Community integration: docling-skill for Codex and Claude Code

[realraelrr]

Hi Docling team,

I built an external community integration on top of Docling: docling-skill, an agent-oriented ingestion skill/package for Codex, Claude Code, and similar LLM agent runtimes.

Repository: https://github.com/realraelrr/docling-skill
Latest release: https://github.com/realraelrr/docling-skill/releases/tag/v0.1.2

What it adds around Docling:

• A manifest-first agent contract so agents inspect extraction quality before reading Markdown.
• Standard sidecar outputs: source.md, source.docling.json, source.images.json, source.manifest.json, and source.meta.json.
• Stable image placeholders such as [[image:picture-p3-0]] with a sidecar map for multimodal agent workflows.
• Quality states for downstream routing: good, salvaged, and failed_for_agent.
• A single SKILL.md source that can be installed into both Codex and Claude Code via symlink.

The project intentionally stays a thin external workflow layer, not a Docling fork or official distribution. It depends on the official Docling package and currently tracks docling>=2.93.0,<2.94.0.

I tested v0.1.2 with Docling 2.93.0 on PDF, DOCX, and HTML sample conversions, plus the project test suite. The outputs were agent-ready and kept source.docling.json as the authoritative structured recovery artifact.

I am sharing this here in case it is useful to other users experimenting with Docling inside coding-agent workflows. Feedback is welcome, especially if there are better conventions for exposing Docling outputs to agent runtimes.

If this belongs somewhere else, I am happy to move or reframe it.