Polyglot PMD

polyglot-pmd is a reusable Python implementation of the PMD 0.1 polyglot Markdown notebook specification in spec.md. It parses and validates plain-text .pmd documents, resolves their dependency graph, runs every cell in an isolated process, and renders a self-contained HTML report.

PMD cells execute arbitrary code with the privileges of the user running pmd. Review a document before running, testing, or rendering it.

Install

From this checkout:

python -m pip install -e ./pmd-impl
pmd check pmd-impl/example.pmd --graph
pmd run pmd-impl/example.pmd --fresh
pmd run pmd-impl/example.pmd --verbose --out-dir pmd-outputs
pmd test pmd-impl/example.pmd
pmd render pmd-impl/example.pmd --to html

Once published:

python -m pip install polyglot-pmd

Python 3.10 or newer is required. The package depends only on PyYAML and markdown-it-py. Language interpreters used by a document must also be installed.

Library API

from pathlib import Path

from pmd_notebook import Runner, load, render_html, validate

document = load("analysis.pmd")
diagnostics = validate(document)
errors = [item for item in diagnostics if not item.startswith("warning:")]
if errors:
    raise ValueError("\n".join(errors))

result = Runner().run(document, fresh=True)
if not result.ok:
    failed = [cell for cell in result.cells if cell.status == "failed"]

page, render_result = render_html(document, result)
Path("analysis.html").write_text(page, encoding="utf-8")

The public API exports parse, load, validate, closure, topological_order, graph_lines, Runner, Cache, execute, and render_html, plus the corresponding result dataclasses.

Context Bindings

Each cell receives PMD_CELL_OUT and PMD_CTX_FILE. Built-in engines add these bindings:

Shell reads print JSON, so a stored string includes JSON quotes. SQL uses an isolated in-memory SQLite database. Override commands under frontmatter engines.<language>.command; custom engines still receive the two environment variables but must provide their own context helpers.

Write .png, .jpg, .jpeg, .svg, .csv, .md, or any other attachment under PMD_CELL_OUT. The HTML renderer embeds all files and makes no network requests. Python also receives display.markdown, display.csv, display.image, and display.file convenience methods.

Dependency outputs

Every cell receives PMD_DEP_OUTPUTS, a JSON object mapping transitive dependency cell IDs to temporary output directories. Python cells also receive an outputs helper:

chart = outputs.path("make-chart", "chart.png")
all_files = outputs.files("make-chart")

Paths are read-only by convention and remain available for the duration of the run. Cached dependency attachments are reconstructed before downstream cells start, so tests behave the same with warm and cold caches.

Use pmd run --out-dir PATH or pmd test --out-dir PATH to retain attachments after the run. Files are exported under PATH/<cell-id>/. Use --verbose to print successful cells' captured stdout and stderr in addition to statuses.

Caching

Successful dependency results are cached under PMD_CACHE_DIR, or ~/.cache/polyglot-pmd by default. Keys include source, attributes, engine command, and the resolved transitive context. --fresh bypasses reads. A cell named by --cell always executes; only its dependencies may come from cache. Context itself remains scoped to one invocation.

External files are not inferable from arbitrary source code. Declare them in frontmatter so their content hashes participate in every cell's cache key:

inputs:
  - data/games.parquet
  - config.json

Paths resolve relative to the .pmd document. Files and directories are supported; directory fingerprints include every contained file. Missing declared inputs fail before any cell executes.

Portable Engine Commands

Frontmatter engine commands expand environment variables and the {document_dir} placeholder. Relative executable paths also resolve from the document directory:

engines:
  python:
    command: "{document_dir}/.venv/Scripts/python.exe"

On POSIX, the corresponding command would normally end in .venv/bin/python. This selects an existing environment; PMD still does not provision packages.

Optional Workbench

Run python server.py and open http://localhost:8765. The local workbench is not installed as part of the Python package.

Publishing to PyPI

1. Replace package author metadata if desired and choose the final project URL.
2. Run python -m pip install -e ".[dev]".
3. Run pytest and python -m build.
4. Check artifacts with python -m twine check dist/*.
5. Upload to TestPyPI, install-test the wheel, then upload to PyPI.

python -m twine upload --repository testpypi dist/*
python -m twine upload dist/*

PDF and .ipynb are optional PMD render targets and are intentionally not implemented. The CLI refuses them clearly instead of silently losing content.