LT2MD 1.3.2 — Figures as text, Markdown repair, long-PDF conversion

This is the first GitHub Release for LT2MD, shipping the second version of the project.

The early version was built around one idea: when a PDF contains a figure, do not leave an empty image link. Have the agent look at the figure and write what it shows as text beside the caption. That idea is still the core, but this version strengthens and supplements nearly everything around it.

"Looking at the figure" is not OCR. OCR only lifts the characters printed inside an image. LT2MD asks a vision model to understand what the image contains: the objects in a diagram, how labels, arrows, and axes correspond in a physics or mathematics figure, the way boxes connect in a flowchart. The description records only what is directly visible, with no conclusion added beyond the source. That written understanding sits next to the caption, so every later read happens in plain text, by people or by text-only models, without sending the image again.

Figures are understood, then written as text

Flowcharts, diagrams, charts, and physics and mathematics figures are read into a passage of text right next to each caption, covering the objects, labels, arrow directions, subfigures, and the relationships between them. No matter how many times the Markdown is read later, the images never have to be sent to a model again. A text-only model is enough, which is fast and cheap.

It repairs Markdown converted before

Many people already have a folder of Markdown produced from PDFs by other tools, marred by typos, missing figures, broken formulas, or mismatched page numbers: too useful to throw away, too risky to trust. LT2MD treats that old Markdown as a candidate, not as fact. It goes back to the PDF and checks it block by block, re-recognizing what is wrong, missing, or out of order, keeping correct passages and adding descriptions to figures that have none. On delivery it reports which pages were touched and what was wrong, so every change can be checked.

Long PDFs stay complete end to end

It works in small batches, each saved and recoverable, so even a few-hundred-page PDF comes out consistent. After each batch, a fresh pass rereads the source pages and checks the candidate, focusing on formulas, figures, cross-page paragraphs, and illegible characters. Only the genuinely risky spots get a third read: low-resolution areas, dense formulas, multi-panel figures, and cross-page seams. Every paragraph, formula, and figure is labeled with its PDF page. Scanned pages, math formulas, and multi-column layouts are restored in reading order, and when a glyph is unclear LT2MD adds a transcription note with the page and exact point of doubt instead of guessing.

Also

• The only dependencies are Python and two page-rendering libraries (pypdfium2 and Pillow). They are not OCR engines, so there is no need to install Tesseract. It runs on CPU and works on Windows, Linux, and macOS, with a one-line install.
• Correctness comes from the workflow and three static validator scripts. The model only needs to understand images, so a small vision model is enough.
• Instructions embedded inside a PDF are treated as document content, not as commands to the agent. The local scripts do not transmit files; before any page is sent to a remote visual service, LT2MD stops and asks for explicit permission.
• Licensed under AGPL-3.0. The core of this skill is entirely crafted by Codex.

About this release

This is the first GitHub Release.

• Core workflow. This is the main part of the release. PDFs are converted in small saved batches with an independent review pass, a risk-driven third read, a separate format review, and static validation as final gates, so even very long documents stay complete and consistent. Every output block carries its PDF page anchor, and each figure is understood and written as text beside its caption. The same source-grounded checks also run in reverse to repair a Markdown file converted earlier: it is checked against the PDF block by block, and anything wrong, missing, or out of order is re-recognized. The 1.3.x entries below only document, package, and patch this workflow.
• 1.3.0 rewrote both the English and Simplified Chinese READMEs from the ground up, documenting the long-document batch-and-review workflow, figure-understanding-as-text, and the repair workflow, and added cross-platform install instructions and use cases.
• 1.3.1 added the Codex skills directory to the manual install paths and ignored the local agent workspace.
• 1.3.2 backfilled the changelog for 1.3.0 and 1.3.1.

To install, send this sentence to your agent.

Install this skill for me: https://github.com/libnyx/LT2MD