Metamodel reference

Metamodel — template types

Each section below is one template.<subType>. The Attributes table lists the subtype's own + concern-contributed attributes (provider-tagged); universal documentation attributes are omitted here (see providers.md). Allowed children lists the structural child rules with their cardinality (min..max, * = unbounded).

template.base

Abstract base template — the shared root subtype for the fourth pillar (FR-004, ADR-0011). A template is a typed payload bound to either a rendered text artifact (prompt/output) or a tool-call envelope. The base carries no attrs of its own; concrete subtypes add their reference + governance attrs. Not authored directly: a template.base node fails to load (ERR_ABSTRACT_SUBTYPE_AUTHORED) — this subtype is a registry anchor concrete subtypes inherit from, never a node in a document.

Owning provider: metaobjects-core-types

Rules: A single MetaTemplate node class backs every template subtype (no subType→class dispatch); the subtype selects which reserved-attr set applies. @format (a closed-enum ATTRIBUTE, never a subtype) drives the render engine's escaper/whitespace behavior, so a new format costs one escaper + one enum value rather than a new subtype + cross-language port.

Attributes

No subtype-specific attributes.

Allowed children

No structural children.

template.output

An output / serialization template (FR-004): every rendered artifact other than an LLM prompt — a document (email, export, docs, config) or an email. Carries the generic reference + governance attrs and the @kind + email part-refs. OUTBOUND ONLY (ADR-0052) — it renders, and generates no parser.

Owning provider: metaobjects-core-types

Rules: output is either a document (@kind="document" or absent → renders @textRef in @format to one string) or an email (@kind="email" → renders subject + html + optional text to a structured EmailDocument). The cross-field presence rule is enforced in the loader's validateTemplatePayloadRefs pass: document requires @textRef; email requires @subjectRef AND @htmlBodyRef (with @textBodyRef optional) and carries NO @textRef. @format is a closed enum keyed by the render engine's escaper. template.output is OUTBOUND ONLY (ADR-0052): it emits a render helper and nothing that reads a model's reply — the parser-on-receipt, the tolerant extract and the FR-010 response-format fragment all belong to template.prompt @responseRef.

When to use: You render a document/email/serialized output from typed data. Declare an output template so the {{fields}} are drift-checked against the payload VO at build time. To parse a model's reply instead, declare @responseRef on the template.prompt that elicits it — never a template.output.

Attributes

Allowed children

No structural children.

template.prompt

An LLM-targeted renderable prompt template (FR-004). Carries the generic reference + governance attrs plus the LLM overlay (@maxTokens / @requiredSlots / @model / @responseRef). Its renderable body is required via @textRef. A prompt declaring @responseRef also owns the INBOUND half (ADR-0052): the parser-on-receipt, the FR-010 response-format fragment (@promptStyle), and the reply syntax (@responseFormat).

Owning provider: metaobjects-core-types

Rules: prompt requires @payloadRef (the typed payload it renders against) AND @textRef (the body text, provider-resolved at render time — enforced in the loader's validateTemplatePayloadRefs pass, not at the attr layer where @textRef is relaxed to optional so template.output email can omit it). @format is the syntax of the rendered PROMPT body, a closed enum keyed by the render engine's escaper; @responseFormat (ADR-0053) is the syntax of the model's REPLY. @responseRef (optional) names the response value-object the prompt expects, drives typed LLM-call trace derivation, and is the sole gate on the inbound codegen tier — which keys on its presence, never on a format value.

When to use: You are sending text to an LLM. Declare a prompt template with a typed payload so the prompt is versioned, drift-checked against its fields, and cache-stable — instead of string-building it in code. Add @responseRef when you also want the reply parsed into a typed shape.

Attributes

Allowed children

No structural children.

template.toolcall

A vendor-agnostic LLM tool-call envelope (ADR-0011). Unlike prompt/output it has NO renderable text body — the body IS the structured output schema resolved via @payloadRef. This is why toolcall is its own subtype rather than template.output + @toolName. Does NOT inherit the generic attrs.

Owning provider: metaobjects-core-types

Rules: toolcall requires @toolName (the wire tool name surfaced to the LLM) AND @payloadRef (the output value-object the tool produces). @textRef is intentionally absent — a tool-call has no renderable text. Core keeps toolcall vendor-agnostic; vendor wire details (retry semantics, fallback shapes, parallel invocation, cache hints) are added by consumer providers via registry.extend(TYPE_TEMPLATE, "toolcall", { attributes: [...] }). The LLM-facing tool description reuses the @description documentation common attr (added to every type by docProvider), so it is not redeclared here.

Attributes

Allowed children

No structural children.