Paper Search

Paper Search

Overview

This skill exposes three discovery commands from the JushenRenji repo as a JSON-stdout CLI you can call from opencode / Bash:

All three return the same JSON contract:

{
  "ok": true,
  "count": 10,
  "papers": [
    {
      "arxiv_id": "2410.24164",
      "title": "...",
      "abstract": "...",
      "url": "https://arxiv.org/abs/2410.24164",
      "submitted_date": "2026-01-08",
      "upvotes": 0,
      "github_repo": null,
      "project_page": null,
      "source": "arxiv_venue"
    }
  ]
}

On error: {"ok": false, "error": "<msg>", "papers": []} and exit code 1.

Cross-conda invocation pattern

You MUST run the command inside the paperagent conda env from the repo root:

source ~/miniconda3/etc/profile.d/conda.sh
conda activate paperagent
cd ~/Projects/VlogCutter/JushenRenji
python -m src.discovery_sources.cli <subcommand> [args...]

The CLI auto-applies the network workaround when JSR_NETWORK_PROFILE=gsjts is in the env (clash proxy + IPv4-only — needed on the GSJts box). Set it once via export JSR_NETWORK_PROFILE=gsjts if you hit network failures from this host.

Tool 1: arxiv_search

Search arXiv recent papers by a topic phrase + cs categories + recency window.

Input

Bash example

python -m src.discovery_sources.cli arxiv-search \
  --query "diffusion policy" \
  --days 14 \
  --limit 20 \
  --categories "cs.RO,cs.LG"

Tool 2: arxiv_by_venue

Search arXiv accepted-paper notes by venue + year. Best-effort — arxiv co: (comments) field is not normalized: authors write Accepted to RSS 2025, To appear at NeurIPS 2025, RSS 2025, RSS'25, etc. We OR all four typical variants. Recall is not guaranteed; precision is high.

Input

Known venue strings

Use these exact strings (no extra punctuation):

RSS         (Robotics: Science and Systems)
NeurIPS     (or "NeurIPS" — note no space inside)
CoRL        (Conference on Robot Learning)
ICLR
ICML
ICRA
IROS
CVPR
ICCV
ECCV
RSS Workshop   (matches "RSS Workshop 2025" comments)

Non-canonical strings (e.g. "rss") still work but recall is worse.

Bash example

python -m src.discovery_sources.cli arxiv-by-venue \
  --venue RSS --year 2025 --limit 20

# Also include "submitted to ..." (less precise)
python -m src.discovery_sources.cli arxiv-by-venue \
  --venue NeurIPS --year 2025 --limit 50 --include-submitted

Tool 3: hf_daily_papers

Pull HuggingFace Daily Papers leaderboard (today's hot papers, with upvotes / github_repo / project_page).

Input

Bash example

python -m src.discovery_sources.cli hf-daily --limit 20

Typical scenarios

"Find RSS 2025 accepted papers that have GitHub"

python -m src.discovery_sources.cli arxiv-by-venue \
  --venue RSS --year 2025 --limit 50 \
  | jq '.papers[] | select(.github_repo != null) | {title, arxiv_id, github_repo}'

(GitHub repo is rarely populated for arxiv-only entries — for repos use arxiv_search and post-filter, or pull HF Daily.)

"Find recent latent-action papers in cs.RO"

python -m src.discovery_sources.cli arxiv-search \
  --query "latent action" --days 7 --limit 30 --categories "cs.RO"

"Today's HF Daily papers, top 10"

python -m src.discovery_sources.cli hf-daily --limit 10 \
  | jq '.papers[] | {title, upvotes, github_repo}'

Output contract details

Each paper object in papers[]:

Error handling

The CLI never crashes on network failures. On error, stdout is still valid JSON:

{"ok": false, "error": "RuntimeError: simulated network down", "papers": []}

Exit code is 1 when ok=false, 0 otherwise. Always parse stdout JSON before reading exit code.

Limitations

• co: (comments) field on arxiv is best-effort. Some accepted papers don't list the venue in their abstract/comments. If you need an authoritative list, use the conference's official accepted-papers page (out of scope).
• HF Daily only returns papers HF has indexed; the upvote count is "live" so re-running gives different ordering.
• The CLI assumes the paperagent conda env is active (uses feedparser, requests, pyyaml).