@milehimikey/em 1.13.1 → 1.14.0

package/plugin/shared/reference/methodology.md

@@ -0,0 +1,329 @@
+<!-- GENERATED FILE from .claude/skills/event-modeling-shared/reference/methodology.md by scripts/build-plugin.ts — do not hand-edit; run npm run docs:generate -->
+# Event Modeling — Methodology Reference
+
+This is the authoritative reference for the **7 steps** and **4 patterns** of Event Modeling
+(Adam Dymitruk's method). Consult it while running any phase of the `event-modeling` skill.
+Do not invent domain facts — use the Socratic prompts to extract them from the user.
+
+---
+
+## Core idea
+
+An event model is a **timeline of state changes** told as a story. Information only moves
+three ways:
+
+1. **Into** the system via a **command** (a request to change state) → produces an **event**.
+2. **Out of** the system via a **read model / view** (a projection of past events) → shown on a **UI** or consumed by an automation.
+3. There is **no other way** for information to flow. Every box on the diagram is one of:
+   UI screen, command, event, read model, or automation/processor.
+
+Concretely, that leaves exactly six legal connections, and `em` infers only these:
+`ui → command`, `command → event`, `event → read model`, `read model → ui`,
+`read model → reaction`, `reaction → command`. Every other pair is a rule violation, and an
+explicit `arrow` writing one is a validation error. The two worth naming, because they are the
+tempting shortcuts:
+
+- **`command → read model`** — the CQRS violation. A write is only ever visible to a reader
+  through the event it recorded. Put the event between them.
+- **`read model → command`** — reads never drive a write directly. A processor or translation
+  watches the read model and issues the command that sits in the same slice.
+
+If you find yourself wanting an arrow the patterns don't allow, the model is missing an
+**element**, not an arrow.
+
+**Automations and translations are not exceptions to rule 1.** They are *reactions* — a processor
+or an adapter that wakes up, decides something must change, and **issues a command** to do it. They
+never record an event themselves. A reaction box always points at a **command in its own slice** —
+the same shape a `ui` already uses in State Change — and that command produces the event. If you
+ever see a translation or automation wired straight to an event, the model is wrong.
+
+Events are **immutable past-tense facts** ("Order Placed", "Payment Captured"). They are the
+spine of the model. Everything else hangs off the events.
+
+The unit of delivery is a **slice**: a thin vertical cut through the swimlanes that delivers
+one of the four patterns. A slice is what a developer implements and tests in isolation.
+
+---
+
+## The 4 patterns
+
+Every slice is exactly one of these. In `em`, each maps to a specific shape (see `em-dsl.md`).
+
+### 1. State Change (Command pattern)
+**UI → Command → Event.** A user (or automation) submits a command; the system validates it
+against invariants and records one or more events.
+- **A State Change never travels alone, and it's joined at both ends.** Something must *trigger*
+  the command — the `ui` it's issued from, or (patterns 3 and 4) the reaction that triggers it,
+  also in this slice — and its event needs a read model that projects it. So the unit of work is a State Change
+  slice *plus* the State View slice that reads its event. Write them together; a command nothing
+  points at is a write nobody can start, an unread event is a write nobody can see, and
+  `em validate` warns on each.
+- `em` shape (one element per line — there is no one-line slice form):
+  ```em
+  slice "Do The Thing" {
+    ui X @Persona
+    command Y
+    event Z @Context
+  }
+  slice "See The Result" {   # the State View slice that reads Z — pattern 2, but required here
+    view V from "Z"
+    ui Screen @Persona
+  }
+  ```
+- This is where **invariants** live — the command is rejected if a rule is violated.
+- Socratic prompts: *"What does the user do here? What request are they making? What must be
+  true for it to succeed? What fact gets recorded when it does? What gets rejected and why?"*
+
+### 2. State View (View pattern)
+**Event(s) → Read Model → UI.** Past events are projected into a read model that a screen
+displays. Read-only; changes no state.
+- `em` shape: `slice { view V from "Event A", "Event B"  ui Screen @Persona }`
+- **Headless / API systems:** still use `ui`/`persona` — declare a persona per external
+  caller/role (e.g. `IntegratorAPI`) and treat its `ui` boxes as API calls, not screens. Shape:
+  `slice { view V from "..."  ui Read V @IntegratorAPI }`, same as any State View slice.
+  `translation` stays reserved for genuine reactions/external-system boundaries (pattern 4), not
+  for a synchronous request/response API call.
+- **Repeat read models across the timeline.** A read model is drawn fresh in **every** slice where
+  it is read — after the events that update it, before the actions that consume it — so the diagram
+  shows state flowing left-to-right (the information-completeness staircase). The same read-model
+  name recurring is intentional and **renders cleanly**. Declare every instance after the first with
+  **`view V again from "..."`**: `again` instances are exempt from the duplicate-name warning even
+  when something references them, and each reference resolves to the right instance. (A plain repeat
+  only stays warning-free while nothing references it by name, and resolves to the *first*
+  declaration when something does.) **Wire each event to a read model
+  exactly once:** a repeated instance's `from` lists only the **new** events since the previous
+  instance (not cumulative) — otherwise an event draws a duplicate arrow to the same read model at
+  every repeat. (An event may still feed several *different* read models, once each.)
+- **Instances are never connected to one another.** There is no arrow between two appearances of one
+  read model, and an explicit one is a validation error. The repeat is an ergonomic device for
+  showing the same read model at successive points in time — continuity is implied by the shared
+  name, and the events arriving at each instance are what show it changing. An arrow between them
+  would say the read model feeds itself.
+- **Place each repeat immediately after its feeding event (span-1).** The cleanest staircase puts a
+  read-model instance right after each event that updates it, sourcing **only that one adjacent
+  event** — every `event → read model` arrow is then short and forward-flowing. A read model sitting
+  far from its source events (e.g. a "list/queue" read placed early but fed by late events) draws
+  long arrows that sweep across the diagram. The renderer routes them *around* intervening boxes
+  rather than through them, so they no longer *look* like a forbidden read→read link — but the
+  distance itself is the problem: the arrowhead lands columns away from the event that produced it,
+  so **the write slice reads as dangling**, as if nothing consumed its event. A reader has to trace
+  the line across the diagram to see otherwise. The fix is never a `view → view` edge; it's to
+  repeat the read model next to each event that feeds it, and to keep a sub-flow that detours into
+  another context together rather than parking it at the end of the model.
+- Socratic prompts: *"What does the consumer need to see to make their next decision? Which past
+  events provide that information? Is this a screen or an API read? What's the freshness/consistency
+  expectation?"*
+
+### 3. Automation (Processor pattern)
+**Read Model → Processor → Command → Event, the last three together in one slice.** The system
+reacts on its own: a processor watches a read model (a "to-do list" of work) and **issues a
+command** — it never records an event directly.
+- `em` shape: the read model in the slice before, then the reaction with its command and event —
+  plus the read slice that consumes the event — three slices in all:
+  ```em
+  slice "Todo" {
+    view Todo from "..."
+  }
+  slice "Do It" {           # processor, command, and event all share this slice
+    processor P from "Todo"
+    command C
+    event E @Context
+  }
+  slice "Todo — done" {
+    view Todo again from "E"   # every event needs a reader; here it clears the to-do list
+    ui Work Board @Persona     # ...and every read model needs a consumer
+  }
+  ```
+  The processor never records the event directly — the command it triggers does, and that event
+  happens to live in the processor's own slice now because the command does. A processor with no
+  command in its slice (and no explicit arrow to one) is a validation warning: a decision the
+  system never acts on. Forgetting the third slice leaves `E` unread, which is also a warning.
+- **Naming:** name the read model after the pending work — `Payments To Process`,
+  `Orders To Fulfill`, `Pending Approvals` — never after the triggering event. The event is what
+  happened; the view is what's left to do. Reusing the event's name also collides in the shared
+  element namespace: `from` references then resolve by element kind, and `em validate` flags the
+  duplicated name.
+- Socratic prompts: *"What should happen without a human? What condition triggers it? What work
+  list does the processor watch? What command does it fire? What if it fails or retries?"*
+
+### 4. Translation
+**Boundary crossing → command → event, the reaction and what it triggers together in one slice.**
+An adapter translates data across a boundary (an external system, or another bounded context) into
+the model's own language. Like an automation, a translation is a *reaction*: it **triggers a
+command**, never records an event directly. Two independent questions shape which form it takes —
+don't conflate them:
+- **Trigger source:** does the input come from outside the model, or from the model's own state
+  pushed back out?
+- **Durable artifact:** is there a queryable, persisted thing behind the trigger (a queue, topic,
+  or log), or is it an ephemeral call with nothing to query afterward?
+
+The `em` shape only tracks the second question — trigger source doesn't change it:
+- **No durable artifact:** `external input → translation → command → event`. No internal `from`.
+  Typically (not necessarily) externally triggered — a bare webhook call with nothing persisted
+  behind it.
+- **Durable artifact:** `read model → translation → command → event`. The translation reads a
+  **view** via `from`. This shape is the same whether the view was filled by the model's own event
+  (internally triggered, e.g. pushing data out) or by an external system's persisted queue
+  (externally triggered, e.g. a webhook whose inbound message is stored first for retries,
+  ordering, or audit) — an externally triggered translation backed by a real queue is
+  architecturally closer to this case than to the no-artifact one above.
+- **Event legitimacy:** a received external message counts as a legitimate domain `event` when
+  it's scoped to the context/lane whose fact it represents, not by who committed it — the same
+  move as an Anti-Corruption-Layer boundary event in DDD — and stays legitimate as long as it
+  doesn't leak into the model's own domain vocabulary.
+- `em` shape: exactly like an automation — the read model (durable-artifact form only) in the
+  slice before, then the reaction with its command and event, plus the read slice for the event:
+  ```em
+  slice "Boundary" {
+    view Source from "..."     # durable-artifact form only; omit if there's nothing persisted
+  }
+  slice "Record It" {          # translation, command, and event all share this slice
+    translation T from "Source"   # omit `from` entirely for the no-durable-artifact form
+    command C
+    event E @Context
+  }
+  slice "Source — recorded" {
+    view Source again from "E"   # every event needs a reader
+    ui Status Screen @Persona    # ...and every read model needs a consumer
+  }
+  ```
+  The translation never records the event directly — the command it triggers does.
+- Socratic prompts: *"What boundary are we crossing, and which way? What outside system or context
+  feeds us (or do we feed)? In what format? How do we know its data is trustworthy/complete? What
+  internal **command** does it trigger, and what event does that command record?"*
+
+---
+
+## The 7 steps
+
+Run them in order. The skill groups them into phases: **discover = 1–4**, **model = 5–7**,
+and a dedicated **slice** phase deepens step 6 to implementation-ready specs.
+
+### Step 1 — Brainstorm Events  *(discover)*
+List the domain **events** as past-tense facts, unordered at first. Go wide; capture every
+state change anyone can think of. No commands, no UI yet — just facts.
+- Prompts: *"What are all the things that happen in this process? Say each as something that
+  already occurred. What changed when that happened?"*
+- **Is it an event?** Events here are *business* facts, not technical ones. Test: if this fact
+  failed to occur, would it wake the CEO at 3am? Yes → event. No → not an event.
+  - **Yes:** Order Created, Item Added to Order, Order Shipped, Order Delivered, Order
+    Cancelled, User Registered, Item Provisioned.
+  - **No — derived values** (a read model computes these from other events; they aren't
+    independent facts): Order Total Calculated, Order Tax Calculated, Discount Applied.
+  - **No — activity, not business state** (telemetry, not something the domain decides on):
+    User Clicked Button, User Viewed Page, User Logged In.
+- Model the process the business **needs**, not how the current system happens to work. Existing
+  system behavior is a common source of fake events — screens and tables leaking in as facts.
+  (**Extract mode inverts this**: in the `em:discover` skill's `extract` phase the
+  existing system *is* the subject and this caution is deliberately suspended — see the
+  "Extract" section below and `em:discover`'s own `reference/extract.md`. The
+  derived-value and telemetry filters still apply there.)
+- Output: a flat list of candidate `event` names.
+
+### Step 2 — The Plot / Storyboard  *(discover)*
+Put the events in **timeline order** to form the narrative. Identify the **personas** (actors)
+and sketch the **UI** screens that move the story forward. This is the storyboard.
+- Prompts: *"In what order do these happen? Who is on screen at each step? What screen are they
+  looking at? What's the happy-path story start to finish?"*
+- Output: ordered events + `persona` list + `ui` screens per step.
+
+### Step 3 — Inputs (Commands)  *(discover)*
+For each event, identify the **command** that causes it (State Change pattern). Name the intent,
+not the outcome ("Place Order", not "Order Placed").
+- Prompts: *"What action produces this event? Who or what issues it? Could it be refused?"*
+- Output: `command → event` pairs.
+
+### Step 4 — Outputs (Read Models)  *(discover)*
+Identify the **read models / views** each UI and automation needs (State View pattern). Wire
+each view to the events that feed it.
+- Prompts: *"What information does this screen show? Which past events supply it? Does any
+  automation need a work list derived from events?"*
+- Output: `view` elements with their `from "Event"` sources.
+
+### Step 5 — Swimlanes & Apply the Patterns  *(model)*
+Organize elements into swimlanes: one **persona** row per actor, one **context** (bounded
+context / aggregate) row per event family. Classify every slice as one of the **4 patterns** and
+wire them correctly — especially **share the slice between every automation or translation and the
+command it triggers** (the read model it watches, if any, stays in the slice before). A translation
+or automation never records an event directly; it triggers a command.
+- Prompts: *"Which events belong together as one consistency boundary (aggregate)? Who owns this
+  data? Is this slice a state change, a view, an automation, or a translation?"*
+- Output: a structurally complete model with swimlanes and pattern-correct slices.
+
+### Step 6 — Elaborate Scenarios  *(model first pass, slice deep pass)*
+For each slice, write **Given / When / Then** scenarios and surface **invariants**, **critical
+fields**, and **alternate / error flows**. In the `model` phase do a light first pass; the
+dedicated `slice` phase writes the full rich spec (see `../templates/slice.md`).
+- Prompts: *"Given what starting state, when this command/trigger fires, then what event(s)
+  result? What must always be true? What are the failure paths? Which fields are essential and
+  what are their rules?"*
+- **Existing codebase:** before finalizing field names/types, check adjacent real sources
+  (OpenAPI specs, DB migrations, existing DTOs/event classes in sibling contexts) rather than
+  guessing — mirrors `extract.md`'s whole-system grounding, applied per-slice.
+- Output: rich slice docs linked into the `.em` via `note "slices/<name>.md"`.
+
+### Step 7 — Evaluate Completeness  *(model)*
+Walk the whole model with stakeholders and check for loose ends:
+- Every **command** has something that **triggers** it — a `ui` in its slice, or the reaction that
+  triggers it, also in its slice. A command nothing points at is a write nobody can start.
+  `em validate` warns.
+- Every **command** produces at least one **event**.
+- Every **view** has at least one source event, **and a consumer** — a `ui` (headless: the `ui`
+  tagged to the API-caller persona), or a reaction watching it. Every *instance* of a repeated
+  view, not just the last one.
+  A read model nothing displays is information projected out and then dropped. `em validate` warns.
+- Every **event** is **read by a read model** — a `view` naming it in `from` (any `again` instance
+  counts), or a `view` with no `from` sitting in its slice. A reaction consuming it does **not**
+  count: reactions read views, not events. An event nothing projects is a write nobody can see, so
+  there was no point recording it. `em validate` warns on each one.
+- Every **UI** is reachable and leads somewhere.
+- Every connection is one of the six legal pairs (`ui → command`, `command → event`,
+  `event → read model`, `read model → ui`, `read model → reaction`, `reaction → command`). No
+  arrow between two instances of one read model.
+- Automations **and** translations are wired correctly (reaction → command → event, all in one
+  slice — never a reaction wired straight to an event, and never a reaction with no command at
+  all); translations cover every external input.
+- Run `em validate` and resolve all errors and warnings. It catches illegal `arrow` kind pairs
+  (error), unread events (warning), and a reaction with no command in its slice (warning) — a
+  reaction wired straight to an event with no command anywhere is exactly what that last check
+  catches, so a clean `em validate` run does cover this case.
+- Prompts: *"Is there any event nobody reads — and if the business genuinely never looks at it, why
+  are we recording it? Any command that just happens, with nobody and nothing asking for it? Any
+  screen with no way in or out? Any command that records nothing? Any
+  external system we haven't translated? Any translation or automation that records an event instead
+  of triggering a command?"*
+
+---
+
+## Extract: current-state models of existing systems
+
+The `extract` phase derives a model from an **existing** system instead of a greenfield
+conversation. It is discover's as-is sibling: its confirm-and-clarify rounds converge the same
+step 1–4 outputs (events, timeline, commands, read models) — *derived from the system*, then
+confirmed with the user — after which the model enters `model` (steps 5–7) as usual.
+
+The stance inversion, stated authoritatively: **extraction captures how the system behaves
+today — faithfully, warts included — and never invents future or desired state.** Step 1's
+anti-current-system caution is suspended for the duration; the "is it an event?" filter is
+not — derived values and telemetry are still rejected even when the current system emits them.
+Unknown or ambiguous current behavior is parked, never guessed — mirrored in the state file's
+Open Questions, and on the element itself with `issue "text"` when it should also show up on the
+rendered diagram (preferred over a bare `# TBD` comment, which is invisible once rendered).
+Desired-state improvements belong to the `model` and `slice` phases, after the as-is picture is
+agreed.
+
+Procedure, source modes (event-driven vs. procedural synthesis), and the round-by-round loop:
+the `em:discover` skill's own `reference/extract.md`.
+
+---
+
+## Socratic stance (applies throughout)
+
+- Ask **one focused question at a time**; never assume a domain fact — extract it.
+- Prefer "why", "what if", "who", "what must always be true", "how do you know" over yes/no.
+- Mirror the model back after each increment and **re-render** so the team sees it evolve.
+- Park unresolved questions in the state file instead of guessing.
+- Name things crisply: events past-tense, commands imperative, views as the thing-shown.
+- **Live workshop, one chat interface:** route through a single human proxy who relays questions
+  to the room and reports answers back; attribute every answer/decision in the Decisions log to
+  the named participant who gave it (see the state file's Participants section).

package/plugin/shared/reference/operating-principles.md

@@ -0,0 +1,222 @@
+<!-- GENERATED FILE from .claude/skills/event-modeling-shared/reference/operating-principles.md by scripts/build-plugin.ts — do not hand-edit; run npm run docs:generate -->
+# Operating principles, preconditions, and project layout (every phase)
+
+Shared by every `em` plugin skill. Read this before doing any work in any phase — it's
+the discipline and mechanics that don't change no matter which phase you're facilitating.
+`methodology.md` (the 7 steps + 4 patterns) and `em-dsl.md` (DSL syntax and validation rules) in
+this same directory are the other two shared references; read them too before doing real work —
+they, together with this doc, are the source of truth every phase skill builds on. Templates
+live in `templates/`, a sibling of this `reference/` directory. References to `docs/*.md`
+throughout these skills mean the
+[em repository's docs](https://github.com/milehimikey/em/tree/main/docs) — they are not
+vendored alongside these skills, with one exception: the frontmatter schema doc is vendored
+right here as `reference/slice-doc-schema.md` (kept in sync with the em repo's
+`docs/slice-doc-schema.md` by `em`'s own `docs:generate`/`docs:check`), so every
+`slice-doc-schema.md` reference elsewhere in this bundle points there instead of at `docs/`.
+
+## Operating principles
+
+- **Socratic, one question at a time.** Ask focused questions; use `AskUserQuestion` for tight
+  multiple-choice decisions. Never assume a domain fact — extract it. Prefer "who/why/what-if/
+  what-must-always-be-true/how-do-you-know" over yes-no.
+- **Don't guess — park it.** Unresolved items go into the Open Questions list in
+  `.event-modeling.md`, not into invented model content. Note the source, blocker, and revisit
+  date when known — cheap now, useful when you come back to it.
+- **Happy path first; branches belong to slicing.** Steps 1-7 build the **happy-path spine**.
+  Alternate, unhappy, and exception-path events (rejections, removals, cancellations,
+  expirations, declines) are **not** enumerated as a separate discover/model task — they
+  surface during the `slice` phase as each slice's alternate/error flows and rule-boundary
+  scenarios, and any new branch events get added to the `.em` then. Never list "draft the
+  branch events" as a discover/model to-do or ask the user to do it before slicing.
+- **Reflect and re-render.** After each meaningful increment, update the `.em` file, re-render,
+  and show the user what changed. Encourage running the live view (`watch`) so a team can follow.
+- **Keep the `.em` structural; put depth in `note` docs.** The diagram holds flow; invariants,
+  fields, and scenarios live in `slices/*.md` linked via `note "slices/<name>.md"`.
+- **Headless/API models still use `ui`.** If the system is headless (clients call an API, not
+  screens), declare a persona per external caller/role (e.g. `IntegratorAPI`) and treat its `ui`
+  boxes as API calls, not screens: writes are `ui → command → event` and reads are
+  `read model → ui`, exactly like any other State Change/State View slice — no separate trigger
+  slice. `translation` stays reserved for genuine reactions and real external-system boundaries,
+  not a synchronous request/response call. Internal-only commands/views with no public route carry
+  no `ui` at all — they follow the ordinary Automation pattern instead. **Repeat read models**
+  across slices so the timeline flows left-to-right — put each repeat **right after the event that feeds it**, sourcing
+  only that one adjacent event, so every arrow is short (a read model far from its source events
+  draws a long arrow whose head lands columns away, making the *write* slice read as dangling —
+  keep a sub-flow that detours into another context together, not parked at the end of the model).
+  Declare every instance after the first with
+  **`view X again from "..."`** — `again` instances are exempt from the duplicate-name warning even
+  when referenced, and each reference resolves to the right instance. A reaction's command is
+  wired by same-slice presence, exactly like `ui → command` — no positional/adjacency rule to get
+  right (see `em-dsl.md`).
+  See `methodology.md` (State View) and `em-dsl.md`.
+- **Never connect two instances of one read model.** The repeat is a timeline device, not a flow:
+  continuity is implied by the shared name, and the events arriving at each instance are what show
+  it changing. An arrow between instances says the read model feeds itself, and is an error.
+- **Every slice is a COMPLETE pattern, never a half-slice.** A State Change is
+  `ui → command → event`; a State View is `event → read model → ui` (or `→ reaction`). So: a
+  command needs something that **triggers** it (the `ui` it's issued from, or the reaction that
+  triggers it, also in this slice), its event needs a **reader**, and that read model needs a
+  **consumer** of its own. A command nothing points at is a write nobody can start; a reaction
+  that triggers no command is a decision the system never acts on; an event nothing projects is
+  a write nobody can see; a read model nothing displays is information dropped on the floor.
+  Write the whole chain together rather than sweeping up dangling ends later. Reactions don't
+  count as readers of *events* — they read views. **Every instance** of a repeated read model
+  needs its own consumer: if you repeat a view next to an event just to keep the arrow short,
+  bring its screen along, or don't add the instance.
+- **Only six connections are legal:** `ui → command`, `command → event`, `event → read model`,
+  `read model → ui`, `read model → reaction`, `reaction → command`. Reaching for anything else —
+  above all `command → read model` (the CQRS violation) or `read model → command` — means the model
+  is missing an **element**, not an arrow.
+- **Validate continuously.** Run `em validate` and fix errors/warnings as you go (see DSL ref).
+- **Save state at the end of every session** so work resumes cleanly. Also log a Usage log
+  line: `em state log-usage <model>.em --phases <phase1,phase2,...>` computes and dedupes the
+  diagnostic *categories* that fired for you and appends the canonically-formatted line — never
+  hand-run `em validate --json` and format the line yourself. This is the team's only usage
+  signal today (`docs/usage-data.md`) — keep it cheap and habitual, not a task to skip.
+
+## Write scope and evidence (MIL-236)
+
+These bind every design session — `discover`, `model`, `slice` — whichever skill runs it,
+including sessions started from bare `/em:event-modeling`.
+
+- **Write scope: one model directory.** The session is bound to the model directory it started
+  in: `dirname(<model>.em)`, which holds `slices/`, `model-versions/`, `conformance/`,
+  `contracts/`, `.event-modeling.md` and `README.md`. Never write under another model's
+  directory or to another model's contract file (`<model dir>/contracts/<model key>.tsp`). When
+  you need a neighbour, read only its public surface, read-only: the elements with
+  `public: true` in `em export <neighbour>.em`, or its committed contract file. Never open the
+  neighbour's slice docs or state file.
+- **Seam change request instead of a cross-model edit.** When resolving an open question needs
+  another model's `public` element changed (a field added, renamed, retyped; an element made
+  `public`), record it in THIS model's `.event-modeling.md` `## Decisions log` as one dated
+  bullet, exactly this shape (`em changelog` reads only dated bullets):
+
+  ```
+  - YYYY-MM-DD: seam change request → <producerKey>:<kind>.<slug> — <what> — <why>
+  ```
+
+  `<producerKey>:<kind>.<slug>` is the same ref a `consumes` clause uses (`kind` is `event` or
+  `view`; write `command` for a public command). Leave the open question open with
+  `blocked on: <producerKey> owners (seam change request YYYY-MM-DD)`, and tell the user to take
+  the request to the producer's owners. Never edit the other side.
+- **Out of scope during design — never write these:**
+  - generated contract artifacts: `contracts/*.tsp`, any OpenAPI or JSON-schema output;
+  - SDD/spec artifacts: `specs/`, `.specify/` (read `.specify/memory/constitution.md` when a
+    slice needs the house rules; never write it);
+  - implementation source: `src/`, and any path under the model's `Code roots:` bullet;
+  - `plugin/` (the generated skill plugin);
+  - a draft slice doc's `version:` (see the version rule below).
+
+  One exception, and only this one: when the session marks an element `public` (or changes a
+  field of a `public` element), run `em api generate <model>.em` before handing slices to
+  ratification. The API-first gate (`em validate --slice-ready`) refuses a public-touching
+  slice whose contract is missing or stale, so the regenerated `contracts/<model key>.tsp` is
+  committed with the model edit. Never hand-edit it.
+- **Evidence rule.** The `.em` model and the slice docs are the evidence for every modeling
+  decision. Never open implementation source, generated contracts, or `specs/` to settle a
+  modeling question; ask the user. Code is consulted only in `extract`
+  (`em:discover`) and `conform` (`em:conform`), where reading it is the
+  point.
+- **Version rule: never bump a draft's `version:` on edit.** A `draft`/`reviewed` doc is edited
+  in place. Check the release state with `em status <model>.em --json`: the model is
+  **pre-release** when its `modelVersion` entry has `design: null` and
+  `slices.byStatus.implemented === 0`. Pre-release: never run `em slice reratify`; edit
+  `draft`/`reviewed` docs in place and leave `version:` as it is. A pre-release doc that is
+  already ratified (`ready-to-implement` with `ratifiedBy`) carries a human sign-off: STOP,
+  record the needed change as an open question, and let the ratifier decide. Released (either
+  condition false): changing an `implemented` or ratified doc goes through `em slice reratify`
+  (see `em:design`).
+- **Pre-commit check (STOP on any finding).** Stage the session's files, then run both:
+  1. `em system scope --staged` from the repository root. Any `seam-crossing` error → STOP:
+     unstage the other model's files and record a seam change request instead.
+  2. `git diff --cached --name-only`. Any path outside the bound model directory, or any
+     out-of-scope path above (`contracts/*.tsp` except the regenerated exception, OpenAPI /
+     JSON-schema output, `specs/`, `.specify/`, `src/`, code roots, `plugin/`), or a draft
+     doc whose `version:` line changed (`git diff --cached -- <doc>`) → STOP: unstage it and
+     tell the user what was touched and why it is out of scope.
+
+## Preconditions (run first)
+
+1. Check the tool: `em --version`. If missing, tell the user to run `npm i -g @milehimikey/em`
+   and stop until installed.
+2. Check this vendored skill bundle's own currency (MIL-219): read this file's `em-version:`
+   frontmatter stamp (or any sibling skill's SKILL.md — they're stamped together) and compare it
+   to `em --version` from step 1. **If the installed em is ahead by a minor version or more
+   (1.9.x → 1.11.x, or any major bump), STOP** — don't proceed with discover/model/slice/
+   implement/conform/review work on this stamp. Tell the user the vendored bundle is stale and to
+   run `em skill sync` (refresh the bundle only) or `em upgrade` (the bundle plus the model
+   repo's other mechanical migrations, docs/upgrading.md) before continuing. A patch-only gap
+   (1.11.2 → 1.11.4) is never worth stopping over — proceed normally. This matters because a
+   stale bundle silently follows old rules: an agent reading superseded preconditions/gates
+   can hand back work that looks right by last week's contract and wrong by this week's.
+3. Locate the model. Look for an existing `<dir>/.event-modeling.md` and `*.em` in the working
+   directory, or — for a multi-model project laid out per the convention below — one level down
+   inside a `models/` subfolder (each model gets its own `models/<slug>/` directory; check each
+   candidate for its own `.event-modeling.md`/`*.em` pair). If found, run `em state read <dir>`
+   to read the state file's mechanical fields (model path, phase, step, last updated, last
+   conformance, last review) as JSON — don't parse the bullets by hand. If not, and the phase
+   needs one, ask the user for the model name and where to create it (a lone model goes at the
+   project root; one of several goes under `models/`, per "Multi-model projects" below).
+4. If you're not sure which phase applies, or the user gave no phase, use `em state read <dir>`'s
+   `phase`/`step` to resume the recorded phase/step — or invoke the `event-modeling` skill, which
+   does exactly this and routes to the right skill. If no model exists, propose starting
+   `discover` (greenfield, `em:discover`) or `extract` (modeling an existing system,
+   also `em:discover`).
+5. Populate the state file's Participants section at session start. For a live workshop, ask for
+   a single human proxy to relay questions to the room, and attribute every answer/decision in
+   the Decisions log to a named participant.
+
+## Project layout these skills create
+
+```
+<model-name>/
+  <model-name>.em          # the model (slice docs linked via note "...")
+  <model-name>.svg         # render target for em watch
+  README.md                # overview + slice index (from templates/model-readme.md)
+  .event-modeling.md       # resumable state (from templates/state.md)
+  slices/<slice-name>.md   # one rich slice doc per slice (from templates/slice.md)
+  conformance/<date>-report.md
+                           # conform-phase reports (from templates/conformance-report.md)
+  <model-name>-asis.em     # conform-phase scratch model, regenerated per run — git-ignore this
+```
+
+When creating a new model, run `em scaffold <model-name>` — it creates `<model-name>/` and
+writes all three starter files in one step (`<model-name>.em`, `README.md`,
+`.event-modeling.md`), filling the Model Name/model-name placeholders and the state file's
+mechanical fields (model path, `Current phase: discover`, `Current step: 1`, dates) for you, so
+there's nothing left to hand-copy or fill in. Refuses if `<model-name>/` already exists; pass
+`--force` to overwrite. You'll
+usually replace the scaffolded `.em`'s content as the discovery conversation builds up the real
+model — it starts out as the same starter model `em init` writes, titled from `<model-name>`.
+`slices/*.md` and `conformance/` aren't part of `em scaffold`; they come from the `slice` and
+`conform` phases (`em:design` and `em:conform`). The `conformance/`
+directory and `<model-name>-asis.em` only appear once the `conform` phase runs; `em conform-scope
+--seed-asis` creates `<model-name>-asis.em` and gitignores `*-asis.em` for you (see
+`em:conform`'s `reference/conform.md`) — it's scratch, never committed.
+
+### Multi-model projects
+
+Slice docs resolve at a fixed path — `slices/<slug>.md`, sibling of the `.em` file that
+declares the slice — with no model namespace anywhere, so **one directory per model** is the
+whole guardrail against two models' slice names colliding. For more than one model, nest each
+one's directory (the exact layout above) under a shared `models/` parent instead of scattering
+them at the project root:
+
+```
+models/
+  checkout/
+    checkout.em
+    slices/...
+  fulfillment/
+    fulfillment.em
+    slices/...
+```
+
+Run `em scaffold <model-name> --under models` for each model — it creates `models/<slug>/`
+directly rather than a two-step `cd models && em scaffold <model-name>`. See
+[docs/cli.md, "Multi-model projects"](https://github.com/milehimikey/em/blob/main/docs/cli.md#multi-model-projects)
+for the full rationale, and why slice keys stay unqualified (no `<model>/<slice>` prefix) —
+directory isolation alone is sufficient. Two `.em` files sharing a directory with colliding
+slice keys don't go undetected: `em status`/`em catalog`, the two commands that ever compile
+more than one model in a run, raise a `cross-model-slice-doc-collision` warning naming both
+files — treat that warning as a sign the models need separating, not a false positive.