Setup Review

For arXiv URLs (if the user provides one containing arxiv.org):

• Extract the arXiv ID
• Check if an HTML version exists at https://arxiv.org/html/{id}
• Prefer the HTML version over PDF for better parsing quality

Create .review/config.json:

{
  "paper_repo": "<url-or-local>",
  "paper_commit": "<sha>",
  "manuscript_path": "<path-to-main-file>",
  "manuscript_format": "tex|md|html|docx|pdf",
  "has_code": true|false,
  "has_data": true|false,
  "code_paths": ["code/", ...],
  "data_paths": ["data/", ...],
  "models": ["claude-sonnet-4-6", "gpt-4o", "gemini-2.5-pro"],
  "reviewed_at": "<ISO timestamp>"
}

Step 3: Parse manuscript

Run the parse_document.py script to convert the manuscript to one-sentence-per-line markdown.

Look for the roboree scripts path in the system prompt (it will say "roboree scripts are at: <path>"). Use that path:

python <r2-scripts-path>/parse_document.py \
  --input <manuscript-path> \
  --output-dir .review/parsed/ \
  --symbol-table .review/symbol_table.json

The script handles:

• LaTeX: native section splitting (preserves raw LaTeX, no pandoc)
• HTML/docx/PDF: markdown conversion
• Sentence splitting (one sentence per line, preserving math blocks)
• Symbol table extraction (via edsl, or regex fallback)

Step 4: Create review structure

mkdir -p .review/chunks .review/holistic

Verify that .review/issue_spotters/ was created by the r2 CLI launcher. If not, check the system prompt for the issue spotters path and copy them manually.

List the available issue spotters and tell the user they can customize them before running /broad-sweep:

ls .review/issue_spotters/

Create a .gitignore that excludes .env:

echo ".env" > .gitignore

Step 5: Write overview.json

Write .review/overview.json summarizing what was found:

{
  "manuscript_format": "<format>",
  "sections_parsed": <N>,
  "symbols_extracted": <N>,
  "has_code": true|false,
  "has_data": true|false,
  "issue_spotters": ["<list of spotter names>"]
}

Step 6: Commit

git add -A
git commit -m "review: setup for <repo>@<sha>"

Report what was found: manuscript format, whether code/data exist, number of sections parsed, number of symbols extracted, and the list of issue spotters available.