# @wrnexus/viewer

Complete implementation of the
[file viewer plan](../../docs/plans/2026-09-08-file-viewer-design.md).

Add `@wrnexus/viewer` to an app's dependencies. Its package plugin discovers the component,
stylesheet and browser runtime automatically. Explicit plugin configuration can import
`viewerPlugin` from `@wrnexus/viewer/plugin`; do not register it twice.

```wrn
import Viewer from "@wrnexus/viewer/components/Viewer.wrn"

<Viewer src={file.url} name={file.name} />
```

Optional props: `mime`, `size` (byte count as a string), `class`, and `theme` (`auto`, `light`,
`dark`), `files` (an array of source descriptors), and `richPdf` (opt-in canvas rendering). Pinned themes use the host's generated theme tokens. The `/file-viewer` page in
`examples/basic-app` provides local samples for every supported group.

For an upload preview before any network request, mount from the browser:

```ts
import { mountViewer } from "@wrnexus/viewer/runtime";

const dispose = mountViewer(container, { blob: file, name: file.name });
// Before replacing the preview or removing your component:
dispose();
```

## What renders

**Tier 1 — no dependencies.** Native image formats with zoom, rotation, reset and
scroll/drag panning; PDF and audio/video through native controls; UTF-8 text and code
through the block editor's lazy read-only CodeMirror mount, with theme tokens, line numbers
and a wrap toggle. HTML/XML use XML highlighting; other languages use plain mode. A text-only
`<pre>` remains usable if the editor cannot load.

**Tier 2 — in-package parsers, still no runtime dependencies.**

| Files          | Preview                                                            |
| -------------- | ------------------------------------------------------------------ |
| `.csv`, `.tsv` | Windowed table (RFC 4180 quoting, delimiter detection), raw toggle |
| `.json`        | Lazy collapsible tree, raw toggle                                  |
| `.xml`         | Lazy element tree with attributes, raw toggle                      |
| `.svg`         | Sanitised markup inside a sandboxed frame, source toggle           |
| `.tar`, `.gz`  | Entry listing with sizes; gzip is expanded in the browser          |

Each Tier 2 renderer is registered per extension and sits _in front of_ the source view for
its kind, so a parse failure — malformed JSON, a `.csv` that is really prose, an unreadable
archive — falls through to the highlighted source rather than erroring.

**Tier 2 with an optional dependency.** Declared as optional peer dependencies and never
installed for you. Each is served as its own runtime asset, contributed only when the package
resolves on the server, and fetched by the browser the first time such a file is opened — an
app that never views a spreadsheet neither builds nor ships `xlsx`.

| Files           | Package      | Preview                                                 |
| --------------- | ------------ | ------------------------------------------------------- |
| `.xls`, `.xlsx` | `xlsx`       | Sheet tabs over the same windowed table as CSV          |
| `.zip`          | `fflate`     | Entry listing; the archive is never inflated            |
| `.docx`         | `mammoth`    | Semantic HTML in a sandboxed frame — layout is not kept |
| `.pdf`          | `pdfjs-dist` | Opt-in page rasterisation with page count (see below)   |

`.docx` sits _behind_ the Tier 3 converter: LibreOffice's PDF keeps the page layout, which
mammoth cannot, so mammoth is the fallback when no converter is reachable. Rich PDF is opt-in
because the native frame costs nothing:

```ts
import { registerRenderer } from "@wrnexus/viewer";
import { richPdfRenderer } from "@wrnexus/viewer/renderers/pdf-rich";

registerRenderer(richPdfRenderer());
```

pdf.js is served from its own published build rather than bundled — putting its ESM entry
through a bundler breaks the worker handshake and `getDocument()` never settles. Page
rendering is driven by `requestAnimationFrame`, which does not run while the document is
hidden, so each page render is bounded by a timeout (20s by default) and a page that will not
draw degrades to the native frame rather than hanging.

**Office documents.** Everything below reads in the browser with no binary on the host. All of
it recovers content and structure, not a faithful page layout, so wherever a converter is
reachable its PDF is preferred and these are the rung beneath it — the exception is
spreadsheets, where the data matters more than the layout and the table view comes first.

| Files                     | Read by             | Preview                                    |
| ------------------------- | ------------------- | ------------------------------------------ |
| `.pptx`, `.ppsx`, `.potx` | in-package + fflate | Slide-by-slide reader with presenter notes |
| `.odt`                    | in-package + fflate | Headings, paragraphs and lists             |
| `.odp`                    | in-package + fflate | Slide-by-slide reader                      |
| `.ods`                    | SheetJS, then ODF   | Sheet tabs over the windowed table         |
| `.rtf`                    | in-package          | Text, with a Source toggle — no dependency |
| `.docx`                   | mammoth             | Semantic HTML in a sandboxed frame         |
| `.xls`, `.xlsx`           | SheetJS             | Sheet tabs over the windowed table         |

`.doc` and `.ppt` are the legacy OLE compound formats and stay converter-only: there is no
honest way to read them in a browser. With LibreOffice installed every format in this table
converts to PDF instead.

**Tier 3 — server-side conversion, opt-in.** `.doc`, `.docx`, `.odt`, `.rtf`, `.ppt`,
`.pptx`, `.xls` and `.xlsx` convert to PDF; `.psd`, `.ai`, `.tif` and `.tiff` to PNG; `.avi`
and `.mkv` to a poster frame; `.zip`, `.rar` and `.7z` to a listing. Disabled by default —
see Conversion below.

HTML opens in a sanitized, fully sandboxed preview with a Source toggle. Markup masquerading as another format stays inert source. Anything with no renderer and no converter displays
a metadata/download fallback, and native playback errors do the same.

## Sources and streaming

Image, media and PDF renderers are `streaming`: for an `http(s)` source the element is
pointed at the original URL, so the browser fetches and decodes it, playback starts without
buffering the whole file here, and the request is governed by the app's own CSP rather than
by a `blob:` URL (which `default-src 'self'` blocks). A local `File`/`Blob` still renders
from an object URL served under its _sniffed_ type.

The PDF renderer uses an iframe. The framework now defaults to `frame-ancestors 'self'`
and `X-Frame-Options: SAMEORIGIN`, so an app can embed its own files. Apps can retain full
framing denial with `security.frameOptions: "DENY"`; explicit CSP directives still win.

## Conversion

```ts
import { viewerPlugin } from "@wrnexus/viewer/plugin";
import { localBinaryConverter } from "@wrnexus/viewer/convert";

plugins: [viewerPlugin({ convert: { converters: [localBinaryConverter()] } })];
```

That mounts `GET /api/viewer/capabilities`, `/preview`, `/thumb` and `/text`. Without the
`convert` option no routes exist and no external process is ever run.

The client asks `capabilities` once per page and only requests a conversion for an extension
the server actually advertises, so a host without the binaries costs one small request and
every affected file keeps its download card. `localBinaryConverter` probes for `soffice`,
`magick`, `ffmpeg` and `7z` on first use; **a missing binary is not an error**.

Converted output is cached in-process, keyed by stable source identity, extension, operation, converter name
and version, and returned with `X-Viewer-Cache: hit|miss`. The source resolver runs before every cache lookup,
so an expired or withdrawn signed link cannot read cached output. Unused source streams are
cancelled on cache hits. Responses are private and no-store.

Apps that store files privately implement the contract themselves:

```ts
viewerPlugin({
  convert: {
    converters: [myGotenbergConverter],
    // Apply your own access checks here; the default accepts same-origin URLs only.
    resolveSource: async (request, { src, name }) => ({
      body: await openStream(src, name),
      identity: objectRevision(src),
    }),
    maxInputBytes: 32 * 1024 * 1024,
  },
});
```

### Converter safety

`soffice`, `magick` and `ffmpeg` are large C/C++ parsers with a long CVE history, pointed at
untrusted files. Every run gets a hard wall-clock timeout, a private scratch directory as its
only writable path, a stripped environment (no proxy or credential variables), no stdin, and
a cap on the bytes read back. Input is size-capped before it reaches a binary. Converter
error messages stay server-side — they can carry file paths — and the client sees only a 422.
The default source policy accepts same-origin URLs and fetches them without credentials, so
the endpoint cannot be used to reach a private network or to read anything the caller could
not already read anonymously.

## Security

Uploaded files are hostile input, so nothing here trusts a filename.

- Classification uses the extension, declared/server MIME, and then up to 4 KiB of signature
  bytes. Extension overrides cannot bypass conflicting signatures, and executable names or
  signatures always fall back.
- Nothing from a file ever reaches `innerHTML`. Text, tables and trees are built as DOM nodes
  with `textContent`.
- The SVG frame is its own document, so theme tokens and `currentColor` cannot reach it and
  its base background is opaque. The host's resolved text and surface colours are copied into
  the frame by value, which keeps themed artwork legible in light and dark.
- SVG is sanitised to an element/attribute allowlist — no `script`, no `foreignObject`, no
  event handlers, and only same-document `#fragment` references — _and_ rendered inside an
  `iframe` with `sandbox=""` under `default-src 'none'`. Neither layer is trusted alone.
- A file whose bytes look like active markup is shown as inert source regardless of its name,
  including a `.json` or `.png` that carries a `<script>` tag.

## Resolution and limits

`resolve(source)` returns a renderer and ordered fallback list. Signature checks select a
renderer; the browser decoder still validates the complete file.

For remote sources the resolver requests HEAD before a bounded Range read. A server without
HEAD/Range support can still work: the response stream is cancelled at the sniff limit.
Complete preview reads enforce a byte cap even if Content-Length is absent or inaccurate,
and the downloaded body is reclassified before rendering. A single network chunk can be
larger than the cap; excess bytes are never retained as a preview. HTTP(S), blob and data
URLs are supported; remote servers must allow browser CORS access. A kind with no renderer
at all reports that no preview exists rather than a size limit.

Defaults: text/code 2 MiB, SVG 2 MiB, CSV/JSON/XML 8 MiB, images/PDF 32 MiB, audio/video and
archives 64 MiB. Delimited files are additionally capped at 200,000 parsed rows, and only the
visible window of rows is kept in the DOM. Remote native media streams from its original URL; local blobs remain bounded. Disposal aborts pending requests, destroys
editors and revokes object URLs. Automatic mounts observe source changes and framework
navigation.

## Uploader-backed sources

Core is source-agnostic — `src` or `blob` is all it needs, and `@wrnexus/uploader` is an
optional peer dependency. Enabling `uploads` lets a component name a storage key instead:

```wrn
<Viewer key={file.key} store="reports" />
```

```ts
import { viewerPlugin } from "@wrnexus/viewer/plugin";

plugins: [
  viewerPlugin({
    uploads: {
      stores: ["media", "reports"],
      secret: process.env.VIEWER_FILE_SECRET!,
      ttlMs: 5 * 60 * 1000,
      authorize: async (request, { store, key }) => canRead(await currentUser(request), store, key),
    },
  }),
];
```

That mounts `GET /api/viewer/source` and `GET /api/viewer/file`. The browser exchanges the key
for a descriptor, and only then loads anything:

- A **public** object answers with its own URL — no token, nothing signed, nothing to leak.
- A **private** object answers with a short-lived signed link to the file route, and only
  after `authorize` agrees. Signing uses the uploader's own `createSignedFileToken`.

The file route serves `Content-Disposition: inline`, `X-Content-Type-Options: nosniff` and
`Cache-Control: private, no-store` — a per-viewer link must never enter a shared cache — and
answers HEAD and byte ranges so audio and video can seek. An expired, forged, or
withdrawn-store link is answered exactly like a missing one.

There is deliberately **no default access check**: a private store without both a `secret` and
an `authorize` hook is refused rather than served. `stores` is an allow-list; anything outside
it is refused before a key is even looked at, and keys that could escape their store
(`..`, absolute paths, control characters) are rejected outright.

The MIME in the descriptor comes from the uploader's own extension table, not a second copy of
it here — and the viewer still sniffs the bytes before choosing a renderer.

For conversion, `uploadedConversionSource()` makes converted output inherit exactly the access
check of its input: a converter can only open what a valid signed link already grants.

```ts
import { uploadedConversionSource } from "@wrnexus/viewer/uploader";

viewerPlugin({
  convert: { converters: [localBinaryConverter()], resolveSource: uploadedConversionSource() },
  uploads: {/* … */},
});
```

## Where the preview appears

Inline in a card is the default. Three other placements are built in.

**Over the page.** `openViewer` puts a file in a native `<dialog>`, which takes the top layer,
traps focus, closes on Escape and restores focus to whatever opened it. Closing disposes the
viewer, so a video stops and object URLs are revoked rather than leaking behind the overlay.

```ts
import { openViewer } from "@wrnexus/viewer";

const preview = openViewer({ src: file.url, name: file.name });
// preview.close() when you need to close it yourself.
```

Nothing needs wiring for the common case — any element carrying `data-viewer-open` opens the
file it names:

```wrn
<button type="button" data-viewer-open data-src={file.url} data-name={file.name}>Preview</button>
```

It reads the same attributes the component takes (`data-src`, `data-key`, `data-store`,
`data-name`, `data-mime`, `data-size`, `data-files`), so a storage key or a whole filmstrip
opens the same way.

**Covering the page.** `fullScreen` mounts the viewer over the page with a Close control that
returns it inline. It is a layered overlay rather than the Fullscreen API, so it needs no user
gesture and can be the state a viewer starts in. Toggling the attribute later re-mounts it.

```wrn
<Viewer src={file.url} name={file.name} fullScreen={true} />
```

**A whole page.** `layout="page"` drops the card chrome and fills the container, for a route
given over to one file.

```wrn
<Viewer src={file.url} name={file.name} layout="page" />
```

## Toolbar actions

Every control is an icon button carrying its name in `aria-label`, in a visually
hidden span for screen readers, and in a CSS tooltip on hover and focus. The icons are drawn
as SVG nodes from a small built-in set rather than markup strings, because the framework
enables Trusted Types in production where `innerHTML` is refused.

**Print** is offered on any preview that actually rendered — slides, tables, documents,
images and PDFs — and withheld on the download card, where there is nothing to print. It
prints the preview rather than the page: the viewer is marked and `viewer.css` hides
everything else, while a same-origin PDF frame is asked to print itself. Opt out with
`<Viewer print={false} />` or `mountViewer(host, source, { print: false })`.

Apps add their own actions, either globally or per viewer:

```ts
import { registerViewerAction } from "@wrnexus/viewer";

const unregister = registerViewerAction({
  id: "open-in-tab",
  label: "Open in new tab",
  icon: "preview", // a name from ICONS; falls back to one from the label
  placement: "actions", // or "controls", beside the renderer's own buttons
  when: ({ source, kind }) => kind === "pdf" && Boolean(source.src),
  run: ({ source, content }) => window.open(source.src, "_blank", "noopener"),
});
```

```ts
mountViewer(host, source, {
  actions: [{ id: "send", label: "Send for review", run: ({ source }) => share(source) }],
});
```

Actions are offered after every render, so `when` sees the renderer and kind that actually
drew, and a re-render (a Raw/Table toggle, say) never leaves duplicates behind. `run` receives
the `content` element the renderer drew into, so an action can read what is on screen instead
of fetching the file again. An action is offered on the download card too — only its `when`
decides — so "report this file" still works for something nothing can preview.

## Custom renderers

```ts
import { registerRenderer } from "@wrnexus/viewer";
import { mountViewer } from "@wrnexus/viewer/runtime";

const unregister = registerRenderer({
  kind: "text",
  exts: [".log"],
  tier: 2,
  component: "LogViewer",
  maxInlineBytes: 1024 * 1024,
  load: () => import("./log-parser"),
  async mount({ parent, file, button }) {
    parent.textContent = await file.text();
    button("Reload", () => location.reload());
    return () => parent.replaceChildren();
  },
});
```

Every renderer — built-in and app-defined — receives the same `ViewerMountContext`: the
`parent` container, the `file` blob, the `source` descriptor, the resolved `kind`, a `button`
factory for toolbar controls, an `objectUrl` helper whose URLs are revoked on disposal, a
`degrade` callback that abandons the renderer for the download card, and an abort `signal`.
`component` selects a presentation template when one exists; the mount adapter determines rendering. A renderer without a
`mount` can never be displayed, and the ladder moves to the next candidate.

Bundle registration and manual mounting in the same browser entry. New registrations take
precedence and return an unregister function; isolated `RendererRegistry` instances can be
passed to `resolve`. Failed optional loads move to the next renderer. Custom adapters are
trusted application code and must handle file contents safely.

The parsers are exported for reuse: `parseDelimited`, `detectDelimiter`, `sanitizeSvg`,
`parseTar`, `isTar`, `gzipName`, `gunzip`.

## Source composition

The example accepts both signed uploader objects and anonymous static files:

```ts
import { firstOf, sameOriginSource } from "@wrnexus/viewer/convert";
import { uploadedConversionSource } from "@wrnexus/viewer/uploader";

resolveSource: firstOf(uploadedConversionSource(), sameOriginSource());
```

A resolver returns undefined when it cannot handle a source. Errors stop composition. The
same-origin resolver refuses viewer API links, cross-origin URLs, embedded credentials and
redirects. Uploader identities use the store and object key rather than the rotating token.
For mutable custom sources include a revision in `identity`; otherwise replacement bytes
may remain cached until TTL expiry. The URL remains the fallback identity.

## Presentation components

`ViewerShell` composes `ViewerToolbar`, `DownloadCard`, and renderer templates:
`ImageViewer`, `VectorViewer`, `PdfViewer`, `CodeViewer`, `SheetViewer`, `DocViewer`,
`MediaViewer`, and `ArchiveViewer`. Copy a component into `app/components` with the same
name to override its presentation. Keep its `data-viewer-*` hooks. The browser mounts the
existing safe adapters into `data-viewer-mount`, preserves custom classes, and restores the
shell on disposal. Components are presentation shells; source handling stays with Viewer.

## Filmstrip, waveform and extracted text

```wrn
<Viewer files={[{src: '/photo.png', name: 'photo.png'}, {src: '/report.pdf', name: 'report.pdf'}]} />
<Viewer src="/report.pdf" name="report.pdf" richPdf={true} />
```

Manual mounting accepts the same source array, including local blobs. Up to 500 files are
shown; one preview is active. Arrow keys, Home and End move selection. Switching disposes
media, editors, worker work, object URLs and pending fetches. Image thumbnails use the source;
convertible files request `/api/viewer/thumb` sequentially with a 2 MiB limit. Missing or
non-image thumbnails keep the filename button. No converter is required to use a filmstrip.

Audio has a **Show waveform** action. It decodes on demand, caps the encoded read at 8 MiB,
and accepts known durations up to ten minutes; unsupported or larger files retain native
playback. The waveform combines all channels and includes an accessible seek slider.

Convertible sources have an **Extract text** action. The exported `extractText(source,
{baseURI, signal})` returns at most 2 MiB of plain text for search/index consumers. The action
shows the same text inline and reports unavailable extraction without breaking the preview.
Client conversion and optional-module helpers accept the host base URI for embedded documents
and tests, instead of requiring the global document.

Rich PDF success was verified in ordinary Chrome on 2026-09-09: the canvas visibly painted
text and the renderer reported one completed page. Timeout fallback remains for hidden tabs.

## Accepted limits

The conversion cache is per-process rather than shared, and byte
ranges on a signed link are served only for buffered objects — a streaming driver answers the
whole body, so seeking a large private video re-reads it.

Run `bun test packages/viewer`, `bun run --cwd packages/viewer typecheck`.