event-modeling（事件建模）

Event Modeling

Value: Communication -- event modeling is a structured conversation that surfaces hidden domain knowledge and creates shared understanding between humans and agents before any code is written.

Purpose

Teaches the agent to facilitate event modeling sessions following Martin Dilger's "Understanding Eventsourcing" methodology. Produces a complete event model (actors, events, commands, read models, automations, slices) that drives all downstream implementation. The model lives in docs/event_model/.

Practices

Two-Phase Process: Discovery Then Design

Never jump into detailed workflow design without broad domain understanding first. Phase 1 maps the territory; Phase 2 explores each region.

Phase 1 -- Domain Discovery. Identify what the business does, who the actors are, what major processes exist, what external systems integrate, and which workflows to model. Ask these questions of the user; do not assume answers. Output: docs/event_model/domain/overview.md.

Phase 2 -- Workflow Design. For each workflow, follow the 9-step process. You MUST follow references/nine-steps.md for the full methodology. Design one workflow at a time. Complete all 9 steps before starting the next workflow. Output: docs/event_model/workflows/<name>/overview.md plus individual slice files in slices/.

The Prime Directive: Not Losing Information

Store what happened (events), not just current state. Events are immutable past-tense facts in business language. Every read model field must trace back to an event. If a field has no source event, something is missing from the model.

Event Design Rules

1. Name events in past tense using business language: OrderPlaced, not PlaceOrder or CreateOrderDTO
2. Events are immutable facts -- never modify or delete
3. Include relevant data: what happened, when, who/what caused it
4. Find the right granularity -- not DataUpdated (too broad) and not FieldXChanged (too narrow)
5. Commands depend on user inputs and the event stream, not read models. Read models serve views and automations only.
6. Events record domain facts (true on any machine). Runtime context (file paths, hostnames, PIDs, working directories) does not belong in event data.

The Four Patterns

Every event-sourced system uses these patterns. Each pattern maps to one vertical slice.

1. State Change: Command -> Event. The only way to modify state. A command may produce multiple events as part of a single operation.
2. State View: Events -> Read Model. How the system answers queries. When the domain supports concurrent instances, use collection types in read model fields, not singular values. Commands derive their inputs from user-provided data and the event stream — never from read models. No ReadModel → Command edges should appear in diagrams. If a command needs to check whether something already happened (e.g., idempotency), it checks the event stream, not a read model. Read models represent meaningful domain projections. Infrastructure preconditions ("does directory exist?", "is service running?") that are implicit in the command's execution context do not need their own read model.
3. Automation: Event -> Read Model (todo list) -> Process -> Command -> Event. Background work triggered by events. Requires all four components: triggering event, read model consulted, conditional process logic, and resulting command. If there is no read model and no conditional logic, it is NOT an automation — it is a command producing multiple events. Must have clear termination conditions.
4. Translation: External Data -> Internal Event. Anti-corruption layer for workflow-specific external integrations. Generic infrastructure shared by all workflows (event persistence, message transport) is NOT a Translation — it is cross-cutting infrastructure that belongs outside the event model.

Required Layers Per Slice Pattern

Each slice pattern implies a minimum set of architectural layers. A slice is not complete until all required layers are implemented and wired together.

• State View: infrastructure (read events/data from store) + domain (projection/query logic) + presentation (render or return result to caller) + application wiring (connect layers end-to-end)
• State Change: presentation (accept user input or external request) + domain (command validation and business rules) + infrastructure (persist resulting events/data) + application wiring (connect layers end-to-end)
• Automation: infrastructure (detect triggering condition — timer, external event, threshold) + domain (policy/decision logic) + infrastructure (execute resulting action — send message, write data, call service) + application wiring (connect trigger to policy to action)
• Translation: infrastructure (receive from external system) + domain (mapping/transformation logic) + infrastructure (deliver to target system) + application wiring (connect inbound adapter to mapper to outbound adapter)

When decomposing a slice, verify that your acceptance criteria and task breakdown cover every required layer. A slice that only implements domain logic without presentation or infrastructure is incomplete — it is a component, not a vertical slice.

GWT Scenarios

After workflow design, generate Given/When/Then scenarios for each slice. These become acceptance criteria for implementation.

Command scenarios: Given = prior events establishing state. When = the command with concrete data. Then = events produced OR an error (never both).

View scenarios: Given = current projection state. When = one new event. Then = resulting projection state. Views cannot reject events.

Critical distinction: GWT scenarios test business rules (state-dependent policies), not data validation (format/structure checks that belong in the type system). If the type system can make the invalid state unrepresentable, it is not a GWT scenario.

You MUST use references/gwt-template.md for the full scenario format and examples.

Application-Boundary Acceptance Scenarios

Every vertical slice MUST include at least one GWT scenario defined at the application boundary:

• Given: The system is in a known state (prior events, seed data, configuration)
• When: A user (or external caller) interacts through the application's external interface — the specific interface depends on the project (HTTP endpoint, CLI command, message queue consumer, UI action, etc.)
• Then: The result is observable at that same boundary — a response, output, rendered state change, emitted event, etc.

A GWT scenario that can be satisfied entirely by calling an internal function in a unit test describes a unit-level specification, not a slice acceptance criterion. Slice acceptance criteria must exercise the path from external input to observable output.

Acceptance Test Strategy

Where the application boundary is programmatically testable — HTTP endpoints, CLI output parsing, headless browser automation, message queue assertions, API contract tests, etc. — write automated acceptance tests that exercise the full GWT scenario from external input to observable output. These tests provide fast feedback and serve as living documentation of slice behavior.

Where automated boundary testing is not feasible (complex GUI interactions, hardware-dependent behavior, visual/aesthetic verification), document what the human should manually verify: the specific steps to perform and the expected observable result. This manual verification checklist becomes part of the slice's definition of done.

Slice Independence

Slices sharing an event schema are independent. The event schema is the shared contract. Command slices test by asserting on produced events; view slices test with synthetic event fixtures. Neither needs the other to be implemented first. No artificial dependency chains between slices.

Model Validation

After GWT scenarios are written, validate the model for completeness:

1. Every read model field traces to an event
2. Every event has a triggering command, automation, or translation
3. Every command has documented rejection conditions (business rules)
4. Every automation has a termination condition
5. GWT Given/When/Then clauses do not reference undefined elements

When gaps are found, ask the user to clarify, create the missing element, and re-validate. Do not proceed with gaps remaining.

Facilitation Mindset

You are a facilitator, not a stenographer. Ask probing questions. Challenge assumptions. Keep asking "And then what happens?" after every event, every command, every answer. Use business language, not technical jargon. Do not discuss databases, APIs, frameworks, or implementation during event modeling. The only exception: note mandatory third-party integrations by name and purpose.

Do:

• Follow all steps in order -- the process reveals understanding
• Ask "And then what happens?" relentlessly
• Use concrete, realistic data in all examples and scenarios
• Design one workflow at a time
• Ensure information completeness before proceeding
• Ask "Can there be more than one of these at the same time?" for read model fields
• Verify automations have all four components before labeling them as such

Do not:

• Skip steps because you think you know enough
• Make architecture or implementation decisions during modeling
• Write GWT scenarios for data validation (use the type system)
• Design multiple workflows simultaneously
• Proceed with gaps in the model

Enforcement Note

• Standalone mode: Advisory. The agent follows the nine-step methodology by convention.
• Pipeline mode: Gating. Incomplete models (missing GWT scenarios, undefined automations) block slice decomposition.

Hard constraints:

• Do not proceed with gaps in the model: [RP]

Constraints

• "MUST follow nine-steps.md": Following the nine steps means executing each step's specific activities and producing its specific outputs. It does not mean reading the reference and claiming "I followed the spirit." Each step has defined outputs -- produce them.
• "Do not design multiple workflows simultaneously": This includes starting "discovery" for Workflow 2 while Workflow 1's steps are incomplete. Discovery IS design. If you're gathering information about a future workflow, you're designing it.
• Facilitation vs. stenography: Facilitation means asking questions that help the domain expert discover things they haven't articulated yet. It does not mean asking leading questions that guide toward your preferred answer. The test: could the expert's answer genuinely surprise you? If not, you're leading, not facilitating.

Verification

After completing event modeling work, verify:

• Domain overview exists at docs/event_model/domain/overview.md with actors, workflows, external integrations, and recommended starting workflow
• Each designed workflow has docs/event_model/workflows/<name>/overview.md with all 9 steps completed
• All events are past tense, business language, immutable facts
• Every read model field traces to a source event
• Every event has a trigger (command, automation, or translation)
• Automations have all four components (event, read model, conditional logic, command)
• Read model fields use collection types when domain supports concurrent instances
• No cross-cutting infrastructure modeled as Translation slices
• GWT scenarios exist for each slice (inline in docs/event_model/workflows/<name>/slices/*.md) with concrete data
• GWT error scenarios test business rules only, not data validation
• Slices sharing an event schema are independently testable (no artificial dependency chains)
• No gaps remain in the model after validation

If any criterion is not met, revisit the relevant practice before proceeding.

Dependencies

This skill works standalone. For enhanced workflows, it integrates with:

• domain-modeling: Events reveal domain types (Email, Money, OrderStatus) that the domain modeling skill refines
• tdd: Each vertical slice maps to one TDD cycle
• architecture-decisions: Event model informs architecture; ADRs should not be written during event modeling itself
• task-management: Workflows map to epics, slices map to tasks

Missing a dependency? Install with:

npx skills add jwilger/agent-skills --skill domain-modeling