PROJECT: action-code-style — deterministic black formatting for Python in MyST/markdown

Start by reviewing the .github/actions/code-style-checker/ implementation and test suite referenced in closed PR #240, then compare the architecture questions with issues #268 and #266. Establish the standalone repository and its relationship to myst-lint before implementation. Done means the documented extraction, black formatting, check/fix modes, configuration, and listed tests work in a pilot repository.

Summary

Build a standalone GitHub Action that runs black over Python code embedded in MyST {code-cell} directives and standard markdown fenced code blocks, keeping lecture code PEP8-compliant and consistently formatted. The tool is deterministic (no AI), cheap, and safe to run automatically.

This issue captures the essence of the closed Copilot PR #240 and the original request in #221, plus the architecture decision reached in discussion (a separate repo, not folded into action-style-guide), so the feature can be cleanly re-implemented once it has a home.

Background

• #221 requested automatic black style checks on Python contained in MyST code-cell directives and standard fenced markdown code blocks.
• #240 (Copilot) implemented a code-style-checker action inside meta (.github/actions/code-style-checker/), comment-triggered via @quantecon-code-style. It was never reviewed/merged and is structured the old (in-meta) way, contrary to the migration of actions into individual repos (#244). It has been closed in favour of this documented project.
• This is a real, still-unmet need: no current QuantEcon repo provides deterministic black formatting of code embedded in markdown. action-style-guide is adjacent but fundamentally different (see below).

Architecture decision: separate action-code-style, not part of action-style-guide

The two share the phrase "code style" but differ on every axis that matters for packaging:

Axis	black formatting (this project)	action-style-guide
Mechanism	Deterministic, fixed rules	AI (Claude), non-deterministic, confidence-scored
Dependencies	black only — no API key, free	LLM API key, cost per run
Auto-run	Safe on every PR / push	Comment-triggered only (cost + noise)
Cadence	Tiny, stable utility	Large, evolving (prompt versioning, rules DB)
Audience	Any MyST/Jupyter Book project	QuantEcon-opinionated

Decisive points:

1. Invocation policy differs — black wants to be cheap and always-on (a normal CI check); the AI style guide must stay comment-triggered to avoid being intrusive/expensive. One repo cannot cleanly serve both runtime profiles.
2. Coupling a stable formatter to a fast-moving AI tool creates release friction — every prompt/rule change would churn a repo that also houses a formatter that rarely changes.
3. Small, single-purpose repos match the #244 grain (action-link-checker, action-check-warnings, …).

The one genuine overlap is the MyST/markdown code-extraction parser (both need to pull Python out of {code-cell} and fenced blocks). If duplication becomes annoying, factor that into a small shared helper rather than merging two tools with opposite runtime profiles.

➡️ Home: a new standalone QuantEcon/action-code-style repo. (The standalone-vs-centralised question is being worked through more broadly in #266 — the final home should be consistent with whatever that lands on.)

Relationship to myst-lint (#268)

#268 proposes myst-lint, a deterministic MyST quality tool (wraps markdownlint + MyST rules: directive/role validation, math-delimiter matching, code-block delimiters, code-cell validation). It shares this project's exact design philosophy (deterministic, no API cost, fast, CI/pre-commit) but a different concern:

• myst-lint (#268) validates markdown structure (headings, delimiters, math, directives) — it checks, doesn't reformat code.
• action-code-style (this, #320) reformats Python code inside blocks with black (PEP8).

Open question to resolve before building: one tool or two? They could ship as separate siblings, or black-formatting could be a module/rule-set within myst-lint (one deterministic "MyST quality" tool with both a structure-lint pass and a code-format pass). Decide this jointly with #268 rather than in isolation.

Scope / requirements

(capturing the essence of #221 + #240)

• Parse md files and extract Python from both MyST {code-cell} directives and standard fenced ```python blocks
• Only format blocks tagged python, python3, ipython, ipython3; skip others (e.g. julia) with an informative log line
• Run black over each extracted block and write the formatted code back into the file, preserving the surrounding markdown/directive structure exactly
• Configurable inputs:
  • files — explicit list and/or glob patterns (e.g. lecture/**/*.md)
  • toggle MyST code-cell scanning (default on)
  • toggle standard markdown block scanning (default on)
  • python-languages (default python,python3,ipython,ipython3)
  • black-args (e.g. --line-length=88)
• Design intent: deterministic and cheap → support always-on / push-triggered CI usage, in addition to an optional comment trigger (@quantecon-code-style)
• Check mode (non-mutating): report diffs / fail CI without committing, so it can run as a status check
• Fix mode: commit formatted results, per-file, with a message like [{filename}] applying black changes to code (per #221)
• Tests: unformatted MyST+markdown, already-formatted (no-op), non-python (skip), glob expansion

Reference implementation

#240 (closed) is a usable starting point — it already had a working line-by-line MyST/markdown parser, language detection, glob handling, and a test suite. Re-implement cleanly in the new repo rather than porting the Copilot bulk wholesale.

Related

• #221 — original feature request (closed, superseded by this project)
• #240 — closed Copilot PR; reference implementation
• #244 — migration of actions into individual repos (the grain this follows)
• #268 — myst-lint, sibling deterministic MyST quality tool; resolve "one tool or two?" jointly
• #266 — mono-repo vs centralised actions vs hybrid (informs the home decision)
• QuantEcon/action-style-guide — adjacent AI style tool; intentionally kept separate
• QuantEcon/actions — build/preview/publish actions monorepo

Suggested next steps

1. Create QuantEcon/action-code-style (+ a test-action-code-style fixture repo, matching the action-style-guide pattern)
2. Port/clean the parser + black runner from #240
3. Implement both check mode (always-on CI) and fix mode (comment trigger / commit)
4. Pilot on one lecture repo, then roll out