Lintel on GitHub Copilot. Generated from
skills/generate-xlsx/SKILL.md; edit the canonical file, then runli-copilot init.
- Resource root:
../../..from this skill's base directory (the Lintel source withbin/,lib/,skills/). Write plans, state and evidence into the working repository's.claude/tree, never into the resource root.- Skill-relative paths: this skill's own
scripts/,references/anddata/folders (and a<base>that the workflow defines as its own directory) mean../../../skills/generate-xlsx/in the Lintel source, not this generated folder.${LINTEL_SKILLS_DIR:-skills}means the skills root,../../../skills. Abin/li-runstep runs in the working repository, so use$LINTEL_SKILLS_DIR/generate-xlsx/there.- Shell steps: run Bash snippets with Bash (Git for Windows'
bash.exeon Windows, neverSystem32\bash.exe). Save a snippet to a temporary.shfile and runbash "<resource root>/bin/li-run" <file>; it preparesLINTEL_SOURCE_ROOT,LINTEL_REPO_ROOTand the profile context.- Tools: Read=
view, Write=create, Edit=edit, Bash=bash/powershell, Grep=grep, Glob=glob, AskUserQuestion=ask_user, TodoWrite=the plan checklist, Task or a named role=taskwith that custom agent, WebFetch=web_fetch.- Other Lintel workflows are native skills: invoke
/li-<name>rather than reading their files. Named roles such asCodeReviewerare custom agents.
/generate-xlsx
Create an editable workbook for estimates, capacity models, comparison tables,
schedules or trackers. Standalone --brief needs no shared design-spec or domain
envelope. Reusable format methods belong here; actual customer data does not.
Source facts, formulas, actual engine results, saved caches and rendered layout
are different evidence layers.
Inputs and retained entry paths
--brief <path|inline>: complete source, including assumptions and evidence; retain inline text before hashing.--from-pipeline <run-dir>: use existing input admission with--format xlsx, external input context, original package/leaves and live profile. Feed the full admitted content into the existing workbook procedure, preserving Inputs/Calculation/Summary/Sources and formula/cache requirements. Do not invent aper_format.xlsxlayout or call source admission recalculation. Shared artifact acceptance remains distinct from this input-only join.--out <path>: explicit new owned.xlsxpath, default<brief-stem>.xlsxin the selected working directory. Refuse unapproved replacement.--update-data <source>: retain the existing workbook/source identity, apply only the selected update, then recollect calculation/cache/reopen evidence. A remembered total or an old receipt cannot cover changed input.--template <path|name>: explicit template or a name within the verified configured brand directory. No personal template scan.--use-defaults: explicit neutral blank/template route when policy allows.--customer-share: apply the current profile's actual customer-facing requirements to labels, notes, disclosures and the complete source.
Capability and source preflight
Discover actual tool schemas and permissions before invoking them. Prefer a native workbook writer/read API when available. The native workbook procedure describes an observed Copilot canvas route; it is not a guarantee for every client or formula.
The existing declared library alternative is openpyxl, or an explicitly selected available Node workbook library. A library's ability to write a formula is not proof it calculates it. Do not install automatically/globally, start another application, use macros, fetch external workbooks or bypass a denied permission. Missing writer/calculation/persistence/render capability leaves the dependent action blocked; independent source preparation can continue.
Use the accepted P05/P07 fidelity procedure: explicit source/output roots, verified profile reference, required-policy bridge, original work/leaf identity where supplied, and immutable QA obligations before observations. Keep formula/source/config changes content-bound. Neutral mode does not invent currency, retention, brand colors, compliance gates or a voice score.
Workbook composition
- Read the entire source. Inventory tables, units, input dates/versions, formulas, claims, citations, reasoning and material limitations. Distinguish measured inputs, assumptions, unknowns and calculated projections. Do not replace unknown input with zero or infer actual cost/utilization from a manifest.
- Separate responsibilities. Keep Inputs, Calculation, Summary and Sources/Methods content, using separate Sources and Methods sheets when useful. One sheet per substantive table remains a useful organization, not a quota. Preserve long-form explanations in text cells or a delivered linked source; source references and limitations must stay discoverable from the workbook.
- Use editable cell types. Quantities and rates are numeric with units;
dates/currency/rounding follow the actual source. Treat literal formula-looking
text such as
=1+1as text unless it is explicitly a formula. Never promote imported prose to a formula. Preserve complete paragraphs, cell labels, provenance and source-backed table values. - Use formulas for derived results. Retain SUM/PRODUCT/AVERAGE and relevant cross-sheet dependencies rather than replacing them with hardcoded totals. Explicitly review references/ranges, rounding and missing-input behavior. Conditional formats, validation and charts are useful when available and requested; color alone cannot convey a result or substitute for a limitation.
- Use accepted specialist methods read-only. CostAnalyzer supplies cost scope/date/currency/commitment/uncertainty discipline; CapacityPlanner supplies load/measurement/bottleneck reasoning. Use actual available delegation only when needed, otherwise label the builder's own pass honestly. No agent body or shared domain-result interface is changed by this format procedure.
Recalculation, persistence and actual inspection
Collect observations from the same exact workbook revision. Read formula text and typed values from the actual provider, change one controlled input, and verify dependent results change. Independently inspect persisted OOXML formula and cached value/type, then reopen a saved copy through the permitted application/API. Edit/read back a cell and verify retention. Preserve the original artifact and clearly identify the new version after any edit.
The release's small synthetic case is quantity 3 times 12.5, expected 37.5, then quantity 4, expected 50, with a dependent cross-sheet SUM. These expected values are an independent test oracle, never constants written to formula cells or caches as a substitute for actual engine behavior. A later input change or edit requires fresh observations.
Native observations on 2026-09-22 showed correct live results and dependency
recalculation but empty persisted <v/> formula caches. That is a real useful
native calculation path with a blocked persistence gate, not total XLSX
support. Writing Python-computed or copied values into the cache is not an
authorized workaround. No unlisted recalculate/save/export operation is assumed.
Check live errors and unsupported behavior even when file creation succeeded: division by zero, unsupported function, circular dependency, missing sheet/range, empty/unknown result and stale cached value cannot pass. Missing-sheet references were observed to return an empty value in this host, so empty is not zero or proof that the formula succeeded.
Inspect the rendered grid/print pages and any charts through an actual available,
permitted renderer: clipping, #### displays, long-text wrapping, row/column
dimensions, table headers, units, color/contrast and print-page boundaries.
Model dimensions or a wrapText flag do not prove legibility. The observed canvas
has no explicit grid-render/export action; missing rendering stays unverified.
Never start Excel/LibreOffice or another Office application without its own
authority and permission.
Read-only formula/cache integrity check
scripts/check_xlsx.py consumes only the exact selected local workbook and an
explicit expectation file. It reads all worksheets, including hidden sheets,
retains the literal/formula distinction and checks cached types/values and exact
source text. It does not evaluate formulas, publish caches, modify the workbook,
render anything or establish release clearance.
python <trusted-source>\skills\generate-xlsx\scripts\check_xlsx.py \
--root <owned-artifact-root> --workbook output.xlsx --expect expectations.json
Use the host's native path/line-continuation syntax. The paths after --workbook
and --expect are relative to the explicit owned root. A minimal expectation:
{
"cells": [
{"sheet": "Inputs", "cell": "B2", "formula": null, "type": "number", "value": 4},
{"sheet": "Calculation", "cell": "B2", "formula": "=Inputs!B2*Inputs!B3", "type": "number", "value": 50},
{"sheet": "Summary", "cell": "B2", "formula": "=SUM(Calculation!B2)", "type": "number", "value": 50}
],
"required_text": ["<exact source-backed limitation>"]
}
Include every formula and all relevant input/source cells, not just a favorable
sample. The numbers above are expectation data, not executed results. Missing
cache is unverified; stale values, changed formulas/types, unselected formulas
and lost source text fail. Omitted formula type and explicit t="normal" are
equivalent ordinary formulas. Shared/array/data-table types, unknown types or
other uninterpreted formula attributes remain unverified. An empty-string
formula cache is still ambiguous, not guessed to be a computed empty result.
Before decoding cells, the checker requires a valid [Content_Types].xml,
unambiguous declarations for every part, and matching declared kinds for the
workbook, worksheets, shared strings and supported style/theme relationships.
Every internal relationship is checked, including unconsumed ones: its owning
part must exist, its target must resolve inside the package to a present part,
and its declared kind must be consistent. Missing/malformed metadata, duplicate
declarations and traversal/directory/query/fragment targets are refused. This is
a bounded canonical package-URI reader; percent-escaped targets require another
explicitly supported reader, not silent URI reinterpretation.
Forbidden VBA/macro-enabled, OLE/embedded-package and external-link mechanisms are rejected by content-type/relationship declarations as well as the retained filename guards. Renaming an inert or executable payload cannot erase those declarations. This checks declared package integrity; it is not a general malware scanner, full OOXML schema validator or permission to open untrusted content.
Shared strings need the actual sst namespace/root and valid string entries.
Cell types and single formula/value/inline payload combinations are validated
before any missing-value branch. Legitimate blank cells, shared/inline strings,
rich-text runs and literal formula-looking text are retained. The native
phoneticPr formatting metadata is accepted without inventing phonetic text;
unsupported phonetic annotations remain explicit errors.
DTD/entity declarations in the XML parts decoded by this reader are refused through the existing parser's declaration hook after encoding recognition, including UTF-8 and UTF-16LE/BE. Ordinary non-DTD UTF-16 and declaration-looking literal text remain readable. No external entity resolution, custom XML framework or calculation engine is introduced. The unchanged 64 MiB input/total-expanded, 16 MiB part and 10,000-entry bounds refuse oversized input without truncation.
Exit 0 means only the declared persisted-integrity checks passed; exit 3 means failed/unverified integrity and exit 2 is invalid/unreadable/unsupported input. Record this result as P05 evidence alongside native calculation, edit/reopen and render observations. Do not replace P05's control schema with this local diagnostic.
Reusable patterns
Follow the reusable pattern consumer contract. This remains the working workbook provider; patterns do not change its status. Resolve, or verify a supplied attachment, before composing sheets. Mandatory clauses (for example required sheets or columns) shape the composition, and each is reported by clause ID with the inspection evidence above, or as unverified when it was not inspected.
Status and handoff
- DONE: source retention, real calculation, required persisted caches, editable reopen and all requested rendered/configured controls verified.
- DONE_WITH_CONCERNS: required observations complete, only advisory concerns.
- BLOCKED: a required observation fails, errors or remains unverified. Retain the editable/source artifacts with exact remaining gates.
- NEEDS_CONTEXT: missing source, output authority or unresolved data semantics.
Report exact paths/hashes, original work/profile references, provider/tool/actions, formula/input changes, actual typed live results, persisted-cache inspection, reopen/editability, rendered coverage and unresolved boundaries separately. Keep qa-report.json and original P05 receipts; never replace historical failures with a later passing-looking summary. PDF/Visio and shared theme/pipeline binding remain their own releases, not prerequisites invented for this standalone route.