Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
import {
FILE_REFERENCE_MAX_COUNT,
type FileReferenceV1,
buildGfsFileReference,
classifyBytes,
} from '@clerum/gfs-interaction-policy'
import type { ComposerGlobalFileReference, ComposerReferenceAttachment } from '../uiTypes'
/**
* Builds the structured FileReference v1 for one Global Files selection
* (#666). The picker lists no media type and no bytes: classification falls
* back to the name, and a zero-byte selection classifies as empty text;
* mcp-host resolves the reference against GFS before the model sees it.
*/
function globalFileReference(reference: ComposerGlobalFileReference): FileReferenceV1 {
const built = buildGfsFileReference({
drive: reference.drive,
resourceId: reference.resourceId,
gfsUri: reference.gfsUri,
version: reference.version,
name: reference.label,
declaredMediaType: null,
byteLength: reference.bytes,
classification: classifyBytes({
bytes: new Uint8Array(0),
totalByteLength: reference.bytes,
declaredMediaType: null,
filename: reference.label,
}),
})
if (!built.ok) {
throw new Error(`Global file reference is invalid (${built.code}): ${built.message}`)
}
return built.value
}
export function buildComposerFileReferences(
references: ComposerReferenceAttachment[]
): FileReferenceV1[] {
const globalFiles = references.filter(
(reference): reference is ComposerGlobalFileReference => reference.type === 'global_file'
)
if (globalFiles.length > FILE_REFERENCE_MAX_COUNT) {
throw new Error(`A message can reference at most ${FILE_REFERENCE_MAX_COUNT} files.`)
}
return globalFiles.map(globalFileReference)
}