iWork 2013+

There are three different styles of iWork files:

1. The macOS applications generate ZIP files which contain the metadata and special .iwa files which hold the file data.

2. iCloud persistence on macOS is a folder based structure containing an Index.zip file which is "similar" to the macOS standalone file structure.

3. The web iCloud editors export ZIP files which contain an Index.zip file similar to iCloud persistence. Note that this is literally a ZIP file within a ZIP file

The Index.zip file has an identical structure to an actual file generated by the macOS applications, so the discussion is applicable to all file styles.

The ZIP container holds a number of Mac binary "property list" files (.plist) which can be safely ignored or blanked. It also can hold preview images that can be safely ignored.

Inspector

https://sheetjs.com/tools/iwa-inspector/ is a web-based tool for analyzing iWork archives. Source code: https://git.sheetjs.com/sheetjs/iwa-inspector

File Structure

The iWork file (.KEY, .NUMBERS, .PAGES) is a ZIP file containing a number of .iwa entries. The primary entrypoint is /Index/Document.iwa.

TSPersistence.framework handles the byte-level operations for the files.

.iwa files are sequential blocks of compressed data. Each "block" starts with a 4-byte header consisting of a 0 byte followed by the compressed length (stored as a 3-byte little-endian integer)

Each block follows the Snappy compressed format as described in the format description from the snappy repo. iWork apps do not expect a particular compression level, and it is possible to create the equivalent of a "STORED" block.

Protocol Buffers

Most of the data is stored in Protocol Buffer ("protobuf") wire messages.

The iWork apps (Keynote, Numbers, Pages) include embedded Protocol Buffers definitions as part of the file format processors.

The Protocol Buffer definitions can be dumped with an automated script:

deno run -A https://oss.sheetjs.com/notes/iwa/otorp.ts /Applications/Numbers.app out/

Note that some fields marked as required have been changed to optional in later versions. File parsers should assume all fields are optional.

App-Specific Definitions

The listed definitions only appear in one app:

Keynote

• KNArchives.proto
• KNArchives.sos.proto
• KNCommandArchives.proto
• KNCommandArchives.sos.proto

Numbers

• TNArchives.proto
• TNArchives.sos.proto
• TNCommandArchives.proto
• TNCommandArchives.sos.proto

Pages

• TPArchives.proto
• TPCommandArchives.proto
• TPCommandArchives.sos.proto

The other files are common across the apps.

Data Storage

The decompressed data is a series of chunks.

Each chunk starts with a length stored in a Base 128 varint, followed by a protobuf packet of type .TSP.ArchiveInfo.

The .TSP.ArchiveInfo message contains a number of .TSP.MessageInfo messages (tag 2). Each MessageInfo has a length field (tag 3, type uint32) for the actual message body. The data for the message bodies are stored immediately after the ArchiveInfo, in the same order as the MessageInfo parts.

The message type from the MessageInfo (tag 1, type uint32) corresponds to a dynamic registry spread across the embedded frameworks. The actual message data is a protobuf packet.

Dynamic Registry and Message Types

The .TSP.Reference type acts as a pointer, referencing another message. The references do not include message type info, so readers and writers must be aware of the message types and their interpretations.

Each framework is responsible for registering message types with the master registry by sending specific messages like setMessageType. Some relevant message types for Numbers files are listed below:

All referenced types must be registered, but ancillary types do not need to be registered. For example:

message .TST.TableInfoArchive {
  // ...
  required .TSP.Reference tableModel = 2;
  // ...
}

message .TST.TableModelArchive {
  // ...
  required .TST.DataStore base_data_store = 4;
  // ...
}

The reference in field 2 from .TST.TableInfoArchive is expected to be of type .TST.TableModelArchive so the latter must be registered.

.TST.DataStore is the type of field 4 from .TST.TableModelArchive. Since it is not referenced indirectly, the message type does not have to be registered.

Discovering the registry mapping

There are two approaches to discovering the registry mapping:

1. Search in the frameworks for places where setMessageType is sent.

2. Run the program in a debugger and inspect [TSPRegistry sharedRegistry]:

The latter approach was publicly documented by Sean Patrick O'Brien. A fully automated script runs on Intel Macs:

deno run -A https://oss.sheetjs.com/notes/iwa/dump_registry.ts

Document Structure

The iWork apps start from the DocumentArchive message, whose definition varies between formats.

Numbers

The root message (type 1) has the following structure:

message .TN.DocumentArchive {
  repeated .TSP.Reference sheets = 1;
  // ...
}

The message referenced in field 1 (type 2) has the following structure:

message .TN.SheetArchive {
  required string name = 1;
  repeated .TSP.Reference drawable_infos = 2;
  // ...
}

name is the name of the worksheet. Each worksheet can contain multiple tables. The messages referenced in field 2 (type 6000) refer to .TST.TableInfoArchive

Table Storage in iWork

Protobuf Structure

Table structure is shared across iWork apps. The protobuf definitions are identical. The root element for tables is the .TST.TableInfoArchive:

message .TST.TableInfoArchive {
  required .TSP.Reference tableModel = 2;
  // ...
}

The message referenced in field 2 (type 6001) has the following structure:

message .TST.TableModelArchive {
  required .TST.DataStore base_data_store = 4;
  required uint32 number_of_rows = 6;
  required uint32 number_of_columns = 7;
  // ...
}

message .TST.DataStore {
  required .TST.TileStorage tiles = 3;
  required .TST.TableRBTree rowTileTree = 9;
  // ...
}

message .TST.TileStorage {
  message .TST.TileStorage.Tile {
    required uint32 tileid = 1;
    required .TSP.Reference tile = 2;
  }
  repeated .TST.TileStorage.Tile tiles = 1;
  // ...
}

The message referenced in the tiles (type 6002) has the following structure:

message .TST.Tile {
  repeated .TST.TileRowInfo rowInfos = 5;
  // ...
}

message .TST.TileRowInfo {
  required uint32 tile_row_index = 1;
  required uint32 cell_count = 2;
  required bytes cell_storage_buffer_pre_bnc = 3;
  required bytes cell_offsets_pre_bnc = 4;
  optional bytes cell_storage_buffer = 6;
  optional bytes cell_offsets = 7;
  // ...
}

Each .TST.TileRowInfo message holds the data and property references for a single row in the table.

The placement of rows in the table is governed by the tree in field 9:

message .TST.TableRBTree {
  message Node {
    required uint32 key = 1;
    required uint32 value = 2;
  }
  repeated .TST.TableRBTree.Node nodes = 1;
}

Each node key is a row offset and value is an index into the tile array. For larger tables, tiles are generally expected to hold 256 rows.

Data Storage

Non-numeric data values are stored in lists referenced from .TST.DataStore:

message .TST.DataStore {
  required .TSP.Reference stringTable = 4;
  optional .TSP.Reference formulaErrorTable = 12;
  optional .TSP.Reference rich_text_table = 17;
  // ...
}

iWork uses a "shared string table" like Excel. Excel stores both plaintext and rich strings in the same table, while iWork has two separate tables.

Cell Storage

The cell offset fields are an array of 16-bit integers that describe offsets within the respective storage buffers. 0xFFFF indicates that the column index for the given row is not included.

Old Storage

The "pre-BNC" storages are specified in fields 3 and 4

Versions

The first byte of the storage is the version number. There are three known versions for the old storage: V1 (0-1), V3 (2-3), and V4 (4).

Field Mask

A bitmask is stored at offset 4, describing which fields are in the cell:

The size of the bitmask is 2 bytes in V1 and 4 bytes in V3 and V4.

Cell Type

The cell type is stored at byte offset 2 in V1 / V3 and offset 1 in V4:

Fields

V1 fields start at offset 8, while V3 and V4 start at offset 12. Fields are enumerated in order that the data appears in the storage.

New Storage

The "BNC" ("post-BNC"?) storages are specified in fields 6 and 7

Versions

The first byte of the storage is the version number. At the time of writing, the only version is V5 (5).

Field Mask

A bitmask is stored at offset 8, describing which fields are in the cell:

Cell Type

The cell type is stored at byte offset 1:

Fields

Fields start at offset 12. The fields are in the same order as the bit flags.

Spreadsheet Features

Hyperlinks

Like HTML, hyperlinks are attached to rich text fragments. Unlike Excel, there is no cell-level hyperlinks in iWork tables.

The .TST.DataStore includes a reference to the rich text table:

message .TST.DataStore {
  optional .TSP.Reference rich_text_table = 17;
  // ...
}

The referenced message is of type 6005 (.TST.TableDataList):

message .TST.TableDataList {
  message ListEntry {
    required uint32 key = 1;
    required uint32 refcount = 2;
    optional .TSP.Reference rich_text_payload = 9;
    // ...
  }
  enum ListType {
    RICH_TEXT_PAYLOAD = 8;
    // ...
  }
  required .TST.TableDataList.ListType listType = 1; // 8 = RICH_TEXT_PAYLOAD
  repeated .TST.TableDataList.ListEntry entries = 3;
  // ...
}

Each rich text payload is of type 6218 (.TST.RichTextPayloadArchive):

message .TST.RichTextPayloadArchive {
  required .TSP.Reference storage = 1;
  // ...
}

The storage is of type 2001 (.TSWP.StorageArchive):

message .TSWP.StorageArchive {
  repeated string text = 3;
  optional .TSWP.ObjectAttributeTable table_smartfield = 11;
  // ...
}

message .TSWP.ObjectAttributeTable {
  message ObjectAttribute {
    required uint32 character_index = 1;
    optional .TSP.Reference object = 2;
  }
  repeated .TSWP.ObjectAttributeTable.ObjectAttribute entries = 1;
}

The styling is similar to XLS: the styling metadata is stored separately, with pointers indicating the starting character. If there is a hyperlink, an entry will point to a message of type 2032 (.TSWP.HyperlinkFieldArchive):

message .TSWP.HyperlinkFieldArchive {
  optional string url_ref = 2; // <-- encoded URL string
  // ...
}

Misc

Determining File Type

The root DocumentArchive message fields vary between formats.

In the 13.0 apps, the required fields are:

// Keynote optional fields 4
message .KN.DocumentArchive {
  required .TSA.DocumentArchive super = 3;
  required .TSP.Reference show = 2;
}

// Numbers optional fields 1, 3, 7, 9 - 12
message .TN.DocumentArchive {
  required .TSA.DocumentArchive super = 8;
  required .TSP.Reference stylesheet = 4;
  required .TSP.Reference sidebar_order = 5;
  required .TSP.Reference theme = 6;
}

// Pages optional fields 2 - 7, 11 - 14, 16, 17, 20, 21, 30 - 50
message .TP.DocumentArchive {
  required .TSA.DocumentArchive super = 15;
}

Pages is the only format to use and require field 15. Keynote requires field 2, a field that does not appear in Numbers.