Agent Skill

Extract and configure

Discover the installed interface with afpdf --help, afpdf --docs, and afpdf config show. Before the first extraction on a machine, run afpdf engine status; if the library is not available, run afpdf engine install once, which fetches the pinned PDFium build for this platform, verifies it, and probes it. Use afpdf extract INPUT; document facts, organized article structure, and continuous text are returned by default in AFDATA events. Read diagnostics and each page’s status before trusting text: a capped, failed, or scanned page is reported, never hidden.

• Add --set output.basic=true for per-page raw text, glyph styles, and geometry.
• Add --set extract.image_mode=embedded for encoded bitmap payloads, or render_pages for full-page renders. Bytes appear once in images as base64.
• Use --set extract.processing.remove_page_furniture=false to keep furniture.
• Use config show/get/set/unset and dot paths for durable settings; extraction’s --set KEY=VALUE is invocation-only. The default file is .afpdf/config.json in the current directory; reads do not create it.

Isolated execution is the default. The worker memory limit is enforced on Linux only and defaults off elsewhere; do not set it nonzero on another platform. --stdout-file/--stderr-file work on Unix only; on Windows use --output-to or shell redirection. Report missing text layers and failed pages; there is no implicit OCR or external model.

Develop or integrate

Read README.md for design principles, docs/configuration.md for settings, and docs/design.md for the engine choice, result model, and limitations. Native extraction uses the engine feature; pure organization can rerun from BasicDocument without I/O. IsolatedParser requires the afpdf executable in the worker environment, without native linkage in its caller; its EngineOptions name the PDFium library the worker binds.

Every extraction rule needs a real-page fixture, and engine changes are accepted only after diffing old and new builds over the consumer corpus with every class of change explained. Build and run the engine inside a container. Services that ingest PDFs from the web use the isolated (child process) mode.

When implementing extraction, focus on article reading order, natural paragraphs without layout-induced line breaks, page-furniture removal, and source locations. Keep images, captions, tables, formulas, and footnotes distinguishable from prose. Indexing and semantic interpretation belong to callers.

Keep extraction in the Rust library and CLI behavior at the executable boundary. Discover current commands with afpdf --help and afpdf --docs. Use the fixed scripts/test.sh entrypoint for checks; prerequisites and modes are in CONTRIBUTING.md. After changing the CLI registry, regenerate its reference with scripts/test.sh docs.