v0.2.1

Two rendering bugs, both of which made a correctly imported deck look wrong, and
the ts-pptx 4 upgrade. Nothing about the model or the editable surface changes,
so a document rendered by 0.2.0 still parses.

Fixed

• A group no longer offsets its children twice. Every grouped shape was drawn
  at twice its distance from the origin, because the renderer put the group's own
  translate on the wrapper and then rendered the children against the slide.
  The further a group sat from the origin the further its contents left it: a row
  of KPI cards across a slide ends up with some cards off the slide and the rest
  on top of the body text. The model has always said a child's placement is
  slide-absolute and that a renderer "never composes a transform"; this makes the
  renderer obey it. A table keeps its own transform, because unlike a group it is
  what its cells are positioned against.

• Percentage line spacing is written as the multiple it is. A paragraph asking
  for 105% got lines 16.8px apart whatever size its text was, so a large title
  painted its lines on top of each other. PowerPoint's spcPct is a multiple of
  each line's own font size; a CSS percentage resolves once, against the
  paragraph, before the runs inherit it, and the size lives on the runs. It is
  unitless now, which is what the inherited-spacing branch beside it already did
  and for the same stated reason. An exact spcPts spacing still emits px.

Changed

• Requires @shbernal/ts-pptx 4. The read model consolidated every colour
  onto one ColorRef, renamed Part.bytes to Part.originalBytes, and made
  importSlide refuse a source page whose hyperlink leaves the set being
  imported. Nothing here renders differently for it. Two recorded forms move
  because upstream now reads more than it did: the coverage baseline gains
  text.color.inherited on the two corpus decks that build their text as
  paragraph arrays, and the canonical model gains company.

v0.2.0

Fixed

• A table cell's two diagonals are modeled and drawn. a:lnTlToBr and
  a:lnBlToTr — PowerPoint's Diagonal Down/Up Border, and how a cell is struck
  out — were dropped on import with no FidelityNote, which is neither modeled,
  carried nor warned. They were also not a loss that had to happen: the read model
  decodes both (CellBorders.tlToBr / .blToTr) and the write API takes them back
  (TableCellProps.diagonal), so both sides of @shbernal/ts-pptx could express a
  diagonal and only this pipeline could not. TableCell.borders now carries all
  six lines and the renderer draws them, the two diagonals last, across the whole
  spanned region of a merged cell.

• The document's language reaches the emitted deck. The heuristic lane parses
  <html lang> and assigned it to pptx.lang and to theme.lang. The writer has
  neither: both landed as own properties on the instance and nothing wrote them
  into the package, so every deck came out en-US whatever the page said. lang
  is a run option, so it now travels on each addText call and on each table
  cell — a table-level one does not reach the cells, which is checked against the
  emitted XML rather than read off the type.

• A gradient stop and a table cell's edge keep their colour's transform list.
  Both arrived pre-flattened — a scheme token and a painted hex with nothing
  between them — so transforms: [] meant "the reader had nothing to say", which
  is the same shape as "the deck stated none". A lumMod-darkened stop was
  indistinguishable from one that was not, and re-authoring either against a
  different theme silently stopped tracking it. The cell edge lost more than the
  list: its colour carried no alpha, so a transparent rule was imported opaque
  and drawn as one. Both now decode through the same colorOf a solid fill uses,
  on ResolvedColors that arrived in @shbernal/ts-pptx 3.6.0
  (ts-pptx#26).

• An inherited italic is painted. italic was the one run property with no
  resolved counterpart, so a run taking its slant from a master's a:defRPr i="1"
  rendered upright. Paint-only — props is what emit reads, so the deck always
  round-tripped, which is exactly why nothing caught it. Run.resolvedItalic
  (ts-pptx#27) closes it, and
  ResolvedRunProperties now carries italic beside bold.

• A cell spanning rows is drawn as tall as the rows it covers. renderTable
  measured a spanning cell's two dimensions two different ways: the column arm
  summed the widths of the columns the cell covers, and the row arm multiplied
  this row's height by the span count. Those agree only when every row is the
  same height, and a header row above body rows — the shape most tables have — is
  where they do not: a two-row cell was drawn as twice the header rather than as
  the header plus the body. Both arms sum now. Paint only, so nothing here reached
  the island and no oracle number moved.

Changed

• IR_VERSION is 8. TableCell.borders gained the two diagonals as required
  fields, so a document rendered by an older build carries no key to read them
  from. The parser refuses the mismatch and says to re-render, which is the true
  answer; nothing migrates.

• @shbernal/ts-pptx ^3.6.0 is the floor, up from ^3.2.0. Two of the
  fixes above are upstream's (ts-pptx#26,
  ts-pptx#27), so the release that
  carries them is the one this reads them from.

• A delivered deck goes out as bytes. deliverDeck's blob and default
  download modes wrote the package to base64 and decoded it straight back, so a
  routinely multi-megabyte archive was encoded over 32 KB chunks and decoded again
  a character at a time, in the browser, to arrive where the writer started. Both
  take the writer's toBytes() now, and only output: 'base64' pays for an
  encode. That method is optional on the type — opts.pptxFactory may hand back a
  mock, or a writer predating it — so the base64 path stays as the fallback.

Added

• Five more preset geometries, and connector line ends. round2DiagRect,
  round2SameRect, rightArrow, rightArrowCallout and wedgeRectCallout are
  resolved from their formulas instead of falling back to the obvious box, and a
  modeled a:headEnd / a:tailEnd is drawn as a stroke-relative SVG marker,
  which is how OOXML defines an end's @w and @len in the first place. Both
  additions are on the fidelity gate.

• cell-diagonal joins the corpus as a primitive-tier deck: three cells, one
  with a single diagonal, one with both, one with none. The distinction is only
  testable from a generated deck if the deck states it three ways — a lane that
  carried the first cell's rule onto every cell would agree with a corpus that
  used one.

• color-transform joins the corpus as a primitive-tier deck, holding both
  halves of the flattening above: a themed gradient stop with a transform beside
  one without, and a table whose rules state an opacity. transparency is the one
  colour transform the write API can author, which is what makes the distinction
  reachable from a generated deck at all. layout-placeholder's master now states
  italic for body level 1, so the corpus contains a run whose slant is written down
  nowhere but the master.

• cloneIr(ir) — a deep copy of a RenderIr, exported beside emuOf and
  inchesOf. It exists because the two spellings the repo was using are not
  equivalent in general: JSON.parse(JSON.stringify(...)) drops
  undefined-valued keys and cannot carry a Uint8Array, and they agree on this
  model only because of its JSON contract. A consumer holding an IR needs the
  same guarantee, and the one that keeps working if that contract ever moves is
  structuredClone.

• EMU_PER_POINT joins EMU_PER_INCH on the public surface. Both are the
  model's stated units, and a consumer reading a widthPt beside a Box needs
  the second one to make sense of the pair. It was already in the package, one
  directory down in the renderer.

• eachNode(nodes, visit) and eachTextBody(nodes, visit) — the two
  traversals of a node tree, exported beside the model they walk. eachTextBody
  is the answer to "where does text live in a node tree", stated once and
  exhaustive over RenderNode['kind'], so a new kind is a compile error rather
  than an omission wherever the question is asked. eachNode visits a group
  before its children and lets a consumer prune a subtree by returning false,
  which is what "this node has no box, so nothing inside it was drawn" needs.

v0.1.2

Fixed

• An SVG picture is drawn as its SVG, not as the raster fallback beside it.
  A vector picture is two parts in OOXML: the art hangs off the asvg:svgBlip
  extension, and a:blip/@r:embed holds a fallback for readers that cannot draw
  vectors. The importer took the fallback — and @shbernal/ts-pptx writes that
  fallback as a 1×1 transparent PNG, so every icon in a deck rendered as one
  invisible pixel. Nothing warned, because a picture that resolves to a real part
  looks resolved: the SVG parts arrived in the asset manifest and no node
  referenced them. Found by the first outside consumer, on a deck where 28 of 49
  slides lost every glyph.

  The same change picks up the SVG-only form, where a:blip/@r:embed is
  legitimately absent (PowerPoint's Insert → Icons, and a plain SVG insert).
  Those pictures previously took the dropped arm and rendered as nothing at
  all.

  The raster fallback is still carried in the manifest, so a consumer that cannot
  draw SVG still has something to fall back to.

Added

• picture-svg joins the corpus as a primitive-tier deck. The gap above was
  not a subtle mapping error; it was a construct the corpus did not contain, and
  a ledger cannot report on a construct nobody generates. The fidelity ledger now
  carries upstream's image.svg note, which states that the vector part is
  carried and the source's own fallback is regenerated rather than kept.

v0.1.1 — published from CI, with provenance

The library is unchanged from v0.1.0. What changed is how it reaches you.

Changed

• Releases are published from CI, through npm's trusted publishing. The
  workflow mints a short-lived OIDC token that npm exchanges for a credential
  scoped to this package and that one workflow file — no long-lived token exists
  to be stolen — and the package is stamped with provenance, so the npm page
  links the tarball to the commit and the run that built it. 0.1.0 was published
  from a laptop and carries none; this is the first version you can verify that
  way.
• Publishing is triggered by publishing the GitHub release rather than by
  pushing the tag, keeping the deliberate half of a release a human act.

---

Changelog ·
npm ·
Docs

v0.1.0 — first public release

First public release.

pnpm add pptx-html

Run the loop in your browser →

Added

• The loop. importDeck, renderDeck, parseDeck and emitDeck — .pptx
  to a paint model, to HTML, back to the model, back to .pptx. Anything the
  model does not express is carried across as its original XML rather than
  approximated, and anything that can be neither modeled nor carried is warned
  about.
• Invariant R and the oracle that gates it. A generated corpus of 26 decks
  runs the full loop on every CI build, compared under a normalized read model
  rather than byte-for-byte. The per-construct fidelity ledger is snapshotted so
  coverage cannot move without a visible diff.
• The JSON island. renderDeck embeds the model beside the visible SVG and
  parseDeck reads that, never getComputedStyle — which is what lets the
  visible channel approximate freely without any of it reaching the emitted deck.
  A document whose integrity hashes do not match is refused, not downgraded.
• A declared editable surface. project names what a rendered document may
  change and freeze is its complement; both are exported, so the question "what
  may I safely edit in this HTML?" is answerable without reading the renderer.
  It is run text, bold / italic / underline / strike / sizePt /
  color, a paragraph's align / bullet / marginLeftPt / indentPt, and
  deleting a node. A property is in only if a write-API option can author every
  one of its values, including the absent one — everything else is detected as
  drift rather than interpreted.
• Lane reporting. parseDeck states per slide how it read the document
  (exact, reconciled, drifted, heuristic) instead of leaving a caller to
  infer success from the fact that a file appeared.
• A preview drawn from the model, never rasterized. The renderer resolves
  twelve preset geometries exactly from ECMA-376's Annex D definitions and marks
  the rest as an obvious box rather than guessing an outline, and it draws the
  layout's and master's non-placeholder shapes as chrome — visible, and
  editable nowhere, because an edit to it would belong to every slide bound to
  that layout.
• The heuristic lane. convertDeck / convertSlide infer a model from a
  rendered DOM, for HTML this library did not produce. Browser-only,
  explicitly best-effort, and sharing no code path with the loop.
• A public site, at https://shbernal.github.io/pptx-html/, deployed from the
  same workflow that gates the library — the publish hangs off needs: validate,
  so a site whose round-trip oracle is failing cannot go out.
  • A playground that runs importDeck → renderDeck → parseDeck → emitDeck in
    the visitor's browser. The sample decks are the oracle's own corpus rather
    than decks written for the page, editing goes through EDITABLE_SURFACE and
    nothing else, the lane is reported per slide, and the round-tripped .pptx
    is downloadable beside the original. Nothing is uploaded anywhere.
  • The fidelity ledger, generated from the oracle's coverage reporter over
    the corpus, so the page cannot claim more than CI knows. Every row links to
    that deck running in the playground.
  • docs/ published as written. site/docs/ is a generated mirror; the
    tracked design record stays the source of truth, and the generator fails
    rather than warns when the two drift.
• A runnable example. examples/round-trip.mjs runs all four legs, edits a
  run, and re-imports the emitted deck to show the edit arrived. CI runs it,
  because it is the only thing here that imports the package by name and so the
  only check that the exports map works the way a consumer will use it.

Notes on scope

• The guarantee covers decks written by @shbernal/ts-pptx, which is what the
  corpus is made of. Decks authored in PowerPoint are a deliberate second tier
  and are not yet gated.
• There is no rasterizer, on purpose. A flattened slide is the one output that
  can never re-enter the loop — see
  docs/decisions.md.
• ESM only, Node >=24. @shbernal/ts-pptx ^3.2.0 is the one runtime
  dependency.

---

Changelog ·
npm ·
Docs