pentoolopen source Menu

Release notes

What changed.

Release history synced from GitHub, with repository notes available as a build-time fallback.

November 12, 2026

Pentool v0.28.0

View on GitHub ↗

Pentool v0.28.0

Studio selections now show their real shape and can be painted, photos can be edited in place, and two new agent tools arrive: pentool watch and pentool report. The library is also split into a pentool-core crate that compiles to WebAssembly. No new dependencies, no model, no network; .pen files and existing flags and JSON fields are unchanged (new fields are additive).

New

  • pentool watch DOC: one JSON line per change with the actor and the ids of added, changed and removed objects, for agents that react to each other's edits.
  • pentool report: build cited, charted reports from a JSON spec and CSV, TSV or JSON data files (bar, column, segmented, table, stat tiles, callouts), with numbered citations, a generated Sources page, auto-flow pagination and check --strict. See Reports.
  • Studio selection you can see: every selection result carries a bounded preview coverage grid (at most 256 cells on the long edge). Studio tints the real shape and draws its true outline instead of the bounding box.
  • Studio selection brush (A): paints into the selection with the shared quickmask engine action; Alt or subtract mode removes.
  • Studio Move selection tool: drags the selected pixels; Alt copies.
  • Studio edits photo pixels: selecting an image and switching to a paint or selection tool (or pressing Edit photo pixels) rasterizes it in place, so marquee, wand, brush and Move selection work on photos.
  • Studio shortcuts: Ctrl+C and Ctrl+J copy the selection to a layer, Ctrl+X cuts it, Ctrl+Shift+I inverts it, and Delete erases the selected pixels.

Fixed

  • The rasterizer clamps over-alpha pixels, so a premultiplied pixel can no longer exceed its alpha.

Internal

  • pentool-core crate: the document and scene model, renderer, exporters, transactions, libraries and the collaboration op-log now live in crates/pentool-core, which builds for wasm32-unknown-unknown. The pentool crate keeps the CLI, server, AI, extensions and networked collaboration, and re-exports every core module under its old path. Output bytes are unchanged. See Architecture.

Not included

  • A cutout mask source and --preview, a non-destructive mask from a selection, moving, transforming or nudging the selection outline in Studio, and a live-browser check of the selection brush stroke preview. These stay open in docs/roadmap/v0.28.0/TASKS.md.

Compatibility

Existing flags, JSON fields and .pen files behave as before. The selection preview field is additive. Rust code that depends on the pentool library keeps its module paths.

October 23, 2026

v0.27.0

View on GitHub ↗

Pentool v0.27.0

A focused feature release on selection, cutout and crop, so an agent (or a person in Studio) can take a photo, select a subject, remove the background, crop, and build a mood board. No new dependencies, no model, no network; all algorithms are deterministic. Guide: Selecting and cutting out.

New

  • Wand for photographs: --sample seed averaging, --gradient neighbor-step flooding and --leash on select-wand and fill. A noisy gradient sky selects as one region (about 0.99 IoU on the synthetic test) where tolerance alone leaked or stopped partway.
  • raster DOC select-refine: snaps a rough selection to image edges with a deterministic geodesic path race, --band, --core and --fg/--bg hints.
  • raster DOC cutout ID --id NEW: one step from a box, polygon, wand or stored selection to a new transparent layer, working on raster layers and image nodes without touching the source. --refine, --invert, --shrink, --feather, --decontaminate.
  • Shared dispatcher: POST /api/raster and raster batch now accept select-refine, select-modify, select-save/load/delete, lift, move-pixels, transform-pixels and paste.
  • Selection queries: raster-selection (active or saved), invert, grow, shrink and feather; selection crop --node ID crops one image.
  • Studio: wand sample/gradient, selection modifier buttons, Refine edge, and Copy/Cut to layer in the Raster panel.

Also new

  • Lab wand distance: --perceptual on select-wand, fill and cutout measures colour difference in CIE Lab (so lightness matters more than the largest channel).
  • select-modify smooth-edge: rounds a jagged outline but keeps a soft edge.
  • raster DOC delete-pixels ID: erases the selected pixels (also in batch and /api/raster).
  • Studio canvas crop tool: drag eight handles, aspect presets, Enter/Esc; the pixels stay put and the frame and crop window change as one undoable edit. Works on image nodes with Fit: fill and no rotation or scale. "Delete selected" and "Smooth edge" buttons in the Raster panel.

Not included

  • A cutout mask source, --preview on cutout (use the global --preview), moving or transforming the selection outline in Studio, keyboard nudge, and learned subject detection.
  • cutout is not in the shared dispatcher; it needs the document's folder to bake image nodes.

Compatibility

Existing flags, JSON fields and .pen files behave as before; wand defaults are unchanged. The new decontaminated_pixels, undecided_pixels and similar fields are additive.

October 23, 2026

v0.26.0

View on GitHub ↗

Pentool v0.26.0

A stabilization release: about 70 bug fixes, each with a regression test, plus a small set of CLI conveniences approved with them. No new .pen fields or dependencies. Several commands print less or report more than before; see "Output changes" if you parse their JSON.

Output changes

  • import no longer prints the full fidelity.items array. Without a written --report it prints items_omitted and a report_hint (#204).
  • linked embed and relink results omit the embedded payload (#173).
  • --json error message now carries the full cause chain, not only the outermost context (#150).
  • paint-plan rejects unknown pass args and missing required args at parse time; fill passes now require x and y (#132).
  • paint-palette ignores pixels with alpha below 16, and share is a fraction of the scope (#235, #143). paint-score luma matches the protocol (#133).
  • serve answers unknown routes and wrong methods with JSON problem bodies (#35).
  • package-verify reports files not listed in the manifest as unlisted (#28).
  • image analyze focal suggestion uses pixel centers (#275, #386).
  • paint-diff reports layers_added, layers_removed and journal entry counts (#181). stroke-field accepts the same 1-512 cell range as paint-constraints preview (#180).

New

  • export DOC - with --format streams SVG/PNG/PDF to stdout (#52).
  • batch accepts inline JSON, @file or stdin operations (#89).
  • new takes --background, --name and --if-revision with --overwrite (#48, #387); diff --visual --out (#51); fonts lists embedded faces (#53).
  • info --revision and a history revision field (#54, #83).
  • object set addresses grouped children by ID (#13); asset property set works without --target (#65).
  • scene3d: --dry-run/--if-revision accepted after any subcommand, documented INPUT/ID and camera flags; inspect reports product forms instead of panicking (#213, #148).
  • Studio: layer context menu moves a layer forward or backward (#392); AI generation shows live progress and Studio requests time out instead of freezing.

Fixes

  • Paint protocol: scopes distinguish layer, group and node (#134); constrained strokes record mask hashes (#140); protect_score (#142); empty-mask preview and unequal-size scoring are rejected (#135, #141); id/description changes show in plan diffs (#138).
  • object set validates colors and path data on v3 documents (#19, #18).
  • Errors name the fix: compositing needs migrate --target 6 (#8); missing-file, unsafe-path and ICC errors keep their codes and hints (#207, #73, #355, #241); analyze names a malformed --samples or --query (#42).
  • Studio: stale "unsaved" state, tablet-width panel overlap, mask operations and AI run state fixes (#305, #400, #282, #374).
  • Many more CLI, library, package, photo and documentation corrections; see the commit log since v0.25.0.

Known gaps

Still open: Ed25519 operation signing and stale-base rejection for agents, --cell 0 handling (#180), remaining scene3d flag help (#148), and the issues tracked on GitHub.

October 22, 2026

v0.25.0 — Studio AI

View on GitHub ↗

Pentool v0.25.0

Optional image generation and editing in Studio, plus the accumulated internal 0.24.2 bug fixes. No new document fields or dependencies.

Studio AI

  • Connect and validate providers using credential references; browser forms never receive an API key. Provider settings require the collaboration owner role.
  • Review requests offline, then explicitly confirm provider, model, host, call count, prompt hash and source hashes before contacting a paid model.
  • Generate images, preview candidates, accept into the selected layer, replace a selected image, or discard a run. Runs survive reloads and remain bounded to 16.
  • Accept with revision guards and source/lock checks using the CLI transaction engine. Each acceptance is one undoable document-history step. Selected accepted images show their provider/model provenance in the AI panel.
  • Gemini image variations and background removal; local magenta key-out is offline.
  • Cancel between provider calls and before storing candidates. A request already sent may still be billed and finishes or times out before the worker exits.
  • PENTOOL_AI=off disables operations and hides the AI entry points.

Bug fixes

Rendering and export

  • Parent-page artwork now renders. Items inherited from parent pages previously appeared in parent resolve but were missing from PNG, SVG, PDF, and Studio previews. They now render beneath the page's own layers, respecting overrides, detached items, and hidden parent layers.
  • Hidden objects stay hidden in exports (#396). In v4 documents, nodes marked visible: false and descendants of hidden groups are now excluded from PNG, SVG, and PDF output. Layer visibility already worked.
  • Spot-only composite previews keep their CMYK destination (#397). color separate --composite now retains the configured or explicit CMYK output destination, so spot colors render over process paper correctly.

Editing and document safety

  • Spread facing can be switched on and off. spread <doc> facing <spread> true|false now parses the positional value correctly; it was previously treated as a flag.
  • Negative geometry values work directly (#394). Positional numbers for layer-geometry translate, rotate, scale, transform, hit, move-anchor, and set-handle accept negative values without requiring a -- separator.
  • Batch edits reject invalid paints instead of silently losing them. put-shape no longer drops non-string fill/stroke values, and put-path rejects paint objects with neither fallback nor ref. Invalid requests fail with an explanation and leave the document unchanged. Gradient styles use {"fallback": {...}}.

Server, extensions, and provider setup

  • Server startup errors are accurate (#160). serve checks that the document exists before binding and does not announce that it is listening after a failure. A busy port reports [conflict] with a --port hint.
  • Unreadable resources get consistent diagnostics (#207). import and extension add/run report [missing-resource] for unreadable input files.
  • Extensions validate inputs before execution (#205, #206). extension run rejects an unknown --page with [missing-resource] and non-object --args with [invalid-input] before running the module.
  • Pasted API keys are rejected safely (#357). ai connect --credential-env accepts an environment-variable name, not the key itself. Invalid values return invalid-credential-env, save nothing, and never echo the supplied value.

CLI help corrections

  • layout-grid guide-remove --help explains that a guide ID and --page are mutually exclusive (#169).
  • layout-grid step-repeat --help documents the count range of 1–1000 (#388).
  • raster select-load --help lists the accepted selection modes (#380).
  • shape --help lists the required geometry for each shape kind (#399).
  • benchmark --help documents its layer and object ranges (#350).

These include the fixes tracked internally as 0.24.2; that version was not published separately.

Limitations

Masked inpaint/outpaint and spending budgets remain deferred. Model availability, pricing and support depend on the configured provider. CPU rendering remains normative. Placed native table rendering remains a known gap.

Validation

  • Three-platform CI, Studio tests, browser core and binary budgets.
  • Local verification: 1,731 Rust tests and 49 Studio tests; optimized CLI discovery and PNG/SVG/PDF exports; mock-provider Studio AI and spot-only composite smoke checks; live browser accept/undo and AI-disabled checks.
  • Release archives verified against SHA-256 checksums before publication.
October 20, 2026

v0.24.1

View on GitHub ↗

Pentool v0.24.1

Patch release. No format, CLI, or output changes.

Fixes

  • Studio page thumbnails of a multi-page v3–v5 document all showed the first page. The server's PNG cache keyed on the document JSON, which holds every page whichever is selected, so the first page rendered was served for every page. The cache key now includes the selected page. Only pentool serve used that cache; CLI export and preview were not affected.
  • Studio created new documents at format v11 and offered "Upgrade format" only up to v11, although v12 has shipped. New Studio documents now start at v12.
  • Studio rendered v10–v12 documents on the server instead of in the browser. The in-browser renderer handles them byte-identically, so 3D, bindings and gradient documents get the fast local path that v6–v9 already had. A test now fails if either Studio constant falls behind the newest format.

Known limitation

pentool new still creates a v4 document. Formats v6 and later export SVG as an embedded raster and PDF as raster pages (a nine-page deck: 48 KB vector PDF at v4, 25 MB at v12), so starting the CLI at the newest format would change released output. Vector SVG/PDF export for compositing documents is the prerequisite for making the format invisible.

October 20, 2026

v0.24.0

View on GitHub ↗

Pentool v0.24.0 — Agent previews and a faster CLI

Plan and acceptance criteria: docs/roadmap/v0.24.0/TASKS.md. Exported bytes are unchanged; new flags and JSON fields are additive.

Speed

The agent loop: add a shape, then look at the whole page at 1024 px, ten steps. Milliseconds per step on GitHub's CI runners, against the published v0.23.0 binaries (scripts/bench-agent-loop.py). Every image is byte-identical to v0.23.0's export.

Documentv0.23.0 edit+exportv0.24.0 --previewv0.24.0 session
Linux, 55 layers352129 (2.7x)110 (3.2x)
Windows, 55 layers274142 (1.9x)119 (2.3x)
macOS, 55 layers412216 (1.9x)155 (2.7x)
Linux, small4435 (1.3x)18 (2.4x)
Windows, small5338 (1.4x)22 (2.4x)
macOS, small8579 (1.1x)38 (2.2x)

A single uncached export of a large page is about 1.2x faster (multithreaded blur, gradients and blending); most of the gain comes from fewer processes and reusing unchanged layers.

Changes

  • Agent loop (add a shape, then look at the whole page at 1024 px; Windows, ms per step, v0.23.0 edit+export / v0.24.0 --preview / v0.24.0 session): small document 498 / 361 / 98, 55-layer document 780 / 423 / 300. Every image is byte-identical to v0.23.0's export. scripts/bench-agent-loop.py reproduces it, and CI runs it on all three OSes.
  • --preview on every mutating command and its dry run: one call edits the document and returns a PNG of the changed page, rendered by the normal export renderer at 1024 px on the long edge. The JSON result gains a top-level preview object with path, page, width, height, scale, revision (always the revision the same call wrote) and dry_run. --preview-size PX (1 to 8192) sets the long edge, --preview-out PATH chooses the file, --page chooses the page (otherwise the first page that changed), and PENTOOL_PREVIEW=1 turns previews on for a whole session. Images go to .pentool/preview/ beside the document, which keeps the 32 newest. A failed preview is reported as preview.error and never fails an edit that already committed. Without previews every command prints exactly the same bytes as before; the Studio server never renders CLI previews.
  • Previews zoom in on the change: when the edited or added objects cover less than half the page, the image shows them plus padding at up to 8x and the result adds focus (ids, x, y, width, height) and a 256 px whole-page thumbnail. The focused image is a crop of the normal render at the reported scale, so every pixel matches export --scale. --focus ID (repeatable) picks the objects and --no-focus shows the whole page; an unknown ID is reported as preview.error.
  • pentool preview DOC renders the document as it is on disk, read-only, with the same preview object (no dry_run): the whole page, or the --focus objects with a thumbnail; --page, --max-size PX and --out PATH as above.
  • --annotate draws object IDs and bounds on the preview image (never on exports or the thumbnail) and adds annotations with labels (id, color, box in preview pixels) and truncated. It labels every object on a whole-page preview, or the focused objects, at most 50, and implies --preview.
  • --compare REV|FILE|before renders the same view of a history revision, a .pen file or the state before the change, and adds compare with changed_pixels, changed_pct, changed_bbox and, when pixels changed, a diff image (changed pixels in red). It implies --preview.
  • export --max-size PX picks the export scale so the longest page edge is at most PX pixels (1 to 8192; with --all-pages the largest page decides), instead of the agent guessing --scale. The result adds max_size and the chosen scale. It cannot be combined with --scale, and a bound that would need a scale below 0.1 fails with [limit-exceeded] before anything is written. Exports without it print the same bytes as before.
  • Previews and exports of composite (v6) documents keep finished layer surfaces in .pentool/cache/render-* beside the document, so the next CLI call re-renders only the layers that changed (warm preview of a 55-layer 1200x675 landscape: 335 ms to 209 ms). Output bytes are identical with and without it. Entries are checksummed (a damaged one is a miss), the directory is bounded to 256 MB by least recent use and is replaced when the binary changes. --no-cache or PENTOOL_CACHE=0 turns it off.
  • pentool session --jsonl runs one command per stdin line (a JSON array of arguments, or {"id", "args"}) and prints one JSON reply per line (ok, result or text, error, and the id), keeping fonts and layer surfaces warm between lines. Commands use the same transaction, lock and revision-guard path as separate processes and write identical documents. On the 55-layer landscape (fast build, Windows): tree 4 ms per line against 27 ms per process; edit plus whole-page preview 89 ms per line.
  • Gradient fills, blur effects and layer blending run on up to 16 threads (std::thread::scope, no new dependency). Every row is computed by the same arithmetic as on one thread, so exports are byte-identical for any thread count; PENTOOL_THREADS=N sets it and 1 forces one thread. Uncached export of the 55-layer landscape at scale 2: 1315 ms to 1064 ms.
  • Faster start for renders. Layers without text no longer build a font database, and text layers share one instead of rebuilding it per layer (in-process time of a scale-0.1 landscape export: about 50 ms to 37 ms). On Windows the installed-font list is kept in the per-user cache (%LOCALAPPDATA%\pentool\fonts.bin) and used only while the listing of font files in the font folders is unchanged (other files there are not even looked at): loading it takes about 5 ms against 27–77 ms for a warm scan and about 340 ms for a cold one. Linux and macOS keep scanning: Linux font folders come from fontconfig's configuration, and on macOS checking the system asset folders cost more than the scan in CI. Output bytes are unchanged.

Full Changelog: https://github.com/agkomyint/pentoolgg/compare/v0.23.0...v0.24.0

October 19, 2026

v0.23.0

View on GitHub ↗

Pentool v0.23.0 — Fast Studio on heavy pages

No format, CLI or dependency changes. Rendered output is byte-for-byte unchanged: 212 browser-core renders of 106 fixtures (two scales each, with the new cache cold and warm) and the native PNG export comparison match v0.22.0 exactly.

Speed

Measured on one machine with the same document: a 55-layer landscape of blurred cloud and hill shapes, rendered by the browser core (WebAssembly).

Renderv0.22.0v0.23.0Faster
First render, 0.64 scale~1150 ms~390 ms~3x
Re-render after moving one layer, 0.64 scale~900 ms~100–125 ms~8x
First render, 1.0 scale~2800 ms~850 ms~3.3x
Re-render after moving one layer, 1.0 scale~2400 ms~250–300 ms~9x

Dragging a selected object no longer waits for a render at all (see below).

Changes

  • Studio: the local (WebAssembly) preview now renders in a background worker. Previously each render ran on the page's main thread, so on a heavy page (for example many blurred layers) every edit or drag froze clicks, scrolling and typing for seconds until the preview finished. A newer canvas render now also replaces an older one that has not started yet, so fast edits do not queue up stale previews, and the page on screen renders ahead of thumbnails.
  • Studio: dragging an object no longer waits for a render. Selecting an object prepares the page below it, the object, and the page above it at the displayed size; a drag only slides the object between the other two, so it follows the pointer at once whatever the number of layers. Parts that did not change (the page around an object that was just moved) are reused.
  • Browser core: finished shape layers (fills, paths, rectangles, ellipses, lines, including their blur effects) are kept between renders, up to 96 MiB. An edit re-renders only the layers that changed. Resource limits are accounted exactly as if every layer were rendered again, so the same documents succeed or fail with the same errors.
  • Browser core: built with WebAssembly 128-bit SIMD, which every current browser supports. Results are identical lane by lane.
  • Blur effects render several times faster: the blur runs over rows in memory order with two buffers instead of copying the whole image for every channel and pass. A blur at full opacity now takes the blurred pixels directly.
  • Gradient fills render several times faster: the gradient geometry is read once per fill instead of once per sample, and only the pixels the fill can cover are visited.
  • Fully transparent and fully opaque pixels skip premultiplication arithmetic.
October 19, 2026

v0.22.0 — Layers and live documents

View on GitHub ↗

Layers in Studio now work the way Photoshop users expect: real adjustment layers with proper controls, clipping masks, layer masks, blend modes, and copy and paste of all of it between documents. An open Studio tab also follows the file live, shows who changed what, and keeps a version history you can restore from.

No new .pen fields, CLI commands or dependencies. Version 6 validation accepts one more stack shape (a clipped adjustment inside a clipping group, see Engine); documents that use it do not validate in 0.21.x and earlier.

Highlights

  • Adjustment layers as real layers. Add any of 13 adjustments from the Layers panel, edit them with sliders, curves and gradients, clip them to one layer or let them affect everything below, and copy them into another image.
  • Clipping masks like Photoshop. Alt-click or Alt-drop to clip; dragging a layer into or out of a clipping group joins or releases it.
  • Layer masks. Mask a layer with a shape and edit density, feather and invert.
  • Live documents. Changes from the CLI, an agent, another tab or a collaborator appear in an open tab without reloading, and your unsaved edits are kept.
  • Version history. See recent editors, preview any stored version and restore it as an undoable step.
  • Photo to canvas. A Photos strip in the Layers dock places photos, with their edits, by drag and drop.

Studio

Layers, adjustment layers and masks

  • Layers panel: per-kind icons, an eye on every row, a clip arrow and indent for clipped layers, an underlined clipping base, an fx badge and a mask chip. A bar above the list sets the selected layer's blend mode (16 modes, plus Pass through for groups), Opacity and Fill. A toolbar below it adds an adjustment layer, clips, groups, duplicates and deletes. The right-click menu adds Create/Release clipping mask, Mask with selected shape, Invert mask and Delete mask.
  • New adjustment layer (13 kinds: levels, curves, brightness/contrast, hue/saturation, vibrance, exposure, color balance, black & white, channel mixer, gradient map, invert, posterize, threshold) inserts directly above the selected layer, outside its clipping group, and is selected. A version 4 or 5 document is upgraded to version 6 first, as one undoable step.
  • Properties shows real controls for each kind instead of a JSON field: sliders, a curves editor (click to add, drag, double-click to remove), gradient stops and channel mixer outputs, with live preview and one undo step per change. "Affects" chooses all layers below or only the layer below (clipped). A Mask section edits density, feather, invert and link.
  • Clipping works as in Photoshop: Alt-click a row, or Alt-drop a layer onto another, to clip it to the layer below. Releasing a clipped layer releases the ones above it. Dragging a layer into a clipping group joins it, dragging one out releases it, and moving a base re-clips its layers to the layer now below them, so reordering never leaves an invalid stack.
  • Copy and paste keep adjustment and clipped layers working across documents: pasted IDs are remapped together with the scopes and clipping that refer to them, referenced images and mask resources travel with the copy, and pasting into a version 4 or 5 document upgrades it instead of failing at save. Adjustments paste in place without the usual offset. Duplicating a clipping base keeps its clipped layers attached to the original.

Live document, presence and history

  • An open Studio tab updates by itself when the served file changes, whether the edit came from the CLI, an agent, another tab or a collaboration peer. Unsaved edits in the tab are merged with the incoming change instead of being overwritten; there is no "Changed on disk" banner to dismiss.
  • Remote changes are pointed out on the canvas with the editor's name and colour, a short activity toast, and avatars of recent editors in the header.
  • A Version history panel lists recent changes with who made them, previews any stored version, and restores it as a new step that can be undone.

Photo and pages

  • An "Add to canvas" button in the Photo toolbar and panel places the selected photos, with their edits, on the active layer and opens Design. Several photos are placed with a small offset; when the button is disabled its tooltip says why. Later Photo edits update the placed photo.
  • The Layers dock has a Photos tab with a thumbnail strip of the catalog. Drag thumbnails onto the canvas to place them where they are dropped, or onto a page thumbnail to place them in the middle of that page. Double-click a thumbnail to add it to the current page.
  • The page strip at the bottom has a compact mode (the chevron on the right) that shows page name chips instead of thumbnails. Studio remembers the choice.

Engine

  • A clipped adjustment (scope targets naming only the clipping base) no longer ends the clipping group, so layers above it can stay clipped to the same base, as in Photoshop. Hiding a clipped adjustment no longer releases the layers clipped above it at render time.

Server

  • New GET /api/document/events server-sent event stream that reports the shared document's revision on connect and on every change. Works with or without --collab (view access required when collaboration is on).
  • New GET /api/history/version and POST /api/history/restore for previewing and restoring stored versions. See docs/PROTOCOL.md.
  • Error responses (400 {"error": ...}) now include the cause chain, so a failed /api/scene batch says why, for example operation 0 failed (page default, layer -, group -, object base): [missing-mask] mask nope missing instead of only the operation context.

Fixes

  • Photo: placing a photo on a page that had not been saved yet failed with [missing-resource] page page-N not found. Studio now saves the canvas first.
  • Photo: a project with no photos yet opens with an empty grid instead of [missing-resource] the document has no photography catalog, so the first JPEG, PNG or DNG can be imported from Studio.
  • Studio: reordering a clipped layer or a clipping base could leave a stack that failed to save with invalid-clipping.
  • tree, search and geometry output round bounds to 3 decimals, so values such as 28.799999999999997 are reported as 28.8.
  • paint-plan - refuses stdin with a clear [invalid-request] message instead of a confusing missing-file error.

Not in this release

  • Painting directly on a layer mask (masks come from a shape, an alpha or an image), one combined layer stack per page, and the healing fix described in docs/roadmap/architecture/layer-model.md need a format or output decision and are planned for a later minor release.

Full Changelog: https://github.com/agkomyint/pentoolgg/compare/v0.21.1...v0.22.0

October 19, 2026

v0.21.1

View on GitHub ↗

Patch release. No format, CLI, or output changes.

Performance

Faster compositing of pages with many objects. Studio previews and PNG/PDF exports render roughly three to four times faster, and output bytes are unchanged.

Page (Windows x64 release binary)768 px before → after2048 px before → after
50 rectangles323 → 143 ms1955 → 535 ms
50 text objects296 → 93 ms1891 → 419 ms
10 text objects70 → 23 ms416 → 120 ms
  • The rasterized node surface is converted to straight alpha in place; transparent and opaque pixels, which convert to themselves, are skipped.
  • Blending skips pixels where the source is fully transparent under separable sRGB modes, and full-opacity layers skip the opacity pass.
  • Exhaustive unit tests pin both shortcuts to the general path, and benchmark renders were compared byte for byte against v0.20.1.
October 18, 2026

v0.21.0

View on GitHub ↗

A Studio usability release: selection, direct manipulation, and navigation now work the way designers expect from desktop editors. It also carries every fix from the unreleased 0.20.1 patch (docs/release-v0.20.1.md). No .pen format, CLI, or dependency changes; Studio edits go through the existing server routes.

Studio

  • Multi-selection. Shift-click adds or removes objects, Ctrl/Cmd+A selects everything on the page, and a selection of several objects moves, resizes, aligns, and groups together.
  • Accurate picking. Clicks hit the drawn shape instead of its bounding box. Clicking selects the outermost group first; double-click steps into a group. Hovering outlines the object under the pointer.
  • Smart guides. Dragging snaps to the edges and centers of nearby objects and the page, with a badge showing position and size.
  • Context menu and clipboard. Right-click the canvas for common commands; copy, cut, paste, and duplicate objects with the usual shortcuts.
  • Command palette. Ctrl/Cmd+K searches every command; ? opens the shortcut sheet.
  • Layers. Drag to restack, nest into a group, or move to another layer; double-click to rename; hovering a row highlights the object on the canvas and the reverse. Locked layers and legacy documents keep their existing rules.
  • Text in place. Double-click a text object to edit it directly on the canvas.
  • Start screen. Canvas presets grouped as Social, Screen, and Print, a custom size with an orientation swap, a thumbnail card to continue the open project, and drag-and-drop to open a .pen file.

Fixes

  • Bursts of previews (page strip, thumbnails, zoom) no longer fail with [limit-exceeded] preview workers busy: Studio keeps its requests under the server's worker limit and retries briefly when the workers are busy.
  • The examples/v015-glass-crystal model and material files keep their line endings in Windows checkouts, so the shipped example's pinned SHA-256 matches and it exports on Windows.

Notes

  • Server previews of pages with many objects are slower than single-object pages because each object is composited at page size; a bounded-region renderer is planned. Run the release binary, not a debug build, for interactive use.
  • In-browser ("Local render") previews still cover format v6-v9 documents only; newer documents render on the server.

Full Changelog: https://github.com/agkomyint/pentoolgg/compare/v0.20.0...v0.21.0

October 17, 2026

v0.19.1

View on GitHub ↗

What's Changed

Full Changelog: https://github.com/agkomyint/pentoolgg/compare/v0.19.0...v0.19.1

October 16, 2026

Pentool v0.16.1

View on GitHub ↗

Bug fixes

  • Restore browser raster rendering in WebAssembly builds and embed the WASM renderer in release binaries.
  • Align the live raster stroke preview with committed paint, including transformed raster layers, brush shape, spacing, and stabilized stroke endpoints.
  • Fix wasm32 compilation for native-only scene cache and cancellation helpers.

Verification

  • Rust test suite passed on Linux, Windows, and macOS.
  • Release workflow raster paint smoke tests passed on all three platforms.
  • WebAssembly raster render smoke test passed.

Full changelog: https://github.com/agkomyint/pentoolgg/compare/v0.16.0...v0.16.1

October 16, 2026

Pentool 0.16.0

View on GitHub ↗

Pentool 0.16.0 — mockup bindings, UV tools, and packaging (draft/reference)

New

  • Live material bindings: 3D materials can bind to .pen pages and nodes; editing the source updates dependent mockups, with bounded deterministic render caching and structured diagnostics (missing source/slot, cycles) that leave documents unchanged.
  • Mockup presets and editable product forms: 8 parametric starters (including cylinder-container) plus UV island layout, projection and surface wrapping.
  • Product-design / mockup CLI with stable JSON reports, bounded batch rendering, and 3D carried through color and print (PDF/X draft, separations).
  • Packaging: parametric dielines (straight/reverse-tuck cartons, sleeve, mailer, pouch, tube and bottle labels), flat and folded views kept in sync, panel-linked artwork, pack CLI (inspect, preflight, preview, export, export-bundle, verify-bundle), and a Studio packaging workspace.
  • .pen format v11 with schema (docs/pen-format-v11.schema.json) and pentool migrate --target 11.

Fixed

  • Straight-tuck carton dust flaps no longer collide with the lid in the flat and fully folded states.
  • scene3d uv on a dieline preview now refuses with a clear error instead of panicking.

Known limitations (please read)

  • Packaging output is draft / reference only. It is not manufacturing-certified: no score compression, cutter conformance, or structural validation. Confirm with your printer before production.
  • Straight-tuck dust flaps still overlap the lid at mid-fold (fold 0.5), and the mailer lid overlaps its glue joint when folded.
  • Bottle-label wraps do not conform to tapered bottles; pouch geometry is a simplified visual mockup.
  • The rebuilt Studio packaging workspace had no live browser QA on this build.
  • Roadmap tasks 8, 8a, 9, 11, 13 and 14 remain open (see docs/roadmap/v0.16.0/TASKS.md).

Verification

CI green on Linux, Windows and Apple Silicon macOS for the tagged commit; local Windows gate (fmt, clippy -D warnings, all tests, gpu check, release build, packaging smoke) passed. Checksums in SHA256SUMS.txt.

  • Folded 3D preview (pack preview --fold) can map bound node artwork onto a panel as a squeezed strip; the flat SVG/PNG export is correct. Planned fix in a patch release.
October 13, 2026

Pentool v0.14.0

View on GitHub ↗

Pentool v0.14.0

Professional page layout. Documents with layout use .pen format v9; earlier versions keep working and migrate explicitly. Existing CLI flags and JSON fields keep their behavior.

Layout engine

  • Spreads, parents, sections and page numbering; stories with threaded frames; a shaping engine and paragraph composer; paragraph, character, object, table and cell styles with regex styles; baseline grids, guides and alignment.
  • Text wrap and anchored objects; tables with merged cells and repeating header rows; TOC, index, cross-references, links and bookmarks.
  • .penbook books with continuous numbering, synchronization, book-wide TOC and index, preflight and packaging; data merge and adaptive layout.

Interchange and output

  • IDML import and export (features without an exact equivalent fail the export unless --allow-approximation); tagged PDF and EPUB 3 (reflowable and fixed); layout preflight.

Studio

  • Layout workspace: pages panel, story editor, frame threading, tables, wrap/anchor controls, styles of all five kinds with property editing, TOC, index and links, preflight, overlays (frames, threads, overset, margins, columns, baseline grid, guides, hidden characters), loaded-cursor placement of text files, and an opt-in InDesign key set. Every edit is one POST /api/layout op through the Rust planners and one undo step.
  • Local WASM preview core and a queued, non-blocking paint pipeline with immediate stroke feedback.

Fixes

  • Pages whose canvas size is stored as a float (as IDML import writes it) now render and export.
  • JSON table actions use the same defaults as the CLI.

Known limits

  • Recomposition is whole-document (616 ms cold on 500 pages, ~20 ms cached); per-story-range incremental recomposition is not implemented.
  • Loaded-cursor placement of images and on-canvas anchor handles are not built.
  • Golden composition on all four targets, fuzz targets for IDML/EPUB/tables, cancellation at every import/compose/export stage and book crash-recovery tests are not complete (see docs/roadmap/v0.14.0/TASKS.md).
October 11, 2026

Pentool v0.12.1

View on GitHub ↗

Pentool v0.12.1

A Studio-focused patch release on top of v0.12.0. The .pen format is unchanged (still up to version 8); no existing document or CLI flag changes behavior.

Studio

  • The Classic editor is retired. Photo and raster paint panels now live in Pentool Studio, the only browser UI. /classic returns 404.
  • New documents start at the newest format (v8), so images, compositing, photo and print work without a manual upgrade.
  • The header shows the document format and an Upgrade button (POST /api/migrate, the same migrations as pentool migrate). Upgrading is undoable and nothing is written until Save.
  • Paint has an Add paint layer button (POST /api/raster action add), so a new document can be painted without the CLI.
  • Moving an object follows the pointer immediately: the object and the page behind it are server-rendered once on grab and the object moves as a layer.

Server

  • Identical preview renders are answered from a bounded, digest-keyed cache. Exports and benchmarks are unaffected.
  • The edit-to-preview debounce dropped from 120 ms to 30 ms.

Build

  • cargo build --profile fast is a local-testing profile with parallel codegen and incremental builds. Release artifacts keep the release profile.
October 9, 2026

Pentool v0.11.1

View on GitHub ↗

Pentool v0.11.1

Bug fixes for v0.11.0, found by developing a real camera DNG (colorchart.dng from syoyo/tinydng: Canon EOS 550D, Magic Lantern). No format changes; documents written by v0.11.0 open unchanged.

Fixes

  • raw add no longer refuses DNGs whose BaselineExposure (tag 50730) is the rational 0/0. Magic Lantern writes the tag this way when it is unset. The tag is now read as absent, and the result carries a warnings list naming it. A zero denominator in any tag the develop depends on, such as ColorMatrix, is still refused.
  • Page export no longer develops each photo node from scratch on every render. Page renditions now go through the photo cache (.pentool/cache/photo/), keyed by the develop settings and size, so nodes that show the same variant share one develop and repeat exports reuse it. The cached output is byte-identical to a fresh render. With three nodes of the 1734×1156 test DNG, export took 2.8 s before this change, 0.9 s on the first run and 0.3 s on repeat runs.
  • photo place DOC PHOTO --layer LAYER [--variant ID] [--id ID] [--x --y] [--width --height] [--fit] [--position X,Y] [--dry-run] [--if-revision] adds a photo node through a transaction. Version 0.11.0 documented photo nodes but gave no CLI to create one. Missing size comes from the developed size, keeping the aspect ratio. Missing photos or variants, missing layers, locked layers and IDs already in use are refused, and the document is left unchanged.
  • Clearer errors:
    • a grain effect without a seed suggests {"amount": 25, "seed": 1};
    • a boolean monochrome points to monochrome.enabled;
    • crop given inside geometry points to the top-level crop group.

Documentation

  • docs/photography-v1.md no longer lists photo import or photo local add, which were never implemented. Imports go through raw add, and local adjustments are made with raw develop --set 'local=[...]', for which there is a worked example.
  • The CLI section now says which commands take PHOTO --variant, which take positional selections, and which take PHOTO/VARIANT.
  • The monochrome, crop and grain seed rows now match validation.
  • The page placement section covers photo place and the rendition cache.

Known limitation

Pentool trusts the black level a DNG declares. In the reported file, the declared level is 2056, but the masked optical-black pixels measure about 2047–2050. This crushes red in deep shadows, so the darkest patches show a green/cyan cast. Estimating the black level from masked areas would be a new feature, so it is not part of a patch release.

October 9, 2026

Pentool v0.11.0

View on GitHub ↗

Professional photography and RAW development. Pentool now develops DNG RAW files non-destructively in a bounded, deterministic Rust engine, organizes and synchronizes a shoot, and delivers it through output recipes, from the CLI or the new Photos studio in the browser editor. The specification is docs/photography-v1.md; the roadmap is docs/roadmap/v0.11.0/TASKS.md.

Format

  • .pen format version 7 adds a photography catalog (sources, photos, variants, snapshots, recipes, collections) and a photo page node. Schema: docs/pen-format-v7.schema.json.
  • The first photo command upgrades a document to v7 in the same undoable transaction. pentool migrate --target 7 upgrades v1–v6 losslessly; --target 6 is refused while photo content exists. Older builds refuse v7 documents clearly and without modifying them.
  • Existing sRGB documents keep their appearance.

RAW and color

  • Bounded DNG ingestion with pentool's own decoder: CFA (RGGB, BGGR, GRBG, GBRG) and LinearRaw, 8–16-bit integer and 16/24/32-bit float samples, uncompressed, lossless JPEG and deflate, strips or tiles, and opcode lists. Other RAW containers (CR2, NEF, ARW, X-Trans and so on) are refused with [unsupported-capability] and a suggestion to convert to DNG. Sources stay immutable and content-addressed, embedded or by verified relative path (raw add, raw info).
  • 16-bit storage and an f32 scene-linear working representation with no accidental 8-bit or sRGB round trip.
  • Camera profiles (embedded, .dcp, dual-illuminant interpolation) and white balance by as-shot, temperature/tint, sampled neutral or suggestion (photo profile add, raw develop).
  • Wide-gamut and HDR RGB delivery: sRGB, Display P3, Adobe RGB, Rec. 2020 and ProPhoto, 8 or 16 bits, perceptual or relative-colorimetric mapping, tagged PNG with ICC/cICP, HDR tone mapping and inspection (photo render, photo inspect).

Development

  • Lens and geometry: distortion, vignetting, chromatic aberration, perspective (upright), leveling, rotation and constrained crop, with lens profiles.
  • The development stack: exposure, auto tone, highlights/shadows, whites/blacks, curves, clarity, texture, dehaze, vibrance/saturation, HSL, color grading, monochrome, grain and vignette.
  • Detail: defective pixels, luminance and color noise reduction, moiré reduction, defringe and capture sharpening.
  • Local adjustments with linear, radial, range and painted (photo mask paint) masks. Depth masks are refused until a depth source exists.
  • HDR and panorama merges into derived, non-destructive DNG sources with alignment, deghosting, projection, seam blending, attribution and cancellation (photo merge-hdr, photo merge-pano, optional --external).

Workflow and delivery

  • Variants (virtual copies), snapshots and settings sync with included or excluded groups (photo variant, photo snapshot, photo settings sync).
  • Local organization without a database: ratings, picks, labels, keywords, stacks, collections, a query language, contact sheets and compare (photo search, photo rate, photo keyword, photo stack, photo collection, photo contact-sheet, photo compare).
  • Metadata privacy: a deliberate EXIF/IPTC/XMP subset and export policies for copyright, keywords, GPS, identity, serials and timestamps (photo metadata, photo privacy-report, photo describe, --metadata), and a privacy warning from package pack.
  • Output recipes and batch delivery: built-in web-gallery, social, archive-master and photo-lab, plus document recipes; sizes, print dimensions, frames, RGB space, depth, PNG/JPEG/TIFF, naming and output sharpening. Exports are staged and leave nothing behind on failure or cancellation (photo export, photo recipe).
  • The Photos studio in the browser editor: grid and filmstrip, loupe, compare, culling keys, histogram with clipping warnings, clipping/gamut/mask overlays, development panels, crop, radial/linear/brush masks, copy/sync, variants, snapshots and keywords. Every preview is a Rust render and every edit is a revision-guarded, undoable transaction.
  • photo nodes are drawn on pages: page export (PNG, SVG, PDF) and the editor canvas show the developed variant as sRGB, placed with fit and position. Each render develops the photo at full size, so pages holding many large photos redraw slowly in the editor.
  • A disposable preview cache in .pentool/cache/photo/ (photo preview, photo cache clear, serve --no-cache, PENTOOL_PHOTO_CACHE_BYTES). It never enters packages and page exports never read it.

Quality

  • tests/v0110_conformance.rs: hostile DNG, profile, lens and PNG mutation sweeps; pinned develop and merge digests; rollback of about thirty failing commands; the cache contract; photo-node page rendering; and offline reproduction of every export from an installed package.
  • Cancellation of decode, merge and export is tested at every checkpoint.
  • Fixed before release: a hostile .dcp whose table sizes overflowed could panic; it is now refused with [limit-exceeded].
  • pentool benchmark --photo measures ingest, culling, development, previews and batch export. Measurements are in docs/performance.md.

Building

Cargo.toml now declares rust-version = "1.85", the oldest toolchain the release is built and tested with. CI checks Clippy on current stable.

Out of scope

CMYK/Lab, spot inks, separations, soft proofing for print and PDF/X are planned for v0.12.0. HDR display preview in the editor is not included.

Full Changelog: https://github.com/agkomyint/pentoolgg/compare/v0.10.2...v0.11.0

October 8, 2026

v0.10.2

View on GitHub ↗

Pentool v0.10.2

Bug-fix release for the v0.9.0 audit (PEN-090-001..018) and the v0.10.0 audit (PEN-010-001..002). It also carries the fixes prepared as 0.10.1, which was never published. No new features and no .pen format change. Behaviour that was previously unsafe now fails with a structured error and leaves the document byte-for-byte unchanged.

Browser editor and docs

  • The Layers panel no longer throws on v6 scene layers (no paths/texts arrays), so raster documents load and the Raster Canvas panel lists their layers (PEN-010-001; regression test tests/web/layers-v6.test.mjs, run in CI).
  • The v0.10.0 roadmap status now records the shipped release (PEN-010-002).

Document safety

  • pentool new refuses to replace an existing file; --overwrite records the previous document in history so pentool undo restores it (PEN-090-014).
  • Canvas width/height must be integers in 1..=16384 for new, canvas and page add (PEN-090-015).
  • Invalid canvas backgrounds, fill/stroke fallback colors (grammar: none, #RGB, #RGBA, #RRGGBB, #RRGGBBAA) and blank layer/object IDs are rejected when a v4+ command would introduce them; already-invalid documents stay editable (PEN-090-016, PEN-090-017).
  • export help and errors name PDF (PEN-090-018).

AI tooling

  • Endpoints reject credential-like query parameters and fragments and never echo them (001). PENTOOL_AI=off hides ai from --help (002). Capability errors no longer recommend a provider that lacks the capability (003). --dry-run enforces PENTOOL_AI_ALLOW_HOSTS (004). Opaque background-removal results fail with keyout-failed and coverage is recorded (005). Duplicate candidates are reported (requested/unique/warnings) and all usage is kept (006). Windows credential files must be under the user profile and outside the project (007). Oversized credential files keep code limit-exceeded with a fix (008). Malformed limit variables fail closed (009). --replace rejects stale sources (--allow-stale-source overrides) (010), runs are scoped per document (011), --replace honors layer/node locks (012), and --default-model provider/model is accepted (013).

Runs created by v0.10.0 and earlier in the shared .pentool/ai-runs/ folder are not listed by v0.10.2; regenerate if you still need them.

October 8, 2026

v0.10.0

View on GitHub ↗

Raster paint (v0.10.0)

Deterministic raster editing inside .pen v6: kind:"raster" layers with content-addressed 256-px tiles, replayable stroke journals and checkpoints.

  • Brush engine (hard/soft/pixel/calligraphic/textured), pen/mouse/touch input normalization, presets
  • Eraser, smudge, blur, sharpen, dodge/burn, clone stamp (incl. other layers, below, composite), healing and patch healing
  • Flood fill and magic wand, marquee/lasso/quick-mask selections, layer operations (trim, resize, merge down, rotate, flip)
  • Editor: POST /api/raster, Raster canvas panel, rulers, guides, snapping, navigator
  • CLI: every operation plus pentool raster DOC batch OPS.json; fuzz, golden and benchmark coverage

Known limits: per-stroke cost grows on very long sessions (see docs/performance.md); tilt/twist capture and animated selection edges are not in the editor yet.

Full Changelog: https://github.com/agkomyint/pentoolgg/compare/v0.9.0...v0.10.0

October 7, 2026

v0.9.0

View on GitHub ↗

Optional BYOK image models (pentool ai)

  • Zero-touch setup for agents: ai setup --from-env, connect, doctor [--check], resolve. Keys are stored as references, never values; raw --api-key is rejected.
  • ai generate (Gemini, OpenAI-compatible) and ai edit / background removal (Gemini). Every model call needs --allow-model-call; --dry-run never connects.
  • Reviewable runs: ai run list|show|accept|discard, with --replace to swap pixels of an existing image node and an ai provenance field. Prompts and keys are never written to the document.
  • ai keyout: offline flat-colour background removal, no model.
  • Nothing runs unless you configure it; PENTOOL_AI=off disables it. See docs/ai.md.

Fixes

  • B16: effects budget on large canvases.
  • Centred text now anchors inside its box (Text.width).
  • import on v4+ documents gives a clear unsupported-version error.

Not in this release

Inpaint/outpaint with masks, recipes/regenerate, editor UI, AI SDK bridge and external adapters, OS keyring, spend cap, catalog refresh.

October 7, 2026

v0.8.1

View on GitHub ↗

Pentool 0.8.1 — bug-fix release for 0.8.0

  • Tokens: design-token changes now apply to PNG/SVG/PDF exports of documents that contain images (B01).
  • Batch: batch --upsert / "mode":"replace" replaces existing IDs for put-shape, put-text, put-path, put-image (B02).
  • Images: image add --mask and --mask-fill-rule (B03).
  • PDF: v6 documents keep a selectable text layer (B04).
  • Performance: adjustments cost scales with the target node, not the canvas; faster v6 compositing (B05, B14).
  • Effects/masks: effects render on masked images; preset paste keeps the target's mask (B06, B07).
  • Gradients: default geometry spans the node; out-of-node geometry is rejected (B08).
  • CLI: compact linked report (B09), gradient-overlay defaults (B10), clearer preset save error (B11), effect list without an effect ID (B12).
  • Docs: new docs/params-reference.md; unknown-parameter errors list valid keys (B13).

Full changelog: https://github.com/agkomyint/pentoolgg/compare/v0.8.0...v0.8.1

October 6, 2026

Pentool 0.8.0

View on GitHub ↗

Pentool 0.8.0 adds professional, deterministic compositing workflows while preserving the single-binary, offline-first editor model.

Highlights:

  • Layer blend modes, clipping stacks, isolated/pass-through groups, and linear-light compositing.
  • Reusable masks, non-destructive adjustments, effects, transforms, gradients, patterns, and appearance presets.
  • Selection, histogram, linked-asset collection, inspection, packaging, and composite benchmarking workflows.
  • Browser editor panels for compositing and image import, including bare pentool serve documents.
  • Development .pen v6 schema, deterministic fixtures, and a two-page photographic campaign example.
  • Backward-compatible v1-v5 loading and unchanged released v4/v5 contracts.

Every archive was built and smoke-tested on its target platform. Verify downloads with SHA256SUMS.txt.

Full changelog: https://github.com/agkomyint/pentoolgg/compare/v0.7.2...v0.8.0