broad-sweep

name	broad-sweep
description	Phase 1 — Run the chunks x issue_spotters cross-product to find candidate issues. Uses edsl for parallel LLM review.
user-invocable	true
allowed-tools	Bash, Read, Write, Glob, Grep

Phase 1: Broad Sweep

Run from the repo root. Requires .review/config.json, .review/parsed/, .review/symbol_table.json, and .review/issue_spotters/ from Phase 0.

How it works

The broad sweep runs a cross-product of chunks x issue_spotters:

• Chunks: The parsed manuscript is split into overlapping ~800-word windows
• Issue spotters: Named markdown files in .review/issue_spotters/ describing what to look for

Each (chunk, spotter) pair is one scenario in the edsl survey. The result is either:

• issue.json — the spotter found something in that chunk
• no_issue.json — the spotter checked and found nothing

Execute the broad sweep

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

python <r2-scripts-path>/broad_sweep.py \
  --parsed-dir .review/parsed/ \
  --symbol-table .review/symbol_table.json \
  --spotters-dir .review/issue_spotters/ \
  --output-dir .review/chunks/ \
  --config .review/config.json

Output structure

.review/chunks/
  <chunk-id>/
    chunk.md                        # The chunk text
    mathematical_errors/
      issue.json   OR  no_issue.json
    undefined_symbols/
      issue.json   OR  no_issue.json
    logical_gaps/
      no_issue.json
    ...

Each issue.json contains:

{
  "title": "<issue title>",
  "spotter": "<spotter_name>",
  "quoted_text": "<relevant quote>",
  "description": "<what's wrong>",
  "source_section": "<section file>",
  "chunk_index": 0,
  "created_at": "<ISO timestamp>"
}

Resumability

The script checks for existing output files before running. If interrupted, re-running it will skip already-processed (chunk, spotter) pairs.

After completion

Commit:

git add -A
git commit -m "review: broad sweep — <N> issues from <total> (chunk x spotter) pairs"

Report: number of issue spotters, number of chunks, total matrix size, and how many issues were found.