Turn any technical book PDF into a Claude Code skill — ready to study, reference, and use while you work.
Python agent-skillsai-agentsbook-to-skillcontext-engineeringdocument-processingedtechknowledge-baseknowledge-managementllmpdf-to-markdownragself-studystudy-tools
fix(evals): keep text before the first chapter heading reachable (#286)
yi111cf06be13b86c
diff --git a/tests/evals/test_paper_flat.py b/tests/evals/test_paper_flat.pyindex 0481ebe..6f91ed5 100644--- a/tests/evals/test_paper_flat.py+++ b/tests/evals/test_paper_flat.py@@ -70,6 +70,24 @@ def test_no_heading_uses_one_full_source_chunk(tmp_path):assert (tmp_path / "pack" / "chunks" / "chunk-01.md").read_text(encoding="utf-8") == "Plain synthetic source.\n"+def test_text_before_the_first_heading_stays_reachable(tmp_path):+ text = "Preface with a unique safety constraint of 42 degrees.\n\nChapter 1: Only\n\nBody.\n"+ source = tmp_path / "source.txt"+ metadata = tmp_path / "metadata.json"+ source.write_text(text, encoding="utf-8")+ metadata.write_text(json.dumps({"title": "Prefaced", "description": "Preface fixture.", "chunks": [{"description": "Preface and chapter.", "summary": "One chunk."}]}), encoding="utf-8")+ build(tmp_path, source, metadata)+ chunks = sorted((tmp_path / "pack" / "chunks").iterdir())+ assert len(chunks) == 1+ assert chunks[0].read_text(encoding="utf-8") == text+ assert "42 degrees" in chunks[0].read_text(encoding="utf-8")+++def test_chunk_payloads_concatenate_back_to_the_source(tmp_path):+ text = "Preface constraint 42 degrees.\n\nChapter 1: Observe\n\nMeasure.\n\nChapter 2: Respond\n\nApply.\n"+ assert "".join(paper_flat._chunks(text)) == text++def test_pack_has_no_child_skill_hierarchy(tmp_path):build(tmp_path)files = sorted(diff --git a/tools/evals/paper_flat.py b/tools/evals/paper_flat.pyindex 96c92bb..062f812 100644--- a/tools/evals/paper_flat.py+++ b/tools/evals/paper_flat.py@@ -22,7 +22,15 @@ def _chunks(source: str) -> List[str]:starts = [index for index, line in enumerate(lines) if _CHAPTER.match(line.rstrip())]if not starts:return [source]- return ["".join(lines[start:end]) for start, end in zip(starts, starts[1:] + [len(lines)])]+ chunks = ["".join(lines[start:end]) for start, end in zip(starts, starts[1:] + [len(lines)])]+ prefix = "".join(lines[: starts[0]])+ if prefix:+ # Fold any text before the first heading (preface, abstract, held-out+ # constraints) into the first chunk instead of dropping it, so reading+ # every chunk still reaches the whole source and the chunk count keeps+ # matching the supplied metadata.+ chunks[0] = prefix + chunks[0]+ return chunksdef _metadata(path: Path, chunk_count: int) -> Dict[str, Any]: