linexcel 0.7.0

linexcel 0.7.0

Release backfillée (NOU-11) — 27 commits depuis v0.6.0. Notes : section CHANGELOG ci-dessous.

[0.7.0] — 2026-08-02

Added

• Token accounting: result.token_usage tallies every document() and
  document_workbook() call, exposing input_tokens, output_tokens,
  total, requests and a readable str(). Recovered from the abandoned
  feat/agnostic-ai branch, whose provider abstraction was superseded

• Counts are read from the provider when it reports them — Gemini's
  usage_metadata and the OpenAI-compatible usage block — so the figure
  matches what is billed. They are approximated only as a fallback, and
  TokenUsage.estimated then flags the tally

• aidoc.estimate_tokens(), the fallback estimator, counts CJK characters
  individually instead of letting \w+ swallow a spaceless Japanese or Chinese
  sentence as a single word

• Optional usage= accumulator on aidoc.document_nodes() and
  aidoc.document_workbook(), and a UsageReportingProvider protocol that
  custom providers may implement to report real counts. Return types are
  unchanged, so existing callers are unaffected

• Seven more languages for language=, covering both the AI prompts and the
  viewer interface: es, de, it, pt, nl, ja, zh (with en and
  fr, nine in total). The set stays a closed allowlist — language selects a
  stored system prompt and reaches the generated JavaScript, so free-form input
  would be a prompt-injection and interpolation vector

• linexcel.i18n, single source of truth for the language list and the viewer
  interface strings, which used to be hand-maintained inside the JS template.
  The suite asserts i18n.UI_STRINGS and the two aidoc prompt registries
  cover the same languages and the same keys

• Translation provenance is stated in the README, the AI guide and the
  linexcel.i18n docstring: en and fr are hand-written, the other seven
  were produced with AI assistance and are not natively reviewed

• Multi-provider AI support: Google Gemini (default), OpenAI-compatible endpoints (Ollama, vLLM, LM Studio), and custom callables

• base_url and provider parameters on document() and document_workbook()

• Environment variables: LINEXCEL_AI_BASE_URL, LINEXCEL_AI_MODEL

• Optional dependency: pip install linexcel[openai]

• README: multi-provider documentation section

• linexcel.__version__, read from the installed distribution metadata

• max_workers on document() to tune AI request concurrency

Fixed

• Workbook screenshots now work on Windows and macOS, not only Linux.
  save_screenshots() looked for LibreOffice and pdftoppm on PATH alone,
  which neither the Windows nor the macOS installer extends, so the renderer
  reported itself as missing on machines where it was installed. Both binaries
  are now also looked up in their standard install directory, including
  winget's package store for Poppler
• On Windows the renderer invokes soffice.com rather than soffice.exe:
  soffice.exe detaches and returns immediately, so the conversion was reported
  as finished before the PDF existed and the run failed with "LibreOffice did
  not produce a PDF"
• Screenshot rendering uses a throwaway LibreOffice user profile
  (-env:UserInstallation). A LibreOffice already open on the desktop owns the
  default profile, and the headless process would exit successfully having
  converted nothing
• save_screenshots() reports which of the two binaries is missing and gives
  the install command for the running platform, instead of naming both and
  assuming Debian
• The screenshot pane hard-coded its heading and its Page N buttons in French,
  so every report in the eight other languages was partly French. Both now go
  through linexcel.i18n, which gains a page string
• The workbook-context example in the documentation iterated
  result.workbook_context as if it were keyed by sheet name; per-sheet context
  lives under its sheets key
• linexcel.PackageNotFoundError was never meant to be public; the
  importlib.metadata import is now private
• provider= now accepts a plain callable, as the documentation always claimed;
  previously only objects exposing a generate method worked
• VBA call edges were never emitted: the procedure lookup was keyed on
  Module.Name while the call scanner reports unqualified names. Calls now
  resolve in the calling module first, then across modules, and an ambiguous
  name stays unresolved instead of pointing at an arbitrary module
• VBA call detection is case-insensitive throughout, matching the language:
  procedure lookups are keyed on the lowercased name, so two modules declaring
  Taux and taux no longer capture each other's calls nor defeat the
  ambiguity guard, and a function's own return assignment (Taux = 1) is not
  mistaken for a call
• A dotted name is a VBA call only when the qualifier is a module
  (Module1.Taux, now resolved exactly); .Value or .Count is member access
  and no longer produces an edge to a similarly named procedure
• document() no longer discards every successful card when a single node
  fails; failures are reported through a UserWarning and AiDocError is
  raised only when all nodes fail
• to_html() no longer fails on a LineageResult built without the workbook
  bytes; the sheet-preview tab is simply omitted
• Viewer template placeholders are substituted in a single pass. Chained
  replacements let __TITLE__ and __LANG__ be rewritten inside the
  already-injected graph, so a cell, sheet name, or comment containing either
  literal was silently replaced by the title — and a title ending in a
  backslash truncated the embedded JSON, blanking the viewer
• The cells interface string was used twice by the viewer but defined in no
  locale, so reports fell through to the raw key ("1 000 cells" in French).
  The unused sheets_summary_title key was dropped

Changed

• Breaking: render_html(), to_html() and save_html() reject a
  language outside ("en", "fr") with ValueError. It was previously
  interpolated raw into the generated JavaScript, where it could terminate the
  string literal; unknown values used to fall back to English at runtime
• pytest now collects the src/ doctests, which testpaths used to exclude
• Generated validation artifacts are no longer tracked in the repository
• AI data-handling notes describe every provider, not just Gemini
• save_screenshots(timeout=) defaults to 180 seconds instead of 60. A first
  headless run builds the LibreOffice profile from scratch, which alone can
  exceed the old budget on Windows
• linexcel.insights exposes find_libreoffice() and find_pdftoppm(), so a
  caller can check the renderer is available before analyzing a workbook
• linexcel.i18n gains a page string, used by the screenshot page switcher

[0.3.0] - 2026-07-14

Added

• Localization: en (default) and fr language support for AI documentation and UI
• Sidebar worksheet overview in HTML report
• Sheet-specific screenshot embedding in HTML report
• Workbook-level AI documentation (document_workbook())
• SECURITY.md, THIRD_PARTY_NOTICES.md

Changed

• Refined AI data handling documentation and security notes

[0.2.2] - 2026-07-10

Changed

• Improved formula decomposition and step evaluation
• Refined dependency graph edge resolution for named ranges

[0.1.0] - 2026-07-10

Added

• Initial release
• Formula extraction via formualizer (Rust engine)
• Stretched pattern grouping via R1C1 canonicalization
• Dependency graph: cells, ranges, defined names, VBA procedures
• Step-by-step composite function evaluation
• Standalone HTML viewer (Cytoscape.js, fully offline)
• Optional AI documentation via Google Gemini
• LibreOffice screenshot rendering
• PyPI trusted publishing