xstate-v5（状态 v5）

xstate-v5 (Strict) Skill

Build, refactor, and review XState v5 state machines (TypeScript and React) using a strict, strongly-typed ruleset.

Source of truth: references/xstate-v5-rules.md (this repo, resolved from the skill root). If anything in this skill conflicts with the rules file, follow the rules file.

Inlined Non-Negotiable Contract (Agent Must Follow)

These requirements are intentionally duplicated here so an agent can comply without first loading external references:

• Design first, code second:

- Produce planning artifacts before implementation (state inventory, event catalog, transition table, async/actor map, acceptance-test scenarios). - Include a boundary/decomposition decision record (what is split vs orchestrated, and why).

• No god-machine architecture:

- Unrelated domains (auth, toasts, navigation, workflows, transport/retry policy) must not be fully modeled in one machine. - A single machine is acceptable only as a thin orchestration root or tightly coupled app-shell parallel regions.

• No state-mirroring context flags:

- Do not duplicate mode in context booleans (isLoading, isAuthenticated, etc.) when state.value already represents the mode. - If a temporary migration flag exists, document rationale and removal plan.

• XState v5 strictness:

- Use setup({...}).createMachine({...}). - Implement actions/guards/actors in setup(...). - Send event objects only. - Avoid forbidden v4 patterns (interpret, Machine, cond, send, pure, choose, etc.).

• Enforce explicit async and actor boundaries:

- Request/response: invoke with onDone and onError. - Long-lived collaborators: spawnChild/stopChild and explicit routing via sendTo.

Fail-closed rule:

• If any detail is ambiguous, load references/xstate-v5-rules.md before writing or reviewing machine code.
• If that file is missing/unreadable, stop and report the issue instead of guessing.

When To Use

Use this skill when the task involves:

• XState v5 machines/actors (xstate, @xstate/react)
• Refactoring XState code to be more type-safe and idiomatic v5
• Migrating XState v4 patterns (Machine/interpret/cond/send/pure/choose/etc.) to v5 equivalents
• Planning/designing statecharts prior to implementing them as XState v5 machines
• Designing actor boundaries (invoke vs spawnChild) and React integration via createActorContext

Hard Requirements (Enforced)

• Prefer setup({...}).createMachine({...}) for all machines.
• Implement all actions, guards, and actors in the setup({...}) object.
• Prefer passing event-derived data via typed params to actions/guards.
• Never send string events. Always send event objects: actor.send({type: '...'}).
• Forbid XState v4 legacy APIs and patterns listed in references/xstate-v5-rules.md.
• Do not model unrelated domains (auth, notifications/toasts, navigation, data lifecycle, etc.) inside one machine, except a thin orchestration machine.
• Do not mirror state in context booleans (isLoading, isAuthenticated, etc.) when state.value already represents that mode.

Rules Table Of Contents

• Statechart Design: 0 Statechart Design (Before You Code)
• Core Principles: 1 Machine Creation Pattern, 2 Type Safety First
• Setup Object Rules: 3 Implementation Placement, 4 Parameter Typing
• Event Handling Rules: 5 Event Type Safety with assertEvent, 6 Event Object Requirement
• Deprecated Pattern Prevention: 7 Forbidden v4 Patterns, 8 Modern v5 Equivalents
• Context and Input Rules: 9 Context Initialization, 10 Context Updates
• Invoke and Actor Rules: 11 Invoke Configuration, 12 Actor Spawning
• Type Helper Rules: 13 Type Helpers
• Testing Rules: 14 Type-Safe Testing (14.1 Deterministic Actor Tests, 14.2 Model-Based Testing, 14.3 Graph Utilities)
• XState React Rules: 15 Shared State via createActorContext
• Best Practices Summary: 16 Code Organization, 17 Performance Considerations, 18 Error Handling, 19 Documentation
• Enforcement Rules for AI Agents: 20 Mandatory Patterns, 21 Quality Assurance, 22 When Docs Are Ambiguous

Quick Start

Minimal typed machine skeleton:

import { setup, assign } from "xstate"

type Ctx = { count: number }

type Ev = { type: "inc" } | { type: "add"; amount: number } | { type: "reset" }

type Input = { initialCount?: number }

export const counterMachine = setup({
  types: {
    context: {} as Ctx,
    events: {} as Ev,
    input: {} as Input,
  },
  actions: {
    inc: assign({ count: ({ context }) => context.count + 1 }),
    add: assign({
      count: ({ context }, params: { amount: number }) => context.count + params.amount,
    }),
    reset: assign({ count: ({ input }) => input.initialCount ?? 0 }),
  },
}).createMachine({
  id: "counter",
  context: ({ input }) => ({ count: input.initialCount ?? 0 }),
  on: {
    inc: { actions: "inc" },
    add: {
      actions: {
        type: "add",
        params: ({ event }) => ({ amount: event.amount }),
      },
    },
    reset: { actions: "reset" },
  },
})

React shared state skeleton:

import { createActorContext, shallowEqual } from "@xstate/react"
import { type SnapshotFrom } from "xstate"
import { counterMachine } from "./counterMachine"

const CounterCtx = createActorContext(counterMachine)

export const CounterProvider = CounterCtx.Provider
export const useCounterSelector = CounterCtx.useSelector
export const useCounterActorRef = CounterCtx.useActorRef

const selectCount = (s: SnapshotFrom<typeof counterMachine>) => s.context.count
export const useCount = () => useCounterSelector(selectCount)

const selectCtx = (s: SnapshotFrom<typeof counterMachine>) => s.context
export const useCounterContext = () => useCounterSelector(selectCtx, shallowEqual)

Workflow (What To Do Each Time)

1. Read references/xstate-v5-rules.md first.
2. Identify the target surface area:

- Plain machine/actor (non-React) - React integration (createActorContext) or local useActor - Migration from v4 patterns

1. Design the statechart first:

- See references/xstate-v5-rules.md#0-statechart-design-before-you-code. - Write down: state inventory, event catalog (payload + source), transition table, async/actor boundaries, and acceptance-test scenarios. - Record a boundary/decomposition decision before type design: - Which concerns belong in separate actors/machines (based on lifecycle, owner, and failure mode). - Which concerns are intentionally orchestrated at top-level (and why). - Any exception note if keeping logic in one machine. - Default: do this unless the user explicitly asks to skip planning.

1. Design types first:

- types.context, types.events, and types.input (and types.output if applicable) - Use zod schemas only if the codebase already uses Zod or the user asks for it.

1. Put implementations in setup({actions, guards, actors}):

- Actions/guards take typed params when they need event-derived data. - Only read event-specific fields inside implementations when necessary; use assertEvent then.

1. Choose async boundaries:

- Prefer invoke with typed input for request/response flows. - Use spawnChild/stopChild for long-lived child actors.

1. Use v5 runtime APIs:

- createActor(machine) (not interpret) - raise / sendTo (not send action) - enqueueActions (not pure/choose)

1. Validate:

- tsc passes; no any leaks in params. - No string event sends. - No v4 forbidden imports or config keys (cond, withContext, withConfig, etc.). - Prefer waitFor for async actor tests and xstate/graph (createTestModel, getShortestPaths, etc.) for model/graph-driven test generation.

When A Monolith Is Acceptable

Use a single top-level machine only when it is a thin orchestration root that coordinates child actors or tightly coupled parallel app-shell concerns. Even then, unrelated domain logic should live in separate actors/machines and communicate via explicit events.

Review Checklist (Use When Auditing PRs)

• All machines are setup(...).createMachine(...).
• Machine has a clear statechart plan (states/events/transition table) or the PR description includes it.
• No domain-smell: one machine has one responsibility or explicit orchestration scope.
• No state-mirroring context flags without documented rationale.
• No implementations inside types.
• actions/guards read event fields only via params or assertEvent.
• No interpret, Machine, withConfig, withContext, cond, send, pure, choose.
• invoke has onError (and onDone when appropriate).
• React usage prefers createActorContext at module scope; selectors are stable and use shallowEqual for objects.

Common Fixes

• Event payload used inside an action:

- Prefer: compute payload in params and make the action implementation depend on params. - Otherwise: add assertEvent(event, 'someType') inside the implementation.

• v4 cond:

- Rename to guard and move guard implementation into setup({guards: {...}}).

• v4 interpret:

- Replace with createActor(machine) and .start() where needed.

Notes

If official docs feel ambiguous, follow the rules file guidance: confirm behavior against XState v5 source, issues/discussions, or a minimal reproduction/type-test rather than guessing.