dorfteich/apps/api/src/import-export/import.service.ts

import { BadRequestException, ForbiddenException, Injectable } from '@nestjs/common';
import { ConversionJob, User } from '@prisma/client';
import { ConversionJobView, editorSchema, markdownToDoc } from '@dorfteich/shared';
import { Node } from 'prosemirror-model';
import { PinoLogger } from 'nestjs-pino';

import { FilesService } from '../files/files.service';
import { PagesService } from '../pages/pages.service';
import { docToState } from '../pages/yjs-content';
import { PrismaService } from '../prisma/prisma.service';

import { ImportProcessor } from './import.constants';
import { ConversionError, PandocConverter } from './pandoc.converter';
import { ConversionJobService } from './conversion-job.service';

/** pandoc source format per accepted upload extension (ADR 0009). */
const IMPORT_FORMATS: Readonly<Record<string, string>> = {
  docx: 'docx',
  odt: 'odt',
};

/** Job `kind` per source format, so the worker can route the job to the import
 * pipeline (a plain byte→byte conversion has a different kind, #62/#65). */
function importKind(format: string): string {
  return `import_${format}`;
}

/**
 * Writer format for the structural pass. GFM matches our editor schema
 * (tables, task lists, strikethrough); `-implicit_figures` keeps images as
 * inline `![alt](src)` instead of wrapping them in a figure with a duplicated
 * caption; `-raw_html` stops pandoc emitting raw HTML for constructs it can't
 * represent (our Markdown parser runs with `html:false` and would drop it
 * anyway — per ADR 0009 we preserve structure, not layout).
 */
const IMPORT_MARKDOWN_FORMAT = 'gfm-implicit_figures-raw_html';

/**
 * The two-pass conversion at the heart of import, extracted so tests can run it
 * over the fixture corpus against a real sidecar. pandoc-server is stateless and
 * hands back a document's media no other way, so we first inline every image as
 * a `data:` URI (`source → html`, embed-resources), then produce clean
 * structural Markdown that still carries those data URIs inline at the right
 * positions (`html → gfm`, no wrapping).
 */
export async function convertImportedDocument(
  converter: PandocConverter,
  sourceFormat: string,
  input: Buffer,
): Promise<string> {
  const html = await converter.convert({
    from: sourceFormat,
    to: 'html',
    input,
    standalone: false,
    embedResources: true,
  });
  const md = await converter.convert({
    from: 'html',
    to: IMPORT_MARKDOWN_FORMAT,
    input: html.output,
    standalone: false,
    wrap: 'none',
  });
  return md.output.toString('utf8');
}

/** Minimal shape of a ProseMirror document as JSON — enough to walk it for
 * image nodes and the leading heading without pulling in prosemirror types. */
interface PmNode {
  type: string;
  attrs?: Record<string, unknown>;
  content?: PmNode[];
  text?: string;
  marks?: unknown[];
}

interface DecodedImage {
  buffer: Buffer;
  extension: string;
}

/**
 * Imports `.docx`/`.odt` documents as new pages (ADR 0009, issue #63). Enqueue
 * is synchronous and cheap (validate + persist a job on the #62 queue); the
 * heavy conversion runs out of band in {@link run}, invoked by the shared
 * {@link ConversionWorker} for import-kind jobs so it inherits the queue's
 * locking, retry, and restart-survival.
 *
 * Pipeline: the pandoc sidecar is stateless and will not hand back a
 * document's embedded media any other way, so we convert in two passes —
 * `docx/odt → html` with `embed-resources` inlines every image as a `data:`
 * URI, then `html → gfm` produces clean structural Markdown with those data
 * URIs still inline at the right positions. We parse that to an editor
 * document, store each embedded image as a pond file (with quota accounting)
 * and rewrite its reference to the file id, derive the title from a leading
 * top-level heading (else the file name), and create the page from the
 * resulting Yjs state.
 */
@Injectable()
export class ImportService implements ImportProcessor {
  constructor(
    private readonly prisma: PrismaService,
    private readonly jobs: ConversionJobService,
    private readonly converter: PandocConverter,
    private readonly files: FilesService,
    private readonly pages: PagesService,
    private readonly logger: PinoLogger,
  ) {
    this.logger.setContext(ImportService.name);
  }

  /** Validate and enqueue an import; the returned job id is polled via
   * `GET /jobs/:id` until it reports the created `resultPageId` (#62). */
  async enqueue(
    user: User,
    pondId: string,
    file: { buffer: Buffer; originalname: string },
  ): Promise<ConversionJobView> {
    const extension = fileExtension(file.originalname);
    const format = extension ? IMPORT_FORMATS[extension] : undefined;
    if (!format) {
      throw new BadRequestException({ code: 'import_unsupported_format' });
    }
    const job = await this.jobs.enqueue({
      ownerId: user.id,
      pondId,
      kind: importKind(format),
      from: format,
      to: 'page',
      input: file.buffer,
      sourceName: file.originalname,
      standalone: false,
    });
    return this.jobs.viewOf(job);
  }

  /**
   * Run one import job to completion (called by the worker). Throws a
   * {@link ConversionError} on failure so the worker applies the queue's
   * retry/fail policy; on success it records the created page on the job.
   * Media stored during a failed attempt is rolled back so a retry does not
   * leak files or double-count quota.
   */
  async run(job: ConversionJob): Promise<void> {
    if (!job.pondId) {
      throw new ConversionError('conversion_failed', false, 'import job has no pond');
    }
    const user = await this.prisma.user.findUnique({ where: { id: job.ownerId } });
    if (!user) {
      throw new ConversionError('conversion_failed', false, 'import job owner is gone');
    }

    const rawMarkdown = await convertImportedDocument(
      this.converter,
      job.sourceFormat,
      Buffer.from(job.input),
    );

    // Media is stored before the page exists (the page's state references the
    // file ids); track what we create so a later failure can be rolled back.
    const storedFileIds: string[] = [];
    try {
      const markdown = await this.storeEmbeddedImages(rawMarkdown, user, job.pondId, storedFileIds);
      const json = markdownToDoc(markdown).toJSON() as unknown as PmNode;
      const { title, doc } = this.splitTitle(json, job.sourceName);
      const state = docToState(Node.fromJSON(editorSchema, doc));

      const page = await this.pages.createWithState(user, job.pondId, title, state);
      await this.files.linkAttachmentsToPage(storedFileIds, page.id);
      await this.prisma.conversionJob.update({
        where: { id: job.id },
        data: { status: 'SUCCEEDED', resultPageId: page.id, errorCode: null },
      });
      this.logger.info(
        { jobId: job.id, pageId: page.id, pondId: job.pondId, images: storedFileIds.length },
        'audit: document imported',
      );
    } catch (error) {
      await this.rollbackMedia(user, storedFileIds);
      throw error;
    }
  }

  /**
   * Store each embedded image (an inline `data:` URI, from the embed-resources
   * pass) as a pond file and rewrite the Markdown reference to the stored file
   * id. Done on the Markdown text, before parsing, because the editor's parser
   * only admits a whitelist of `data:` image types (png/jpeg/gif/webp) and
   * would leave any other inline as a huge literal base64 string. base64 has no
   * `)`, so the image regex is unambiguous. An image whose bytes the upload
   * pipeline rejects (e.g. a vector `image/x-emf` Word can embed) is dropped
   * rather than failing the whole import — structure over layout (ADR 0009); a
   * pond that runs out of storage fails the import.
   */
  private async storeEmbeddedImages(
    markdown: string,
    user: User,
    pondId: string,
    storedFileIds: string[],
  ): Promise<string> {
    const image = /!\[([^\]]*)\]\((data:[^)\s]+)\)/g;
    const uris = new Set<string>();
    for (const match of markdown.matchAll(image)) uris.add(match[2]!);
    if (uris.size === 0) return markdown;

    // Dedup identical images so a document that repeats one stores it once.
    const resolved = new Map<string, string | null>();
    let index = 0;
    for (const uri of uris) {
      resolved.set(uri, await this.storeImage(uri, user, pondId, ++index, storedFileIds));
    }
    return markdown.replace(image, (_whole, alt: string, uri: string) => {
      const fileId = resolved.get(uri);
      return fileId ? `![${alt}](${fileId})` : '';
    });
  }

  private async storeImage(
    uri: string,
    user: User,
    pondId: string,
    index: number,
    storedFileIds: string[],
  ): Promise<string | null> {
    const decoded = decodeDataUri(uri);
    if (!decoded) return null;
    try {
      const view = await this.files.upload(user, pondId, {
        buffer: decoded.buffer,
        size: decoded.buffer.length,
        originalname: `import-${index}.${decoded.extension}`,
      });
      storedFileIds.push(view.id);
      return view.id;
    } catch (error) {
      if (error instanceof ForbiddenException && errorCode(error) === 'quota_exceeded') {
        throw new ConversionError('quota_exceeded', false, 'pond storage exhausted during import');
      }
      // Unsupported image type / rejected bytes — drop this one, keep going.
      this.logger.warn({ pondId, code: errorCode(error) }, 'import: dropped an image');
      return null;
    }
  }

  /**
   * Use a leading top-level heading as the page title (and remove it from the
   * body so it is not duplicated), else fall back to the upload file name
   * without its extension, else a generic title.
   */
  private splitTitle(root: PmNode, sourceName: string | null): { title: string; doc: PmNode } {
    const content = root.content ?? [];
    const first = content[0];
    if (first && first.type === 'heading') {
      const headingText = textOf(first).trim();
      if (headingText) {
        return { title: headingText, doc: { ...root, content: content.slice(1) } };
      }
    }
    const fallback = baseName(sourceName) || 'Imported document';
    return { title: fallback, doc: root };
  }

  private async rollbackMedia(user: User, fileIds: string[]): Promise<void> {
    for (const id of fileIds) {
      try {
        await this.files.remove(user, id);
      } catch (error) {
        this.logger.warn({ fileId: id, code: errorCode(error) }, 'import: media rollback failed');
      }
    }
  }
}

/** Concatenated text of a node's inline content (headings have no nesting). */
function textOf(node: PmNode): string {
  if (typeof node.text === 'string') return node.text;
  return (node.content ?? []).map(textOf).join('');
}

const DATA_URI = /^data:(image\/[a-zA-Z0-9.+-]+);base64,(.+)$/;
const EXTENSION_BY_MIME: Readonly<Record<string, string>> = {
  'image/png': 'png',
  'image/jpeg': 'jpg',
  'image/gif': 'gif',
  'image/webp': 'webp',
  'image/svg+xml': 'svg',
};

function decodeDataUri(value: string): DecodedImage | null {
  const match = DATA_URI.exec(value);
  if (!match) return null;
  const mime = match[1]!;
  const buffer = Buffer.from(match[2]!, 'base64');
  if (buffer.length === 0) return null;
  return { buffer, extension: EXTENSION_BY_MIME[mime] ?? 'bin' };
}

/** Lowercased extension without the dot, or '' when there is none. */
function fileExtension(fileName: string): string {
  const dot = fileName.lastIndexOf('.');
  return dot >= 0 ? fileName.slice(dot + 1).toLowerCase() : '';
}

/** File name without its extension. */
function baseName(fileName: string | null): string {
  if (!fileName) return '';
  const dot = fileName.lastIndexOf('.');
  return (dot > 0 ? fileName.slice(0, dot) : fileName).trim();
}

/** The `code` from a Nest HttpException response body, if any. */
function errorCode(error: unknown): string | undefined {
  if (error instanceof BadRequestException || error instanceof ForbiddenException) {
    const body = error.getResponse();
    if (body && typeof body === 'object' && 'code' in body) {
      const code = (body as { code?: unknown }).code;
      return typeof code === 'string' ? code : undefined;
    }
  }
  return undefined;
}