dorfteich/apps/api/src/import-export/obsidian-vault.ts

import { MAX_PAGE_DEPTH, slugify } from '@dorfteich/shared';
import { unzipSync } from 'fflate';

/**
 * Obsidian vault transform (issue #116): pure functions from "uploaded ZIP"
 * to an import plan — no DB, no DI, fully unit-testable. The orchestration
 * (issue #117) feeds the plan into the page/file/label services.
 *
 * The core impedance mismatch: Obsidian links reference note NAMES
 * (`[[Meine Notiz]]`, `[[folder/Note#Heading|shown]]`), Dorfteich wikilinks
 * resolve by SLUG within a pond — and the shared markdown parser reads
 * `[[target]]` verbatim. So every rewrite here is a text pre-pass over the
 * note Markdown, done before `markdownToDoc` ever sees it.
 */

export class VaultError extends Error {
  constructor(
    readonly code: 'import_vault_invalid_zip' | 'import_vault_too_large',
    message: string,
  ) {
    super(message);
  }
}

/** Zip-bomb ceiling for the unpacked vault (the ZIP itself is capped by the
 * upload limit). Matches the plugin-package model. */
export const MAX_VAULT_UNPACKED_BYTES = 256 * 1024 * 1024;

/** Raster formats that become inline images. SVG deliberately stays an
 * attachment — it is never served inline (security.md §Uploads). */
const IMAGE_EXTENSIONS = new Set(['png', 'jpg', 'jpeg', 'gif', 'webp']);

/** Placeholder scheme the orchestrator replaces with real file ids after
 * uploading: `![alt](vault-asset:<vault path>)`. */
export const ASSET_PLACEHOLDER_PREFIX = 'vault-asset:';

export interface VaultNote {
  /** Vault-relative path without the `.md` extension. */
  path: string;
  /** Basename without `.md` — what Obsidian links refer to. */
  name: string;
  markdown: string;
}

export interface VaultAsset {
  path: string;
  name: string;
  data: Uint8Array;
  /** Lowercase extension without the dot. */
  extension: string;
}

export interface Vault {
  notes: VaultNote[];
  assets: VaultAsset[];
}

export type FrontmatterMode = 'strip' | 'preserve';

export interface PlannedContainer {
  /** Unique key: the merged folder path. */
  key: string;
  parentKey: string | null;
  /** Human title — a merged folder reads like `c/d`. */
  title: string;
  slug: string;
}

export interface PlannedNote {
  key: string;
  parentKey: string | null;
  title: string;
  slug: string;
  /** Rewritten Markdown: links → final slugs, tags stripped, frontmatter
   * handled, asset embeds → `vault-asset:` placeholders. */
  markdown: string;
  /** Tag paths (nested tags split at `/`), e.g. [['projekt'], ['a','b']]. */
  tags: string[][];
  /** Image assets referenced by this note (vault paths). */
  imageAssets: string[];
  /** Non-image assets referenced by this note (become page attachments). */
  attachmentAssets: string[];
}

export interface VaultImportPlan {
  /** Topologically ordered (parents before children). */
  containers: PlannedContainer[];
  notes: PlannedNote[];
  /** Vault paths of every asset actually referenced by some note. */
  referencedAssets: Set<string>;
}

/** Rejects absolute paths and `..` traversal (zip-slip), plugin-package model. */
function safeRelativePath(rawPath: string): string {
  const path = rawPath.replace(/\\/g, '/').replace(/^\/+/, '');
  if (path.split('/').some((segment) => segment === '..')) {
    throw new VaultError('import_vault_invalid_zip', `Illegal path in archive: ${rawPath}`);
  }
  return path;
}

/**
 * Repairs a ZIP entry name that fflate decoded as Latin-1 (#127). fflate
 * honors only the ZIP UTF-8 flag (bit 11), which common archivers omit —
 * the UTF-8 bytes of "Fußball" then arrive as one mojibake char per byte
 * and end up in page titles. Latin-1 decoding is byte-lossless (charCode ==
 * byte), so when every char fits a byte we re-read them as UTF-8; names the
 * flag DID mark as UTF-8 contain either multi-byte chars (> 0xFF, left
 * alone) or precomposed Latin-1 chars whose bytes are invalid UTF-8 (the
 * decoder throws → keep the original). Finally NFC-normalize: macOS zips
 * store umlauts decomposed (NFD), which breaks slugify's ä→ae digraphs,
 * wikilink matching, and duplicate-basename detection.
 */
function decodeZipName(raw: string): string {
  let name = raw;
  // eslint-disable-next-line no-control-regex
  if (/^[\u0000-\u00ff]*$/.test(raw)) {
    const bytes = Uint8Array.from(raw, (char) => char.charCodeAt(0));
    try {
      name = new TextDecoder('utf-8', { fatal: true }).decode(bytes);
    } catch {
      // Genuinely Latin-1 (or already plain ASCII) — keep as decoded.
    }
  }
  return name.normalize('NFC');
}

/**
 * Unpacks the vault ZIP into notes (`.md`) and assets (everything else),
 * enforcing the unpacked-size ceiling incrementally. Hidden housekeeping
 * directories (`.obsidian/`, `.trash/`, any dot-directory) are skipped —
 * they hold app config, not content.
 */
export function parseVaultZip(
  zip: Uint8Array,
  maxUnpackedBytes: number = MAX_VAULT_UNPACKED_BYTES,
): Vault {
  let entries: Record<string, Uint8Array>;
  try {
    entries = unzipSync(zip);
  } catch {
    throw new VaultError('import_vault_invalid_zip', 'The file is not a valid ZIP archive');
  }

  const notes: VaultNote[] = [];
  const assets: VaultAsset[] = [];
  let unpacked = 0;
  for (const [zipName, bytes] of Object.entries(entries)) {
    const rawPath = decodeZipName(zipName);
    if (rawPath.endsWith('/')) continue; // directory entries
    const path = safeRelativePath(rawPath);
    const segments = path.split('/');
    if (segments.some((segment) => segment.startsWith('.'))) continue;
    unpacked += bytes.length;
    if (unpacked > maxUnpackedBytes) {
      throw new VaultError('import_vault_too_large', 'The unpacked vault is too large');
    }
    const basename = segments[segments.length - 1]!;
    if (basename.toLowerCase().endsWith('.md')) {
      const noteName = basename.slice(0, -3);
      notes.push({
        path: path.slice(0, -3),
        name: noteName,
        markdown: new TextDecoder().decode(bytes),
      });
    } else {
      const dot = basename.lastIndexOf('.');
      assets.push({
        path,
        name: basename,
        data: bytes,
        extension: dot >= 0 ? basename.slice(dot + 1).toLowerCase() : '',
      });
    }
  }
  if (notes.length === 0) {
    throw new VaultError('import_vault_invalid_zip', 'The archive contains no Markdown notes');
  }
  // Deterministic plan regardless of zip entry order.
  notes.sort((a, b) => a.path.localeCompare(b.path));
  assets.sort((a, b) => a.path.localeCompare(b.path));
  return { notes, assets };
}

export interface Frontmatter {
  /** The note body with the frontmatter handled per mode. */
  markdown: string;
  /** Tag paths extracted from `tags:`/`tag:` (nested split at `/`). */
  tags: string[][];
}

function parseTagValue(value: string): string[][] {
  // `tags: a, b` · `tags: [a, b]` · quoted variants.
  const inner = value.trim().replace(/^\[/, '').replace(/\]$/, '');
  return inner
    .split(',')
    .map((tag) => tag.trim().replace(/^["'#]+|["']+$/g, ''))
    .filter(Boolean)
    .map((tag) => tag.split('/').filter(Boolean));
}

/**
 * Handles a leading YAML frontmatter block. Only the `tags:` forms Obsidian
 * uses are interpreted (scalar list, inline array, block list) — everything
 * else is opaque. `strip` drops the block; `preserve` re-emits it as a
 * fenced yaml code block so nothing is lost but nothing leaks as prose.
 */
export function extractFrontmatter(markdown: string, mode: FrontmatterMode): Frontmatter {
  const match = /^---[ \t]*\r?\n([\s\S]*?)\r?\n---[ \t]*(?:\r?\n|$)/.exec(markdown);
  if (!match) return { markdown, tags: [] };

  const block = match[1]!;
  const body = markdown.slice(match[0].length);
  const tags: string[][] = [];
  const lines = block.split(/\r?\n/);
  for (let i = 0; i < lines.length; i += 1) {
    const line = lines[i]!;
    const keyMatch = /^(tags?)[ \t]*:[ \t]*(.*)$/i.exec(line);
    if (!keyMatch) continue;
    const value = keyMatch[2]!.trim();
    if (value) {
      tags.push(...parseTagValue(value));
      continue;
    }
    // Block list form: consume the following `- item` lines.
    for (let j = i + 1; j < lines.length; j += 1) {
      const item = /^[ \t]+-[ \t]*(.+)$/.exec(lines[j]!);
      if (!item) break;
      tags.push(...parseTagValue(item[1]!));
    }
  }

  const preserved = mode === 'preserve' ? `\`\`\`yaml\n${block}\n\`\`\`\n\n` : '';
  return { markdown: preserved + body, tags };
}

/**
 * Splits Markdown into alternating segments outside/inside code so the tag
 * and link passes never touch fenced blocks or inline code. Yields the
 * outside segments to `transform` and stitches everything back together.
 */
function transformOutsideCode(markdown: string, transform: (segment: string) => string): string {
  const lines = markdown.split('\n');
  const out: string[] = [];
  let inFence = false;
  let fenceMarker = '';
  for (const line of lines) {
    const fence = /^([ \t]*)(```+|~~~+)/.exec(line);
    if (fence && !inFence) {
      inFence = true;
      fenceMarker = fence[2]![0]!;
      out.push(line);
      continue;
    }
    if (fence && inFence && fence[2]![0] === fenceMarker) {
      inFence = false;
      out.push(line);
      continue;
    }
    if (inFence) {
      out.push(line);
      continue;
    }
    // Outside fences: protect inline code spans.
    const parts = line.split(/(`[^`]*`)/);
    out.push(parts.map((part) => (part.startsWith('`') ? part : transform(part))).join(''));
  }
  return out.join('\n');
}

export interface InlineTags {
  markdown: string;
  tags: string[][];
}

/**
 * Extracts inline `#tag` / `#nested/tag` occurrences and removes them from
 * the text (Stefan's decision: ALL Obsidian tags become labels). Headings are
 * safe — `#` followed by whitespace never matches; pure-number "tags" (like
 * anchors or issue refs) are ignored, matching Obsidian's own rule.
 */
export function extractInlineTags(markdown: string): InlineTags {
  const tags: string[][] = [];
  const result = transformOutsideCode(markdown, (segment) =>
    segment.replace(/(^|[\s(])#([\p{L}\p{N}_/-]+)/gu, (whole, prefix: string, tag: string) => {
      if (!/\p{L}/u.test(tag)) return whole; // needs at least one letter
      tags.push(tag.split('/').filter(Boolean));
      return prefix;
    }),
  );
  // Collapse the whitespace runs left behind by removed tags.
  const cleaned = transformOutsideCode(result, (segment) =>
    segment.replace(/[ \t]{2,}/g, ' ').replace(/[ \t]+$/g, ''),
  );
  return { markdown: cleaned, tags };
}

/** How a `[[target]]` resolves: to a note, an asset, or nothing. */
export interface LinkResolver {
  /** Final slug for a note target, or null when unresolvable. */
  noteSlug(target: string): string | null;
  /** Vault path of an asset target, or null. */
  assetPath(target: string): { path: string; extension: string } | null;
}

/**
 * Builds the vault-wide resolver: case-insensitive basename matching with
 * full-path precedence, duplicates resolved to the lexicographically first
 * vault path (documented rule; Obsidian's "closest" has no meaning without
 * a source-file anchor when the whole vault imports at once).
 */
export function buildResolver(vault: Vault, noteSlugByPath: Map<string, string>): LinkResolver {
  const notesByName = new Map<string, string>(); // lower name → path (first wins)
  const notesByPath = new Map<string, string>(); // lower path → path
  for (const note of vault.notes) {
    const lowerName = note.name.toLowerCase();
    if (!notesByName.has(lowerName)) notesByName.set(lowerName, note.path);
    notesByPath.set(note.path.toLowerCase(), note.path);
  }
  const assetsByName = new Map<string, VaultAsset>();
  const assetsByPath = new Map<string, VaultAsset>();
  for (const asset of vault.assets) {
    const lowerName = asset.name.toLowerCase();
    if (!assetsByName.has(lowerName)) assetsByName.set(lowerName, asset);
    assetsByPath.set(asset.path.toLowerCase(), asset);
  }

  return {
    noteSlug(target: string): string | null {
      const lower = target.toLowerCase().replace(/\.md$/, '');
      const path = lower.includes('/')
        ? (notesByPath.get(lower) ?? notesByName.get(lower.split('/').pop()!))
        : notesByName.get(lower);
      return path ? (noteSlugByPath.get(path) ?? null) : null;
    },
    assetPath(target: string): { path: string; extension: string } | null {
      const lower = target.toLowerCase();
      const asset = lower.includes('/')
        ? (assetsByPath.get(lower) ?? assetsByName.get(lower.split('/').pop()!))
        : assetsByName.get(lower);
      return asset ? { path: asset.path, extension: asset.extension } : null;
    },
  };
}

export interface RewrittenLinks {
  markdown: string;
  imageAssets: string[];
  attachmentAssets: string[];
}

/**
 * The link pass (fence-aware): rewrites every Obsidian link form to what the
 * shared markdown parser understands.
 * - `[[Name]]`, `[[Name|Display]]`, `[[Name#Heading]]`, `[[folder/Name]]` →
 *   `[[final-slug|display or original name]]` (heading stripped).
 * - Unresolvable note targets → `[[slugify(target)|target]]` — a phantom
 *   link, exactly like typing a dead wikilink by hand.
 * - `![[image.png]]` / `![](relative/path.png)` on a vault image →
 *   `![name](vault-asset:path)` for the uploader to resolve.
 * - `![[doc.pdf]]` (non-image asset) → `*doc.pdf*` text; the file becomes a
 *   page attachment (paperclip panel), which has no inline node.
 * - `![[Other Note]]` (note embed) → a plain wikilink (no transclusion).
 */
export function rewriteLinks(markdown: string, resolver: LinkResolver): RewrittenLinks {
  const imageAssets = new Set<string>();
  const attachmentAssets = new Set<string>();

  const rewriteTarget = (rawTarget: string, rawDisplay: string | undefined): string => {
    const target = rawTarget.split('#')[0]!.trim();
    if (!target) return rawDisplay ?? rawTarget; // pure heading link — keep text
    const display = rawDisplay?.trim() || target.split('/').pop()!;
    const slug = resolver.noteSlug(target) ?? (slugify(target.split('/').pop()!) || 'page');
    return `[[${slug}|${display}]]`;
  };

  const result = transformOutsideCode(markdown, (segment) => {
    let out = segment;
    // Embeds first — `![[…]]` would otherwise be eaten as `!` + wikilink.
    out = out.replace(/!\[\[([^[\]\n]+)\]\]/g, (_whole, inner: string) => {
      const [target, display] = splitPipe(inner);
      const asset = resolver.assetPath(target.split('#')[0]!.trim());
      if (asset) {
        if (IMAGE_EXTENSIONS.has(asset.extension)) {
          imageAssets.add(asset.path);
          const alt = display ?? asset.path.split('/').pop()!;
          return `![${alt}](${ASSET_PLACEHOLDER_PREFIX}${asset.path})`;
        }
        attachmentAssets.add(asset.path);
        return `*${display ?? asset.path.split('/').pop()!}*`;
      }
      // A note embed (transclusion) degrades to a normal link.
      return rewriteTarget(target, display);
    });
    // Standard markdown images with a relative vault path.
    out = out.replace(/!\[([^\]]*)\]\(([^)\s]+)\)/g, (whole, alt: string, src: string) => {
      if (/^[a-z][a-z0-9+.-]*:/i.test(src)) return whole; // absolute URL/data:
      const asset = resolver.assetPath(decodeURIComponent(src));
      if (!asset) return whole;
      if (IMAGE_EXTENSIONS.has(asset.extension)) {
        imageAssets.add(asset.path);
        return `![${alt}](${ASSET_PLACEHOLDER_PREFIX}${asset.path})`;
      }
      attachmentAssets.add(asset.path);
      return `*${alt || asset.path.split('/').pop()!}*`;
    });
    // Plain wikilinks.
    out = out.replace(/\[\[([^[\]\n]+)\]\]/g, (_whole, inner: string) => {
      const [target, display] = splitPipe(inner);
      return rewriteTarget(target, display);
    });
    return out;
  });

  return {
    markdown: result,
    imageAssets: [...imageAssets],
    attachmentAssets: [...attachmentAssets],
  };
}

function splitPipe(inner: string): [string, string | undefined] {
  const pipe = inner.indexOf('|');
  if (pipe < 0) return [inner, undefined];
  return [inner.slice(0, pipe), inner.slice(pipe + 1)];
}

/**
 * Maps every note's folder path to a container chain of at most
 * `maxContainerLevels` (the deepest levels merge into one container whose
 * title keeps the joined path, e.g. `c/d`). Returns the ordered container
 * list (parents first) and the container key each note hangs under.
 */
export function planFolders(
  vault: Vault,
  maxContainerLevels: number,
): {
  containers: { key: string; parentKey: string | null; title: string }[];
  parentKeyByNote: Map<string, string | null>;
} {
  const containers = new Map<string, { key: string; parentKey: string | null; title: string }>();
  const parentKeyByNote = new Map<string, string | null>();

  for (const note of vault.notes) {
    const folders = note.path.split('/').slice(0, -1);
    if (folders.length === 0 || maxContainerLevels <= 0) {
      parentKeyByNote.set(note.path, null);
      continue;
    }
    const kept = folders.slice(0, Math.max(0, maxContainerLevels - 1));
    const merged = folders.slice(Math.max(0, maxContainerLevels - 1));
    const chain = merged.length > 0 ? [...kept, merged.join('/')] : kept;

    let parentKey: string | null = null;
    let keyPrefix = '';
    for (const title of chain) {
      keyPrefix = keyPrefix ? `${keyPrefix}/${title}` : title;
      if (!containers.has(keyPrefix)) {
        containers.set(keyPrefix, { key: keyPrefix, parentKey, title });
      }
      parentKey = keyPrefix;
    }
    parentKeyByNote.set(note.path, parentKey);
  }

  // Map insertion already guarantees parents-before-children (each chain is
  // walked top-down), and iteration order is insertion order.
  return { containers: [...containers.values()], parentKeyByNote };
}

/**
 * Assigns final slugs to every container and note: `-2/-3` suffixing against
 * the existing pond slugs and everything planned so far. Deterministic:
 * containers (path order) before notes (path order).
 */
export function planSlugs(
  entries: { key: string; base: string }[],
  existingSlugs: ReadonlySet<string>,
): Map<string, string> {
  const taken = new Set(existingSlugs);
  const result = new Map<string, string>();
  for (const entry of entries) {
    const base = slugify(entry.base) || 'page';
    let slug = base;
    for (let n = 2; taken.has(slug); n += 1) slug = `${base}-${n}`;
    taken.add(slug);
    result.set(entry.key, slug);
  }
  return result;
}

/**
 * The whole transform: ZIP bytes + options → an import plan the orchestrator
 * (issue #117) can feed into the page/file/label services.
 *
 * `mountDepth` is the 1-based depth of the chosen parent page (0 for the
 * pond root): containers + note level must fit inside MAX_PAGE_DEPTH.
 */
export function planVaultImport(
  zip: Uint8Array,
  options: {
    frontmatterMode: FrontmatterMode;
    existingSlugs: ReadonlySet<string>;
    mountDepth: number;
  },
): VaultImportPlan {
  const vault = parseVaultZip(zip);

  // Levels available below the mount page, minus one for the notes.
  const maxContainerLevels = Math.max(0, MAX_PAGE_DEPTH - options.mountDepth - 1);
  const { containers, parentKeyByNote } = planFolders(vault, maxContainerLevels);

  // Reserve every slug up front — containers first, then notes; the note
  // key is its vault path.
  const slugEntries = [
    ...containers.map((c) => ({ key: `container:${c.key}`, base: c.title.split('/').pop()! })),
    ...vault.notes.map((note) => ({ key: `note:${note.path}`, base: note.name })),
  ];
  const slugs = planSlugs(slugEntries, options.existingSlugs);

  const noteSlugByPath = new Map<string, string>();
  for (const note of vault.notes) {
    noteSlugByPath.set(note.path, slugs.get(`note:${note.path}`)!);
  }
  const resolver = buildResolver(vault, noteSlugByPath);

  const plannedNotes: PlannedNote[] = vault.notes.map((note) => {
    const frontmatter = extractFrontmatter(note.markdown, options.frontmatterMode);
    const inline = extractInlineTags(frontmatter.markdown);
    const links = rewriteLinks(inline.markdown, resolver);
    return {
      key: `note:${note.path}`,
      parentKey: parentKeyByNote.get(note.path)
        ? `container:${parentKeyByNote.get(note.path)!}`
        : null,
      title: note.name,
      slug: noteSlugByPath.get(note.path)!,
      markdown: links.markdown,
      tags: dedupeTags([...frontmatter.tags, ...inline.tags]),
      imageAssets: links.imageAssets,
      attachmentAssets: links.attachmentAssets,
    };
  });

  const referencedAssets = new Set<string>();
  for (const note of plannedNotes) {
    for (const path of note.imageAssets) referencedAssets.add(path);
    for (const path of note.attachmentAssets) referencedAssets.add(path);
  }

  return {
    containers: containers.map((c) => ({
      key: `container:${c.key}`,
      parentKey: c.parentKey ? `container:${c.parentKey}` : null,
      title: c.title,
      slug: slugs.get(`container:${c.key}`)!,
    })),
    notes: plannedNotes,
    referencedAssets,
  };
}

function dedupeTags(tags: string[][]): string[][] {
  const seen = new Set<string>();
  const result: string[][] = [];
  for (const tag of tags) {
    if (tag.length === 0) continue;
    const key = tag.join('/').toLowerCase();
    if (seen.has(key)) continue;
    seen.add(key);
    result.push(tag);
  }
  return result;
}