FR-033: strict per-subtype metamodel + 5-concern provider split + metamodel docs + 5-port fan-out#29
Merged
Merged
Conversation
Brainstormed + approved design for giving Claude Code / LLMs accurate, context-tiered info about what every metadata type/subtype is for and what each attribute means. Net shape: a shared JSON descriptions source (core) embedded per port + in-code descriptions (language/module-specific), attached to each port's LIVE registry; each port dumps "registry-with-descriptions" — ONE artifact that is both the registry-conformance gate target AND the doc-gen input; one neutral Tier-2 engine renders TIERED markdown (INDEX one-liners + per-provider detail pages) wired into the metaobjects-authoring skill. Accurate-by-construction (derived from the live registry), DRY (single source), conformance-gated (every entry described; core byte-identical cross-port). Metaobjects applied to its own metamodel. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…model docs) Reframed per design discussion: instead of a descriptions LAYER on top of code-defined providers, a provider's DECLARATIVE definition becomes JSON data the provider reads to register itself — vocabulary, attr constraints, descriptions, and rule PROSE, all single-sourced. Root fix: the declarative metamodel stops being hand-coded 5× and gated; it's authored once and consistent by construction. Key additions: the data/code boundary (file = declarative + rule prose; imperative validation + codegen stay per-port code — no rules DSL); file layering (global/core in spec/metamodel, language-specific overlays in local port repos, overlay/last-writer-wins merge; a provider reads ≥1 file); registration reads+merges embedded files; registry-conformance extended to verify embed-integrity + description coverage + core identity; doc-gen + agent-context wiring unchanged; provider-by-provider migration sequencing. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…metamodel docs) Adds the FR-033 row linking GH #23 and stamps the design spec with its FR number. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Implementation plan for FR-033. Phase 1 (TS foundation on the field provider) is fully bite-sized + grounded in the real registry/provider APIs (the registry ALREADY carries TypeDefinition.description + AttrSchema.description — the manifest just excludes them, so this is "source from a shared file + emit + gate + doc-gen," not "add the slot"). Phases 2-4 (remaining TS providers, tiered doc-gen + agent- context wiring, cross-port fan-out) are recipe-driven task groups that repeat the proven Phase-1 recipe; each may expand into its own detailed plan when reached. Data/code boundary held throughout: declarative facts + rule prose in JSON; factories, imperative validation, codegen stay code. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Folds in the brainstormed structural-constraint model: a single unified `children` list per type/subtype (attrs are the type:"attr" entries; structural children are field/identity/etc., distinguished by type), uniform min/max cardinality (attr = required/optional, max always 1; structural = full 0..N), a child-side additive `parents` list, purely additive merge (provider union + extends inheritance, never narrowing), and one contradiction pass (six checks: dangling ref, unsatisfiable required child, min>max, closed-set clash, required-cycle, conflicting attr redefinition). Adds Phase 1b plan tasks 7-9 (constraint fields, additive merge, contradiction validator). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ation) Plan Tasks 1+2 + the format half of Task 7. New provider-data.ts: the declarative provider-definition format — ProviderDefinition / TypeDef / ChildDef (the unified child requirement: attrs are type:"attr" entries carrying value-type/default/allowedValues/isArray; structural children are field/identity/... entries; subType accepts value|list|"*"; cardinality min/max; named) + FactoryMap. defineProviderFromData fans an attr child-entry to an AttrSchema (required = min>=1) and a structural entry to a ChildRule, passes parents/rules/example/whenToUse onto the TypeDefinition; throws on missing factory + the attr/cardinality validations (attr max!=1 unless isArray, list-subType attr, min<0, max<min). registry.ts: ChildRule.childSubType widened to string|list + optional min/max/named; AttrSchema/TypeDefinition gain optional rules/example/whenToUse (+ TypeDefinition.parents). The factory stays code — the data/code boundary. 13 new tests; metadata suite 1727/0; typecheck clean. Foundation only — no provider/manifest changes yet (Tasks 4/5/8/9). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…dator Plan Tasks 8+9 (spec §3.1). constraint-merge.ts: mergeConstraints() produces EffectiveConstraints per type.subType — horizontal provider union (a child-side `parents` claim additively folds into the named parent's effective children — the extensibility point) + vertical extends inheritance (additive union of super's children+attrs; TS tracks super as <type>.base today; conflicting redefs left for check #6). Never subtracts. constraint-validate.ts: validateConstraints() runs the six contradiction checks (dangling ref, unsatisfiable required child, bad cardinality, closed-set clash, required-child cycle via colored DFS, conflicting attr redefinition), each → ERR_INVALID_METAMODEL_CONSTRAINT (new code, added to errors.ts + the shared ERROR-CODES.json allowlist). Wired into bootstrap: composeRegistry(providers, {validate:true}) and MetaDataLoader's default registry — so a contradictory provider set fails at load. The library's own metamodel is contradiction-free (no check loosened). 12 new tests; metadata suite 1739/0; full server suite green; typecheck clean. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
First real provider conversion under FR-033. The field provider's declarative metamodel definition (per-subtype attr constraints + descriptions + rule prose) is externalized to spec/metamodel/field.json as a ProviderDefinition and embedded at build, replacing the hand-coded field-schema.ts composition. - Authored spec/metamodel/field.json: faithfully externalizes commonFieldAttrs + the currency/enum special attrs (descriptions copied verbatim) and the 15-entry subtype list, with a one-line description per subtype + rules prose where a complex cross-cutting rule applies (currency wire-format, enum member pattern, object @storage). Attrs use the unified children shape (type:"attr" entries); per-subtype dataType matches MetaField's internal FIELD_DATA_TYPE map. - Added scripts/generate-embedded-metamodel.ts (globs spec/metamodel/*.json, repo-root discovery, no hardcoded paths) emitting field-definition.embedded.ts (FIELD_DEFINITION typed object literal); gated by a deep-equal drift test. - Rewired core-types.ts field loop to defineProviderFromData(FIELD_DEFINITION, FIELD_FACTORIES), post-assigning the unchanged fieldRules so registered childRules stay byte-identical (the JSON carries attrs only). Trimmed field-schema.ts to the still-shared normalizeAttr (used by object.value too). - Made ChildDef.subType optional in provider-data.ts so a polymorphic attr (@default) can omit its value-type, matching AttrSchema's optional valueType. - Added a completeness check asserting the composed registry registers, per subtype, exactly the pre-FR-033 attr name-set + valueType + required + dataType. Manifest unchanged: fixtures/registry-conformance/expected-registry.json stays byte-identical (the manifest excludes description/default/allowedValues/childRules; attr name/valueType/isArray/required sets are preserved). 1758 metadata tests pass (was 1739 + 19 new), workspace typecheck clean. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… + coverage gate (TS) ManifestType/ManifestAttr now carry description + optional rules/example/whenToUse, and each ManifestType carries the structural constraint graph: children (from childRules — childType/childSubType/childName + optional cardinality min/max/named, canonically sorted) and optional parents (the FR-033 child-side placement claim, sorted). description is required + non-empty on every type and attr; a coverage gate (registry-conformance.test.ts) hard-fails CI on any empty/missing description (mirrors ADR-0023). childSubType may be a string, "*", or a string[]; cardinality keys are emitted ONLY when defined on the rule (max: null = unbounded emitted, undefined omitted) — legacy wildcard rules fabricate no cardinality. expected-registry.json regenerated — the diff is PURELY ADDITIVE (every pre-existing field — type/subType, attr name/valueType/isArray/required — is byte-unchanged; only new keys were added; +1167 lines). A committed regen script (scripts/regen-expected-registry.ts, repo-root-discovery, no hardcoded paths) drives Phase 2 regeneration. README + exclusions-file header updated: type/attr description + the constraint graph are now first-class manifest properties (the COMMON_ATTR_DUP carve-out for the per-type attr NAMED "description" is unchanged — distinct mechanism). Growing the canonical RED-flags the other four ports' (Java/C#/Python/Kotlin) registry-conformance until Phase 4 — the intended, documented intermediate state, same lifecycle as the FR-032 sweep. The TS reference is correctly advanced; the ports reconcile later. TS reference: 1760/1760 metadata tests pass, typecheck clean. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Second provider conversion under FR-033, mirroring the field conversion (a72f8bb). The attr type family's declarative metamodel definition is externalized to spec/metamodel/attr.json as a ProviderDefinition and embedded at build, replacing the hand-coded probe-and-register loop in core-types.ts. - Authored spec/metamodel/attr.json: provider metaobjects-core-types, one TypeDef per attr subtype (9: base/string/int/long/double/boolean/class/ properties/filter — stringarray stays unregistered) with a real one-line description (replacing the generic "Attribute of type X") and the per-subtype dataType probed off the subtype's MetaAttr class (base/string/class→string, int→int, long→long, double→double, boolean→boolean, properties/filter→object). No children — attrs are leaf value-type vocabulary (no childRules, no attrs). - Generated core/attr/attr-definition.embedded.ts (ATTR_DEFINITION) via scripts/generate-embedded-metamodel.ts; gated by a deep-equal drift test (test/attr-definition-embed.test.ts). - Rewired core-types.ts attr block to defineProviderFromData(ATTR_DEFINITION, ATTR_FACTORIES). ATTR_FACTORIES is a loop over ATTR_SUBTYPES constructing via attrClassFor(subType); each class computes its own dataType internally, so no setDataType. childRules are empty for attr, so NO post-assign is needed (unlike field, which carries the validator/view/attr/origin wildcards). - Added a completeness check (test/attr-definition-completeness.test.ts) asserting, per subtype, empty attributes, empty childRules, and dataType == the pre-FR-033 probe. Regenerated fixtures/registry-conformance/expected-registry.json: the diff touches ONLY the 9 attr.* type-level description fields (dataType is excluded from the manifest; attr name/attrs/children sets are unchanged). The other four ports (C#/Java/Kotlin/Python) go RED on registry-conformance until their Phase-4 reconciliation — the same documented intermediate state as the field conversion. 1772 metadata tests pass (was 1760 + 12 new), metadata typecheck clean. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…(embedded) Phase 2 provider conversion mirroring the landed field (a72f8bb) and attr (e3e33c4) conversions. The validator type family's declarative definition (6 subtypes — base/required/length/regex/numeric/array — vocabulary + per-subtype attr constraints + real descriptions) is externalized to spec/metamodel/validator.json, embedded at build into VALIDATOR_DEFINITION. - spec/metamodel/validator.json: 6 TypeDefs with real subtype descriptions replacing the generic "Validator (X)"; @min/@max (int, optional) on base/length/numeric/array, +@pattern (string, optional) on regex, none on required. Attr descriptions copied verbatim from the old validator-schema.ts. No dataType (validators carry no typed value). - core/validator/validator-definition.embedded.ts: generated embed + a deep-equal drift gate (test/validator-definition-embed.test.ts). - core-types.ts: rewired the validator block to defineProviderFromData( VALIDATOR_DEFINITION, VALIDATOR_FACTORIES) with the subType→class factory reusing VALIDATOR_CLASS_MAP; the structural childRules (= [wildcard(attr)]) are kept byte-identical by post-assigning the constant validatorRules array after lowering (the wildcard can't be a type:"attr" JSON entry). Removed the now-dead VALIDATOR_ATTRS_MAP import and deleted the fully-superseded validator-schema.ts. - test/validator-definition-completeness.test.ts: per-subtype asserts the registered attr name-set + valueType + required + childRules == [wildcard(attr)] + no dataType, hardcoded from the pre-conversion VALIDATOR_ATTRS_MAP. - expected-registry.json regenerated: the diff touches ONLY the 6 validator.* descriptions; the children/attrs blocks (the attr wildcard + @min/@max/@pattern) were already in the Task-5 canonical and stay byte-identical. metadata suite 1788 pass / 0 fail; typecheck + build clean. Other ports go RED until Phase 4 (expected). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…mbedded) Phase 2 provider conversion mirroring the landed validator (58c0eda), field (a72f8bb), and attr (e3e33c4) conversions. The identity type family's declarative definition (3 subtypes — primary/secondary/reference, NO base — vocabulary + per-subtype attr constraints + real descriptions) is externalized to spec/metamodel/identity.json, embedded at build into IDENTITY_DEFINITION. - spec/metamodel/identity.json: 3 TypeDefs with real subtype descriptions replacing the generic "Identity (X)". All three carry the required, isArray @fields (min:1/max:1/isArray:true); primary adds @generation (allowedValues increment/uuid/assigned), secondary adds @unique (boolean), reference adds @references (required) + @Enforce (boolean). Attr descriptions copied verbatim from the old identity-schema.ts. No dataType (identities carry no typed value). - core/identity/identity-definition.embedded.ts: generated embed + a deep-equal drift gate (test/identity-definition-embed.test.ts). - core-types.ts: rewired the identity block to defineProviderFromData( IDENTITY_DEFINITION, IDENTITY_FACTORIES) with the subType->class factory reusing IDENTITY_CLASS_MAP (primary->MetaPrimaryIdentity, secondary->MetaSecondaryIdentity, reference->MetaReferenceIdentity, default->MetaIdentity). The structural childRules (= [wildcard(attr)]) are kept byte-identical by post-assigning the constant identityRules array after lowering (the wildcard can't be a type:"attr" JSON entry). Removed the now-dead identityFieldsAttr/IDENTITY_ATTRS_MAP imports and deleted the fully-superseded identity-schema.ts (core-types.ts was its only importer). - test/identity-definition-completeness.test.ts: per-subtype asserts the registered attr name-set + valueType + isArray + required + allowedValues + childRules == [wildcard(attr)] + no dataType, hardcoded from the pre-conversion IDENTITY_ATTRS_MAP. - expected-registry.json regenerated: the diff touches ONLY the 3 identity.* descriptions; the children/attrs blocks (the attr wildcard + @fields/ @generation/@unique/@references/@Enforce) were already in the Task-5 canonical and stay byte-identical. metadata suite 1798 pass / 0 fail (baseline 1788 + 10 new); typecheck + build clean. Other ports go RED until Phase 4 (expected). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
….json (embedded) Phase 2 provider conversion mirroring the landed validator (58c0eda), identity (3eb610d), field (a72f8bb), and attr (e3e33c4) conversions. The relationship type family's declarative definition (4 subtypes — base/association/aggregation/ composition, all backed by the single MetaRelationship node class with NO subType->class dispatch — vocabulary + the 7 shared attr constraints + real descriptions + the complex M:N rules prose) is externalized to spec/metamodel/relationship.json, embedded at build into RELATIONSHIP_DEFINITION. - spec/metamodel/relationship.json: 4 TypeDefs with real subtype descriptions replacing the generic "Relationship (X)". All four carry the SAME 7 attrs, copied VERBATIM from the old relationship-schema.ts (relationshipAttrs): @Cardinality (string, open — no allowedValues), @objectref, @through, @sourceRefField (string, optional), @Symmetric (boolean), @onDelete / @onUpdate (string, allowedValues cascade/set-null/restrict/no-action). Attr descriptions copied verbatim. No dataType (relationships carry no typed value). Each TypeDef carries a top-level `rules` string capturing the genuinely-complex M:N model (@through junction MUST declare two identity.reference children; FK fields are DERIVED from those references, never restated; @sourceRefField = directed self-join, @Symmetric = undirected self-join, mutually exclusive) plus the per-subtype lifecycle semantics (composition cascade / aggregation set-null / association restrict) — documentation prose; enforcement stays in code. - core/relationship/relationship-definition.embedded.ts: generated embed + a deep-equal drift gate (test/relationship-definition-embed.test.ts). - core-types.ts: rewired the relationship block to defineProviderFromData( RELATIONSHIP_DEFINITION, RELATIONSHIP_FACTORIES) with the subType->class factory built over RELATIONSHIP_SUBTYPES mapping every subtype to MetaRelationship. The structural childRules (= [wildcard(attr)]) are kept byte-identical by post-assigning the constant relationshipRules array after lowering (the wildcard can't be a type:"attr" JSON entry). Removed the now-dead relationshipAttrs import and deleted the fully-superseded relationship-schema.ts (core-types.ts was its only importer; relationship-constants.ts kept). - test/relationship-definition-completeness.test.ts: per-subtype asserts the registered attr name-set + valueType + required + allowedValues match the pre-conversion relationshipAttrs, childRules == [wildcard(attr)], no dataType. - expected-registry.json regenerated: the diff touches ONLY the 4 relationship.* descriptions AND adds `rules` to each; the children/attrs blocks stay byte-identical to the Task-5 canonical. metadata suite 1810 pass / 0 fail (baseline 1798 + 12 new); typecheck + build clean. Other ports go RED until Phase 4 (expected). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ded) 4 subtypes (base/entity/value/projection): verbatim @discriminator/@discriminatorValue on all + @normalize on value, real descriptions, and ADR-0028 taxonomy rules prose (entity owns data; value pure shape no identity/source; projection derived read-only, omits relationship+template). Per-subtype structural childRules are post-assigned in code (entity adds template; projection drops relationship+template; base/value carry relationship, no template) — the completeness test pins both attrs and childRules per subtype, and the regenerated canonical's object.* children blocks are byte-identical (only descriptions changed + rules added). Frees and deletes the last consumer of core/field/field-schema.ts. RED-flags non-TS ports' registry-conformance until Phase 4. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ded) Generalizes scripts/generate-embedded-metamodel.ts (concept→dir map covering persistence/presentation/template; computed relative provider-data import) — the 6 existing embeds regenerate byte-identical. Converts origin (4 subtypes) via defineProviderFromData + post-assigned childRules; completeness check; canonical regenerated (origin.* descriptions only). RED-flags non-TS ports until Phase 4. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ded) Phase 2 provider conversion mirroring the landed origin (bed24df) and validator (58c0eda) conversions. Converts ONLY the CORE source registration (the bare source shells — base/rdb — that coreTypesProvider registers with NO own attrs) to read from the externalized spec/metamodel/source.json, embedded at build into SOURCE_DEFINITION. - spec/metamodel/source.json: 2 TypeDefs (base, rdb) with real descriptions replacing the generic "Source (X)" + ADR-0007 rules prose on rdb (source declares where an object's data lives; rdb is the relational paradigm subtype; physical name is @table not the structural name; read-only-ness is derived from @kind; multi-source via @ROLE with exactly one primary per object). children == EMPTY (no own attrs) and no dataType — the @table/@kind/@role/@schema/@parameterRef attrs on source.rdb are NOT here; they are contributed by a SEPARATE provider (dbProvider, persistence/db) via registry.extend, which is UNTOUCHED by this conversion. - persistence/source/source-definition.embedded.ts: generated embed + a deep-equal drift gate (test/source-definition-embed.test.ts). The 7 existing embeds regenerate byte-identical. - core-types.ts: rewired the source block to defineProviderFromData( SOURCE_DEFINITION, SOURCE_FACTORIES) over SOURCE_SUBTYPES (all -> MetaSource); the structural childRules (= [wildcard(attr)]) are kept byte-identical by post-assigning the constant sourceRules array after lowering. No dead imports to remove (the old source block used no schema import). - test/source-definition-completeness.test.ts: composes with CORE ONLY (composeRegistry([coreTypesProvider])) so the dbProvider extend does not add attrs, and asserts per subtype: own attributes == [], childRules == [wildcard(attr)], no dataType. - expected-registry.json regenerated: the composed canonical includes dbProvider, so source.rdb's attrs block (table/view/kind/role/schema/...) stays byte-identical — the diff touches ONLY the source.base + source.rdb descriptions (+ the rdb rules prose). metadata suite 1846 pass / 0 fail; typecheck + build clean. Other ports go RED until Phase 4 (expected). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ded) Phase 2 provider conversion mirroring the landed validator (58c0eda) and origin (bed24df) conversions. The layout type family's declarative definition (2 subtypes — base/dataGrid — vocabulary + the 6 dataGrid attr constraints + real descriptions) is externalized to spec/metamodel/layout.json, embedded at build into LAYOUT_DEFINITION. - spec/metamodel/layout.json: 2 TypeDefs with real subtype descriptions replacing the generic "Layout (X)" (base = abstract object-level UI surface; dataGrid = a metadata-driven data grid). The 6 dataGrid attrs are copied verbatim from the old layout-schema.ts (dataGridLayoutAttrs): @pageSize (int), @defaultSortField (string), @defaultSortOrder (string, allowedValues asc/desc), @filterable (boolean), @filter (valueType filter), @columns (string, isArray:true) — all optional. base carries no attrs. No dataType. - presentation/layout/layout-definition.embedded.ts: generated embed + a deep-equal drift gate (test/layout-definition-embed.test.ts). - core-types.ts: rewired the layout block to defineProviderFromData( LAYOUT_DEFINITION, LAYOUT_FACTORIES); every subtype maps to the single MetaLayout class (layouts carry no per-subtype behavior). The structural childRules (= [wildcard(attr)]) are kept byte-identical to the pre-FR-033 registration by post-assigning them after lowering (the wildcard can't be a type:"attr" JSON entry). Deleted the now-dead layout-schema.ts; only core-types.ts imported dataGridLayoutAttrs. - test/layout-definition-completeness.test.ts: per-subtype asserts the registered attr name-set + valueType (incl. @filter→"filter") + isArray (incl. @columns→true) + required + allowedValues (asc/desc) match the pre-conversion dataGridLayoutAttrs; childRules == [wildcard(attr)]; no dataType. - expected-registry.json regenerated: the diff touches ONLY the 2 layout.* descriptions; the dataGrid attrs/children blocks were already in the Task-5 canonical and stay byte-identical. metadata suite 1854 pass / 0 fail (baseline 1846 + 8 new); typecheck clean. The 8 existing embeds regenerated byte-identical. Other ports go RED until Phase 4 (expected). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Phase 2 provider conversion mirroring the landed field (a72f8bb), validator (58c0eda) and layout (6fb2e11) conversions. The view type family's declarative definition (13 subtypes — base/text/textarea/date/month/hotlink/dropdown/radio/ checkbox/number/password/hidden/web/currency — vocabulary + real per-subtype descriptions + the single @Locale attr on view.currency) is externalized to spec/metamodel/view.json, embedded at build into VIEW_DEFINITION. - spec/metamodel/view.json: 13 TypeDefs with real one-line descriptions replacing the generic "View (X)" — including the 11 generic presentation controls that are manifest-excluded (PRESENTATION_ONLY) but still REGISTERED, so every subtype carries a faithful description for the embed + docs. Only view.currency carries an attr: @Locale (string, optional, default "en-US", description copied verbatim from the old view-schema.ts); the other 12 have no children. No dataType (single MetaView class, no typed value). - presentation/view/view-definition.embedded.ts: generated embed + a deep-equal drift gate (test/view-definition-embed.test.ts). The 9 existing embeds regenerate byte-identical. - core-types.ts: rewired the view block to defineProviderFromData( VIEW_DEFINITION, VIEW_FACTORIES); every subtype maps to the single MetaView class (views carry no per-subtype behavior). The structural childRules ([wildcard(attr)]) are kept byte-identical to the pre-FR-033 registration by post-assigning them after lowering (the wildcard is not modeled in the JSON, which carries attrs only). Removed the now-dead currencyViewAttrs import + the VIEW_SUBTYPE_CURRENCY import (used only by the old loop) and deleted the fully-superseded view-schema.ts. view-constants.ts kept (VIEW_SUBTYPE_CURRENCY still consumed by registry-manifest-exclusions / codegen / tests). - test/view-definition-completeness.test.ts: asserts all 13 subtypes registered; view.currency has exactly the @Locale attr (valueType "string", required false, default "en-US"); all 12 other subtypes have own attributes == []; childRules == [wildcard(attr)] for every subtype; no dataType. - expected-registry.json regenerated: the diff touches ONLY the view.base + view.currency descriptions — the 11 generic controls are manifest-excluded so their description change is invisible to the canonical; the view.currency @Locale attr block stays byte-identical. metadata suite 1886 pass / 0 fail (was 1854 + 32 new); typecheck + build clean. Other ports go RED until Phase 4 (expected). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…mbedded) Convert coreTypesProvider's template registration to read its declarative definition from spec/metamodel/template.json (embedded at build into TEMPLATE_DEFINITION), lowered via defineProviderFromData. The most attr-heavy provider: 4 subtypes (base/prompt/output/toolcall) with all ~16 attrs externalized verbatim incl. allowedValues + defaults + required flags — @Format (default "text", 7 formats), @PromptStyle (default "guide", 3 styles), @kind (default "document", 2 kinds), required @payloadRef (prompt/output) and required @toolname + @payloadRef (toolcall). Real subtype descriptions replace the old "Template (X)" placeholders, plus FR-004/ADR-0011 rules prose capturing the document-vs-email @textRef/@subjectRef/@htmlBodyRef presence rule and the toolcall no-renderable-text rationale. The factory stays code via TEMPLATE_FACTORIES (every subType -> MetaTemplate); structural childRules post-assigned [wildcard(attr)] byte-identically. The separate templateProvider (which EXTENDS every field subtype with @xmlText) is UNTOUCHED — this conversion covers only the core template registration. The depth-1 embed import is ../provider-data.js and resolves. A completeness check pins every attr (name-set / valueType / isArray / required / default / allowedValues + childRules + no dataType) against the pre-FR-033 TEMPLATE_ATTRS_MAP; a drift gate deep-equals the embed to the JSON. Canonical registry regenerated: the ONLY diff is template.* descriptions + new rules prose (attrs/children blocks byte-unchanged). Deleted the now-unused template-schema.ts (only importer was core-types.ts). RED-flags non-TS ports until Phase 4. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…on.json (embedded) Adds optional commonAttrs to ProviderDefinition + defineCommonAttrsFromData (the common-attr analog of defineProviderFromData). docProvider now registers the 8 universal doc common attrs from data instead of hand-coded commonDocAttrs. Canonical UNCHANGED (verbatim externalization — the commonAttrs block already carried these descriptions since Task 5), so no new cross-port RED. Completes the Phase-2 TS core-provider conversions. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…r split Revises the FR-033 approach from the loose model (attrs lumped in core, code-post-assigned wildcard childRules, bespoke documentation path) to the pre-6.0-Java strict model: per-subtype fail-closed child rules enforced in validate(), attrs scattered across 5 concern providers (core / db / documentation / prompt / ui), only documentation universal. Records the attr-ownership map and the S0..S5 implementation sequence. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…on attrs + structural placement enforcement Unifies the provider apply path (base-composition + a universal *.* entry for common attrs), deletes the bespoke commonAttrs helper, reworks documentation to the *.* universal model, and adds ERR_CHILD_NOT_ALLOWED structural-child enforcement in validate() (a no-op under today's wildcard childRules — the rail strict per-subtype rules will use in S1). Canonical byte-identical; full metadata suite green. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ui/prompt/db) Creates metaobjects-ui (filter/sort field attrs), expands templateProvider into metaobjects-prompt (adds @example/@Instruction + enum extract attrs + the object.value @normalize default to its existing @xmlText), grows metaobjects-db (@storage/@autoset), and drops all of these from core field.json. Pure ownership move at current subtype scope — canonical byte-identical. (Strict per-subtype re-scoping is step B.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
field.json restructured to the strict model: field.base holds the universal core attrs (@required/@readOnly/@default/@unique) + the open validator/view/origin child rules; each subtype extendsBase and adds only its own (@maxlength→string, @precision/@scale→decimal, @Currency→currency, @objectref→object, @values/@provided →enum). The "any attr" wildcard child rule is dropped (strict/fail-closed); childRules no longer post-assigned in code. db @storage→object, @autoset→temporal. Misplaced attrs now rejected (ERR_UNKNOWN_ATTR). Canonical field.* re-scoped; corpus fixtures fixed: auto-set-on-create / auto-set-on-update / auto-set-on-create-and-update (@autoset field.string → field.timestamp). Non-TS ports RED until S5. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…roviders validator/view/layout/source/origin/identity/relationship now register empty childRules (named attrs only, no any-attr catch-all); metadata.root drops its attr wildcard but keeps its open structural child rules. Strict/fail-closed: a structural child under an attr-only type is now ERR_CHILD_NOT_ALLOWED, a misplaced attr stays ERR_UNKNOWN_ATTR. Canonical drops the (attr,*,*) childRule from those types; attrs unchanged. Corpus fixtures fixed: none. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…core provider) object.json restructured to the strict model via extendsBase: object.base holds the intersection child rules (field/identity/validator/layout/source); value adds relationship; entity adds relationship+template; projection inherits base only. The any-attr wildcard is dropped from all four. @discriminator/@discriminatorValue scoped to object.entity only (FR-014 TPH is entity-inheritance). childRules no longer post-assigned in code. Canonical object.* re-scoped; corpus fixtures fixed: none (no fixture placed a discriminator on a non-entity object or a relationship/template under projection). Completes the strict re-scope of all core providers. Non-TS ports RED until S5. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…ads spec/metamodel/db.json
Adds ProviderDefinition.extends ({type, subType: value|list|"*", children}) applied via
registry.extend in the unified apply path, so concern (extend) providers are data-driven
like the core types. Converts dbProvider to read spec/metamodel/db.json (field @column/
@db.indexed/@dbColumnType + @storage→object + @autoset→temporal + source.rdb attrs).
Canonical byte-identical. (ui/prompt follow in S1.5-B; type-attr re-home in S2.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…prompt}.json uiProvider and promptProvider now register their field/object.value attr extensions from data (spec/metamodel/ui.json, prompt.json) via the S1.5-A extends mechanism, replacing the hand-coded registry.extend arrays. All three concern providers (db/ui/ prompt) are now data-driven. Canonical byte-identical. (Type-attr re-home is S2.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…pt; template strict view.currency @Locale + layout.dataGrid attrs move to the ui provider; template.* type attrs (@payloadRef/@textRef/@format/@kind/...) move to the prompt provider — all via data-driven extends. template's any-attr wildcard childRule is dropped (strict completion; toolcall vendor attrs stay extensible via consumer registry.extend). Composed attrs unchanged (canonical attrs byte-identical); only template.* children drop the (attr,*,*) wildcard. Completes the 5-provider concern ownership. Corpus fixtures fixed: extends-view-triple-nest (+metaobjects-ui), flattened-kitchen-sink (+metaobjects-prompt). Non-TS ports RED until S5. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…er index) Generates LLM-readable docs FOR THE METAMODEL from the strict registry: metamodel/ INDEX.md (every type.subType + one-liner + link), types/<type>.md (each subtype's full composed allowed attrs — provider-tagged — + child rules + cardinality), and providers.md (the 5-concern ownership index). New `meta docs --metamodel` surface (needs no user metadata); byte-gated docs fixture. Authoring-skill wiring is a follow-on. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…EX/type pages providers.md's core "Owns" list now applies isExcludedTypeSubType, so the 11 manifest-excluded generic view.* controls no longer appear there (they're absent from INDEX/type pages). All three doc surfaces now share one vocabulary. Fixture regenerated; consistency gated by a new assertion. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…cipe Measured the non-TS gap (Java RED: emits v1 manifest only; golden now carries description/rules/children/parents + strict per-subtype scoping). Records the two load-bearing facts (manifest is provider-agnostic → 5-split is internal Tier-3; descriptions must be byte-identical → ports read the shared spec/metamodel/*.json) + the per-port recipe (embed JSON → data-driven register → grow emitter → byte-match → enforce) and Java→C#→Python sequencing (Kotlin via JVM). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… slots Grows the Java RegistryManifest emitter to the TS reference shape (per-type description/rules/example/whenToUse + per-attr description + the children constraint graph [any-attr wildcard dropped] + parents), and adds the optional model slots (ChildRequirement doc description; TypeDefinition rules/example/ whenToUse/parents) the JSON reader will populate in sub-step B. Emitter structure matches registry-manifest.ts byte layout; RegistryManifestConformanceTest stays RED until S-B sources descriptions + the strict graph from spec/metamodel/*.json. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…escriptions Bundles the 15 shared spec/metamodel/*.json into the metadata module (byte-gated) and adds a reader that sources every type/attr/common-attr description (+ rules/ example/whenToUse) from them onto the Java registry — single-sourced, byte-identical to TS, never hand-copied. The manifest's description fields now match the golden; RegistryManifestConformanceTest stays RED pending the strict children graph + cardinality + per-subtype attr scoping reconciliation (sub-step B2). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…rom JSON Adds cardinality (min/max/named) to ChildRequirement and extends the pre-seal apply hook to replace each type's structural child requirements with the strict graph from spec/metamodel/*.json (extendsBase-composed, any-attr/extra wildcards dropped). The manifest children blocks now byte-match the golden; RegistryManifestConformanceTest stays RED pending per-subtype attr scoping + the ui/prompt providers + metadata.root (B2b). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… registry-conformance GREEN Prunes each type.subType's logical attr requirements to the strict per-subtype set sourced from spec/metamodel/*.json (extendsBase + extends composed), so the manifest attrs blocks byte-match the golden and the loader rejects misplaced attrs (ERR_UNKNOWN_ATTR). Fixes metadata.root (the hand-coded root wrapper) to the strict 4 structural children + "Root metadata document". RegistryManifestConformanceTest is now GREEN — the Java metadata registry byte-matches the cross-port canonical. Corpus fixtures fixed: none (no fixture placed a now-disallowed attr; the loader-test root-attr capability is preserved via the retained any-attr wildcard, which is not a structural child so it does not affect the manifest). Also fixed pre-existing NUL bytes that had been written in place of space separators in two sort-key string literals (SpecMetamodelReader.structKey, RegistryManifest child sort key). The strict allow-list per (type, subType) is composed in SpecMetamodelReader .strictAttrNames exactly as TS composes it: the subtype's own types[].children attrs, plus <type>.base's attrs when extendsBase:true, plus every extends block (db/ui/prompt JSON) whose matcher targets the subtype (exact / list / "*"). A new pre-seal apply pass drops every INCLUDED attr requirement (direct + inherited) not in that set, leaving the classifyPerTypeAttr carve-outs (isArray/isAbstract/extends/implements/isInterface/ object/objectAdapter/description) registered. (The 5 metaobjects-ui/prompt-provider ConformanceTest failures are a separate gate — next sub-step.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…el slots Grows the Python registry_manifest emitter to the TS reference shape (per-type description/rules/example/whenToUse + per-attr description + the children constraint graph + parents) and adds the optional model slots (AttrSchema.description; ChildRule.child_name/min/max/max_is_null/named; TypeDefinition.description/rules/ example/whenToUse/parents) the JSON reader fills in S-B1. test_registry_conformance stays RED until S-B sources descriptions + the strict graph from spec/metamodel/*.json. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… descriptions Embeds the 15 shared spec/metamodel/*.json as Python package data (byte-gated) + adds a reader sourcing every type/attr/common description (+rules/example/whenToUse) from them onto the Python registry pre-seal — single-sourced, byte-identical to TS, never hand-copied. The manifest's description fields now match the golden; test_registry_conformance stays RED pending the strict children graph + cardinality + per-subtype attr scoping (S-B2a/B2b). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…oot → registry-conformance GREEN Extends the pre-seal apply hook to replace each type's structural child_rules with the strict graph + cardinality from spec/metamodel/*.json (extendsBase-composed) and prune its logical attrs to the strict per-subtype allow-list (loader now rejects misplaced attrs); fixes metadata.root (hand-coded root wrapper); adds the ERR_CHILD_NOT_ALLOWED/ ERR_INVALID_METAMODEL_CONSTRAINT error codes. test_registry_conformance is now GREEN — the Python metadata registry byte-matches the cross-port canonical. Corpus fixtures fixed: none. (The 5 metaobjects-ui/prompt-provider conformance failures are a separate provider-composition gate — follow-on.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…tch TS); 5 conformance fixtures green Adds data-driven metaobjects-ui + metaobjects-prompt providers (reading the embedded ui.json/prompt.json) and re-homes @filterable/@sortable/@sortableDefaultOrder + view/ layout type attrs (ui) and @example/@instruction/@xmlText/enum-extract + template type attrs (prompt) out of core into them, matching the TS provider split — so the 5 conformance fixtures that declare those provider ids compose. registry-conformance stays GREEN (manifest is provider-agnostic; composed attr set unchanged); no double-registration. metaobjects-template kept as a back-compat alias. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…pt providers (match TS/Python) Brings Java in line with TS + Python: ui/prompt metadata config now lives in dedicated data-driven metaobjects-ui + metaobjects-prompt providers (reading the embedded ui.json/ prompt.json extends), NOT in core. Renames TemplateTypesMetaDataProvider → metaobjects-prompt (absorbs @xmlText; metaobjects-template kept as a back-compat alias), adds metaobjects-ui. Removes the re-homed attrs from core (no double-registration). registry-conformance + Kotlin stay GREEN (manifest is provider-agnostic); the 5 metaobjects-ui/prompt ConformanceTest fixtures now pass. All three ports now share the same provider split. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…lots Grows the C# RegistryManifest emitter to the TS reference shape (per-type description/ rules/example/whenToUse + per-attr description + the children constraint graph + parents) and adds the optional model slots (ChildRule Min/Max/MaxIsNull/Named; TypeDefinition Rules/Example/WhenToUse/Parents; AttrSchema Rules/Example/WhenToUse — Description already existed) the JSON reader fills in S-B1. RegistryManifestConformance stays RED until S-B sources descriptions + the strict graph from spec/metamodel/*.json. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…criptions Embeds the 15 shared spec/metamodel/*.json as assembly resources (byte-gated) + adds a reader sourcing every type/attr/common description (+rules/example/whenToUse) from them onto the C# registry pre-seal — single-sourced, byte-identical to TS, never hand-copied (overriding the providers' placeholder descriptions). The manifest's description fields now match the golden; RegistryManifestConformance stays RED pending the strict children graph + cardinality + per-subtype attr scoping (S-B2). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…→ registry-conformance GREEN Extends the pre-seal apply hook to replace each type's structural child rules with the strict graph + cardinality from spec/metamodel/*.json (extendsBase-composed) and prune its logical attrs to the strict per-subtype allow-list (loader now rejects misplaced attrs); fixes metadata.root (hand-coded root wrapper); adds the missing error codes if gated. RegistryManifestConformance is now GREEN — the C# metadata registry byte-matches the cross-port canonical. Corpus fixtures fixed: none. (The 5 metaobjects-ui/ prompt ConformanceTests are a separate provider-composition gate — next step.) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
… providers (match TS/Java/Python) Completes the cross-port ui/prompt re-home: C# ui/prompt metadata config now lives in dedicated data-driven metaobjects-ui + metaobjects-prompt providers (reading the embedded ui.json/prompt.json extends), NOT in core. Renames TemplateProvider → metaobjects-prompt (absorbs @xmlText; metaobjects-template kept as a back-compat alias), adds metaobjects-ui. Removes the re-homed attrs from core (no double-registration). registry-conformance stays GREEN (manifest provider-agnostic); the 5 metaobjects-ui/prompt ConformanceTests now pass. All 5 ported engines now share the same provider split. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…pesProvider @xmlText is now registered by PromptMetaDataProvider (the ui/prompt re-home renamed TemplateProvider → metaobjects-prompt). Comment-only. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pins that a misplaced structural child is rejected at strict load: a relationship.association under object.projection (projection's strict child set omits relationship) and a field under an attr-only validator.required (empty strict child set). Java already enforces this eagerly at MetaData.addChild via MetaDataRegistry.acceptsChild against the strict child graph (S-B2a), throwing ERR_CHILD_NOT_ALLOWED — this test pins it against regression. Mirrors the TS child-placement-enforcement test. Additive (test only); registry-conformance unaffected. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Verifies + pins that a misplaced structural child is rejected at strict load (a relationship under object.projection; a field under an attr-only validator.required). Python's S-B2a strict child_rules were registered but the loader's attr-schema pass did NOT consult them for STRUCTURAL placement (only ERR_UNKNOWN_ATTR for misplaced attrs) — the two cases loaded clean. Wires the ERR_CHILD_NOT_ALLOWED check into loader/validation_passes.py (_validate_attr_schema Check 0b): a structural child (own_children — not an @-attr) not admitted by the parent's registered child_rules now raises under strict, mirroring TS attr-schema-validate.ts (shared wildcard / subtype-list match; lax mode stays a no-op; unregistered parent skipped). registry-conformance stays GREEN. Corpus fixtures fixed: none. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
C# had the gap (strict child rules registered but the loader didn't check structural placement). Wires the ERR_CHILD_NOT_ALLOWED check into load validation — a structural child not admitted by the parent's strict child rules now raises (mirroring TS + Python's Check 0b + Java's acceptsChild), and adds ERR_CHILD_NOT_ALLOWED to the ErrorCode enum. Gating test pins the two cases (relationship under object.projection; field under attr-only validator.required) + a lax no-op. registry-conformance stays GREEN; full project 0 failures. Corpus fixtures fixed: none. Completes structural-enforcement parity across all 5 ports. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
FR-033 — provider definitions as declarative data + strict per-subtype metamodel (GH #23)
Makes a provider's declarative metamodel definition (vocabulary + per-subtype constraints + descriptions + rule prose) live in single-sourced JSON (
spec/metamodel/*.json) that every port reads, replacing the hand-coded-5× model. The metamodel becomes a strict, fail-closed allow-list (the pre-6.0 typesConfig model recovered), enforced in each port'svalidate(), with descriptions + the constraint graph emitted into the registry manifest and rendered as tiered LLM docs.TypeScript (reference) — strict re-architecture
*.*). ui/prompt attrs re-homed out of core into their own data-driven providers.type.subTypedeclares exactly its allowed children (attrs + structural) with cardinality; the "any attr" wildcard is gone. Misplaced attr →ERR_UNKNOWN_ATTR; misplaced structural child →ERR_CHILD_NOT_ALLOWED. All child rules live in the JSON (none hand-assigned in provider code), viaextendsBasecomposition.description+rules/example/whenToUse+ thechildren/parentsconstraint graph (coverage-gated).meta docs --metamodelemitsINDEX.md+ per-type pages (full composed allowed attrs, provider-tagged) + a provider-ownership index, byte-gated.Cross-port fan-out — Java, Kotlin, Python, C#
Each engine reads the SAME embedded
spec/metamodel/*.json, registers descriptions + the strict constraint graph + strict per-subtype scoping (single-sourced, byte-identical to TS), grows its manifest emitter to the TS shape, and byte-matchesfixtures/registry-conformance/expected-registry.json. The manifest is provider-agnostic, so the 5-concern split is an internal detail — but each port also has the identical ui/prompt provider split (metaobjects-ui/metaobjects-prompt, with ametaobjects-templateback-compat alias).All five engines now agree:
registry-conformanceGREEN (byte-match the canonical) — TS / Java / Kotlin / Python / C#.metaobjects-ui/metaobjects-promptconformance fixtures pass everywhere.ERR_CHILD_NOT_ALLOWED) parity, each gated by a test.Notes for review
expected-registry.jsongrew (descriptions + constraint graph + strict scoping) — this is the deliberate cross-port contract; every port re-derives its registry from its embedded copy of the shared JSON and is asserted byte-equal.metadata.rootis the one hand-coded root-wrapper exception (not a provider-declared type) in each port.docs/superpowers/specs/2026-06-13-metamodel-self-description-design.md,docs/superpowers/plans/2026-06-14-fr-033-strict-constraint-model.md,docs/superpowers/plans/2026-06-14-fr-033-cross-port-fan-out.md.🤖 Generated with Claude Code