latex-editing

name	latex-editing
description	Edit, review, and clean up LaTeX source files. Covers AI-generated LaTeX corruption patterns, patch-tool pitfalls with escaped characters and macros, fallback editing strategies when patch fails, and systematic QA review for logical coherence and typesetting correctness.
version	1.0.0
author	Hermes Agent
license	MIT
platforms	["linux","macos","windows"]
metadata	{"hermes":{"tags":["latex","editing","patching","cleanup","ai-corruption","typesetting","review","qa"]}}

LaTeX Editing

This skill covers common problems when AI agents edit LaTeX source files: corruption from automated updates, patch-tool failures due to escaping, and systematic review workflows.

---

AI Cleanup — Fixing AI-Generated LaTeX Corruption

When an AI has updated LaTeX files, specific corruption patterns often appear.

Common Corruption Patterns

1. Escaped newlines (\\n): AI sometimes inserts literal backslash-n instead of a newline.

   • Fix: replace \\n with actual newline characters.

2. Non-ASCII symbols: Smart quotes, em-dashes, or Unicode math symbols may be inserted.

   • Fix: replace with LaTeX equivalents (---, ` ``, "" for quotes; \text{...} for Unicode).

3. Broken math environments: AI may split \begin{equation} across lines or mismatch delimiters.

   • Fix: verify every \begin{...} has a matching \end{...}.

4. Over-escaped backslashes: AI may double-escape (\\alpha instead of \alpha).

   • Fix: audit math mode for \\ sequences that should be \.

Verification Steps

1. Run pdflatex or latexmk and check for errors.
2. Inspect the .log file for missing font or encoding warnings.
3. Diff against the last known-good version to spot unintended changes.

---

Patch Pitfalls — Editing LaTeX with the Patch Tool

The patch tool (and Hermes's patch action) often fails on LaTeX because of:

1. Escaped characters: backslashes, braces, and dollar signs confuse fuzzy matching.
2. Macros with parameters: \newcommand{...} blocks may span multiple lines.
3. Whitespace sensitivity: LaTeX ignores most whitespace, but patch does not.

Strategies

• Use replace_all=false (default) and provide extra context lines so the match is unique.
• Quote the old_string carefully: if it contains backslashes, double-escape in the JSON payload.
• Prefer sed or Python for single-line macro replacements when patch keeps failing.
• Break large changes into smaller patches: replace one environment at a time.

---

Handling Patch Failures — Fallback Editing

When standard patch tools fail on LaTeX due to hidden characters, escaping, or complex formatting:

1. Switch to line-numbered replacement: use read_file to find exact line numbers, then use write_file with a Python script to surgically replace the region.
2. Use Python's re module: for pattern-based replacements that don't need exact text matching.
3. Use latexdiff for review: generate a visual diff of old vs new before committing.
4. Preserve a backup: always keep .bak before bulk edits.

Example: Surgical Python Replacement

with open('paper.tex') as f:
    lines = f.readlines()
# Replace lines 45-50
lines[44:50] = ['new content\n']
with open('paper.tex', 'w') as f:
    f.writelines(lines)

---

Review QA — Systematic LaTeX Review

Review LaTeX source files for both logical coherence and correct technical typesetting, specifically detecting AI-generated artifacts.

Checklist

• All citations resolve (\cite{...} keys exist in .bib)
• Figures and tables are referenced before they appear
• Math environments are balanced (no unclosed $, \[, \begin)
• No smart quotes or non-ASCII in math mode
• Section hierarchy is consistent (\section > \subsection > \subsubsection)
• Tables use booktabs rules (\toprule, \midrule, \bottomrule) not vertical lines
• Algorithm environments compile without algorithmicx version mismatches
• No hard-coded figure widths that break two-column layouts

AI Artifact Detection

• Search for TODO, FIXME, or placeholder text left by AI.
• Check for hallucinated citations — verify every \cite against the .bib file.
• Verify that rewritten sections still match the original meaning (semantic drift).