Menu
Phase 1 — Run the chunks x issue_spotters cross-product to find candidate issues. Uses edsl for parallel LLM review.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
| 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 |
Run from the repo root. Requires .review/config.json, .review/parsed/, .review/symbol_table.json, and .review/issue_spotters/ from Phase 0.
The broad sweep runs a cross-product of chunks x issue_spotters:
.review/issue_spotters/ describing what to look forEach (chunk, spotter) pair is one scenario in the edsl survey. The result is either:
issue.json — the spotter found something in that chunkno_issue.json — the spotter checked and found nothingLook 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
.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>"
}
The script checks for existing output files before running. If interrupted, re-running it will skip already-processed (chunk, spotter) pairs.
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.
Phase 6 — Generate the REVIEW.md report by walking the append-only review tree. Optionally create GitHub Issues.
Add a custom issue spotter for the paper review. Describe what to look for and this will create the spotter file.
Phase 3 — Deduplicate confirmed issues by writing a dedup.json mapping file. No files are moved or modified.
Draft author responses for each confirmed issue. Writes response.json next to each issue (append-only).
One-shot holistic referee review of the full paper. Issues are written to .review/holistic/ as append-only issue.json files.
Phase 2 — Investigate candidate issues. Writes investigation.json next to each issue.json with evidence and a confirm/reject verdict.