cubeplexai/cubeplex/xlsx

xlsx

Skill directory

load_skill returned a path field — the sandbox directory holding this skill's scripts, templates, and references. Use it verbatim; never construct the path yourself:

export SKILL_DIR="<the `path` value from load_skill>"

Handle the request in this conversation — don't delegate to subagents: edit integrity depends on one agent owning the file. Write output files to the working directory, never into $SKILL_DIR.

Choosing the write strategy: openpyxl vs XML surgery

Two ways to write an .xlsx, each with a job. Pick by one question: does an input file already exist whose non-cell content must survive?

Gray zone — a financial model built from scratch: openpyxl works and is faster. Reach for XML surgery only when you need byte-level determinism or precise styles.xml index control (see references/format.md).

The one absolute rule: never openpyxl.load_workbook(existing.xlsx) then save() — that round-trip is what corrupts. openpyxl is safe only on a workbook you created in the same script.

openpyxl writes formula strings but no cached results, so an openpyxl-created model (one that uses formulas) MUST be recalculated (libreoffice_recalc.py) and checked (formula_check.py) before delivery — same gate as the XML path. A report that writes computed values directly needs neither step.

Report vs Model: do you need formulas at all?

The Formula-First rule below applies to models — workbooks the user will keep editing, where they change inputs/assumptions and derived cells must recompute. Most "analyze this data and give me a spreadsheet" requests are reports: the numbers are already final, computed by pandas/openpyxl. Forcing live formulas + a LibreOffice recalc onto a report adds a slow, fragile external dependency for zero benefit — the data already knows the answer, and recalc can hang or silently skip cells.

Decide with one question: Will the user open this file and change numbers, expecting other cells to update? If yes → Model (formulas + recalc). If no → Report (write values). When unsure, a report is the safer default — it always opens correctly with no recalc step.

• Report: set each derived cell to its already-computed number (never a formula).

openpyxl: ws["C2"] = round(revenue, 2). XML: <c r="C2"><v>12345.67</v></c> (no <f>). No libreoffice_recalc.py step. Run formula_check.py only when you edited an existing file (to catch XML/structure issues), not to check formulas.

• Model: follow the Formula-First rules below; recalc is mandatory and is the

final step before delivery.

Task Routing

READ — Analyze data (read references/read-analyze.md first)

Start with xlsx_reader.py for structure discovery, then pandas for custom analysis. Never modify the source file.

Formatting rule: When the user specifies decimal places (e.g. "2 decimal places"), apply that format to ALL numeric values — use f'{v:.2f}' on every number. Never output 12875 when 12875.00 is required.

Aggregation rule: Always compute sums/means/counts directly from the DataFrame column — e.g. df['Revenue'].sum(). Never re-derive column values before aggregation.

CREATE — new file (read references/create.md + references/format.md)

Default: build with openpyxl — it unlocks charts, conditional formatting, cover pages, and styling with far less effort than raw XML. Alternative for byte-level determinism / precise style-index control: copy templates/minimal_xlsx/ → edit XML directly → pack with xlsx_pack.py.

Either way, the strategy depends on Report vs Model (see that section):

• Model (user will edit inputs): every derived value MUST be a live Excel formula

(=SUM(B2:B9) / <f>SUM(B2:B9)</f>), never a hardcoded number. Apply font colors per format.md. Recalculate + formula_check.py before delivery.

• Report (final snapshot): skip formulas entirely — write the computed values

directly and deliver immediately. Charts read the values; no recalc needed.

EDIT — XML direct-edit (read references/edit.md first)

CRITICAL — EDIT INTEGRITY RULES:

1. NEVER create a new `Workbook()` for edit tasks. Always load the original file.
2. The output MUST contain the same sheets as the input (same names, same data).
3. Only modify the specific cells the task asks for — everything else must be untouched.
4. After saving output.xlsx, verify it: open with xlsx_reader.py or pandas and confirm the original sheet names and a sample of original data are present. If verification fails, you wrote the wrong file — fix it before delivering.

Never use openpyxl round-trip on existing files (corrupts VBA, pivots, sparklines). Instead: unpack → use helper scripts → repack.

"Fill cells" / "Add formulas to existing cells" = EDIT task. If the input file already exists and you are told to fill, update, or add formulas to specific cells, you MUST use the XML edit path. Never create a new Workbook(). Example — fill B3 with a cross-sheet SUM formula:

python3 $SKILL_DIR/scripts/xlsx_unpack.py input.xlsx /tmp/xlsx_work/
# Find the target sheet's XML via xl/workbook.xml → xl/_rels/workbook.xml.rels
# Then use `edit` to add <f> inside the target <c> element:
#   <c r="B3"><f>SUM('Sales Data'!D2:D13)</f><v></v></c>
python3 $SKILL_DIR/scripts/xlsx_pack.py /tmp/xlsx_work/ output.xlsx

Add a column (formulas, numfmt, styles auto-copied from adjacent column):

python3 $SKILL_DIR/scripts/xlsx_unpack.py input.xlsx /tmp/xlsx_work/
python3 $SKILL_DIR/scripts/xlsx_add_column.py /tmp/xlsx_work/ --col G \
    --sheet "Sheet1" --header "% of Total" \
    --formula '=F{row}/$F$10' --formula-rows 2:9 \
    --total-row 10 --total-formula '=SUM(G2:G9)' --numfmt '0.0%' \
    --border-row 10 --border-style medium
python3 $SKILL_DIR/scripts/xlsx_pack.py /tmp/xlsx_work/ output.xlsx

The --border-row flag applies a top border to ALL cells in that row (not just the new column). Use it when the task requires accounting-style borders on total rows.

Insert a row (shifts existing rows, updates SUM formulas, fixes circular refs):

python3 $SKILL_DIR/scripts/xlsx_unpack.py input.xlsx /tmp/xlsx_work/
# IMPORTANT: Find the correct --at row by searching for the label text
# in the worksheet XML, NOT by using the row number from the prompt.
# The prompt may say "row 5 (Office Rent)" but Office Rent might actually
# be at row 4. Always locate the row by its text label first.
python3 $SKILL_DIR/scripts/xlsx_insert_row.py /tmp/xlsx_work/ --at 5 \
    --sheet "Budget FY2025" --text A=Utilities \
    --values B=3000 C=3000 D=3500 E=3500 \
    --formula 'F=SUM(B{row}:E{row})' --copy-style-from 4
python3 $SKILL_DIR/scripts/xlsx_pack.py /tmp/xlsx_work/ output.xlsx

Row lookup rule: When the task says "after row N (Label)", always find the row by searching for "Label" in the worksheet XML (grep -n "Label" /tmp/xlsx_work/xl/worksheets/sheet*.xml or check sharedStrings.xml). Use the actual row number + 1 for --at. Do NOT call xlsx_shift_rows.py separately — xlsx_insert_row.py calls it internally.

Apply row-wide borders (e.g. accounting line on a TOTAL row): After running helper scripts, apply borders to ALL cells in the target row, not just newly added cells. In xl/styles.xml, append a new <border> with the desired style, then append a new <xf> in <cellXfs> that clones each cell's existing <xf> but sets the new borderId. Apply the new style index to every <c> in the row via the s attribute:

<!-- In xl/styles.xml, append to <borders>: -->
<border>
  <left/><right/><top style="medium"/><bottom/><diagonal/>
</border>
<!-- Then append to <cellXfs> an xf clone with the new borderId for each existing style -->

Key rule: When a task says "add a border to row N", iterate over ALL cells A through the last column, not just newly added cells.

Manual XML edit (for anything the helper scripts don't cover):

python3 $SKILL_DIR/scripts/xlsx_unpack.py input.xlsx /tmp/xlsx_work/
# ... edit XML with `edit` ...
python3 $SKILL_DIR/scripts/xlsx_pack.py /tmp/xlsx_work/ output.xlsx

FIX — Repair broken formulas (read references/fix.md first)

This is an EDIT task. Unpack → fix broken <f> nodes → pack. Preserve all original sheets and data.

VALIDATE — Check formulas (read references/validate.md first)

Run formula_check.py for static validation. Use libreoffice_recalc.py for dynamic recalculation when available.

Financial Color Standard

Key Rules

1. Formula-First for Models, Values for Reports: A model (user will edit inputs) MUST

use live formulas for every derived cell. A report (final snapshot) MUST write the already-computed value directly — no formulas, no recalc. Decide by the "Report vs Model" rule above; do not force formulas onto a report.

1. CREATE → openpyxl by default (charts/formatting made easy); XML template only when you need byte-level control
2. EDIT existing → XML surgery: never openpyxl round-trip on a file you didn't create — it drops VBA/pivots/charts. Use unpack/edit/pack scripts
3. Always produce the output file — this is the #1 priority
4. Validate before delivery (models only): recalculate (if formulas), then formula_check.py exit code 0 = safe. Reports with no formulas skip this.
5. Recalc is the final step — never overwrite the good copy. libreoffice_recalc.py

now verifies that formula cells actually got cached <v> values; if it reports empty <v> cells, that is a HARD FAILURE — do not deliver. In a report workflow there are no formulas to recalc, so this never blocks you. If a model's recalc cannot be made to succeed, ship a report-style file (write the computed values) rather than delivering empty formula cells.

Utility Scripts