[SuruchBoss]

What this changes

Closes #227. Queued right after #222, before #207.

The bug: a constant error cell, which Excel writes as <c t="e"><v>#DIV/0!</v></c> with no formula, opened blank, and nothing said so. cellValueToRaw in excelIO.ts had no branch for ExcelJS's { error } value, so it fell through to "". A total over it then came out as a clean number: =SUM(A1:A3) over #DIV/0!, #N/A, 5 read 5.

The fix (the minimum #227 asks for, until #226):

• Opening keeps the code. cellValueToRaw returns the error's code. The grid holds it as text for now, so it is visible and nothing is lost; it becomes a real error value once An error code typed, pasted or imported into a cell stays text, so formulas that read it go silently wrong #226 lands.
• The open report says so (บอกผู้ใช้ตอนเปิดไฟล์ ว่าไฟล์นี้มีอะไรที่แอปเก็บไว้ไม่ได้ · Tell the user, on open, what in this file will not survive a save #83). A new kind, errorValues, counts constant error cells per sheet, read from the package like every other kind. A formula whose result is an error is not counted, because the app calculates it again. Its line explains that the app shows them as text, so a formula that reads one does not see an error and SUM steps over it, and that the exported file writes them back as error values.
• Export writes them back as errors. A cell whose text is exactly one of Excel's seven error codes (#NULL! #DIV/0! #VALUE! #REF! #NAME? #NUM! #N/A) goes out as t="e". A leading apostrophe or the Text format keeps it text.

For PO: export changes for typed codes too

Writing the codes back as text was the smaller change, but it breaks the file in Excel itself. Open → export, and the file's own =SUM(A1:A3) now reads 5 in Excel too, because #DIV/0! has become text. So export writes error values.

That also covers a code someone types in the app: #N/A typed into a cell used to export as the text #N/A, and now exports as the error. That is what Excel does with #N/A typed into a cell, and what #226 will do in the grid too. The way to keep it as text is the same as in Excel: an apostrophe ('#N/A) or the Text format. Only the exact spelling is matched (#n/a stays text). I have not checked whether Excel accepts other spellings, so I did not guess.

If you'd rather typed codes keep exporting as text until #226, the alternative is to remember which cells came in as errors. That is a new field on the sheet model, and it would have to survive saving, undo and copy. I don't recommend it for a step #226 replaces anyway.

Reproduced on main first

A file written the way Excel writes it (the test checks the XML is t="e" with no <f>):

Cell	In the file	main after opening	this PR after opening	after export, in the file
A1	constant #DIV/0!	blank	#DIV/0! (text)	t="e" #DIV/0!
A2	constant #N/A	blank	#N/A (text)	t="e" #N/A
A3	5	5	5	5
A4	=SUM(A1:A3)	5, nothing said	5, and the report says why	the formula
Open report		nothing	"2 cells hold an error value such as #DIV/0! or #N/A · sheet "ยอดขาย""

SUM still reading 5 in the app is #226's to fix; this PR makes it visible and keeps the file whole.

How it was verified

• npm run verify (all ten gates): all ten gates green on 4088021 (the current head), result below
• excelIO.test.ts (+3):
  • a real t="e" cell opens as its code, not blank;
  • it exports back as t="e" and is still there on a reopen;
  • '#N/A and a Text-formatted #N/A export as text, #REF! typed as Excel spells it exports as the error, and #n/a stays text.
• fileLosses.test.ts (+2):
  • two constant errors on one sheet are reported with that sheet, and a formula erring on another sheet is not;
  • the same count after the file goes out of the app and back in.
• Re-checked against the old code: on 03aa574 all five new tests fail (the importer gives "", export has no t="e", the report is empty). The first assertion, that the test file really holds t="e" with no formula, passes on both.
• Both READMEs: a row in the "Told up front what a file can't keep" table, a bullet on what export does with error codes, the module map and the two test-file rows.
• Recounted after merging main 5439196:
  • 1979 tests in 140 files (1974 on main plus this PR's 5);
  • picture 27 retaken in both languages and checked by eye, blessed after the retake;
  • social card rebuilt and checked by eye.
• Tried on a production build of this branch (1280, both languages): the file above opens with A1 #DIV/0! and A2 #N/A in the grid instead of blank, A4 still reads 5 (An error code typed, pasted or imported into a cell stays text, so formulas that read it go silently wrong #226), and the report shows the new line with its sheet, in Thai and in English. No page errors.

Gate	Result
lint · check:readme · check:screens	pass
check:deps	every advisory fixed or accounted for
test	1990 passed (1979 by the README's static count)
check:mutants	32/32 killed
build · check:bundle	3799/4200 KB, largest chunk 934/1300 KB
check:a11y	119 checks passed
check:e2e	57 flows passed

Anything reviewers should look at closely

• The report reads every sheet's XML now, not only its relationships, to count t="e" cells. A sheet without the string t="e" is skipped after one includes, so a plain file costs one string search per sheet.
• The report's opening sentence disagrees with this line. It says that exporting over the original may lose things, while the error line says export writes them back. For a file whose only item is error values, that reads as a contradiction. The opening sentence is shared by every kind (บอกผู้ใช้ตอนเปิดไฟล์ ว่าไฟล์นี้มีอะไรที่แอปเก็บไว้ไม่ได้ · Tell the user, on open, what in this file will not survive a save #83), so I left it alone; PO may want the error line to say it plainly, or the opening sentence to allow for a kind that export keeps.
• Picture 69 (the report) is unchanged. Its fixture holds no error cells, so it still shows the same lines.
• The branch: this is the session's branch claude/new-session-m6mehm. The work was written on a local branch from main 03aa574, then main 5439196 (pivot: Sum, Average, Min and Max of a group with an error show the error (#222) #228, Pivot Sum/Average/Min/Max skip error values silently instead of showing the error #222) was merged in.

🤖 Generated with Claude Code

https://claude.ai/code/session_013pmWZsbmugfpqtYazrzfjj

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
excel-to-go	Ready	Preview	Oct 4, 2026 4:30pm UTC

[SuruchBoss]

PO review — merged at 4088021 (main 5d28538).

Verify (local, full): green — 1990 tests, mutants 32/32 killed, a11y 119, e2e 57. CI 17/17 green on the same head.

Tried as a user on a production build (errs.xlsx: A1 #DIV/0! and A2 #N/A as constant t="e" cells, A3 5, A4 =SUM(A1:A3)):

• 1280 and 390: A1/A2 show their codes instead of blank, the open report lists "2 cells hold an error value" on sheet "Errs" with the detail line, no page errors, no alerts. At 390 the report fits and reads cleanly.
• Export → read back with ExcelJS: A1/A2 are {error} again, A4 keeps SUM(A1:A3), B1 untouched.
• Typed in the app: #N/A exports as {error:"#N/A"}, '#N/A stays text, #VALUE!x stays text.
• Re-opening the exported file: codes back, report counts 3 (A1, A2 and the typed C1).
• A plain file (hours.xlsx) opens with no report.
• Known limit, as agreed and tracked in An error code typed, pasted or imported into a cell stays text, so formulas that read it go silently wrong #226: A4 reads 5 in the app where Excel reads #DIV/0!.

The two decisions asked for:

1. Typed #N/A now goes out as an error value — accepted. That is what Excel does with a typed #N/A, and the apostrophe / Text format escape still keeps text text. The in-app reading catches up in An error code typed, pasted or imported into a cell stays text, so formulas that read it go silently wrong #226.
2. The report's intro ("…if you export and save over the original, some of them won't be kept") overstates when the only kinds listed are kept on export (error values now, unknown functions already). Not blocking: the detail line under each item says exactly what happens. Opened as a small follow-up so the intro only promises a loss when a listed kind is actually lost.

Next in the queue: #207 → #226 → #224.

---

Generated by Claude Code