The .fig import gave a file one decoded value per byte of its compressed canvas (at least 20M, at
most 128M). Dense legitimate files decode past that, further since every scalar field counts, and
the whole import was refused as "not a valid FigCraft file: .fig file: message: too many values".
The budget is now four values per compressed byte; the floor and both ceilings (128M, 32M on 32-bit
targets) are unchanged, so a small file that unpacks past the floor is still refused and no file
decodes more values than the ceiling allowed before. Generated 21 MB and 15 MB canvases (220,044
and 108,012 layers) and a 49.7 MB one budgeted at the ceiling (530,106 layers) now open whole
without warnings; the 49.7 MB one's test peaks at about 9 GB.
* M8: opening a file contacts no team library server it names until you check it (#1097)
* M8: translate the team library check strings (#1097)
* M8: the Assets panel library test checks what allows a library, without its own fake link (Sonar duplication and coverage)
* M8: cap the people a client keeps track of and the deleted ids a session remembers (#1121)
A hostile host or server could grow a client's presence, seen-peers and who-is-here lists without bound, and a session's deleted-id set only grew. Clients keep at most MAX_PEERS (4,096) others, leaving newcomers out with a warning; max_people is at most that; a session remembers at most 262,144 deleted ids, the lowest forgotten first.
* M8: a max_people above what clients track is lowered with a warning, not refused (#1121)
Refusing it would stop a server whose existing config says more from starting after an upgrade. Config::people() lowers max_people to MAX_PEERS for the core's limits, and the startup warnings name the configured and the used value.
* M8: the live welcome's handling in its own function (Sonar S3776)
* M8: a large selection's overlay and Design panel values are kept for frames where nothing changed (#1571)
The selection's outlines, box, size and slots on the canvas, and the Design panel's instance checks and position values, are kept with the document snapshot (held weakly), the selection and, for the overlay, the view's mapping to the screen and the overlay green, and made again when any changes. Dragging 648 selected layers: overlays 1.99 to 0.86 ms a frame, Design panel 2.54 to 1.40 ms.
* M8: split the position fields into one function per row (Sonar S3776)
The per-address download bucket never held more than its 16 MiB burst, so a larger library or plugin was refused for good. Bucket::take_spread serves one from a full bucket and leaves the debt to refill at the rate before the next download; the rate over time still holds.
The REST import's layer walk (also the .fig import's second stage) took about 4 KB of stack a level, a megabyte at the 256-level cap, and overflowed a 1 MB thread. It now runs on a work list in the same order, fitting groups once their children are in. The .fig mapping stays recursive at about 1.8 KB a level (455 KB at the cap). Both 256-deep imports now run on a 1 MB stack; the cap stays 256 everywhere.
* M8: one depth cap for a whole render; render threads get 8 MiB stacks (#1082)
A pattern fill drew its source as a fresh render from depth 0, so patterns in patterns could stack three renders of 256 levels. The source now goes on from the depth of the layer it fills. A render level takes about 6.5 KB of stack (measured), about 1.7 MB at the cap: the canvas render, preview and save threads and a band pool of their own now get 8 MiB stacks instead of the default 2 MiB.
* M8: SVG and PDF exports carry the depth into pattern sources and embedded renders (#1082)
Each area (document, images, library, versions, branches) keeps its own cap and now also draws from one total per file open, 6 GiB (2 GiB on 32-bit targets), so a file can no longer make the app unpack about 10 GB. Read in that order, the document and the images up to their own cap fit, so a file whose images pass 4 GiB (#1313) keeps as many as before; images left out by the total stay in the file like those past the image cap. What passes the total is skipped (the document refused) with a note.
variable.import and the variables parameter of file.open and file.import read an explicit path
with the uncapped reader, while the sibling variables.json next to a REST file already went
through read_capped with MAX_VARIABLES_FILE. variables_param now uses that capped read for paths
on native builds; the browser keeps its no-file-system error. Inline objects and JSON text, and
access roots, are unchanged. The check reads the size before the file, so it does not bound a
file that grows or is replaced in between. Primary file reads and image.place stay uncapped.
variable.import's help and docs/mcp.md name the limit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: GPT-6.1 Sol <noreply@openai.com>
Co-Authored-By: GPT-6 Astra <noreply@openai.com>
Every pointer move draws about three frames, and the canvas hit-tested the hover on each: a hover where nothing changed since the last (place, modifiers, zoom, file, page, revision, selection, tool and modes) is no longer sent to the tools again. Dragging an auto layout frame laid its contents out again on every move: a frame whose only change is its transform now keeps its own layout (its parent's still runs). ui.inspect reports frameTimes. Measured on a synthetic 1,900-layer file: hover UI pass 0.91 to 0.49 ms a frame and CPU 33% to 22% of a core; normalise while dragging 648 layers 3.8 to 2.1 ms a move.
Variable and collection nodes are left out of the layer tree (variables are read from the node list), so an instance's expansion no longer walks them again for every instance outside the layer cap. A nested instance takes only the override rows addressed below it, found by range, and shares its outer instances' property values and slot contents instead of copying them. Crafted files: 13.8 s -> 21 ms and 40.5 s -> 98 ms; normal files map as before.
Co-authored-by: Brandon Thomas <echelon@gmail.com>
* M9: comment.list lists only the threads that mention a name (#1415)
comment.list takes mentions (a name, any case): only threads where a message, the first or a reply, @mentions it, by the same parser the thread view highlights with and each message's mentions reports; combined with the other filters.
* M9: comment.people lists mentioned names; comment.next and comment.prompt carry mentions (#1415)
comment.people adds every name a message @mentions to those who wrote or reacted (once each in any case). comment.next takes the mentions filter and returns the thread as comment.list lists it; comment.prompt's bounded comments carry each thread's mentions (at most 32, names cut to 80 characters). One mention helper serves them all; builds on #1585.
* M9: comment.people lists people who wrote in two cases once
The participant list is deduplicated in any case too, not only the names added from mentions, in
one pass (CodeRabbit on #1591).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* M8: grid auto-placement on an occupancy bitmap (#1083)
Grid placement tested every cell of a child's span at every position it tried. The taken cells are now a bitmap: spans are checked a 64-bit word at a time, the search jumps past the blocking cell, a row's search ORs the span's rows once, and a column-anchored child looks on from where the last one with the same column and spans found room. Placements are unchanged (checked against the old search on random grids); a crafted 1000 x 1000 grid takes 0.3 ms instead of 192 ms.
* M8: grid occupancy stores rows only once a cell in them is taken
An empty or sparse grid no longer allocates the whole bitmap up front (CodeRabbit on #1593).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Brandon Thomas <echelon@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
A double-click with the pen drew on: its second press added a point on the last one, and a double-click meant to edit a layer drew a stray vector on it. A press on the last anchor now adds nothing; a double-click ends the path (or, when it only started one, edits the layer under it) and Enter ends the path, both going on in vector edit mode with the Move tool. In edit mode the pen's Enter and double-click take the Move tool, the double-click taking back what its first press began. Escape still keeps the pen (#760).
Desktop recovery copies saved DocState::doc, which is the version shown while a version (or a live text preview) is previewed, so a crash during a preview brought the old version back as the unsaved file. What a recovery copy holds is now one shared helper, branches::recovery_copy: the document itself (document_itself(), the same document the unsaved-changes check compares), its history and its branches. The desktop copies and the browser's (#1580) both use it.
bflatastic