v1.2.1

Patch release that freezes the contract hardening and CI fixes on main.

• Adds v1.2.1 package/version pins in project metadata and README install commands.
• Keeps manifest producer.version aligned with the package version used by the source tree.
• Includes the Python 3.10 CI fix for test metadata parsing.
• Verified by GitHub CI on Python 3.10, 3.11, and 3.12.

v1.2.0

Highlights

• Add Docling-native PPTX and common image input support (png, jpg, jpeg, tif, tiff, bmp, webp).
• Add targeted CJK Markdown normalization for agent-facing source.md while preserving raw source.docling.json.
• Add text_normalization and text_integrity quality signals for CJK cleanup, replacement characters, formula placeholders, and residual compatibility glyphs.
• Refine PDF page-quality aggregation so isolated long-document page failures become medium-risk warnings, while short documents and high failed-page ratios remain failed_for_agent/high.
• Refactor non-PDF conversion internals to share the sidecar attempt builder while preserving format-specific routing and PDF remediation behavior.
• Add lightweight project-owned TypedDict contracts and a ruff dev check for import/unused/syntax hygiene.

Verification

• conda run -n docling python -m ruff check . -> passed.
• conda run -n docling python -m pytest -> 126 passed.
• Root, .codex, and .claude skill validators passed.
• Subagent review found no Critical, Important, or Minor issues.

v1.1.0

Release 1.1.0 makes docling-skill's quality contract more honest and useful for agent workflows.

Changes:

• Adds additive quality risk evidence fields to source.manifest.json: risk_level, warnings, gate, limitations, and grouped signals.
• Keeps the existing quality.status, quality.agent_ready, and quality.reasons fields for compatibility.
• Downgrades quality semantics from implicit trust to risk screening: low risk means no hard failure was detected, not semantic fidelity proof.
• Aggregates PDF page-level quality so local page failures raise document risk, and severe page failures can still fail the document.
• Updates CLI summaries to print risk_level.
• Updates the bundled skill and READMEs to instruct agents to inspect status, risk_level, warnings, and signals.
• Tracks Docling 2.96.x and docling-core 2.77.x.

Validation:

• conda run -n docling python -m pytest -q -> 99 passed.
• Root, Codex, and Claude skill validators passed.
• conda run -n docling python -m pip check -> no broken requirements.
• Real document smoke/evaluation covered PDFs and HTML samples from local machine-learning papers; classifications separated low-risk usable documents, medium-risk local page issues, and high-risk OCR-noise slides.

v1.0.3 (superseded)

Superseded by v1.1.0: https://github.com/realraelrr/docling-skill/releases/tag/v1.1.0

This release is retained for existing installs pinned to @v1.0.3, but it is no longer the recommended install target. Use @v1.1.0 for the Docling 2.96 update and the manifest quality risk evidence model.

Original release note:
Release 1.0.3 adds pragmatic spreadsheet ingestion for xls, xlsx, and csv, keeps xlsm and complex workbook cases as manual preprocessing scope, documents source.docling.json as the authoritative spreadsheet artifact, and expands regression coverage and quality gates.

v0.1.2

Changes:

• Update Docling dependency window to docling 2.93.x and docling-core 2.74.1+.
• Add README acknowledgements for the upstream Docling project.
• Simplify and reorganize the English and Chinese READMEs around the reader path: what it does, install, use, agent setup, contract notes, scope, and acknowledgements.
• Update install examples to v0.1.2.

Verification:

• conda run -n docling python -m pytest: 50 passed.
• quick_validate.py passed for Codex and Claude Code skill paths.
• Real PDF, DOCX, and HTML conversions produced good, agent-ready outputs.
• git diff --check passed for the README cleanup.

v0.1.1

Release v0.1.1 adds shared agent framework support and direct setup instructions for Codex, Claude Code, and other LLM agents while keeping the source.* ingestion contract unchanged.

Highlights:

• Add Claude Code allowed-tools metadata while keeping Codex-compatible skill validation.
• Document Codex and Claude Code support in English and Chinese READMEs.
• Add agent-facing setup instructions for cloning, installing, symlink-based skill configuration, validation, and sample conversion checks.
• Ignore .state/ and remove the previously tracked .state/progress.md from the release tree.
• Support both conda module invocation and pip-installed docling-skill CLI in SKILL.md.

Verification:

• quick_validate.py passed for ~/.codex/skills/docling-skill.
• quick_validate.py passed for ~/.claude/skills/docling-skill.
• git diff --check passed.
• conda run -n docling python -m pytest: 50 passed.

v0.1.0

Initial public release of docling-skill.

Highlights:

• agent-first PDF ingestion layer on top of Docling
• markdown, image sidecars, and manifest-based quality gating
• OCR remediation and page-level remediation
• English and Chinese READMEs
• public installation instructions for the v0.1.0 release

Verification:

• PYTHONPATH=src conda run -n docling python -m pytest tests -q
• conda run -n docling python /Users/rael/.codex/skills/.system/skill-creator/scripts/quick_validate.py .