* fix(metrics): repair review triage — correct secret wiring, privacy-safe aggregated issues
- triage-reviews.yml read secrets.GOOGLE_SERVICE_ACCOUNT_JSON, which doesn't
exist; map the real PLAY_STORE_SERVICE_ACCOUNT_JSON secret onto the env var
the script expects.
- triage-reviews.py rewritten to maintain a single sanitized, deduped
"Play Store Review Triage" issue instead of one public issue per review.
The old version leaked reviewer full names and verbatim review text into
public GitHub issues and spammed the tracker. The new version aggregates
actionable (<=3 star) reviews into one issue with rating counts, a
word-frequency theme summary (no quoted sentences), and opaque review_id
references for Play Console lookup. An embedded HTML comment marker
(matching the product-intelligence.mjs pattern) holds the current
actionable review_id set so runs update in place and skip entirely when
nothing changed.
- product-intelligence.yml referenced the nonexistent
SENTRY_PRODUCT_INTELLIGENCE_TOKEN secret, causing the daily cron to fail
silently (#60). Fall back to SENTRY_AUTH_TOKEN when the dedicated
read-only token isn't configured.
- docs/playstore.md: document that Play Console is still the only trusted
source for acquisition/uninstall metrics (product-intelligence.mjs defers
this), and that review-based signals are sourced via the Android
Publisher API through PLAY_STORE_SERVICE_ACCOUNT_JSON.
Closes#61. Refs #60.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NJKAQ6HAikWGQK7PGZ5Y4E
* fix(triage): fail visibly when GOOGLE_SERVICE_ACCOUNT_JSON is missing
Review finding on PR #78: env_client() exited 0 on missing credentials,
so the scheduled workflow would report success while silently doing
nothing — contradicting issue #61's 'missing credentials fail visibly'
done-criteria.
---------
Co-authored-by: engineer <engineer@gray-knight-m1.local>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
scripts/triage-reviews.py was fully written but had no workflow, so it
never ran. Add a daily 07:00 UTC cron (staggered after product-intelligence)
plus workflow_dispatch, with Python 3.12 + the Android Publisher API client
deps the script imports, and GOOGLE_SERVICE_ACCOUNT_JSON / GH_TOKEN passed
through as named secrets.
Also fix a stale doc-string reference: the issue body linked to a
non-existent monitor-reviews.yml; point it at the workflow actually created.
Adds privacy-safe aggregate product intelligence, reviewed/versioned website assets, and a dispatch-only rollout until the dedicated Sentry token is verified. Independent review blockers were fixed in 8bc47e4; app checks, website production build, Android CI, and iOS CI are green.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- docs-site/index.html: add Vercel Analytics + Speed Insights CDN scripts
(only fires on opencode.agentlabs.cc served via Vercel, not GitHub Pages)
- scripts/triage-reviews.py: fetch recent Play Store reviews via Android
Publisher API, create GitHub issues for ≤3★ reviews not yet tracked
Run review triage manually on VM:
DAYS_BACK=7 GOOGLE_SERVICE_ACCOUNT_JSON=... GH_TOKEN=... python3 scripts/triage-reviews.py
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Change permission notification title from `req.permission || 'Permission requested'`
to the user-friendly 'Agent needs approval'; permission type + patterns now appear
in the body (e.g. 'bash: echo hello') for context.
- Add `dedupeKey: `perm-${req.id}`` and `dedupeKey: `question-${req.id}``
(60 s cooldown) to both events so a SSE reconnect after disconnect() clears state
can't fire a second notification for the same pending request.
- Fix stale CUA-test comment that claimed 'Agent needs approval' did not exist;
fallback assertion already matched correct title; update the comment to reflect
the real events.ts behavior.
Closes#39
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Shows amber 'Reconnecting… (attempt N)' banner when SSE is down.
Shows brief green 'Connected ✓' flash on reconnect (useRef transition
to avoid atomic state reset bug where lastDisconnectAt resets with
reconnectAttempts in the same set() call).
Banner disappears automatically when SSE is stable.
Updates CUA scenario to check for both ASCII and Unicode ellipsis.
Closes#42
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
#39 (backgrounded permission notification):
- Event is permission.asked, not permission.requested (events.ts already calls
notify() for it; send() only fires while backgrounded).
- Assert on APP_PACKAGE (the only token guaranteed in every dumpsys record) plus
the actual copy ('Permission requested' / 'A tool needs your approval') instead
of the non-existent 'Agent needs approval' string.
#42 (SSE disconnect banner):
- reconnectAttempts only zeroes after STABLE_CONNECTION_MS (10s) past a healthy
reconnect, and the pending backoff timer can take up to 15s — so the banner can
linger ~25s. Poll up to 40s for dismissal instead of a fixed 15s sleep to avoid
a false 'still showing' failure.
Refs #39#42
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Adds three deterministic helpers that use ADB instead of LLM vision,
so pass/fail cannot be hallucinated:
check_ui_text(text) — uiautomator XML dump + grep
check_notification_drawer(text, timeout) — dumpsys notification poll
simulate_network_drop() / restore_network() — svc wifi/data disable
Plus two new feature-test scenarios with deterministic gating:
sse_disconnect_banner (#42)
- ADB cuts WiFi+data, waits, checks UI XML for 'Reconnecting' text
- LLM visual check is supplementary/informational only
- ADB restores network, checks banner disappears
backgrounded_permission_notification (#39)
- ADB backgrounds app (Home key)
- API sends permission-triggering message
- adb dumpsys notification checked for 'Agent needs approval'
- No LLM involved in pass/fail decision
Also adds background_app() / foreground_app() helpers.
Wires new scenarios into --scenarios catalog alongside LLM ones.
Updates --scenarios help text to document both types.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
connect phase: now verifies the connection appears in the list after saving,
not just that the form was dismissed. Prevents false PASS when the LLM
declares connect done before the entry is actually visible.
_precreate_test_session: when external URL (Tailscale) times out from the
CI runner, fall back to localhost:4096 (the runner-local opencode serve).
This ensures the named-session assertion stays active in standard CI runs
while also working when dispatched against a live external server.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The CUA script runs on the CI runner (host), not inside the emulator.
10.0.2.2 is the emulator's special address for the host — it is only
reachable FROM INSIDE the emulator. Calling it from the runner always
fails, so _precreate_test_session returned None, and the session_list
phase fell back to the weak 'screen visible' assertion instead of the
strong 'pre-created session must appear' check.
Fix: replace 10.0.2.2 → 127.0.0.1 before making the pre-create call.
Localhost URLs (100.x.x.x, custom dev server) pass through unchanged.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
sessions_reload regression phase now runs BEFORE the TypeScript task.
Previously it was gated behind typescript which fails in CI when the model
is unavailable — meaning the actual sessions regression check never ran.
Phase order is now:
connect → session_list (pre-created session required) → new_session
→ sessions_reload (navigate back, list must be non-empty) [CRITICAL]
→ typescript (informational) → verify (informational) → settings
Critical phases: connect, session_list, new_session, sessions_reload.
TypeScript/verify/settings are informational (model availability varies).
CI emulator fixes:
- api-level: 30 → 28 (more stable, boots reliably on ubuntu-latest)
- target: google_apis → default (lighter, no Play Services needed for
sessions regression test, avoids known boot issues with google_apis)
- disable-animations: true (reduces boot overhead)
- emulator-boot-timeout: 600 (explicit, matches action default)
- Switch from --scenarios to --showcase (runs the new structured flow
with _precreate_test_session + sessions_reload phase)
- Clear app state before install (pm clear) for deterministic first-run
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Previously the session_list phase goal said 'The session list may be empty
(no sessions yet) — that is fine' and 'Report done when you can see the
session list screen (even if empty)'. This means an empty sessions list
was treated as a test PASS, so every previous 'fix' was validated against
a test that cannot detect the regression.
Two changes:
1. _precreate_test_session(): calls POST /session via HTTP before the CUA
starts. The session_list phase goal now explicitly names this session and
requires it to be visible — if the app fails to load server sessions,
the phase fails (not passes silently with an empty list).
Falls back gracefully if the server is unreachable at pre-create time.
2. sessions_reload phase (new, critical): after completing the TypeScript
task the test navigates back to the Sessions tab and asserts the list is
non-empty. This catches the other variant of the regression — sessions
vanishing after navigating away from a session and back.
Both phases are now in the critical list, so a failure in either causes the
overall test to report partial/fail instead of success.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
ADB input text with %s escaping didn't trigger React Native onChangeText.
New approach: write text to /sdcard/ file on device, then use shell
command-substitution "$(cat ...)" to pass raw text to input text.
This preserves spaces without %s conversion and avoids quote-escaping
issues that prevented React Native from detecting text changes.
Rewrites android-cua-smoke.py to demonstrate the complete first-run journey
instead of the previous "ping" smoke test. The new structured multi-phase
flow covers: server connection setup, session list, new session creation,
TypeScript hello-world task submission (watching tool calls/file writes),
output verification, and Settings/model-selection screenshot.
Key changes:
- run_onboarding_showcase() orchestrates 6 sequential CUA phases with
per-phase goals, step budgets, and PASS/FAIL phase tracking
- run_cua_step() replaces run_cua() — accepts step_label, action_delay,
saves labeled screenshots (/tmp/cua_<phase>_<step>.png) for debugging
- Global --speed-multiplier flag scales all _sleep() calls (0.5 = 2x faster)
- Showcase is now the default mode; legacy --goal / --scenarios flags retained
for backwards compat and CI regression scenarios
- Tighter action_delay (0.7s) and trimmed history window (14 turns) vs
previous 1.0s / 12 turns
- Phase banner log lines ("STEP N: ...") narrate the video in real time
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01No3k1AEioE4PNUZg12TxQo
* fix(ci): bump opencode Azure apiVersion to support /responses endpoint (#22)
The CUA smoke probe was returning MODEL_CAPABLE=false because opencode's
@ai-sdk/azure provider got 'API version not supported' from Azure on
/openai/v1/responses with apiVersion=2024-08-01-preview. Split into two
envs: keep the CUA driver on 2024-08-01-preview (chat-completions only)
and bump the opencode-side provider config to 2025-04-01-preview, which
supports the new responses API.
Effect: send_message/multi_turn scenarios get included again in CUA smoke
when the probe succeeds.
* ci(cua): bound curl timeouts + diag dump on server-start hang (#22)
Step 11 'Start opencode server' has hung past the 45-min job timeout in
two consecutive runs (27195348071, 27198238465). Local boot of
opencode-ai 1.16.2 with the same heredoc config is healthy in 3s, so
something is wrong specifically on the GH-hosted runner — likely curl
post-loop waiting indefinitely on an unresponsive server.
Adds:
- set -x for command tracing
- --connect-timeout 2 -m 5 on every curl so hangs cannot exceed 5s
- HEALTHY flag + explicit exit 1 (drops the unbounded post-loop curl)
- Periodic dump every 10s: server log tail, ss listening sockets,
process liveness — so we can see WHY the server isn't replying
Pure diagnostics; no behaviour change for the green path.
* fix(ci): use api-version=preview for /openai/v1/responses (#22)
Reproduced the probe failure locally against the same Azure resource:
all date-based api-versions (2024-08-01-preview, 2024-12-01-preview,
2025-01-01-preview, 2025-03-01-preview, 2025-04-01-preview) return:
{"error":{"code":"BadRequest","message":"API version not supported"}}
Only api-version=preview and api-version=v1 succeed (200). This is the
new Azure OpenAI v1 responses-API style; date strings are reserved for
the legacy /openai/deployments/{model}/chat/completions endpoint.
@ai-sdk/azure 3.x already defaults apiVersion to "preview" (per the
type definition: "Custom api version to use. Defaults to `preview`."),
so this aligns the workflow with the SDK default. Probe should now
return MODEL_CAPABLE=true and the send_message scenario will run.
* test(cua): extend send_message and multi_turn waits to 30s
Assistant bubbles can take 15+ seconds to appear after send. Previous
5-second wait was too short and caused false failures even when API
calls succeeded. Re-check screenshots periodically up to 30s total.
* fix(cua): screen-relative send button threshold for #22
The send action's auto-locate filtered for y1 > 2200 and fell back to
hardcoded (996, 2358) — both assume a 1080x2400 panel. The CI emulator
(API 30 google_apis pixel profile) is 1080x1920, so:
- the bottom_buttons filter never matched any clickable element
- the fallback tap landed off-screen
→ 'ping' message never sent, scenario timed out with no bubbles.
Switch to a screen-relative threshold (bottom 25%) and a fallback that
uses get_screen_size() to land in the bottom-right corner regardless of
device resolution. This was masked until now because send_message was
gated by MODEL_CAPABLE=false in earlier CI runs.
Refs: #22
---------
Co-authored-by: dzianisv <dzianis.varabyou@gmail.com>
The CI opencode server had NO LLM provider configured — the server log only
showed "listening", never a model. opencode-ai (released npm pkg) does not read
AZURE_OPENAI_* for its own LLM; it needs an explicit provider in opencode.json +
a default `model`. So send_message/multi_turn could never pass and the gate was
stuck on --only-connect-scenario (UI journey minus the model reply).
- Wire opencode to the same Azure resource the CUA driver uses via a generated
~/.config/opencode/opencode.json (@ai-sdk/azure provider, resourceName derived
from the endpoint secret at runtime, apiKey from env, default model azure/gpt-5.4).
- Add a deterministic REST probe step: create a session + send a prompt and check
for an assistant reply BEFORE the ~30min emulator run, exporting MODEL_CAPABLE.
- Add --scenarios to android-cua-smoke.py to run an explicit named set.
- Emulator step now runs connect_and_verify_sessions + send_message +
verify_session_list when MODEL_CAPABLE=true; falls back to the UI-only journey
(connect + verify_session_list) otherwise, logging the environmental reason.
- Raise --max-steps to 40 so multiple scenarios fit.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* chore: repoint OpenCode links to agentlabs.cc/opencode
agentlabs.cc/opencode and /opencode/privacy are now live (200). Repoint
README, distribution listings (Play/App Store/F-Droid/IzzyOnDroid/iOS),
docs, and in-app privacy links (settings + telemetry consent) from
www.vibebrowser.app/opencode to the canonical agentlabs.cc hub.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(ci): run local opencode server for CUA smoke true-E2E (#15)
GitHub-hosted runners can't reach the Tailscale dev server
(100.108.64.76:4096), so the CUA smoke always failed at session creation.
- Install opencode-ai and run `opencode serve` on the runner host; the
Android emulator reaches it via 10.0.2.2. OPENCODE_URL now points there.
- Healthcheck /global/health before launching the app; dump server log on
failure for diagnosis.
- Add --only-connect-scenario to the smoke script and run just the
connect-and-verify-sessions path in CI: deterministic, needs no model
backend. The scenario now creates a session if the list is empty, so a
fresh server still yields a non-empty list.
This makes the smoke a true E2E and also exercises the #10 sessions-list
rendering path against a real server.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(ci): emulator smoke script is dash, not bash — drop brace-group healthcheck
android-emulator-runner runs the script: block under /usr/bin/sh (dash). The
multi-line `|| { ...; }` healthcheck was a dash syntax error (end of file
unexpected), failing the step before the smoke ran. Replace with a non-fatal
one-line re-check; the server was already health-gated in the prior step.
* docs(tasks): record smoke CI round 1 failure + dash fix
---------
Co-authored-by: engineer <engineer@opencode.ai>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Rename commit missed the Python launcher constant; HEAD still targeted
the old package so the smoke could not find the installed app.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(sessions): use active connection client directly, remove roots filter
Root cause A: loadSessions was calling clientForDirectory(serverHome) which
scoped the session list to /home/azureuser — a different project than the
server's active CWD. Sessions in the current project (e.g. opencode-mobile)
were never returned.
Root cause B: roots:true filtered out sessions that have a parentID (sub-task /
AUTO-REVIEW sessions), hiding valid sessions from the list.
Fix: use connState.client directly (the connection's active directory) and drop
the roots filter so all sessions for that project are visible.
Also adds a verify_session_list CUA smoke scenario that navigates back to the
sessions tab after creating a session and asserts the list is non-empty —
covering the regression path that was previously untested.
* fix(sessions): fetch serverHome in addConnection so loadSessions shows correct sessions
Root cause: addConnection() built the HTTP client but never fetched serverHome
(only loadConnections and setActiveConnection did). When the user adds a new
connection (fresh install / first sign-in), serverHome = null, so loadSessions
fell through to connState.client (the server's CWD). On this dev server the CWD
is the deploy directory — 11 old May-19 sessions that are not the user's recent
work sessions.
Fix: addConnection now fetches currentProject + serverHome via the same
Promise.all as setActiveConnection, before calling set(). This ensures
loadSessions immediately uses clientForDirectory(serverHome) → the global
project → the user's actual recent parent sessions.
Also adds --opencode-url flag to the CUA smoke script, which appends a
connect_and_verify_sessions scenario that reproduces the regression:
python scripts/android-cua-smoke.py --opencode-url http://100.108.64.76:4096
* fix(sessions): recover home scope after fresh connect
Resolve stale deploy-only session list by recovering server home during first load and keeping regression coverage in default Android CUA smoke and CI.
* chore(release): bump version to 0.4.0
- Record each scenario as MP4 via ADB screenrecord in a background thread
- Pull video to /tmp/cua_<scenario>.mp4 after test completes (always)
- Upload to ArchiveBox if ARCHIVEBOX_URL + ARCHIVEBOX_API_KEY env vars set;
gracefully skips when not configured (CI default)
- Upload all artifacts (PNG screenshots + MP4 videos) always, not only on failure
- Pass ARCHIVEBOX_URL/ARCHIVEBOX_API_KEY secrets to workflow (optional)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(security): fail closed on biometric init error
H-03: setting isAuthenticated: true on initialization failure was a
security bypass — any crash during biometric setup granted full access.
Fail closed instead; user sees auth prompt on next open.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(security): use Crypto.randomUUID for connection IDs
H-04: Math.random() is not cryptographically random. Connection IDs are
used as SecureStore key suffixes; switch to expo-crypto randomUUID for
a secure source.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(deps): pin expo-crypto to ~15.0.9
15.0.10 does not exist on npm; ~15.0.9 is the latest stable in the 15.x series compatible with Expo SDK 54.
* feat: add OpenCode Connect coming-soon waitlist card
Adds a discoverable 'OpenCode Connect — Coming Soon' card to the
add-connection quick-connect screen. Users can enter their email and
tap 'Join Waitlist' to send a pre-filled mailto. No backend required.
* fix(cua): detect actual screen dimensions and fix JSON parsing
- Get real screen size via `wm size` instead of hardcoding 1080x2400;
emulator is 1080x1920 so y-coordinates were systematically off
- Extract first JSON object via regex when model returns multiple objects
- Use AZURE_OPENAI_MODEL env var for deployment name (defaults gpt-5.4)
- Add AZURE_DEV_AI_* path for Azure AI Foundry endpoints
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(security): SHA-pin upload-google-play and sanitize notification bodies
M-02: Pin r0adkll/upload-google-play to commit SHA e738b9d (v1.1.5)
to prevent supply-chain hijack via tag mutation.
M-03: Sanitize all push notification bodies — strip control chars,
truncate to 200 chars. Prevents server-supplied strings (error messages,
file paths from permission patterns, session titles) from leaking
unbounded text into the OS notification drawer.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(privacy): add telemetry consent gate for Sentry crash reporting
Sentry was always-on, violating F-Droid anti-feature policy and user
trust norms. Now gated behind explicit opt-in:
- First-launch consent modal (TelemetryConsentModal) shows once on
fresh install; user can Allow or Decline.
- Consent state persisted in expo-secure-store (survives restarts).
- Settings > Privacy section: crash reporting toggle + privacy policy link.
- initSentry() called only after consent granted — not on app start.
Closes#3 (partial)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(config): add real icons and complete iOS/Android app.json config
- Add 1024×1024 app icon, 432×432 adaptive icon foreground, 200×200 splash
- iOS: push notification entitlement (aps-environment: production), speech/
microphone/camera/photo usage descriptions for future features, disable
ITSAppUsesNonExemptEncryption
- Android: adaptive icon with dark background (#0F172A), versionCode: 1
- expo-notifications plugin wired in app.json
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat(dist): add iOS CI workflow, README rewrite, CONTRIBUTING, and LICENSE
- publish-app-store.yml: EAS Build + TestFlight submission; runs on tag/release/
workflow_dispatch; bumps ios.buildNumber from github.run_number
- README: full rewrite — features, install badges, connection guide, contributing
- CONTRIBUTING.md: contribution guide for OSS contributors
- LICENSE: MIT
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* docs(dist): add store listings, strategy, privacy policy, F-Droid/IzzyOnDroid templates
- distribution/strategy.md: monetization strategy (free client + opencode Cloud)
- distribution/play-listing.md: Google Play store copy (name, description, tags)
- distribution/app-store-listing.md: App Store listing copy
- distribution/privacy-policy.{md,html}: GDPR-compliant privacy policy
- distribution/PLAY_CONSOLE_SETUP.md: Play Console setup runbook
- distribution/ios-enrollment-runbook.md: Apple Developer Program enrollment steps
- distribution/SIGNING-KEY-FINGERPRINTS.md: keystore fingerprint for reproducible builds
- distribution/fdroid-submission/: F-Droid metadata template
- distribution/izzyondroid-submission/: IzzyOnDroid submission template
- distribution/whatsnew/: Play Store release notes (en-US)
- distribution/whatsnew-ios/: TestFlight release notes
- distribution/play-graphics/: Play Store screenshot placeholders
- distribution/app-store-graphics/: App Store screenshot placeholders
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(telemetry): handle SecureStore failure + Android back button
- add .catch() on loadTelemetryConsent() so SecureStore rejection
shows the consent modal instead of blocking startup forever
- add onRequestClose={onDecline} to Modal so Android back button
records the decline rather than silently dismissing
- fix catch block in telemetry.ts to not clobber _resolved when
SecureStore read fails mid-session
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(ci): run gradlew clean to prevent stale modules.json duplicate
Sentry Gradle plugin writes modules.json to src/main/assets; cached
build intermediates contain an old copy → mergeReleaseAssets fails
with 'Duplicate resources'. Running clean before assembleRelease
clears the intermediate state.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix(ci): remove android build output cache causing duplicate modules.json
Caching android/app/build/intermediates and android/app/.cxx causes
two issues:
1. Stale modules.json in intermediates → Duplicate resources error
2. .cxx CMake artifacts reference absolute paths → ninja clean fails
Keeping only Gradle distribution cache (~/.gradle) which is safe.
Expo prebuild regenerates android sources fresh each run anyway.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
- Added 'send' action type that uses uiautomator XML to find the rightmost
clickable element in the bottom input bar (the send button)
- Added screen resolution (1080x2400) to LLM context for better coordinate estimation
- Updated system prompt to instruct model to use 'send' action instead of manual tap
- Updated scenarios with clearer step-by-step instructions
- Both send_message and multi_turn scenarios pass reliably
