[cryptotavares]

Description

Updates the Mobile testing/mobile-visual-testing skill for @metamask/client-mcp-core 0.13.0 (release PR #80) and @metamask/device-mcp 0.7.0. With 0.13.0, agents that follow the current skill fail on step 2: mm knowledge-search now returns Error: unknown command 'knowledge-search'.

This is the Mobile counterpart of #210 (Extension visual-testing skill).

Area	Before	After
Knowledge store	"Reuse Knowledge" workflow step; mm knowledge-* in the CLI reference, error recovery and verify loop; --goal / --flow-tags launch flags	Step, commands and flags removed; the CLI reference notes the removal and the test-artifacts/<sessionId>/session.json path
Observations	"Mutating commands can return compact observations"	Global --observe none|diff|full flag, per-category defaults (reads, discovery and cdp return none), stale refs after --observe none, only mutating/batch calls wait for animations, no app state in observations, flat mobile a11y tree without status/nav bars and dev overlay
get-state	Listed both as available and as "E2E-context only, do not use"	Documents the real mobile response: isLoaded, extensionId, source: "stub". No currentScreen and no wallet fields; identify the screen from test IDs
Live-run workarounds	Not documented	iOS: click --testid Done instead of dismiss-keyboard; tap-coordinates for wallet home Buy / Swap / Send / Receive (missing from the iOS 18+ axbridge tree). Android: press-button back to hide the keyboard. Both: wait-for after a slow navigation, because the observation can still show the previous screen
Output	Not documented	One-line JSON when not on a terminal; --pretty / MM_PRETTY; meta.timing on every response; node_modules/.bin/mm to skip Yarn startup
run-steps	includeObservations; batchTimeoutMs without limits	Top-level observe (replaces includeObservations); batchTimeoutMs range 1000–30000 and the 30 s request limit
Timeouts	Not documented	Timed-out calls are cancelled, but an action that already started can still have happened
Android snapshots	auto described as the default	helper-first is the default since device-mcp 0.6.0; auto / instrument / dump semantics and first-snapshot install time updated

Files:

• domains/testing/skills/mobile-visual-testing/repos/metamask-mobile.md: workflow (knowledge step removed, observation guidance), batching, gotchas, Android snapshot modes
• domains/testing/skills/mobile-visual-testing/references/cli-reference.md: new Global Options section; knowledge commands and launch flags removed; get-state, run-steps and cdp updated
• domains/testing/skills/mobile-visual-testing/references/error-recovery.md: knowledge step removed; new failure rows
• CHANGELOG.md: Unreleased entry

Companion Mobile PR: MetaMask/metamask-mobile#38594

Type of Change

• New skill
• Skill improvement/update
• Bug fix
• Documentation update
• Other (please describe):

Checklist

• I have read the CONTRIBUTING.md guidelines
• My skill follows the SKILL_TEMPLATE.md format
• I have tested this skill with an AI agent
• My skill does not contain any secrets, private keys, or sensitive data
• I have added appropriate documentation
• My changes don't break existing skills

Testing

• Behavior statements were checked against client-mcp-core release/0.13.0 (mm --help, MobilePlatformDriver.getAppState) and the installed @metamask/device-mcp 0.7.0 README.
• Live sessions on the Mobile PR branch (core 0.13.0 via yalc, device-mcp 0.7.0, Metro attached) on an iOS Simulator (iPhone 17 Pro) and an Android emulator (API 36): launch, describe-screen, get-state ({"isLoaded":true,"extensionId":"...","source":"stub"}), type (diff observation), get-text (no observation), click --observe none|full, run-steps with observe and batchTimeoutMs, includeObservations and batchTimeoutMs > 30000 rejected, --pretty, cdp / hermes-targets with --metro-port, screenshot, list-testids, cleanup; knowledge-* rejected as unknown commands. Results are in chore(llm-workflow): upgrade to @metamask/client-mcp-core 0.13.0 and @metamask/device-mcp 0.7.0 metamask-mobile#38594.
• Unlocked read-only flows on both devices (unlock, wallet home, account picker, open and close Send, cdp isUnlocked()); nothing was sent or signed.
• Found while testing, all the same before 0.13.0, now documented in the skill:
  • mm cdp does not reuse the launch's --metro-port (it defaults to 8081). Pass --metro-port on each cdp / hermes-targets call.
  • Mobile get-state never returns currentScreen.
  • iOS 18+: the wallet home action buttons are missing from the axbridge tree.
  • iOS: mm dismiss-keyboard fails (idb ui key RETURN rejected).
  • --testid matches a substring (device-mcp fuzzyMatch): on Android wallet home, account-picker taps account-picker-avatar and opens the profile screen.
• yarn lint passes and yarn test passes (73/73).
• yarn audit:skills reports 1 error, in typescript/tsc-blindspots (description length). The same error is on main; this PR adds no new errors.

Additional Context

Merge this together with MetaMask/metamask-extension#47526. Until then, agents in the Extension are on 0.12.0, which still has the knowledge commands.

[pnarayanaswamy]

Is this reference to knowledge store necessary for agents running older versions? We anyway now return an error when trying to reference the knoweldge store.