Contentrain/ai

Minor Changes

• 173326c: feat(mcp)!: enforce the field constraints the schema already accepted

  A project reported that items, accept and maxSize are accepted on a field but
  never enforced — emails: ["not-an-email"] and accept: "image/jpeg" against a
  .webp both produced zero errors. The report was right, and the surface was larger
  than the three properties it named: 4 of 27 field types had any semantic
  validation, three constraints were read by nothing, and none of it blocked a write.

  A constraint that isn't a constraint is worse than no constraint — the author stops
  looking.

  content_save now validates before committing and refuses to write. It ran
  plan → commit → validate → report, so an invalid value landed in git, was
  auto-merged, and the caller learned about it from a string in next_steps while
  status still said "committed". Validation now runs on the pending changes and
  blocks on errors, returning isError and no commit. Warnings still pass — they are
  heuristics, and a legitimate value can sit outside an approximate pattern. Only the
  entries being saved are fatal: a pre-existing bad entry elsewhere in the model does
  not hold up an unrelated save.

  Array items share the scalar rule set. They ran through a parallel type switch
  that knew 10 of the 27 types and checked only typeof, so min/max/pattern/
  options never reached an item, and items given as a FieldDef with a non-object
  type ({type:'array', items:{type:'string', max:50}}) matched no branch at all —
  silently unvalidated, while the type emitter rendered it as real. Items now recurse
  through the same validator, which also closes the integer split where 3.7 was
  rejected inside an array but accepted as a scalar.

  17 types were pure typeof checks. slug now uses the SLUG_PATTERN the
  codebase already owned — every shipped template declares slug: { type: 'slug' },
  so "Hello World!!" used to validate clean. date/datetime are parsed (the same
  check schedule.ts already did for meta), percent is range-checked, and color/
  phone warn. Mechanical rules are errors; heuristics are warnings. email/url
  keep their existing warning severity. rating is deliberately untouched — its scale
  is never declared, so any range would be invented.

  unique works on documents. It was gated on a context only the collection
  validator passed, so it was a no-op exactly where every shipped template declares it.
  On singletons it is now rejected at model_save: the model holds one record per
  locale, so there is nothing to compare against.

  The dead constraints, handled honestly. accept is enforced by extension-sniff
  and says that is what it is. default is coherence-checked at model_save (right
  type, within its own options) but not written into content. maxSize cannot be
  enforced by MCP — it holds a path, never the bytes — so model_save now says so and
  points at the provider, which owns the policy at ingest. The docs claimed all three
  worked; they no longer do.

  model_save rejects what it will not enforce. options on a non-select, items
  on a non-array, accept/maxSize on a non-media field, min > max, and an
  uncompilable pattern are now errors instead of silent no-ops. Nested fields/
  items schemas are validated recursively — they were typed z.unknown() and never
  checked. The field schema is .strict(): a typo'd constraint (requird: true) used
  to be stripped without a word.

  BREAKING CHANGE:

  • content_save rejects content it previously committed. Run contentrain_validate
    before upgrading to see what would now be blocked.
  • model_save rejects models it previously accepted (unknown keys, min > max,
    options on a non-select, unique on a singleton).
  • validateModelDefinition returns { errors, warnings } instead of string[].
  • Array-item type errors carry validateFieldValue's message ("Type mismatch:
    expected string, got number") instead of "must be a string". The field path is
    unchanged.
  • Nested object errors are qualified by their parent (seo.title, not title) —
    a bare name was ambiguous with a top-level field.

  @contentrain/types gains validateSemanticType, validateAccept and
  isMediaType; validateFieldValue now applies semantic and accept rules.

  Studio picks all of this up automatically — its content-validation.ts delegates to
  this validator.