The phone had locked itself on its 5-minute screen timeout while the code was being written. - Refused after the prescribed sleep/wake test.
keyguard showing=true secure=true, exit 1 fromkeyguard,tapandrun. The lock screen was still on screen afterwards. - Refused, dozing phone that only looked unlocked. A dozing phone reports
showing=false. Pressing WAKEUP then revealed the secure keyguard, and the run refused. This is why every acting command wakes first. - Wrong app matched "Search". The phone was sitting on the Android
Settings app.
wait "Search"was satisfied by Settings' own "Search Settings" list entry, so the recipe tapped the Settings search bar. Fixed by adding theforegroundstep, which waits until the recipe's app really is in front before any text is trusted. - Blocked by a biometric prompt. Settings was left showing a
Verify that it's youdialog (Touch the fingerprint sensor/Use PIN), which is drawn over Spotify. The run failed with:text "Search" did not appear within 30s; foreground=com.spotify.music; screen shows: "Cancel", "App logo", "Settings", "Verify that it's you" .... Artemis did not touch the prompt, and it must not: it needs Sam's fingerprint or PIN, exactly like the lock screen. - Still blocked by the same prompt on the next attempt.
What is left
The live end-to-end run of spotify_play (Phase 3 DoD) has not completed. It
needs the phone unlocked and free of the biometric prompt, and then:
./bin/artemis run spotify_play query="Rolling Stones Honky Tonk"
./bin/artemis fg
./bin/artemis dump | grep -i pause
tail -1 logs/runs.jsonl
Phase 4 (song radio) and Phase 5 (maps) have not been started, on instruction.