| name | paper-triage |
| description | Batch paper screening and classification. Triggers on "筛一下", "triage", "哪些值得看", "帮我筛", "batch screen". |
Paper Triage
Flow
-
Context carry-over:
- If user explicitly references existing papers ("筛一下刚才的", "帮我筛这些", "筛这些"): triage those directly →
paper_auto_triage(paper_ids=[...])
- If papers in context but reference is ambiguous: ASK "要筛选刚才找到的这些论文?还是筛选库里最近的未读论文?"
- If no context → default to
paper_auto_triage(top_n=5)
-
Credibility check: For papers in the "important" bucket, call paper_credibility_batch(important_ids) to add credibility signals.
-
Present the three buckets as tables:
重要: | # | 标题 | 评分 | 入选理由 | venue | code | 复现风险 |
待读: | # | 标题 | 评分 | 简评 |
跳过: | # | 标题 | 评分 | 跳过理由 |
结论: 为什么这几篇最值得关注,关联用户 profile 说明
-
[FORK] "这是按你 profile 的分类,同意吗?要调整哪些?"
If user adjusts
Move papers between buckets per user's instruction, then:
- Call
paper_reading_status(important_ids, "important") for confirmed important papers
- Call
paper_reading_status(to_read_ids, "to_read") for to_read papers
If user confirms as-is
Apply the auto-triage result directly:
- Mark important papers as "important"
- Mark to_read papers as "to_read"
- Skip papers get no status change
After marking, auto-save: call paper_save_report(report_type="triage", content=<report markdown>, filename="{topic}-{YYYY-MM-DD}.md") using triage-template.
[CONTEXT-AWARE FORK] — Based on the triage result, suggest next steps:
-
If important bucket has papers with PDFs available:
"已标记完成,筛选报告已保存至 {path}。\n💡 下一步建议:\n1. 批量下载 {N} 篇重要论文的 PDF\n2. 深入分析 [{top_important_title}]({score}分)\n(说编号或告诉我你想做什么)"
-
If important bucket has papers but no PDFs:
"已标记完成,筛选报告已保存至 {path}。\n💡 下一步建议:\n1. 深入分析 [{top_important_title}]({score}分)\n2. 搜索并下载这些论文的 PDF\n(说编号或告诉我你想做什么)"
-
If important bucket is empty but to_read has papers:
"没有特别突出的论文,但有 {N} 篇值得一读。\n💡 下一步建议:\n1. 看看 [{top_to_read_title}] — 评分最高的待读论文\n2. 换个方向搜索\n(说编号或告诉我你想做什么)"
-
If all papers are skip:
"这批论文跟你方向关联不大。\n💡 下一步建议:\n1. 调整 research profile — 可能关键词需要更新\n2. 换个方向搜索\n(说编号或告诉我你想做什么)"
Custom source
If user provides specific paper IDs or says "筛这些":
- Call
paper_auto_triage(paper_ids=[...]) with the specific IDs
Rules
- Only 1 checkpoint (confirm/adjust). Classification is automatic via
paper_auto_triage.
- Don't ask about criteria — use profile-based relevance scores by default.
- Don't ask about source — default to recent unread papers.
- Always use tables, never bullet-point lists for paper results.
- Always include a 结论 section explaining why the classification matters.
- FORK suggestions must reference actual paper titles and counts from the triage result.