@payloadcms/plugin-mcp 3.86.0-internal.ac46214 → 4.0.0-internal.183b315

package/src/utils/schemaConversion/sanitizeEntitySchema.ts

@@ -0,0 +1,529 @@
+import type { JsonSchemaType } from '../../types.js'
+
+/**
+ * Turns the JSON Schema that Payload generates for a collection or global into the input schema for
+ * an MCP create/update tool. In short, it:
+ *
+ * - drops fields a client can't set (`id`, `createdAt`, the draft `_status`, …),
+ * - rewrites Payload-specific field shapes (points, relationships) into plain JSON the model can fill,
+ * - and shrinks the result so it's cheaper for the model to read,
+ *
+ * while keeping every node valid JSON Schema draft 2020-12. Each step below is tagged with why it runs -
+ * **Correctness** (valid input the API accepts), **Size** (equivalence-preserving shrink), or **LLM
+ * ergonomics** (easier for the model to read/fill) - and carries a before/after example on its definition.
+ */
+export const sanitizeEntitySchema = (schema: JsonSchemaType): JsonSchemaType => {
+  // Work on a copy — the caller reuses the original schema elsewhere (e.g. when listing tools).
+  let result = structuredClone(schema)
+
+  // Correctness — drop the fields a client can't set (id, createdAt, updatedAt, draft _status) + collapse nullable types.
+  result = removeManagedFields(result)
+
+  // LLM ergonomics — rewrite point fields from a `[number, number]` tuple into a `{ longitude, latitude }` object.
+  result = pointFieldsToObjects(result)
+
+  // Correctness — a relationship value can be an ID or a populated doc; on input only the ID is valid, so keep that.
+  result = relationshipsToIds(result)
+
+  // Size — strip inert type-gen leftovers that only bloat the schema (`tsType`, block-collision notes).
+  result = removeTypeGenArtifacts(result)
+
+  // Size — where a `const` already pins a value, the sibling `type` is redundant; remove it.
+  result = dropRedundantConstType(result)
+
+  // Size — fold per-collection relationship/upload variants (identical but for `relationTo`) into one `enum`.
+  result = mergeConstDiscriminatedUnions(result)
+
+  // Size — pull any subschema that appears more than once into a single shared `$defs` entry.
+  result = deduplicateIntoDefinitions(result)
+
+  // LLM ergonomics — give `$defs` short, readable names (`Code`, `paragraph`, `node`) so the `$ref`s read nicely.
+  result = shortenDefinitionNames(result)
+
+  return result
+}
+
+/**
+ * Rebuilds a schema bottom-up, calling `visit` on each node after its children. Shared by the
+ * transforms that need to touch every node in the tree.
+ */
+const mapNodes = (node: unknown, visit: (node: JsonSchemaType) => JsonSchemaType): unknown => {
+  if (Array.isArray(node)) {
+    return node.map((child) => mapNodes(child, visit))
+  }
+  if (!node || typeof node !== 'object') {
+    return node
+  }
+  const out: Record<string, unknown> = {}
+  for (const [key, value] of Object.entries(node)) {
+    out[key] = mapNodes(value, visit)
+  }
+  return visit(out as JsonSchemaType)
+}
+
+// Payload sets these on every document, so an MCP client never provides them when creating/updating.
+const PAYLOAD_MANAGED_FIELDS = new Set(['_status', 'createdAt', 'id', 'updatedAt'])
+
+/**
+ * **Correctness.** Removes the fields a client can't set on create/update - `id`, `createdAt`, `updatedAt`, and the draft
+ * `_status` - from every field object's `properties` and `required` (recursing into nested objects and
+ * array items). Along the way it also collapses optional array/object types via {@link collapseNullableType}.
+ *
+ * @example
+ * { properties: { id: { type: 'string' }, tags: { type: ['array', 'null'], items: {} } }, required: ['id', 'tags'] }
+ * → { properties: { tags: { type: 'array', items: {} } }, required: ['tags'] }
+ */
+const removeManagedFields = (schema: JsonSchemaType): JsonSchemaType => {
+  if (schema.properties && typeof schema.properties === 'object') {
+    for (const field of PAYLOAD_MANAGED_FIELDS) {
+      delete schema.properties[field]
+    }
+    for (const key of Object.keys(schema.properties)) {
+      const prop = schema.properties[key] as JsonSchemaType
+      if (!prop || typeof prop !== 'object') {
+        continue
+      }
+      const isRequired = Array.isArray(schema.required) && schema.required.includes(key)
+      collapseNullableType(prop, isRequired)
+      if (prop.properties) {
+        removeManagedFields(prop)
+      }
+      if (prop.items && typeof prop.items === 'object' && !Array.isArray(prop.items)) {
+        removeManagedFields(prop.items)
+      }
+    }
+  }
+
+  if (Array.isArray(schema.required)) {
+    schema.required = schema.required.filter((name) => !PAYLOAD_MANAGED_FIELDS.has(name))
+    if (schema.required.length === 0) {
+      delete schema.required
+    }
+  }
+
+  return schema
+}
+
+/**
+ * Drops the `'null'` from an optional array/object type (`['array', 'null']` → `'array'`, likewise
+ * `'object'`). Payload marks a field optional by unioning its type with `'null'`, but a create/update
+ * tool already conveys "optional" through the `required` list - the client just omits the field - so the
+ * `'null'` is redundant here. Dropping it shows the model a plain `type: 'array'` instead of implying it
+ * should send a literal `null`.
+ *
+ * We never touch the `required` list, so this can't change whether a field is required. And we skip fields
+ * that ARE required (`isRequired`): there a `'null'` is a value the field genuinely accepts, not optionality
+ * encoding. Nullable scalars (`['string', 'null']`) are also left alone - they read fine as `string | null`.
+ *
+ * @example
+ * optional { type: ['array', 'null'] } → { type: 'array' }   (still optional, just no longer null-valued)
+ * required { type: ['array', 'null'] } → unchanged
+ * { type: ['string', 'null'] } → unchanged
+ */
+const collapseNullableType = (schema: JsonSchemaType, isRequired: boolean): void => {
+  if (isRequired || !Array.isArray(schema.type)) {
+    return
+  }
+  const nonNullTypes = schema.type.filter((t) => t !== 'null')
+  if (nonNullTypes.length === 1 && (nonNullTypes[0] === 'array' || nonNullTypes[0] === 'object')) {
+    schema.type = nonNullTypes[0]
+  }
+}
+
+/**
+ * **LLM ergonomics.** Rewrites a point field (stored as a two-number tuple) into a `{ longitude, latitude }` object, which
+ * is far easier for the model to fill in than a positional array. This is input-only and lossless: the
+ * create/update handler converts the object back to the `[longitude, latitude]` tuple Payload stores,
+ * via `transformPointDataToPayload`.
+ *
+ * @example
+ * { type: 'array', items: [{ type: 'number' }, { type: 'number' }] }
+ * → { type: 'object', properties: { longitude: { type: 'number' }, latitude: { type: 'number' } }, required: ['longitude', 'latitude'] }
+ */
+const pointFieldsToObjects = (schema: JsonSchemaType): JsonSchemaType => {
+  if (!schema || typeof schema !== 'object') {
+    return schema
+  }
+
+  const transformed = { ...schema }
+
+  if (transformed.properties && typeof transformed.properties === 'object') {
+    transformed.properties = Object.fromEntries(
+      Object.entries(transformed.properties).map(([key, value]) => {
+        if (!value || typeof value !== 'object') {
+          return [key, value]
+        }
+        const isArrayType =
+          value.type === 'array' || (Array.isArray(value.type) && value.type.includes('array'))
+        const isPointField =
+          isArrayType &&
+          Array.isArray(value.items) &&
+          value.items.length === 2 &&
+          value.items.every((item) => item && typeof item === 'object' && item.type === 'number')
+
+        if (isPointField) {
+          const isNullable = Array.isArray(value.type) && value.type.includes('null')
+          return [
+            key,
+            {
+              type: isNullable ? ['object', 'null'] : 'object',
+              description: value.description || 'Geographic coordinates (longitude, latitude)',
+              properties: {
+                latitude: { type: 'number', description: 'Latitude coordinate' },
+                longitude: { type: 'number', description: 'Longitude coordinate' },
+              },
+              required: ['longitude', 'latitude'],
+            },
+          ]
+        }
+
+        return [key, pointFieldsToObjects(value)]
+      }),
+    )
+  }
+
+  if (
+    transformed.items &&
+    typeof transformed.items === 'object' &&
+    !Array.isArray(transformed.items)
+  ) {
+    transformed.items = pointFieldsToObjects(transformed.items)
+  }
+
+  return transformed
+}
+
+/**
+ * **Correctness.** Reduces relationship/upload fields to the IDs a client actually sends. Payload types the value as
+ * "an ID or the full related document" - but the populated-document form only appears in read responses;
+ * on create/update you always reference a relationship by its ID. So we drop that `$ref` option, leaving
+ * the bare ID. A single remaining target collapses inline (with a description naming the target
+ * collection); several targets become an `anyOf` of IDs.
+ *
+ * @example
+ * { oneOf: [{ type: 'string' }, { $ref: '#/$defs/posts' }] }
+ * → { type: 'string', description: 'The ID of the related "posts" document.' }
+ */
+const relationshipsToIds = (schema: JsonSchemaType): JsonSchemaType => {
+  if (!schema || typeof schema !== 'object') {
+    return schema
+  }
+
+  const processed = { ...schema }
+
+  if (Array.isArray(processed.oneOf)) {
+    const isRelatedDocRef = (option: unknown): option is { $ref: string } =>
+      !!option && typeof option === 'object' && '$ref' in option
+
+    if (processed.oneOf.some(isRelatedDocRef)) {
+      // A relationship value is "an ID, or the populated related document". Keep the ID option(s) and
+      // drop the `$ref`s to the related collections, since a client only ever sends the ID.
+      const idOptions = processed.oneOf
+        .filter((option) => !isRelatedDocRef(option))
+        .map((option) => (typeof option === 'object' ? relationshipsToIds(option) : option))
+      const targetCollections = processed.oneOf
+        .filter(isRelatedDocRef)
+        .map((option) => option.$ref.replace('#/$defs/', ''))
+
+      if (idOptions.length === 1) {
+        delete processed.oneOf
+        Object.assign(processed, idOptions[0])
+        if (targetCollections.length > 0 && !processed.description) {
+          processed.description = `The ID of the related "${targetCollections.join('" or "')}" document.`
+        }
+      } else if (idOptions.length > 1) {
+        delete processed.oneOf
+        processed.anyOf = idOptions
+      }
+    } else {
+      processed.oneOf = processed.oneOf.map((option) =>
+        typeof option === 'object' ? relationshipsToIds(option) : option,
+      )
+    }
+  }
+
+  if (processed.properties && typeof processed.properties === 'object') {
+    processed.properties = Object.fromEntries(
+      Object.entries(processed.properties).map(([key, value]) => [
+        key,
+        typeof value === 'object' ? relationshipsToIds(value) : value,
+      ]),
+    )
+  }
+
+  if (processed.items && typeof processed.items === 'object' && !Array.isArray(processed.items)) {
+    processed.items = relationshipsToIds(processed.items)
+  }
+
+  // Lexical node unions and blocks live under `$defs` and have their own relationship fields, so walk
+  // those too — otherwise their `$ref`s would dangle once we don't bundle the related collections.
+  if (processed.$defs && typeof processed.$defs === 'object') {
+    processed.$defs = Object.fromEntries(
+      Object.entries(processed.$defs).map(([key, value]) => [
+        key,
+        typeof value === 'object' ? relationshipsToIds(value) : value,
+      ]),
+    )
+  }
+
+  return processed
+}
+
+/**
+ * **Size.** Strips type-generation leftovers that bloat the schema without helping the model: the `tsType`
+ * hint (a `json-schema-to-typescript` extension; JSON Schema validators just ignore it, they don't reject it)
+ * and the block-interface-collision note Payload adds to some block descriptions (the
+ * `block-interface-name-collisions` docs link set in `registerBlockInterface`, see configToJSONSchema.ts).
+ * Both are inert here - removing them only shrinks the schema (and spares the model irrelevant noise).
+ *
+ * @example
+ * { type: 'object', tsType: 'SerializedBlockNode', description: 'see …#block-interface-name-collisions' }
+ * → { type: 'object' }
+ */
+const removeTypeGenArtifacts = (schema: JsonSchemaType): JsonSchemaType =>
+  mapNodes(schema, (node) => {
+    delete (node as { tsType?: unknown }).tsType
+    const { description } = node as { description?: string }
+    if (
+      typeof description === 'string' &&
+      description.includes('block-interface-name-collisions')
+    ) {
+      delete (node as { description?: unknown }).description
+    }
+    return node
+  }) as JsonSchemaType
+
+/**
+ * **Size.** Removes `type` whenever a `const` sits next to it — the constant already fixes the value.
+ *
+ * @example
+ * { type: 'string', const: 'paragraph' } → { const: 'paragraph' }
+ */
+const dropRedundantConstType = (schema: JsonSchemaType): JsonSchemaType =>
+  mapNodes(schema, (node) => {
+    if ('const' in node && 'type' in node) {
+      delete (node as { type?: unknown }).type
+    }
+    return node
+  }) as JsonSchemaType
+
+/**
+ * **Size.** Merges the members of a `oneOf`/`anyOf` that are identical except for one `const`-valued property
+ * into a single member with that property as an `enum`. Since the members differ only by that one
+ * constant, the `enum` form accepts exactly the same values - it's just smaller. This folds the
+ * per-collection relationship and upload variants, which differ only in their `relationTo` constant.
+ *
+ * @example
+ * { oneOf: [
+ *   { properties: { relationTo: { const: 'posts' }, value: { type: 'string' } } },
+ *   { properties: { relationTo: { const: 'pages' }, value: { type: 'string' } } },
+ * ] }
+ * → { properties: { relationTo: { enum: ['posts', 'pages'] }, value: { type: 'string' } } }
+ */
+const mergeConstDiscriminatedUnions = (schema: JsonSchemaType): JsonSchemaType =>
+  mapNodes(schema, (node) => {
+    for (const keyword of ['oneOf', 'anyOf'] as const) {
+      const members = node[keyword]
+      if (Array.isArray(members) && members.length > 1) {
+        const merged = mergeMembersByConst(members)
+        if (merged) {
+          delete node[keyword]
+          Object.assign(node, merged)
+        }
+      }
+    }
+    return node
+  }) as JsonSchemaType
+
+/** Returns the merged member for {@link mergeConstDiscriminatedUnions}, or `null` if they can't merge. */
+const mergeMembersByConst = (members: Array<boolean | JsonSchemaType>): JsonSchemaType | null => {
+  const objects = members.filter(
+    (member): member is JsonSchemaType =>
+      typeof member === 'object' && member !== null && typeof member.properties === 'object',
+  )
+  if (objects.length < 2 || objects.length !== members.length) {
+    return null
+  }
+
+  for (const discriminator of Object.keys(objects[0]!.properties!)) {
+    // Each member must pin this property to a `const`...
+    const constValues: unknown[] = []
+    const everyMemberPinsConst = objects.every((member) => {
+      const prop = member.properties![discriminator]
+      if (prop && typeof prop === 'object' && 'const' in prop) {
+        constValues.push(prop.const)
+        return true
+      }
+      return false
+    })
+    if (!everyMemberPinsConst) {
+      continue
+    }
+
+    // ...and be otherwise identical (compare each member with the discriminator removed).
+    const fingerprintWithoutDiscriminator = (member: JsonSchemaType): string => {
+      const { [discriminator]: _discriminator, ...otherProperties } = member.properties!
+      return JSON.stringify({ ...member, properties: otherProperties })
+    }
+    if (new Set(objects.map(fingerprintWithoutDiscriminator)).size !== 1) {
+      continue
+    }
+
+    const uniqueConstValues = [...new Set(constValues)]
+    if (uniqueConstValues.length < 2) {
+      continue
+    }
+
+    // Replace the per-member `const` with a single `enum` of every value.
+    const merged = structuredClone(objects[0]!)
+    const discriminatorProp = merged.properties![discriminator]
+    if (discriminatorProp && typeof discriminatorProp === 'object') {
+      delete discriminatorProp.const
+      discriminatorProp.enum = uniqueConstValues as NonNullable<typeof discriminatorProp.enum>
+    }
+    return merged
+  }
+  return null
+}
+
+// A `$ref` costs ~30 characters once names are shortened, so sharing a subschema only pays off when
+// it's larger than that and appears more than once.
+const MIN_SHARED_SIZE = 64
+
+/** A subschema worth sharing: a standalone object/union/described schema, not a `$ref` or a primitive. */
+const isShareable = (node: unknown): node is Record<string, unknown> => {
+  if (!node || typeof node !== 'object' || Array.isArray(node) || '$ref' in node) {
+    return false
+  }
+  const schema = node as JsonSchemaType
+  return (
+    Array.isArray(schema.oneOf) ||
+    Array.isArray(schema.anyOf) ||
+    Boolean(schema.properties) ||
+    Boolean(schema.items) ||
+    typeof schema.description === 'string'
+  )
+}
+
+/**
+ * **Size.** Replaces any subschema that appears more than once with a single shared `$defs` entry referenced by
+ * `$ref`. Lossless — only the serialized size shrinks. A collection with several rich-text fields, for
+ * example, inlines the same large lexical node schema once per field; this collapses them into one.
+ *
+ * @example
+ * { properties: { billing: address, shipping: address } }
+ * → { properties: { billing: { $ref: '#/$defs/shared_0' }, shipping: { $ref: '#/$defs/shared_0' } }, $defs: { shared_0: address } }
+ */
+const deduplicateIntoDefinitions = (schema: JsonSchemaType): JsonSchemaType => {
+  // Count how often each shareable subschema appears, keyed by its serialized form.
+  const counts = new Map<string, number>()
+  const count = (node: unknown): void => {
+    if (isShareable(node)) {
+      const key = JSON.stringify(node)
+      if (key.length >= MIN_SHARED_SIZE) {
+        counts.set(key, (counts.get(key) ?? 0) + 1)
+      }
+    }
+    if (Array.isArray(node)) {
+      node.forEach(count)
+    } else if (node && typeof node === 'object') {
+      Object.values(node).forEach(count)
+    }
+  }
+  count(schema)
+
+  if (![...counts.values()].some((n) => n >= 2)) {
+    return schema
+  }
+
+  // Replace each subschema seen 2+ times with a `$ref`. The first time we meet one it becomes a shared
+  // entry; we never descend into a stored entry, so shared entries never reference one another.
+  const sharedEntries: Record<string, JsonSchemaType> = {}
+  const nameByKey = new Map<string, string>()
+  const share = (node: unknown): unknown => {
+    if (isShareable(node)) {
+      const key = JSON.stringify(node)
+      if ((counts.get(key) ?? 0) >= 2) {
+        let name = nameByKey.get(key)
+        if (!name) {
+          name = `shared_${nameByKey.size}`
+          nameByKey.set(key, name)
+          sharedEntries[name] = node as JsonSchemaType
+        }
+        return { $ref: `#/$defs/${name}` }
+      }
+    }
+    if (Array.isArray(node)) {
+      return node.map(share)
+    }
+    if (node && typeof node === 'object') {
+      return Object.fromEntries(
+        Object.entries(node).map(([childKey, value]) => [childKey, share(value)]),
+      )
+    }
+    return node
+  }
+
+  const result = share(schema) as JsonSchemaType
+  result.$defs = { ...result.$defs, ...sharedEntries }
+  return result
+}
+
+/**
+ * **LLM ergonomics.** Renames `$defs` entries to short, readable names so the `$ref`s stay legible to the model: a block's
+ * `blockType` (`Code`), a node's `type` (`paragraph`), `node` for a rich-text node union, otherwise the
+ * generated name with its disambiguating hash dropped. A numeric suffix keeps collisions unique.
+ *
+ * @example
+ * { properties: { body: { $ref: '#/$defs/LexicalNodes_9FBEC708' } }, $defs: { LexicalNodes_9FBEC708: {} } }
+ * → { properties: { body: { $ref: '#/$defs/node' } }, $defs: { node: {} } }
+ */
+const shortenDefinitionNames = (schema: JsonSchemaType): JsonSchemaType => {
+  const definitions = schema.$defs
+  if (!definitions || Object.keys(definitions).length === 0) {
+    return schema
+  }
+
+  const shortNameFor = (name: string, definition: boolean | JsonSchemaType): string => {
+    const properties = (typeof definition === 'object' && definition.properties) || {}
+    const constString = (key: string): string | undefined => {
+      const prop = properties[key]
+      return prop && typeof prop === 'object' && typeof prop.const === 'string'
+        ? prop.const
+        : undefined
+    }
+    const members =
+      typeof definition === 'object' ? (definition.anyOf ?? definition.oneOf) : undefined
+    const isNodeUnion =
+      Array.isArray(members) &&
+      members.some((m) => typeof m === 'object' && Boolean(m.$ref || m.properties?.type))
+    return (
+      constString('blockType') ??
+      constString('type') ??
+      (isNodeUnion ? 'node' : undefined) ??
+      name.replace(/_[0-9A-F]{6,}$/i, '')
+    )
+  }
+
+  const prefix = '#/$defs/'
+  const usedNames = new Map<string, number>()
+  const rename = new Map<string, string>()
+  for (const [name, definition] of Object.entries(definitions)) {
+    const shortName = shortNameFor(name, definition)
+    const used = usedNames.get(shortName) ?? 0
+    usedNames.set(shortName, used + 1)
+    rename.set(name, used === 0 ? shortName : `${shortName}${used + 1}`)
+  }
+
+  const result = mapNodes(schema, (node) => {
+    if (typeof node.$ref === 'string' && node.$ref.startsWith(prefix)) {
+      const name = node.$ref.slice(prefix.length)
+      node.$ref = `${prefix}${rename.get(name) ?? name}`
+    }
+    return node
+  }) as JsonSchemaType
+  result.$defs = Object.fromEntries(
+    Object.entries(result.$defs!).map(([name, body]) => [rename.get(name) ?? name, body]),
+  )
+  return result
+}

package/src/utils/toStandardSchema.ts

@@ -0,0 +1,9 @@
+import { fromJsonSchema, type StandardSchemaWithJSON } from '@modelcontextprotocol/server'
+
+import type { ToolInputSchema } from '../types.js'
+
+/** Normalize a tool/prompt input to a Standard Schema. Raw JSON Schema gets wrapped; Standard Schema instances pass through. */
+export const toStandardSchema = (schema: ToolInputSchema): StandardSchemaWithJSON =>
+  typeof schema === 'object' && schema !== null && '~standard' in schema
+    ? (schema as StandardSchemaWithJSON)
+    : fromJsonSchema(schema)

package/dist/collections/createApiKeysCollection.d.ts

@@ -1,7 +0,0 @@
-import type { CollectionConfig } from 'payload';
-import type { MCPPluginConfig } from '../types.js';
-export declare const createAPIKeysCollection: (collections: MCPPluginConfig["collections"], globals: MCPPluginConfig["globals"], customTools: Array<{
-    description: string;
-    name: string;
-}> | undefined, experimentalTools: NonNullable<MCPPluginConfig["experimental"]>["tools"] | undefined, pluginOptions: MCPPluginConfig) => CollectionConfig;
-//# sourceMappingURL=createApiKeysCollection.d.ts.map

package/dist/collections/createApiKeysCollection.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"createApiKeysCollection.d.ts","sourceRoot":"","sources":["../../src/collections/createApiKeysCollection.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAkB,MAAM,SAAS,CAAA;AAE/D,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,aAAa,CAAA;AAKlD,eAAO,MAAM,uBAAuB,gBACrB,eAAe,CAAC,aAAa,CAAC,WAClC,eAAe,CAAC,SAAS,CAAC,eACtB,KAAK,CAAC;IAAE,WAAW,EAAE,MAAM,CAAC;IAAC,IAAI,EAAE,MAAM,CAAA;CAAE,CAAC,iCACtC,WAAW,CAAC,eAAe,CAAC,cAAc,CAAC,CAAC,CAAC,OAAO,CAAC,6BACzD,eAAe,KAC7B,gBAuWF,CAAA"}