Each `stash::get` looks at every defined name, so listing the items
with `keys` and then reading each with `get` took time quadratic in
the number of names. A file with 20,000 hidden `_gc_macro_<i>_0`
names took about ten seconds in a debug build, and the macro list
reads its items this way from data that comes from the file.
`stash::items(wb, prefix)` now groups the names by item in one pass
and returns each item's text by key, and `keys` lists from the same
pass. Both return at most `MAX_ITEMS` (10,000) items, the first keys
in sorted order, so a file with any number of hidden names can't make
a listing unbounded. A test counts the names each listing looks at.
The xlsx writer drops characters XML 1.0 can't carry: controls other
than tab, line feed and carriage return, and U+FFFE and U+FFFF. A
stash item holding one lost it on save without a warning, and so did
a JSON item, since serde_json leaves U+FFFE and U+FFFF unescaped: a
stored macro parameter changed silently.
`chunk_formulas` now writes each such character, and the marker
itself, as the noncharacter U+FDD0 followed by four hex digits, and
`get` turns them back for `_gc_` items. A marker that isn't followed
by four hex digits naming a character is read as it is. An escape
counts as five units against the 255-unit literal limit and is never
split across literals, and the guard on the stored size allows the
longest item even when every character is escaped. Legacy `_sc_`
items are read unchanged.
A menu-style call with no parameters, such as `file.open`,
`data.fromTextCsv` or `home.formatCells` with `{}`, only queues its
dialog and succeeds. The recorder kept it as a step with empty
parameters, and then kept the dialog's confirming call as well, so
playing the recording back ran the command twice, or failed on the
empty step when played back headless. Only `execute_headless` took
the step back, and only for the open and save-as dialogs.
`execute` now compares the queued UI requests before and after a
top-level command and does not record one that queued a dialog. The
special case in `execute_headless` goes (it keeps turning a missing
file into an error), and with it `Recorder::truncate`, which had no
other caller.
A command that opened an undo group and then returned an error (a `?`
between `group_begin` and `group_end`) left the group open for good:
Undo and Redo refused with "finish or stop the running action first",
later edits made no undo steps, and nothing could stop the group.
`execute` now rolls back every group a failed command opened, as
`group_rollback` does, restoring the workbook, selection, undo and
redo history to where each group began. Only a command that succeeds
keeps a group it opened, which is what a debugger's start needs.
`Session::abort_groups(doc)` and the menu-less command
`automate.stopAction` roll back the groups left open on a document,
for a debugger the person abandons or a caller that dropped its
token. The command is not undoable and works while a group is open,
and the Undo refusal now names it.
The touch log keeps at most a million writes; past that, a write that kept a
cell's value is missing from the footprint, so a shift derived from the
block of written cells could split a row the macro writes. The footprint
now says when the log filled up (logFull in macro.check), and a shift is
refused then, as it is when the macro deletes sheets or objects. Regression
test with a million-cell fill in a dry run.
Every nested command a dry run executes is journaled, so a long macro pushes
the journal past its limit and the journal drops its oldest entries. The
restore cut the journal back to its old length, which then kept entries from
the dry run and pointed Record Actions at the wrong place. The session now
counts the entries its journal has dropped (journal_dropped) and truncates
with its size bookkeeping (truncate_journal); the restore keeps the old
length less what was dropped meanwhile and moves the recording's start back
by the same count. Regression test with 12,000 nested steps.
Four fixes from a review of the runner, each with a regression test that
fails without it:
- A shift inserts cells over the block of every written cell, which moves
that block's columns (or rows) from there on. When a bound input lay in
the way, its data moved and sheared while the run kept the old coordinates
and reported success. The shift is now refused, naming the input and its
place. The output is not counted: it is where the macro writes, and the
second footprint catches data a shift moved into it.
- A dry run or a refused run left a clean workbook dirty: view.zoom,
sheet.activate and the view modes move the saved mark to the current
workbook on a clean document, and the rollback put the workbook back but
not the mark. Saved keeps and restores it.
- overwrite: true answers the question however ask was given; with an
explicit onConflict: ask the answer was ignored and the question came back
forever.
- The dry runs and the run share one step budget as they share one clock;
each phase counted its own 100,000 steps before. ran reports the run's
own steps.
Input.col("date").data resolved to the whole column, header included, because
every method cleared the header state. col and cols keep it; the other
methods still clear it, and only data keeps the header row for col("…").
The token test asks for D3:D11, cols(2..3).data and col(2).data.col(1).
Two items under one key, in any case, would leave two hidden names of one
name, of which an xlsx file keeps one, and get would mix their chunks. A
caller's mistake, refused before anything changes. Regression test.
Once a step has been dropped, or the step count is full, push serialized
each later step only to count it; a file.open step holds a whole file. The
checks that need no size come first.
builtin_id already wrote the currency formats 5-8 as custom formats
because readers resolve those ids by locale. The accounting formats
with a currency, 42 and 44, are resolved the same way but were written
as bare ids, so LibreOffice showed a 44 cell holding -0.14 as "- 0.14"
where US Excel shows "$ (0.14)". Treat them like 5-8.
Six fixes from a review of the resolver, each with a regression test that
fails without it:
- A bare Row or Column under a general reference key (to, from, at, area,
source, target, destination, location) stays text: Change Spelling's
replacement word and a pivot field's area were being rewritten into the
current row's A1 range. Row.col(2), {{Row}} and {"$ref": "Row"} still
resolve there, and the bare word still resolves under cell and range keys.
- {{{{ in a string is a literal {{, so text that needs the braces can be
written; an unknown token is still an error.
- Input.data on a range that is only a header row, and Name: of a table
with no data rows, are refused instead of reading the header.
- below(0), below(-3) and right(0) are refused like resize(0, 1) is, instead
of becoming 1.
- A defined name on a sheet whose name holds a dollar sign resolves: the
dollar anchors are stripped from the reference part only.
- quote_sheet is the engine's one rule (gridcraft_core::quote_sheet), which
quotes a sheet named like a cell reference, an R1C1 token or a boolean and
leaves a period unquoted.
docs/macros.md says which keys take a bare token and how to write a literal
brace.
A macro could declare an input called Q1, AB12 or Sum1, and a step written
with plain A1 text, as a recorded one is, would then resolve to the input's
range instead of the cell. check_limits refuses such names; the docs say so.
Regression test.
Found while testing crates/functions/src/text.rs:326-340
Co-Authored-By: Claude Haiku 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Session::grouped ran its action without a panic guard, so a caller outside
a command that caught the panic found the group still open: Undo and Redo
disabled, the partial edits kept, and the next grouped action nested in the
abandoned group. The group is now rolled back and the panic resumed.
Regression test.
- Workbook::macros is a shared list (Macros, a copy-on-write newtype): a
workbook copy made for a cell edit or an undo snapshot holds the same list,
and the undo budget charges a list once a snapshot owns its own copy.
Before, every mutating command copied every step of every macro, and a
hundred undo entries kept a hundred copies the budget never saw.
- Saving to xlsx drops the last save's items and writes the macros in one
pass each (stash::set_many), instead of a scan of every name per macro.
- A hidden _gc_macro_<x>_<k> name with a key this build did not write is
kept on open and on save, as a visible one is, instead of being dropped
without a warning.
- macros::check is the one check for every path that adds a macro (JSON, a
file, built in code): the model's limits and the xlsx size limit.
Regression tests for each.
The depth check refused any step list at the fifth level, so a fourth Repeat
or If could not even hold an empty body: three levels nested, not the four
MAX_MACRO_DEPTH and the docs state. The check now counts the Repeat or If
steps themselves, in the model's limits and in the parser's pass over the
JSON. The test asks for four and refuses a fifth.
The architecture note named formulas and validation only; it now lists the
Paste Special kinds and linked pastes that write formulas, and the kinds that
carry nothing.
A group open before a command and ended or rolled back inside it lost its
undo step when the command then panicked: the restored history held no entry
for the group's changes, so Undo skipped the state before the group. The
panic path now pushes the outermost such group's step from its start to the
restored workbook. The group stays ended, because its token went with the
command. Regression test for a group ended and one rolled back.
The inputs paragraph described how a run binds inputs and the message it
shows, which the runner adds in a later pull request. It now says what the
declaration is for and that this version stores and checks it.
Truncating the undo history to its length at the command's start cannot
bring back an entry that a nested edit.undo popped. The guard now keeps the
undo and redo history as they were and puts both back. The regression test
gains the undo-then-panic case.
The source now carries a note and a column width, and the test checks that
the destination received them while its trust stays Local and no cell is
written.
stash::keys reported a key that get read as missing, because only a later
chunk of it existed. keys now counts chunk 0 and legacy names only, as get
does. Regression assertions in the hostile-names test.
file.open with no file succeeds by queuing a dialog, which execute_headless
turns into an error for callers with no UI; the recording kept the step.
Recorder::truncate forgets it and keeps the byte count right. Regression test.
The panic guard restored the workbook and selection but not the history, so
a nested edit made before the panic left a no-op undo step and the redo
history it had cleared stayed lost. The guard now truncates the steps made
during the command and puts the redo history back. Regression test, and the
architecture note says what the guard restores.
`_gc_watch_00` parsed as chunk 0, so a hidden one was read as the item's first
chunk and a visible one stopped `_gc_watch_0` from being written. Only the
chunk number as `set` writes it counts now. Regression test.
Paste Link writes a formula into every cell whatever `what` says, so a linked
paste of formats from a file now makes the destination FromFile as a paste of
formulas does. Column widths and notes write no cells, linked or not, and
still carry nothing. Regression test for the three kinds.
Typed entry, VALUE and arithmetic now read $(1,234), $ (5) and (5)% as
negative numbers, as Excel 365 does. Parentheses no longer nest: (-5),
((5)), $5% and dates in parentheses stay text. This changes main's
((100)) from 100 to text; the no-recursion guard from #2180 is kept and
its test now asserts the new result.
PMT computed (1 + rate)^nper with powf, which rounds 1 + rate first and
drops the low bits of a small rate. Over 400 synthetic probes that left
PMT up to about 1000 ULPs from Excel (59 bit-exact); the ln_1p/exp_m1 form
is within 4 ULPs (244 bit-exact). A loan schedule guarded by
IF(p - SUM(principal) > 0, ...) turns on those bits: in 3000 synthetic
schedules main disagreed with Excel on the number of live periods in 1382,
the fix in 823.
Also follows Excel at the edges: a rate of -1 or less is #NUM!, a huge
nper gives -pv*rate instead of #NUM!, a tiny rate no longer collapses, and
a small negative rate keeps its payment. IPMT, PPMT, CUMIPMT and CUMPRINC
pick this up through pmt_calc. FV and PV are unchanged: Excel's own FV/PV
sit further from the exact value than the current powf form does.
The tests allow the one-ULP libm difference between platforms.
Session::grouped ended its group after the action, and when the action had closed the group's document (file.close, or file.open taking the place of an untouched new workbook) the group had gone with it, so the action was reported as failed although it had done what it was asked. A group that is gone because its document is gone now counts as ended; a group ended by someone else while the document is open is still an error.
The history limit is applied once the outermost command has finished, so the lengths its bookkeeping compares stay true (round 3). That left the steps nested commands make unbounded until then: a script of up to 100,000 steps, or a subtotal with a step per group, kept every one of them alive, each with the cells it retains, and before this branch every push trimmed the history at once.
Each nested command now bounds the steps made since the latest mark on its document to the undo limit and byte budget, dropping the oldest of them. The marks are the starts of the running commands, the starts of the open groups and the end of the last group: the positions a command or group collapses its steps by when it finishes, so nothing before them moves. What the bound drops is only what the final trim would drop.
Two boundaries of the group rules were wrong. A command that ended a group open before it was treated as grouped whatever it did next: with an unchanged group (no step) its later edit had no undo entry at all, and with nested edits after the end the entry was the nested command's. The group's step now stays the group's, and the command makes one step of its own, under its name, for what changed after the end, from the workbook the group ended at.
A command that edited and then opened a group it leaves open (a debugger's start) left its nested entries in the history, so one Undo after the group ended brought back a state from inside the command. Such a group now takes the command's start as its own, and its step covers the command's earlier edits.
The record of where the last group ended now holds the workbook and selection it ended at, the history length then and whether it made a step; it lives until the outermost command finishes. A rollback records where it ended too, unless the document is back exactly where the last recorded step left it.
The check ran once per chunk, each a scan of every defined name: an item of many chunks in a workbook with thousands of names compared every pair. One pass now parses each visible workbook-scope name and refuses when it is a chunk of this key below the chunk count. The behaviour and its test are unchanged.
The header and the revision row said 2026-10-11, older than the row above them. The document trust paragraph now says how DocState::trust (provenance) differs from DocState::scripts_trusted (per-session consent for stored scripts, #2178).
The buffer kept at most 100,000 steps but each step keeps its full parameters, and a file.open step holds the whole file as base64, so a recording of a few opens of a large workbook kept tens of MB alive. Like the journal, the buffer now also stops at 64 MB of steps serialized as JSON. A step that does not fit is counted as dropped, and once a step is dropped every later one is too, so what the buffer holds is always how the recording began, never a sequence with gaps.
prxnvt