| name | expand-references |
| description | Expand one to three seed papers into nearby, bridge, foundational, methodological, recent, and survey follow-ups |
| context | fork |
| agent | Explore |
| disable-model-invocation | true |
| user-invocable | true |
| allowed-tools | Bash, Read |
| argument-hint | <seed-1> [seed-2] [seed-3] [--negative <paper>] [--pool all-cs|recent] [--limit <n>] [--per-bucket-limit <n>] |
Expand References
Turn one to three seed papers into a structured follow-up reading list.
Use this when the human already has anchor papers and wants the next papers to read.
Arguments
- Positional arguments are the seed papers. Quote multi-word titles.
--negative <paper> may be repeated to push the workflow away from an unwanted cluster.--pool all-cs|recent selects the Semantic Scholar recommendation pool.--limit <n> controls how many raw recommendations are requested before reranking.--per-bucket-limit <n> caps each curated bucket after scoring.
Workflow
- Run
python scripts/run.py ....
- Read
result.closest_neighbors for the immediate next reads.
- Read
result.bridge_papers for papers that connect multiple seeds.
- Read
result.foundational, result.methodological, result.recent, and result.surveys_or_benchmarks for curated slices of the neighborhood.
- If the result is sparse or off-topic, adjust the seed set or add
--negative papers and rerun.
Output
- The script prints the unified JSON envelope described in
output_contract.md.
- The underlying workflow result is
ExpandReferencesResult.to_dict().
result.notes captures dropped records and other execution notes.
When To Escalate
- Fewer than one clear seed paper is available.
- The resolved seeds are obviously duplicates or wrong papers.
- The output is empty even after trying better seeds or a different recommendation pool.
