feat(coding): add Mobile selector-patterns skill

[NicolasMassart]

Description

Opt-in coding/selector-patterns skill with a MetaMask Mobile overlay. Agents get a description-triggered guide for authoring and consuming Redux selectors: named select<Feature><Thing> selectors, createSelector / createDeepEqualSelector, leaf reads of state.engine.backgroundState, useSelector(selectX) in UI, and collocated tests with at least two state variants. Version-gated flag selectors stay on feature-flags. Memoization audits stay on performance.

Install/use pointers were added on the Mobile overlays for coding-guidelines, controller-integration (step 11 rewritten so object/array outputs can use createDeepEqualSelector), and performance.

MCWP-472 asked for an always-on Cursor rule (alwaysApply: true). This repo ships opt-in skills, so discovery is the description plus those install/use pointers.

Fixes: https://consensyssoftware.atlassian.net/browse/MCWP-472

Type of Change

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

Skill Details (if adding a new skill)

Provider Name: MetaMask
Skill Name: selector-patterns
Brief Description: Redux selector authoring for MetaMask Mobile: named select<Feature><Thing> selectors in app/selectors/, createSelector / createDeepEqualSelector, leaf reads of state.engine.backgroundState, and useSelector(selectX) in UI. Use when adding, updating, or asking about selectors.

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

• yarn audit:skills — coding/selector-patterns had no schema errors
• ./tools/install --repo metamask-mobile --include coding/selector-patterns
• Two mobile-skill-consumer product tickets (Contacts address-book count; Onboarding password-set store read). Product diffs were discarded after capture.

Important

Skill-test report from two Mobile consumer runs : skill-test-selector-patterns.local.md

Additional Context

CODEOWNERS: @MetaMask/extension-platform @MetaMask/mobile-platform @MetaMask/core-platform (/domains/coding/, /domains/performance/)

Made with Cursor

[andrepimenta]

Review

Checked out the branch, ran the linter and installer dry-runs, and verified every Mobile claim against a metamask-mobile checkout.

Verified clean

• yarn audit:skills — 54 skills, 0 errors (only the repo-wide "recommended section" warnings, which this skill doesn't trigger).
• Installer gating works: dry-run installs mms-selector-patterns for --repo metamask-mobile, correctly skipped for metamask-extension (no overlay), so the Mobile paths in the base skill.md don't leak.
• Naming claim holds: app/selectors has 528 select* vs 48 get* exports. "Existing get* names stay" is the right call.
• --no-coverage is correct — jest.config.js:58 sets collectCoverage: true by default.
• Paths all check out: app/selectors/util.ts → createDeepEqualSelector, app/selectors/<feature>/index.ts, app/components/UI/<Feature>/selectors/.
• Omitting maturity is fine — installer defaults to stable (tools/install:240) and 34 of 54 existing skills omit it. Content-only change, so no CHANGELOG entry needed.

Blockers

1. The example snippet doesn't compile. In repos/metamask-mobile.md, EMPTY_ADDRESS_BOOK_CHAIN: readonly never[] unions with Object.values(...) to readonly never[] | AddressBookEntry[], which fails at any consumer expecting a mutable array — confirmed with tsc:

error TS2345: Argument of type 'readonly never[] | Entry[]' is not assignable to parameter of type 'Entry[]'.
  The type 'readonly never[]' is 'readonly' and cannot be assigned to the mutable type 'Entry[]'.

Repo precedent avoids this deliberately: const EMPTY_TOKENS: Token[] = Object.freeze([] as Token[]) as Token[] (tokensController.ts:130). The object constant (Readonly<Record<string, never>>) matches EMPTY_TOKENS_BY_ADDRESS and is fine — only the array form needs changing.

2. The flagship example violates the skill's own rules. selectAddressBook returns an object on a plain createSelector and is an identity passthrough of a controller slice. The Factory-choice table forbids the first (object output → createDeepEqualSelector); the Reject list forbids the second ("Identity createSelector on a controller slice"). An agent that internalises the table will flag the example it was told to copy. Either change the factory or give the rule its nuance — a sub-key read plus a stable empty constant is arguably fine.

3. "Copy from app/selectors/addressBookController.ts" points at a file that breaks two Reject rules. The real file uses inline ?? {} and inline return [], and addressBookController.test.ts has one state variant, not the two the skill requires. The overlay then shows a rewritten version of that file as if it were its contents. An agent following the instruction reads the anti-pattern; one copying the block writes something that doesn't match the file, or "fixes" the real file as a drive-by. tokensController.ts genuinely has the pattern — make it the sole exemplar and label the addressBook block as illustrative.

4. The Factory table contradicts the performance skill this PR also edits. mm-selector-memoization.md:111: "createDeepEqualSelector is not free: deep-comparing a huge slice every dispatch can itself be costly. Prefer narrowing the input over deep-equalizing a giant object." The new table makes deep-equal unconditional for object/array/Map/Set output and drops narrowing entirely — the cheaper fix, and the one that matters most in the power-user scenario the perf skill is built around. Suggest: narrow the input first, reach for createDeepEqualSelector when the input slice churns and the payload is small enough to compare.

1 and 2 make agents emit bad code; 3 and 4 make them emit inconsistent code.

Non-blocking

• Circular pointer: inside selector-patterns/repos/metamask-mobile.md — "Engine wiring that needs a first selector: install/use coding/selector-patterns from controller-integration" points at itself. Presumably meant "Engine wiring: controller-integration".
• "Install/use" isn't actionable: no other overlay cross-references a not-yet-installed skill this way, and none of the four pointers carries the command. README documents yarn skills --include coding/selector-patterns --save — worth spelling out at least once. Placing the pointers in three base: true skills is otherwise a sound answer to the always-on gap the description notes.
• selectTokensByChainId doesn't exist. The naming example cites it; the real symbols are selectTokensByChainIdAndAddress / selectTokensByAddress. Agents will grep for it.
• Scope the backgroundState Reject rule to new code. 20 non-test files under app/components/app/util/app/hooks read state.engine.backgroundState today. The Requirements section is correctly scoped to "New and changed selectors"; the Reject list isn't, which invites refactor sprees.

On CONTRIBUTING's "is it a duplicate?" test

Overlap with performance is real — the Reject list is essentially the four deadly patterns from mm-selector-memoization.md. But the authoring-vs-audit split is stated explicitly, the pointers go both ways, and the net-new content (file location, naming, test shape) isn't in the perf skill. It earns its slot. Consider linking to mm-selector-memoization.md rather than restating the anti-patterns — the duplicated list is what will drift.

[NicolasMassart]

Addressed the review on this follow-up commit.

Blockers

1. Array empty constant now matches EMPTY_TOKENS (Object.freeze([] as Token[]) as Token[]).
   2–3. tokensController.ts is the sole copy-from (selectTokensByAddress, selectAllTokensFlat, selectTokensLength, stable empties). The leaf is labeled illustrative. Factory choice allows a sub-key read plus a stable empty constant on createSelector; identity wrapping of a whole churning slice does not.
2. Factory table: narrow the input first; createDeepEqualSelector when the narrowed input still churns and the payload is small enough to compare. Workflow step 4 and Requirements match. Authoring Reject list stays short; audits still point at performance / mm-selector-memoization.md.

Nits

• Overlay intro: Engine wiring → controller-integration.
• yarn skills --include coding/selector-patterns --save on this overlay and the three discovery pointers.
• Naming example: selectTokensByChainIdAndAddress.
• backgroundState Reject scoped to new and changed UI.

[NicolasMassart]

Follow-up on CONTRIBUTING’s “is it a duplicate?” note.

Reject now keeps authoring-only bullets (UI backgroundState, inline useSelector, isEqual band-aids, flag evaluation). Identity wrappers, new refs in result functions, mutation, and huge inputs point at performance (mm-selector-memoization.md) instead of restating that catalog. Factory choice keeps the table and the sub-key + empty-constant nuance; identity-passthrough / copy-before-sort live in the memoization reference.