xlsx skill

Purpose: Give the agent a deterministic, script-first path for creating and sanity-checking .xlsx workbooks. The core operations (CSV → styled .xlsx, force formula recalculation, scan for formula errors, structural OOXML validation) are wrapped in small CLIs so the agent does not have to rewrite openpyxl boilerplate on every task and is never surprised by "formulas are stored but not calculated" (the single most common xlsx bug).

1. Red Flags (Anti-Rationalization)

STOP and READ THIS if you are thinking:

• "I'll just call DataFrame.to_excel and ship it." → WRONG. to_excel writes no styles, no frozen header, and no auto-filter. The result looks amateur. Use csv2xlsx.py.
• "I'll just call pd.DataFrame.from_records(rows).to_excel(out) on my LLM JSON output." → WRONG. Same to_excel styling gap, plus pandas' infer_objects heuristics silently promote mixed-type columns to object/float64 (an int column with one null becomes float64). Use json2xlsx.py — preserves native JSON types, ISO-date auto-coercion, csv2xlsx-style header.
• "I wrote formulas with openpyxl, so the numbers are there." → WRONG. openpyxl stores formulas as strings with no cached value. Every downstream consumer (pandas, charts, external apps) sees None. Run xlsx_recalc.py before shipping.
• "Validation says OK, the formulas must be fine." → WRONG. xlsx_validate.py scans for cached error values. If there are no cached values at all (fresh openpyxl output), the scan is meaningless. Use --fail-empty or recalc first.
• "Leading zeros in my phone-number column vanished; it's fine." → WRONG. It is almost never fine. Excel and pandas both coerce "007" to 7 by default. csv2xlsx.py detects leading-zero columns and keeps them as text; inline code should either pass dtype=str to pandas or set cell.number_format = "@" in openpyxl.
• "I'll just regex-replace \| to extract markdown tables." → WRONG. GFM pipe parsing has 5+ edge cases (escaped pipes \|, optional trailing pipes, separator-row alignment markers :---: / ---: / :---, column-count validation, blockquote skip, fenced-code-block skip) that regex hand-rolls always fail. Use md_tables2xlsx.py. For HTML <table> blocks: same tool — lxml.html is locked with no_network=True, huge_tree=False.

2. Capabilities

• Convert a CSV / TSV to a styled .xlsx with bold header, freeze-first-row, auto-filter, auto column widths, and leading-zero preservation.
• Convert an .xlsx workbook back to CSV (per-sheet / per-region) or JSON (flat / dict-of-arrays / nested tables) (xlsx2csv.py, xlsx2json.py). Thin shims on top of the xlsx_read/ foundation library (xlsx-10.A). Supports --sheet NAME|all, --include-hidden, --header-rows N|auto, --tables whole|listobjects|gap|auto, --merge-policy anchor-only|fill|blank, --include-hyperlinks (JSON: dict-form {"value", "href"}; CSV: markdown-link-as-text [text](url)), --include-formulas, --datetime-format ISO|excel-serial|raw. JSON shapes: single-sheet single-region flat array; multi-sheet dict-of-arrays; multi-sheet multi-region nested {Sheet: {tables: {Name: [...]}}}; single-sheet multi-region flat {Name: [...]}. CSV multi-region requires --output-dir → subdirectory schema <dir>/<sheet>/<table>.csv (sheet names with __ are NOT split — L4 lock). Round-trip with json2xlsx.py (xlsx-2) for Shape 1 / Shape 2 is byte-identical (live in tests/test_json2xlsx.py::TestRoundTripXlsx8::test_live_roundtrip); Shape 3 / Shape 4 are lossy on xlsx-2 v1 consume (full restoration deferred to xlsx-2 v2 --write-listobjects). Cross-cutting parity: cross-3 / cross-4 / cross-5 / cross-7 envelopes; path-traversal guard on --output-dir. Public helpers convert_xlsx_to_csv(input, output=None, **kwargs) and convert_xlsx_to_json(input, output=None, **kwargs) in xlsx2csv2json/__init__.py.
• Convert a JSON / JSONL document (file or stdin -) to a styled .xlsx with the SAME visual contract as csv2xlsx (json2xlsx.py). Three input shapes auto-detected: array-of-objects (single sheet), dict-of-arrays-of-objects (multi-sheet), JSONL (one JSON object per line — .jsonl extension). Preserves native JSON types (int / float / bool / null / str); ISO-8601 date strings auto-coerced to Excel datetime cells; --strict-dates makes timezone-aware datetimes a hard fail. Cross-cutting parity: cross-5 --json-errors envelope, cross-7 H1 same-path guard (exit 6 SelfOverwriteRefused), stdin pipe -. Round-trip contract with xlsx2json.py (xlsx-8) is frozen at skills/xlsx/references/json-shapes.md and is LIVE as of xlsx-8 merge.
• Convert an .xlsx workbook back to Markdown (GFM pipe tables, HTML <table> blocks, or per-table hybrid auto-select) (xlsx2md.py). Thin shim on top of the xlsx_read/ foundation (xlsx-10.A) — symmetric pair to md_tables2xlsx.py closing the xlsx → md → edits → xlsx round-trip. Supports --sheet NAME|all, --include-hidden, --format gfm|html|hybrid (default hybrid), --header-rows N|auto|smart, --memory-mode auto|streaming|full, --hyperlink-scheme-allowlist http,https,mailto (Sec-MED-2 default-enabled), --no-table-autodetect, --no-split, --gap-rows N --gap-cols N, --gfm-merge-policy fail|duplicate|blank, --datetime-format ISO|excel-serial|raw, --include-formulas (HTML only — M7 lock: GFM + formulas → exit 2). Hybrid promotion rules (per-table, first match wins): body merges → HTML; multi-row header → HTML; --include-formulas + formula cell → HTML; headerRowCount=0 (synthetic headers) → HTML (D13). Otherwise GFM. Multi-row <thead> reconstruction from xlsx_read's flat ›-joined headers (D-A11). Hyperlinks via Path C′ (parallel pass through openpyxl) preserve display text — emit as [text](url) GFM or <a href="...">text</a> HTML; blocked schemes emit text-only + warning. Round-trip contract frozen at references/xlsx-md-shapes.md; LIVE round-trip via TestRoundTripXlsx9::test_live_roundtrip_xlsx_md (cell-content byte-identical for non-merged plain tables; sheet-name asymmetry History → History_ is xlsx-3's documented sanitisation, NOT a regression). Cross-cutting parity: cross-3/4/5/7 envelopes; terminal InternalError redaction (R23f). Public helper convert_xlsx_to_md(input, output=None, **kwargs) in xlsx2md/__init__.py.
• Extract markdown tables from a .md document (file or stdin -) into a styled multi-sheet .xlsx (md_tables2xlsx.py). Two table flavors auto-detected: GFM pipe tables (with per-column alignment carried to Excel cell alignment) and HTML <table> blocks (with colspan / rowspan honoured as Excel merged cells). Sheet names derive from the nearest preceding markdown or HTML heading; fallback Table-N; UTF-16-aware truncate to Excel's 31-char limit; case-insensitive workbook-wide dedup via -2..-99 suffix. Default-on numeric + ISO-date coercion (csv2xlsx + json2xlsx parity; leading-zero preservation; aware datetimes → UTC-naive). Cross-cutting parity: --json-errors cross-5 envelope, cross-7 H1 same-path guard (exit 6), stdin pipe -. Pre-scan strips fenced code blocks, HTML comments, indented code blocks, <style> / <script> blocks, and blockquoted tables — so tables inside those regions never reach the parser. HTML parser is lxml.html.HTMLParser(no_network=True, huge_tree=False, recover=True) singleton (defense-in-depth against XXE + libxml2 huge-tree expansion). Public helper convert_md_tables_to_xlsx(input_path, output_path, **kwargs) -> int lives in md_tables2xlsx/__init__.py (mirrors xlsx-2 convert_json_to_xlsx; VDD-multi atomic-token protection inherited).
• Force formula recalculation in an .xlsx via headless LibreOffice, then optionally scan for error cells.
• Scan an .xlsx for formula errors (#REF!, #DIV/0!, #VALUE!, #NAME?, #N/A, #NUM!, #NULL!) without recomputing.
• Add a bar / line / pie chart on a value range with optional categories, title, anchor; stays editable in Excel / LibreOffice.
• Insert an Excel comment (legacy <comment>, optionally with the threaded-comment + personList Excel-365 modern layer) into a target cell, with cross-sheet --cell syntax and a batch mode that auto-detects the xlsx-7 findings envelope. Closes the "validation-агент расставляет замечания" pipeline together with xlsx_check_rules.py (xlsx-7).
• Declarative business-rule validation — xlsx_check_rules.py reads a YAML/JSON rules file alongside an .xlsx and emits a machine-readable findings envelope ({ok, summary, findings}) on stdout; pipes directly into xlsx_add_comment.py --batch - for cell-comment placement. Optional --output writes a workbook copy with a Remarks column + per-severity PatternFill. Hardened DSL (closed AST, no eval/exec), ReDoS-lint reject-list, billion-laughs YAML alias rejection, and a 100K-row × 10-rule perf contract (≤ 30 s wall-clock).
• Unpack and repack .xlsx archives for raw OOXML editing (shared office/ module with the docx skill).
• Structurally validate an .xlsx (relationships, content types, required parts, package-layout allow-list — every ZIP entry must live under [Content_Types].xml, _rels/, xl/, docProps/, or customXml/ per ECMA-376 §11.3.10; catches scratch-file leaks that Excel refuses to open. With --strict, the package-layout warning is promoted to exit 1).
• Reject password-protected and legacy .xls (CFB-container) inputs early in the reader scripts (xlsx_recalc.py, xlsx_validate.py, xlsx_add_chart.py, office/validate.py, office/unpack.py, preview.py) with a clear remediation message (exit 3) instead of a BadZipFile traceback. csv2xlsx.py and office_passwd.py are not gated — the former takes CSV/TSV input (no encryption to detect), the latter is the encryption tool itself.
• Detect macro-enabled inputs (.xlsm, with xl/vbaProject.bin) and warn when the chosen output extension would silently drop the macros (xlsm → xlsx).
• Render any .xlsx/.xlsm/.pdf (or peer-skill .docx/.pptx) into a single PNG-grid preview via preview.py (LibreOffice + Poppler).
• Emit failures as machine-readable JSON to stderr with --json-errors (uniform across all four office skills).
• Set or remove a password on a .xlsx/.docx/.pptx (MS-OFB Agile, Office 2010+) via office_passwd.py — three modes: --encrypt PASSWORD, --decrypt PASSWORD, --check (exit 0 encrypted / 10 clean / 11 missing).

3. Execution Mode

• Mode: script-first.
• Why this mode: Spreadsheet tasks have well-defined inputs and outputs and benefit heavily from deterministic CLIs. Writing the styling code inline produces ugly workbooks every time; delegating to scripts frees the agent to focus on the user's intent.

4. Script Contract

• Commands:
  • python3 scripts/csv2xlsx.py INPUT.csv OUTPUT.xlsx [--delimiter auto|,|;|\t] [--encoding utf-8] [--no-freeze] [--no-filter]
  • python3 scripts/xlsx2csv.py INPUT.xlsx [OUTPUT.csv|-] [--output-dir DIR] [--sheet NAME|all] [--include-hidden] [--header-rows N|auto] [--merge-policy anchor-only|fill|blank] [--tables whole|listobjects|gap|auto] [--gap-rows N] [--gap-cols N] [--include-hyperlinks] [--include-formulas] [--datetime-format ISO|excel-serial|raw] [--memory-mode auto|streaming|full] [--encoding utf-8|utf-8-sig] [--delimiter ,|;|tab|\\t|pipe] [--drop-empty-rows] [--json-errors] — convert .xlsx → CSV. Multi-region requires --output-dir. Output - or omitted → stdout. --drop-empty-rows filters out rows where every value is None/"" (default off; same semantics as the JSON path). For full Excel-RU/EU double-click compatibility pass BOTH --encoding utf-8-sig AND --delimiter ';': the BOM lets Excel detect UTF-8 encoding, and ; is the field separator on RU/EU locales (where , is the decimal separator). Defaults (utf-8 + ,) stay pandas-/jq-friendly. Stdout always plain UTF-8 regardless of flags. The tab alias (or 2-char \\t escape) maps to a real tab; pipe maps to |. Python-helper equivalent: convert_xlsx_to_csv(input, output, delimiter=';', encoding='utf-8-sig') (kwarg names map to CLI flags 1:1).
  • python3 scripts/xlsx2json.py INPUT.xlsx [OUTPUT.json|-] [--sheet NAME|all] [--include-hidden] [--header-rows N|auto|leaf|smart] [--header-flatten-style string|array] [--merge-policy anchor-only|fill|blank] [--tables whole|listobjects|gap|auto] [--gap-rows N] [--gap-cols N] [--include-hyperlinks] [--include-formulas] [--datetime-format ISO|excel-serial|raw] [--memory-mode auto|streaming|full] [--drop-empty-rows] [--json-errors] — convert .xlsx → JSON. Output - or omitted → stdout. --drop-empty-rows filters rows where every value is None/"". --header-rows leaf auto-detects the header band but keeps ONLY the deepest non-empty level per column as the JSON key — solves the layout-heavy-report key bloat where rows 1..K-1 are merged metadata banners (title / customer / contract / project / period) and the real column names sit on row K. Example recipe: xlsx2json.py FILE.xlsx out.json --sheet all --tables auto --header-rows leaf --drop-empty-rows → keys become ["Date", "Hours", "Days", "Specialist", "Position", "Task #", "Description"] instead of 7-level ›-concatenated metadata-prefixed strings. --header-rows smart (xlsx-8a-09 / R11; iter-3 2026-05-13) is the "find-the-data-table" recipe — locates the real header row whether or not merges are present. Uses a type-pattern heuristic: scores each top row by string-ratio + column-coverage + downstream type-stability + depth (max 5.0); when a candidate scores ≥ 3.5 AND len(sample_below) ≥ 2, the library shifts the region past the metadata and treats the candidate row as a 1-row header. smart does NOT defer to merge-based detection — it competes purely on score. On merged-banner fixtures (e.g. A1:C1 "2026 plan" over Q1/Q2/Q3), smart shifts to the sub-header row and produces leaf-like keys ["Q1", "Q2", "Q3"]. Callers needing the multi-level concatenated form ("2026 plan › Q1") must use --header-rows auto (or leaf to take the deepest level only) — smart is non-overlapping with auto/leaf at the OUTPUT shape level. Use smart when auto produces {"": ..., "От": ..., "До": ..., "__2": ..., ..., "__N": ...}-style keys on a workbook whose real headers are buried below a config section (typical: financial-modeling sheets with От/До parameter ranges on rows 1-K + a real data table starting at row K+1; or per-row banner + per-row metadata header rows above the column-name row, e.g. masterdata Timesheet pattern). --memory-mode {auto,streaming,full} (xlsx-8a-11 / R13) exposes the read-mode selection: auto (default) lets the library pick based on file size (_DEFAULT_READ_ONLY_THRESHOLD = 100 MiB); streaming minimises RAM — measured 7.6× RAM reduction PLUS 2.6× wall-clock speedup on 15 MB multi-sheet workbooks (1188 MB / 32 s → 156 MB / 13 s, live subprocess measurement 2026-05-13) at the cost of merge-aware features (overlap detection, multi-row header band via merges, merge-policy fill) becoming no-ops — use on workbooks without merges OR where merge fidelity does not matter; full forces non-streaming so all merge features work but RAM scales ~10× with file size. --include-hyperlinks is incompatible with streaming (hyperlink extraction needs the non-streaming cell object); the combination emits a stderr warning and auto-overrides to full. Practical recipe for large multi-sheet workbooks without critical merges: xlsx2json.py BIG.xlsx out.json --sheet all --tables whole --header-rows smart --memory-mode streaming --drop-empty-rows.
  • python3 scripts/md_tables2xlsx.py INPUT.md OUTPUT.xlsx [--no-coerce] [--no-freeze] [--no-filter] [--allow-empty] [--sheet-prefix STR] [--encoding utf-8] [--json-errors] — markdown tables → multi-sheet xlsx. INPUT is a path or - for stdin.
  • python3 scripts/xlsx2md.py INPUT.xlsx [OUTPUT.md|-] [--sheet NAME|all] [--include-hidden] [--format gfm|html|hybrid] [--header-rows N|auto|smart] [--memory-mode auto|streaming|full] [--hyperlink-scheme-allowlist SCHEMES] [--no-table-autodetect] [--no-split] [--gap-rows N] [--gap-cols N] [--gfm-merge-policy fail|duplicate|blank] [--datetime-format ISO|excel-serial|raw] [--include-formulas] [--json-errors] — convert .xlsx → Markdown. Default --format hybrid auto-selects GFM vs HTML per-table by 4 promotion rules (body merges / multi-row header / formulas+include / synthetic header). Output - or omitted → stdout. Hyperlinks ALWAYS extracted (D5; read_only_mode=False default) — pass --memory-mode=streaming to reduce RAM at the cost of unreliable hyperlinks. --hyperlink-scheme-allowlist defaults to http,https,mailto (Sec-MED-2 default-enabled) — '*' allows all, "" blocks all. --format=gfm + --include-formulas → exit 2 IncludeFormulasRequiresHTML (M7 lock); --format=gfm + body merges + --gfm-merge-policy=fail (default) → exit 2 GfmMergesRequirePolicy. --header-rows=smart runs the xlsx-8a-09 type-pattern heuristic to skip metadata banners; on --header-rows=auto (default), xlsx2md.py auto-falls-back to smart when the initial detection produces non-uniform header depths (D-A25 hot-patch 2026-05-14, surfaces xlsx-8a-09 transparently for the masterdata Timesheet banner-and-metadata pattern; one stderr warning per affected region). Round-trip safe with md_tables2xlsx.py for non-merged plain tables (cell content byte-identical; sheet-name asymmetry History → History_ is xlsx-3's documented sanitisation, see references/xlsx-md-shapes.md §6). Python-helper: convert_xlsx_to_md(input, output, **kwargs) in xlsx2md/__init__.py (--flag=value atomic-token kwargs; mirrors xlsx-2 / xlsx-3 D7 lock).
  • python3 scripts/xlsx_recalc.py INPUT.xlsx [--output OUT.xlsx] [--timeout 120] [--scan-errors] [--json]
  • python3 scripts/xlsx_validate.py INPUT.xlsx [--json] [--fail-empty]
  • python3 scripts/xlsx_add_chart.py INPUT.xlsx --type bar|line|pie --data RANGE [--categories RANGE] [--title TEXT] [--sheet NAME] [--anchor CELL] [--titles-from-data | --no-titles-from-data] [--output OUT.xlsx]
  • python3 scripts/xlsx_add_comment.py INPUT.xlsx OUTPUT.xlsx (--cell REF --author NAME --text MSG | --batch FILE [--default-author NAME] [--default-threaded]) [--threaded | --no-threaded] [--initials INI] [--date ISO] [--allow-merged-target] [--json-errors]
  • python3 scripts/xlsx_check_rules.py INPUT.xlsx --rules RULES.{json,yaml} [--sheet NAME | --all-sheets] [--visible-only] [--json | --human] [--max-findings N] [--summarize-after N] [--require-data] [--ignore-stale-cache] [--strict-aggregates] [--treat-numeric-as-date COL] [--treat-text-as-date COL] [--timeout SECONDS] [--no-strip-whitespace] [--no-table-autodetect] [--no-merge-info] [--output OUT.xlsx [--remark-column auto|LETTER|HEADER] [--remark-column-mode replace|append|new] [--streaming-output]] [--json-errors]
  • python3 scripts/office/unpack.py INPUT.xlsx OUTDIR/
  • python3 scripts/office/pack.py INDIR/ OUTPUT.xlsx
  • python3 scripts/office/validate.py INPUT.xlsx [--strict] [--json]
  • python3 scripts/preview.py INPUT OUTPUT.jpg [--cols 3] [--dpi 110] [--gap 12] [--padding 24] [--label-font-size 14] [--soffice-timeout 240] [--pdftoppm-timeout 60]
  • python3 scripts/office_passwd.py INPUT [OUTPUT] (--encrypt PASSWORD | --decrypt PASSWORD | --check) — pass - as PASSWORD to read it from stdin.
  • All scripts above accept --json-errors to emit failures as a single line of JSON on stderr ({v, error, code, type?, details?}). The schema version v is currently 1; argparse usage errors are routed through the same envelope (type:"UsageError").
• Inputs: positional paths; optional flags per command.
• Outputs: a single file at the named output path; office/unpack.py produces a directory tree; validators print a report (or JSON).
• Failure semantics: non-zero exit on missing input, invalid encoding, soffice errors, or formula errors (xlsx_validate.py returns 1 when errors are present). Error detail goes to stderr.
• Idempotency: csv2xlsx.py produces the same workbook for the same input every time. xlsx_recalc.py is idempotent on an already-recalculated workbook. Exception: office_passwd.py --encrypt is intentionally non-deterministic — Office encryption uses a fresh random salt per run.
• Dry-run support: not applicable.

5. Safety Boundaries

• Allowed scope: only the paths named on the command line; never write outside the requested output path.
• Default exclusions: do not fetch data from remote URLs; only read from the provided local file.
• Destructive actions: xlsx_recalc.py rewrites its input in place when --output is omitted — the convention matches how users expect "recalculate this file" to behave, but make it explicit to the user if the file is important.
• Optional artifacts: office/schemas/ is optional; validators fall back to structural checks.

6. Validation Evidence

• Local verification:
  • python3 -m venv .venv && source .venv/bin/activate && pip install -r scripts/requirements.txt — installs openpyxl, pandas, lxml, defusedxml.
  • python3 scripts/csv2xlsx.py examples/fixture.csv /tmp/out.xlsx && python3 scripts/office/validate.py /tmp/out.xlsx — exit 0, validator reports OK.
  • python3 scripts/xlsx_validate.py /tmp/out.xlsx --fail-empty — exit 0 (no formula errors, values are concrete).
  • python3 scripts/xlsx_recalc.py /tmp/out.xlsx --scan-errors --json | jq .ok — prints true if LibreOffice is installed.
• Expected evidence: /tmp/out.xlsx, validator reports, non-empty xlsx_validate.py --json.
• CI signal: python3 ../../.claude/skills/skill-creator/scripts/validate_skill.py skills/xlsx — exit 0.

7. Instructions

7.1 Pick the script before writing inline code

1. Check §10 Quick Reference first. Most common tasks (CSV→xlsx, recalc, error scan) already exist.
2. Drop to inline openpyxl/pandas only when you need something the scripts don't cover (charts, conditional formatting, pivot sources).

7.2 Setup

1. MUST run bash scripts/install.sh once. It creates scripts/.venv/ locally (nothing global), prints a warning for any missing system tool, and is idempotent.
2. External system tools (checked by install.sh, installed manually per project plan §3.3 "внешние инструменты — не бандлятся"):
   • LibreOffice (soffice) — required by xlsx_recalc.py to populate cached formula values that openpyxl cannot compute. macOS: brew install --cask libreoffice. Debian: sudo apt install libreoffice --no-install-recommends. Fedora: sudo dnf install libreoffice. Commands that need it fail with a clear error until it's installed.

7.3 Creating .xlsx from CSV

1. python3 scripts/csv2xlsx.py input.csv output.xlsx covers 80% of cases.
2. Use --delimiter ';' for European exports; --delimiter '\t' for TSV.
3. Preserve leading zeros automatically — the script detects them and keeps the column as text. If you need to force that behaviour for a specific column, pre-process the CSV or write a custom openpyxl pass with cell.number_format = "@".

7.4 Producing a workbook with formulas

1. Write the workbook with openpyxl: cell.value = "=SUM(B2:B100)". Remember the formula is just a string.
2. MUST run python3 scripts/xlsx_recalc.py file.xlsx before shipping. Otherwise consumers that read with data_only=True see None for every formula cell.
3. After recalc, run python3 scripts/xlsx_validate.py file.xlsx. Any #REF! / #DIV/0! is a bug to fix, not a cosmetic issue.

7.5 Validating someone else's workbook

1. Structural: python3 scripts/office/validate.py file.xlsx.
2. Formula errors: python3 scripts/xlsx_validate.py file.xlsx --fail-empty. The --fail-empty flag catches the "formulas never recalculated" case.

7.6 Raw XML editing

Use the shared office/ module (same as docx). Typical reason to drop into XML: tweaking a specific cell style, patching a corrupted sharedStrings.xml, or adding a relationship the user's downstream tool expects.

8. Workflows (Optional)

CSV to a styled, ready-to-ship workbook:

- [ ] Inspect the CSV — check delimiter, encoding, leading zeros
- [ ] `python3 scripts/csv2xlsx.py data.csv out.xlsx`
- [ ] `python3 scripts/office/validate.py out.xlsx`
- [ ] Open in Excel/LibreOffice for a spot-check

Build a formula-driven model:

- [ ] Write workbook with openpyxl (cells + formulas)
- [ ] `python3 scripts/xlsx_recalc.py out.xlsx --scan-errors`
- [ ] Surface any error cells to the user
- [ ] `python3 scripts/xlsx_validate.py out.xlsx`

Audit an incoming .xlsx:

- [ ] `python3 scripts/office/validate.py in.xlsx`
- [ ] `python3 scripts/xlsx_validate.py in.xlsx --fail-empty`
- [ ] If --fail-empty fails, run `xlsx_recalc.py` first then revalidate

9. Best Practices & Anti-Patterns

Rationalization Table

10. Quick Reference

11. Examples (Few-Shot)

Fixture: examples/fixture.csv.

Input — user request:

Convert sales.csv into a presentable .xlsx and make sure the numbers are usable as a pivot source.

Output — agent action:

python3 scripts/csv2xlsx.py sales.csv sales.xlsx
python3 scripts/office/validate.py sales.xlsx
python3 scripts/xlsx_validate.py sales.xlsx --fail-empty

The output has no formulas, so xlsx_validate.py --fail-empty is the right check; all cells carry concrete values.

Input — user request:

The forecast workbook shows #REF! in a bunch of cells. Find them all.

Output — agent action:

python3 scripts/xlsx_recalc.py forecast.xlsx        # ensure cache is fresh
python3 scripts/xlsx_validate.py forecast.xlsx --json

Paste the JSON or the human report back to the user, grouped by error type and sheet.

Input — user request:

Run the validator on timesheet.xlsx and drop a comment on every problematic cell so my team can triage in Excel.

Output — agent action (xlsx-7 → xlsx-6 pipeline; piped envelope):

# xlsx_check_rules.py emits a JSON envelope of findings; xlsx_add_comment.py
# auto-detects the {ok, summary, findings} shape and writes one comment
# per finding (skipping group-findings with row=null).
python3 scripts/xlsx_check_rules.py timesheet.xlsx --rules rules.json --json \
    | python3 scripts/xlsx_add_comment.py timesheet.xlsx timesheet-annotated.xlsx \
        --batch - --default-author "Validator Bot"

The --default-author flag is required for envelope shape (DEP-2); without it the script exits with MissingDefaultAuthor. To attach the comments as Excel-365 threaded comments instead of legacy bubbles, add --default-threaded.

12. Resources

• references/financial-modeling-conventions.md — colour coding, number formats, formula hygiene, drivers layout.
• references/openpyxl-vs-pandas.md — which library for which job, read-mode pitfalls.
• references/formula-recalc-gotchas.md — why cached values matter, engines that recalc, engines that don't.
• scripts/csv2xlsx.py — CSV/TSV → styled workbook.
• scripts/xlsx_recalc.py — LibreOffice-backed formula recalculation + error scan.
• scripts/xlsx_validate.py — fast formula-error scan without recalc.
• scripts/xlsx_add_chart.py — bar / line / pie chart attachment over a cell range; chart stays editable in Excel / LibreOffice.
• scripts/xlsx_add_comment.py — insert an Excel comment (legacy + optional Excel-365 threaded) into a target cell; single-cell mode (--cell) or batch mode (--batch, auto-detects xlsx-7 findings envelope).
• references/comments-and-threads.md — OOXML data model behind xlsx_add_comment.py: part graph, cell-syntax forms, the C1/M-1/M6 pitfalls list (read these before editing the scanner code), v1 honest-scope.
• scripts/xlsx_check_rules.py — declarative business-rule validator. Reads YAML/JSON rules, emits {ok, summary, findings} envelope; pipes into xlsx_add_comment.py --batch -. Closed AST + ReDoS lint + billion-laughs alias rejection. Backed by scripts/xlsx_check_rules/ package.
• references/xlsx-rules-format.md — full SPEC for xlsx_check_rules.py: rule shape, scope vocabulary, check vocabulary, AST safety, output envelope, exit codes, honest-scope catalogue, regression battery anchors.
• examples/check-rules-timesheet.json + examples/check-rules-timesheet.xlsx — worked SPEC §10 example, ready to run end-to-end (validate → pipe → annotate).
• scripts/preview.py — universal INPUT → PNG-grid renderer for .xlsx/.xlsm/.docx/.pptx/.pdf. Byte-identical across all four office skills.
• scripts/office_passwd.py — set / remove / detect password protection on .xlsx/.docx/.pptx via msoffcrypto-tool (MS-OFB Agile, Office 2010+). Byte-identical across the three OOXML skills (not pdf — pdf has its own AcroForm encryption). Pass - as the password to read it from stdin.
• scripts/_errors.py — --json-errors envelope helper (schema v=1).
• scripts/_soffice.py — LibreOffice subprocess wrapper.
• scripts/office/ — OOXML unpack/pack/validate, byte-identical copy from the docx skill (master — see CLAUDE.md §2). Includes deep XlsxValidator (sheet chain, sst+styles index bounds, sheet-name uniqueness, orphan parts).
• references/security.md — Security model & trust boundary (xlsx-8 / xlsx-8a). Canonical statement of trust-boundary assumptions, accepted-risk catalogue, and the parent-symlink + TOCTOU race in _emit_multi_region (deferred code-fix, documented as known-limitation). Read before deploying in shared CI or multi-tenant build farms.

13. Known Limitations & Deferred Refactors

• xlsx_read/ ↔ xlsx_check_rules/ reader duplication (xlsx-10.A, 2026-05-12). The xlsx-10.A foundation library (scripts/xlsx_read/) re-implements a portion of the reader logic also present in xlsx-7's xlsx_check_rules/ package (merge resolution, header detection, value extraction). This is intentional v1 scope to unblock xlsx-8 and xlsx-9 without re-validating xlsx-7's 311-unit / 113-E2E regression battery. The follow-up refactor (xlsx-7 consumes xlsx_read/) is tracked as xlsx-10.B in docs/office-skills-backlog.md. The xlsx-9 owner opens xlsx-10.B within 14 calendar days post-xlsx-9 merge; if missed, the duplication is promoted to permanent technical debt and this bullet is updated accordingly.
• xlsx2md (v1): comments / charts / images / shapes / pivots / data-validation / cell styles / diagonal borders / sparklines / camera objects dropped; rich-text spans → plain-text concat; formulas without cached value emit empty (or data-formula attr in HTML when --include-formulas); hyperlinks always extracted (D5; read_only_mode=False default; --memory-mode=streaming override available); hyperlink scheme allowlist defaults {http, https, mailto} (Sec-MED-2 default-enabled in v1); sheet names emitted verbatim from xl/workbook.xml/<sheets> (xlsx-3 round-trip may sanitise on write-back — documented in references/xlsx-md-shapes.md §6, NOT a regression). See docs/tasks/task-012-*.md for details.