Word MCP Server (FsWordDsl)

Kookerella.FsWordDsl

A typesafe F# DSL for building Word documents, interpreted into calls against the DocumentFormat.OpenXml SDK. The DSL is a plain data model (records/DUs with structural equality) - the interpreter (Writer) compiles it to OOXML, and the reverse transform (Reader) parses an existing .docx back into the same DSL.

This is the WordprocessingML sibling of Kookerella.FsOpenXmlDsl (the Excel/SpreadsheetML one) - same objectives, same round-trip philosophy, translated to Word's own document model. See MAPPING.md for exactly which WordprocessingML features map 1:1, which are approximated, and which aren't modeled yet.

This round-trips in both directions, which most Word libraries don't: they give you an imperative API to build a document from scratch, but no way to turn an existing file back into readable source. Here, Reader parses a real .docx/.docm back into the same DSL, and Document.generateScript goes one step further and renders that model back out as a self-contained script that rebuilds an equivalent file - a decompiler for Word documents, not just a writer. Two more surfaces, Xml.toDocument/Xml.ofDocument (see "## XML" below) and Json.toDocument/Json.ofDocument (see "## JSON" below), do the same translation to/from plain XML or JSON against a real schema - for a caller who'd rather generate or consume data than write code at all.

A fluent C# wrapper (Kookerella.CsWordDsl) sits on top of the F# core for callers who'd rather not touch F# discriminated unions/option types directly - immutable records with With* builders, plus its own CsCodeGen decompiler that renders a Document back out as runnable C# source. See "## The C# wrapper" below.

An MCP server (Kookerella.FsWordDsl.Mcp) exposes the same read/write/decompile capabilities as tools any MCP-compatible AI agent can call directly, and doubles as a plain CLI (fsworddsl-mcp convert/build) for anyone not going through an MCP client at all - see its own README for the full tool list.

Don't have .NET installed? Download a self-contained build of the MCP server for your platform (win/linux/osx, x64/arm64) from the Releases page - unzip and run, no install required.

Layout

• src/Kookerella.FsWordDsl - the library.
  • Units.fs - conversions between points/inches/pixels and the physical units WordprocessingML uses on the wire (twips for page geometry/spacing, EMU for image sizing).
  • Styles.fs - character and paragraph formatting: Color (Rgb, Auto, or a theme-relative Theme color - see ThemeColorKind), HighlightColor (Word's own fixed highlight palette), UnderlineStyle, RunStyle (including small caps/all caps/hidden text), ParagraphAlignment, Indentation, LineSpacingRule, TabStopAlignment/TabLeader/TabStop, ParagraphFormat (including paragraph borders, shading, and custom tab stops), BorderLineStyle, BorderSide, BorderStyle (reused for both paragraph and table borders).
  • NamedStyles.fs - StyleDefinition (paragraph or character, with BasedOn inheritance) and a small BuiltInStyles catalog (normal, heading1/2/3, title, listParagraph, hyperlinkCharStyle).
  • Numbering.fs - NumberFormatKind, ListLevel, NumberingDefinition for numbered/bulleted lists, including multi-level ones (ListLevel isn't limited to one per definition - see Builders.multiLevelNumberedListDef).
  • Hyperlinks.fs - HyperlinkTarget (external URL vs. internal bookmark reference).
  • Protection.fs - EditRestriction and DocumentProtection, document-level (Word has no per-section equivalent of Excel's per-sheet protection).
  • Revisions.fs - RevisionKind/Revision for track changes (Inline.TrackedChange, Paragraph.MarkRevision) - narrowly scoped to inserted/deleted content and paragraph marks, see MAPPING.md for what isn't covered.
  • ContentControls.fs - ContentControlType/ContentControlProps for content controls (structured document tags, w:sdt): plain text, rich text, dropdown/combo box, date picker, checkbox - see MAPPING.md for what isn't covered.
  • PageSetup.fs - PageOrientation, PageSize, PageMargins, SectionBreakType, NoteNumberRestart/NoteNumberingSettings (a section's own footnote/endnote numbering).
  • Tables.fs - TableBorders, VerticalMergeKind, TableCellProps (including a per-cell Margins override), TableStyleRef, TableStyleRegion/ TableStyleDefinition (custom table style definitions - all thirteen of OOXML's conditional-formatting regions), and CellMargins (shared shape for a table's default margins and a single cell's own override).
  • Images.fs - ImageFormat, ImageEntry (raw file bytes plus an on-page size), anchored inline within a run.
  • DocumentProperties.fs - DocumentProperties (Title, Author, Subject, Keywords, Comments, Category, Company) - core document metadata, Document.Properties.
  • Model.fs - the recursive content model: Inline (runs, breaks, images, hyperlinks, bookmarks and comments - both the single-paragraph Bookmark/Comment cases and the cross-paragraph BookmarkRangeStart/End/CommentRangeStart/End markers, simple fields, footnotes/endnotes, TrackedChange for track changes, and InlineContentControl for content controls), Paragraph (including MarkRevision), Block (paragraph, table, or ContentControlBlock - the block-level counterpart to InlineContentControl), TableCell/ TableRow (including RepeatAsHeader)/TableEntry (including CellMargins), HeaderFooterSet, SectionProperties (including BreakType and FootnoteNumbering/ EndnoteNumbering), Section, Document (including Document.VbaProject, a macro-enabled document's raw vbaProject.bin bytes, Document.Properties, and Document.TableStyles).
  • Xml.fs / Xml.xsd - the XML surface: Xml.toDocument/Xml.ofDocument translate a Document to/from an XElement tree, and Xml.schemaSet() loads the paired schema (embedded in the assembly as a resource) for validating either direction. See "## XML" below.
  • Json.fs / Json.schema.json - the JSON surface: Json.toDocument/Json.ofDocument translate a Document to/from a System.Text.Json.Nodes.JsonObject tree. Schema validation is test-suite only, not a public API. See "## JSON" below.
  • Builders.fs - plain functional constructors (section, document, withStyles, withNumbering, withProtection, withVbaProject, withDocumentProperties, withTableStyles, bulletListDef, numberedListDef, multiLevelNumberedListDef) plus DocumentDsl - smart constructors (run, para (with markRevision), hyperlink, bookmark, comment, inserted/deleted (track changes), contentControl/contentControlBlock (content controls), image, footnote, endnote, tableCell, tableRow (with height/repeatAsHeader), table (with style/borders/cellMargins)) with real optional parameters, the Word analog of the Excel repo's SheetDsl.
  • Interpreter/StyleRegistry.fs - shared run/paragraph/border/color conversions plus Document.Styles <-> styles.xml (internal).
  • Interpreter/ImageWriter.fs / ImageReader.fs - an inline image's own DSL <-> DrawingML translation (internal).
  • Interpreter/Writer.fs - DSL -> OOXML (internal).
  • Interpreter/Reader.fs - OOXML -> DSL, the reverse transform (internal).
  • Interpreter/CodeGen.fs - DSL -> F# source text: renders a Document back out as a self-contained .fsx script that rebuilds an equivalent file when run (internal).
  • Api.fs - the public Document.save/saveToStream/load/loadFromStream/ generateScript entry points.
• src/Kookerella.CsWordDsl - the fluent C# wrapper (see "## The C# wrapper" below): immutable records/sealed record closed hierarchies mirroring the F# core's own types one-for-one, DocumentConverter.cs (internal, the two-way F#<->C# translation), DocumentIO.cs (Save/Load, the one place this project does I/O), CsCodeGen.cs (DSL -> C# source text, the C# analog of Interpreter/CodeGen.fs).
• src/Kookerella.FsWordDsl.Mcp - the MCP server (see its own README): DocumentTools.fs (the tool surface, one [<McpServerTool>]-tagged member per tool), Program.fs (dispatches to the MCP stdio server, or to a plain convert/build CLI, depending on argv). Distributed as a dotnet tool (fsworddsl-mcp), same as the Excel sibling's own Kookerella.FsOpenXmlDsl.Mcp.
• tests/Kookerella.FsWordDsl.Tests - one test per feature, each validating the produced file against the OOXML schema (DocumentFormat.OpenXml.Validation.OpenXmlValidator) and asserting an exact round trip back through the DSL. Each test also writes the document it builds to Examples/<test name>/output.docx (checked into the repo), plus script.fsx (regenerates the file - a separate, slower Category=Slow test group actually executes each one via dotnet fsi), document.xml, and document.json - one folder always has four views of the same example.
• tests/Kookerella.CsWordDsl.Tests - DriftGuardTests.cs (a reflection-based tripwire checking the C# wrapper's DU mirrors haven't fallen behind the F# core's own case counts), DocumentTests.cs (targeted round-trip assertions per feature), ExampleTests.cs (reloads the F# suite's own checked-in Examples/*/output.docx fixtures rather than re-authoring every scenario a second time), CsCodeGenTests.cs (actually executes a generated file via dotnet run --file, the C# analog of the F# suite's Category=Slow dotnet fsi group).
• samples/Kookerella.FsWordDsl.Sample - a small console app that builds a document, saves it, and reads it back.

Quick start

open Kookerella.FsWordDsl
open type Kookerella.FsWordDsl.DocumentDsl

let doc =
    document
        [ section
              [ para ([ run "Quarterly Report" ], styleId = "Title")
                para
                    [ run "This report covers "
                      run ("Q1 2026", style = { RunStyle.Default with Bold = true })
                      run ", see the "
                      hyperlink ("full dataset", ExternalUrl "https://example.com/data")
                      run " for details." ] ] ]

doc |> Document.save "report.docx"

// Reverse transform:
let roundTripped = Document.load "report.docx"

document defaults Styles to BuiltInStyles.all, so styleId = "Heading1" (or any other built-in id) just works without registering it first - pipe withStyles afterward to replace or extend that set. run/para/hyperlink/bookmark/comment/image/ tableCell/tableRow/table are DocumentDsl members with real optional parameters (open type Kookerella.FsWordDsl.DocumentDsl brings them into scope unqualified, same as open type SheetDsl does in the Excel repo) - plain F# let bindings can't have optional parameters, which is why this part of the DSL is a type.

A Paragraph's Inlines are naturally several independently-styled runs - rich text (mixed formatting within one paragraph) is first-class, not a documented gap the way Excel's single-uniform-run Text cell is:

para
    [ run "Plain text, "
      run ("bold", style = { RunStyle.Default with Bold = true })
      run ", and "
      run ("colored", style = { RunStyle.Default with Color = Some Color.red }) ]

RunStyle also covers small caps, all caps, and hidden text; ParagraphFormat covers borders (BorderStyle, the same shape used for table borders) and shading:

para
    ([ run "ALL CAPS AND SMALL CAPS" ], format =
        { ParagraphFormat.Default with
            Borders = Some { BorderStyle.None with Bottom = Some { Style = SingleLine; Width = Some 1.0; Color = Some Color.black } }
            Shading = Some(Rgb(0xD9uy, 0xD9uy, 0xD9uy)) })

Custom tab stops (TabStop) sit on ParagraphFormat.TabStops - a right-aligned stop with a dot leader is the classic table-of-contents pattern:

para
    ([ run "Introduction"; Tab; run "1" ], format =
        { ParagraphFormat.Default with TabStops = [ { Position = 288.0; Alignment = RightTab; Leader = DotLeader } ] })

Color also accepts a theme-relative token (Theme) alongside plain Rgb/Auto - since this DSL has no theme part to resolve it against, real Word does that; Fallback is what a themeless reader sees instead, the same "always also write a computed value" convention Word itself follows:

run ("Accent-colored text", style = { RunStyle.Default with Color = Some(Theme(Accent1Theme, (0x1Fuy, 0x49uy, 0x7Duy), None, None)) })

Lists use a (numId, level) reference on the paragraph, resolved against a NumberingDefinition attached to the document - NumberingDefinition.Levels isn't limited to one level, and multiLevelNumberedListDef builds the common correctly-linked outline shape for you:

document
    [ section
          [ para ([ run "First bullet" ], numbering = (1, 0))
            para ([ run "Second bullet" ], numbering = (1, 0)) ] ]
|> withNumbering [ bulletListDef 1 ]

document
    [ section
          [ para ([ run "First topic" ], numbering = (1, 0))
            para ([ run "First subtopic" ], numbering = (1, 1))
            para ([ run "Second topic" ], numbering = (1, 0)) ] ]
|> withNumbering [ multiLevelNumberedListDef 1 3 ]

Tables are built from tableRow/tableCell, with column widths given once for the whole table - a cell without an explicit width falls back to its column's width at write time:

table (
    [ tableRow [ tableCell [ para [ run "Item" ] ]; tableCell [ para [ run "Qty" ] ] ]
      tableRow [ tableCell [ para [ run "Widgets" ] ]; tableCell [ para [ run "12" ] ] ] ],
    [ 200.0; 100.0 ],
    style = TableStyleRef.Default
)

Cell merging - horizontal (GridSpan) and vertical (RestartMerge/ContinueMerge) - are independent and combine on the same cell, matching real Word:

tableCell ([ para [ run "Spans 2 columns" ] ], props = { TableCellProps.Default with GridSpan = Some 2 })

A cell's own margins override the table's default the same CellMargins shape covers both:

tableCell ([ para [ run "Extra padding" ] ], props = { TableCellProps.Default with Margins = Some { CellMargins.Default with Top = Some 8.0; Bottom = Some 8.0 } })

A custom table style (TableStyleDefinition) lives in Document.TableStyles and is applied by name, the same way a built-in like "TableGrid" is - here with a bold white header row on a blue background, an italic last row, and alternating row shading, plus a table-wide default cell margin and a row that repeats on every page:

let corporateStyle: TableStyleDefinition =
    { TableStyleDefinition.Default with
        Id = "Corporate"
        Name = "Corporate"
        FirstRow =
            { TableStyleRegion.None with
                RunFormat = Some { RunStyle.Default with Bold = true; Color = Some Color.white }
                CellShading = Some(Rgb(0x4Fuy, 0x81uy, 0xBDuy)) }
        LastRow = { TableStyleRegion.None with RunFormat = Some { RunStyle.Default with Italic = true } }
        BandedRow = { TableStyleRegion.None with CellShading = Some(Rgb(0xDCuy, 0xE6uy, 0xF1uy)) } }

document
    [ section
          [ table (
                [ tableRow ([ tableCell [ para [ run "Item" ] ]; tableCell [ para [ run "Qty" ] ] ], repeatAsHeader = true)
                  tableRow [ tableCell [ para [ run "Widgets" ] ]; tableCell [ para [ run "12" ] ] ] ],
                [ 200.0; 100.0 ],
                style = { TableStyleRef.Default with Name = "Corporate" },
                cellMargins = { Top = Some 4.0; Bottom = Some 4.0; Left = Some 6.0; Right = Some 6.0 }
            ) ] ]
|> withTableStyles [ corporateStyle ]

TableStyleDefinition also covers FirstColumn/LastColumn, BandedColumn, and the four corner cells (NorthEastCell/NorthWestCell/SouthEastCell/SouthWestCell) - the two regions not modeled are each banding axis's second band, since in practice that's just WholeTable's own background showing through (see MAPPING.md).

Sections carry their own page setup - a document is a sequence of Sections, mapping 1:1 onto real Word section breaks. BreakType is how a section begins relative to the previous one - meaningless (and not written) on the very first section:

let landscape = { SectionProperties.Default with Orientation = Landscape }
document [ sectionWith landscape [ para [ run "A landscape-oriented page." ] ] ]

let continuous = { SectionProperties.Default with BreakType = ContinuousBreak }
document
    [ section [ para [ run "Section 1." ] ]
      sectionWith continuous [ para [ run "Section 2 - no page break from section 1." ] ] ]

Footnotes and endnotes mark a point in a paragraph's own Inlines - content is the note's own body, written to word/footnotes.xml/endnotes.xml with an id Writer assigns automatically (the reference-mark run itself is generated for you, on both ends):

para
    [ run "This claim needs a citation"
      footnote "Smith, J. (2023). A Study of Claims."
      run ", and this one refers to a fuller discussion"
      endnote [ para [ run "See the appendix for the full derivation." ] ] ]

A section's own footnote/endnote numbering (w:footnotePr/w:endnotePr) - None is Word's own default (continuous decimal from 1); here footnotes are lower-roman and restart every page, matching a common legal-document convention:

sectionWith
    { SectionProperties.Default with FootnoteNumbering = Some { Format = LowerRomanFormat; StartAt = None; Restart = RestartEachPage } }
    [ para [ run "Body text."; footnote "A footnote numbered i, ii, iii, ... restarting each page." ] ]

Headers and footers are per-section, with Default/First/Even variants (the titlePg/evenAndOddHeaders flags real Word needs are set automatically):

let footer = { HeaderFooterSet.None with Default = Some [ para [ run "Page "; Field("PAGE", Some "1") ] ] }
sectionWith { SectionProperties.Default with Footer = Some footer } [ para [ run "Body text." ] ]

Comments and bookmarks wrap inline content directly, the common single-paragraph case:

para [ comment ([ run "This figure needs review." ], "Please double check the totals.", author = "Alex") ]

Either spanning more than one paragraph uses two independent markers placed directly in separate paragraphs instead, sharing an id - BookmarkRangeStart/BookmarkRangeEnd for bookmarks, CommentRangeStart/CommentRangeEnd for comments (which carries the comment's own metadata on its Start, since there's no wrapping case here to hang it off - see MAPPING.md on why that id is write-time-only, unlike a bookmark's own name):

document
    [ section
          [ para [ BookmarkRangeStart "Section2"; run "This paragraph starts the bookmark" ]
            para [ run "and this one ends it."; BookmarkRangeEnd "Section2" ] ] ]

document
    [ section
          [ para [ CommentRangeStart("review1", "Alex", None, None, "This section needs review."); run "Comment starts here" ]
            para [ run "and ends here."; CommentRangeEnd "review1" ] ] ]

Track changes (inserted/deleted) wrap inline content the same way, marking it as inserted or deleted under an author and date; a whole inserted or deleted paragraph (rather than just some of its content) uses para's own markRevision instead, for the paragraph's closing mark:

para
    [ run "The quick "
      inserted ([ run "brown " ], "Alex")
      run "fox jumps over the "
      deleted ([ run "lazy " ], "Alex")
      run "dog." ]

para ([ run "This whole paragraph was inserted." ], markRevision = { Kind = Inserted; Author = "Alex"; Date = None })

Content controls (contentControl for run-level, contentControlBlock for block-level) wrap their own currently-displayed content the same way, plus a ContentControlType (see MAPPING.md for the full set - plain text, rich text, dropdown/combo box, date picker, checkbox):

para
    [ run "Client name: "
      contentControl ([ run "Type here" ], PlainTextControl false, alias = "Client Name", tag = "clientName") ]

para
    [ run "Favorite color: "
      contentControl ([ run "Blue" ], DropDownControl([ "Red", "red"; "Green", "green"; "Blue", "blue" ], false)) ]

contentControlBlock([ para [ run "This whole paragraph is a rich-text content control." ] ], RichTextControl, alias = "Notes")

Document-level protection, macros, and core properties are all pipe-friendly, same shape as Excel's own withProtection/withVbaProject:

document [...] |> withProtection { Edit = Some ReadOnlyRestriction; Password = Some "hunter2" }
document [...] |> withVbaProject (System.IO.File.ReadAllBytes("vbaProject.bin"))
document [...] |> withDocumentProperties { DocumentProperties.Default with Title = Some "Quarterly Report"; Author = Some "Kookerella" }

Save the result with a .docm path - Document.save/saveToStream automatically switch the file's own declared content type to Word's macro-enabled kind whenever a VbaProject is present, but real Word also expects the .docm extension to trust and run macros at all.

Regenerating a file as F# source

Given a Document (typically one you just Document.loaded from an existing file), Document.generateScript renders it back out as a self-contained .fsx script that rebuilds an equivalent file when run - a code-generating counterpart to Document.load:

let doc = Document.load "input.docx"

let referenceLines =
    [ "#r \"path/to/Kookerella.FsWordDsl.dll\""
      "#r \"path/to/DocumentFormat.OpenXml.dll\"" ]

let script = Document.generateScript referenceLines "output.docx" doc
System.IO.File.WriteAllText("regenerate.fsx", script)

Running dotnet fsi regenerate.fsx produces output.docx - not byte-identical to the original (zip metadata/timestamps differ) but structurally equivalent through the same round-trip lens every other test in this repo uses. Every scenario under tests/ Kookerella.FsWordDsl.Tests/Examples/ has a committed script.fsx generated exactly this way; the Category=Slow test group actually executes each one via dotnet fsi and checks it reproduces the committed .docx.

The C# wrapper

Kookerella.CsWordDsl is an idiomatic, immutable, fluent C# wrapper over the F# core, for callers who'd rather not touch F# discriminated unions or option types directly. Every F# type has a C# mirror: plain records with With*/factory-method builders for product types, enums for parameterless choices, and sealed record closed hierarchies (abstract record base, private constructor, nested cases) for everything else - the same "sealed hierarchy" pattern the Excel repo's own Kookerella.CsOpenXmlDsl uses for CellValue/ ConditionalFormatRule. Reference Kookerella.CsWordDsl instead of Kookerella.FsWordDsl and never see an FSharpOption:

using Kookerella.CsWordDsl;

var doc = Document.Create(
    Section.Of([
        Block.Paragraph([new Inline.Run("Quarterly Report")], styleId: "Title"),
        Block.Paragraph([
            new Inline.Run("This report covers "),
            new Inline.Run("Q1 2026", new RunStyle { Bold = true }),
            Inline.HyperlinkText("full dataset", new HyperlinkTarget.ExternalUrl("https://example.com/data")),
            new Inline.Run(" for details.")
        ])
    ]));

DocumentIO.Save(doc, "report.docx");
var loaded = DocumentIO.Load("report.docx");

Content controls, tables, track changes, comments, and every other feature the F# core models are covered the same way - see tests/Kookerella.CsWordDsl.Tests/DocumentTests.cs for a worked example per feature. CsCodeGen.Generate is the C# analog of Document.generateScript: it renders a Document back out as a self-contained C# file targeting .NET's "file-based apps" feature (dotnet run --file script.cs), rather than an .fsx script:

var script = CsCodeGen.Generate(["#:project path/to/Kookerella.CsWordDsl.csproj"], "output.docx", loaded);
File.WriteAllText("regenerate.cs", script);

DocumentIO also exposes the F# core's other two ways in and out directly - schema-backed XML/JSON (ToXml/FromXml, ToJson/FromJson) and F# script generation (GenerateFSharpScript, CsCodeGen.Generate's F#-targeting sibling) - so a C# caller never needs its own reference to Kookerella.FsWordDsl to reach any of the F# core's four I/O surfaces from C#.

One design note worth stating explicitly: this wrapper's records use IReadOnlyList<T> properties, and C#'s compiler-synthesized record equality does not deep-compare list contents (two records holding equal-but-distinct list instances compare unequal via plain .Equals()) - the same limitation Kookerella.CsOpenXmlDsl's own records have. Don't rely on whole-Document equality in your own code; compare the specific values you care about, the same way this repo's own DocumentTests.cs does.

XML

Xml.toDocument/Xml.ofDocument (in Xml.fs) are a third way in and out of the DSL, alongside writing F# directly and code generation: plain XML, against a real schema (Xml.xsd, embedded in the assembly). A data-carrying DU case becomes an element named after the case; a parameterless-choice case becomes an attribute value or bare string, matching the convention the Excel repo's own Xml.fs documents.

open System.Xml.Linq

// XML -> Document -> .docx
let doc = XElement.Load "report.xml" |> Xml.ofDocument
Document.save "report.docx" doc

// .docx -> Document -> XML
let xml = Document.load "report.docx" |> Xml.toDocument
xml.Save "report.xml"

A run with direct formatting and a hyperlink, in XML:

<para>
  <run>Visit </run>
  <hyperlink tooltip="Kookerella on GitHub">
    <externalHyperlink>https://github.com/Kookerella-Ltd</externalHyperlink>
    <content>
      <run styleId="Hyperlink">Kookerella on GitHub</run>
    </content>
  </hyperlink>
  <run> for more.</run>
</para>

Xml.schemaSet() loads the compiled schema for validating either direction yourself (XDocument.Validate) - every scenario under tests/Kookerella.FsWordDsl.Tests/Examples/ has a committed document.xml validated against it this way as part of the same test that generates it.

toDocument's output is deterministically ordered (Styles/Numbering/TableStyles sorted by Id, regardless of the order the underlying Document's own lists happen to be in) - paragraph/run content is already real document order and needs no sorting, but these three are ID-referenced catalogs whose own list order carries no meaning, so this is what makes committing document.xml to version control and diffing it across commits actually meaningful: a genuine content change produces a small, isolated diff rather than a spurious one from a catalog getting reshuffled between two otherwise-identical documents.

JSON

Json.toDocument/Json.ofDocument (in Json.fs) are a fourth way in and out of the DSL, alongside writing F# directly, code generation, and XML: plain JSON, for a caller whose tooling speaks JSON rather than XML. The same DU-case conventions apply, in JSON's own idiom (a single-key object for a data-carrying case, a bare string for a parameterless one):

open System.Text.Json.Nodes

// JSON -> Document -> .docx
let doc = JsonNode.Parse(File.ReadAllText "report.json").AsObject() |> Json.ofDocument
Document.save "report.docx" doc

// .docx -> Document -> JSON
let json = Document.load "report.docx" |> Json.toDocument
File.WriteAllText("report.json", json.ToJsonString())

The same hyperlink example as above, in JSON:

{
  "para": {
    "inlines": [
      { "run": { "text": "Visit " } },
      {
        "hyperlink": {
          "target": { "externalHyperlink": "https://github.com/Kookerella-Ltd" },
          "runs": [ { "run": { "text": "Kookerella on GitHub", "styleId": "Hyperlink" } } ],
          "tooltip": "Kookerella on GitHub"
        }
      },
      { "run": { "text": " for more." } }
    ]
  }
}

The same determinism Xml.toDocument has (Styles/Numbering/TableStyles sorted by Id) applies to Json.toDocument's output too, for the same reason: a genuine content change produces a small, isolated diff rather than a spurious one from a catalog getting reshuffled between two otherwise-identical documents.

Unlike XML, .NET has no built-in JSON Schema validator, so Json.schema.json (in the repo) is validated only from this repo's own test suite (via a test-only JsonSchema.Net dependency) rather than exposed as a public API - see Json.fs's own doc comment.

Building and testing

dotnet build
dotnet test --filter "Category!=Slow"
dotnet run --project samples/Kookerella.FsWordDsl.Sample

The default loop above skips the slow Category=Slow tests, which actually invoke dotnet fsi on every generated Examples/*/script.fsx (multi-second process startup each). Run those explicitly, after the fast suite has populated the .fsx files at least once:

dotnet test --filter "Category=Slow"

Plain dotnet test (no filter) runs both groups.

The C# wrapper's own suite has no fast/slow split - CsCodeGenTests.cs shells out to dotnet run --file itself, so a single run already covers the C# analog of the F# suite's slow group:

dotnet test tests/Kookerella.CsWordDsl.Tests