Configuration

All configuration options for the GitHub Action and local CLI.

Action Inputs¶

LLM Model¶

The action uses Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) with extended thinking:

• Extended thinking lets the model reason internally before responding, eliminating false positives (0% FP rate)

• Requires temperature=1.0 (Anthropic constraint for extended thinking)

• Thinking budget: 10,000 tokens

• Minimal rule-agnostic prompt (~40 lines) + detailed rules (120–235 lines) per category

Rule Types¶

The style guide uses a three-tier type system:

rule — Actionable ✅¶

Clearly actionable rules that are automatically applied. These are mechanical, objective checks (32 rules):

• Writing: One sentence per paragraph, capitalization, bold/italic usage

• Mathematics: Transpose notation (\top), matrix brackets, sequence notation

• Code: Unicode Greek letters, package installation placement

• Figures: Caption formatting, figure naming, axis labels, line width

• References: Citation style, cross-references

• Links: Internal vs cross-series linking syntax

• Admonitions: Exercise/solution syntax, nested directive tick counts

style — Advisory 💡¶

Subjective guidelines that require human judgment (13 rules):

• Writing clarity and conciseness

• Logical flow between sections

• Visual element opportunities

• Figure size decisions

migrate — Code Modernization 🔄¶

Legacy patterns that should be updated (4 rules, JAX and code categories only):

• tic/toc → quantecon.Timer() context manager

• %timeit → quantecon.timeit() function

• NumPy in-place operations → JAX functional updates

• Implicit random state → explicit JAX PRNG key management

rule types are automatically applied. style and migrate types are reported as suggestions.

Sequential Category Processing¶

The action processes categories one at a time, feeding the updated document from each category into the next:

1. Category 1 (e.g., Writing): Reviews original document → finds violations → applies fixes → updated document

2. Category 2 (e.g., Math): Reviews updated document → finds violations → applies fixes → updated document

3. Continue for all requested categories...

This ensures:

• All fixes are applied without conflicts

• Later categories see changes made by earlier categories

• More coherent and complete final output

Trade-off: Sequential processing is slower than parallel (8 sequential API calls vs 8 parallel), but produces more reliable results.

Review Modes¶

Single Mode¶

Reviews a single lecture file. Triggered by issue comments:

@qe-style-checker lecture_name
@qe-style-checker lecture_name writing,math

Bulk Mode¶

Reviews all lectures in the lectures-path directory. Typically used with scheduled workflows:

- uses: QuantEcon/action-style-guide@v0.7
  with:
    mode: 'bulk'
    lectures-path: 'lectures/'

PR Creation¶

When create-pr is true (default), the action:

1. Creates a new branch: style-guide/{lecture}-{timestamp}

2. Commits all applied fixes with detailed messages

3. Opens a PR with:

   • Summary of all changes by category

   • Style suggestions as a PR comment

   • Labels: automated, style-guide, review