| name | deep-dive |
| description | Deep analysis of a single paper. Triggers on "分析这篇", "deep dive", "这篇论文讲了什么", paper ID or arXiv ID input. |
Deep Dive Analysis
Flow
- Resolve paper_id from context: If the user refers to a paper by index (e.g. "第3篇", "分析上面那篇"), resolve it from the papers discussed earlier in this conversation. If user gives an explicit paper ID or arXiv ID, use that directly.
- Call
paper_show(paper_id) to get full paper details
- Call
paper_profile() to understand user's research context
- Check for full text: Try
paper_sections(paper_id) — if it returns sections, mark analysis_basis = "full_text". If not parsed but a PDF exists, call paper_parse(paper_id) to parse it first. If no PDF available, mark analysis_basis = "abstract".
- Extract structured profile: Call
paper_extract(paper_id) to get task/method/dataset/baseline/metric data.
- Show tables (if available): If analysis_basis == "full_text", call
paper_tables(paper_id) to show extracted tables.
- [FORK] "全面分析，还是关注某个角度？（方法/实验/跟你的关联/可信度）"
- Generate analysis based on user's choice (or full analysis by default). If full text is available, use
paper_ask(paper_id, question) for specific questions.
Source Annotation Rule
IMPORTANT: Every analysis output MUST start with a source annotation:
- If analysis_basis == "full_text": show [基于全文] at the top
- If analysis_basis == "abstract": show [基于摘要] at the top
This helps the user understand the depth and reliability of the analysis.
Analysis Template (for AI generation)
When generating the analysis, use tables for structured data:
- 核心信息表: | 字段 | 值 | (title, authors, venue, year, citations)
- 方法: key techniques, architecture, loss function. If full text available, drill into method section.
- 实验结果: use table | 指标 | 本文 | Baseline1 | Baseline2 | — pull from extracted profile when available
- 优劣势对比: use table | 维度 | 本文 | 相关工作1 | 相关工作2 |
- 可信度评估: call
paper_credibility(paper_id) — show venue tier, code availability, reproducibility risk
- 结论与建议: 研究价值判断 + 与用户研究方向的关联 + 是否值得深入跟进 + 关键参考文献
After analysis
Auto-track: call paper_note_add to save note to workspace (this is internal tracking, not file export).
If first_use is true in the response, tell the user: "（首次自动记录）我把分析笔记保存到了工作区（.paper-agent/），方便以后查阅。如果不需要自动记录，告诉我'不要记录'就行。"
If first_use is false, just briefly say: "已记录到工作区。"
Auto-save: call paper_save_report(report_type="analysis", content=<analysis markdown>, filename="{paper_id}.md") to persist the analysis as a file.
[CONTEXT-AWARE FORK] — Based on the analysis result, suggest next steps:
Analyze the paper's characteristics and suggest accordingly:
-
If paper has high citation count (>50) or is a seminal work:
"分析笔记已保存至 {path}。\n💡 下一步建议：\n1. 追踪引用链 — 看看谁在跟进这个方向\n2. 找类似方法的论文做对比\n（说编号或告诉我你想做什么）"
-
If paper has code available:
"分析笔记已保存至 {path}。\n💡 下一步建议：\n1. 做实验计划 — 这篇有代码，可以复现/改进\n2. 追踪引用链看后续工作\n（说编号或告诉我你想做什么）"
-
If paper's method is novel (novelty_claim is strong):
"分析笔记已保存至 {path}。\n💡 下一步建议：\n1. 看看这个方法能不能用到你的课题 — 生成 research ideas\n2. 找同方向的论文做对比\n（说编号或告诉我你想做什么）"
-
Default:
"分析笔记已保存至 {path}。\n💡 下一步建议：\n1. 追踪引用链\n2. 找 [{paper_method_family}] 方向的更多论文\n（说编号或告诉我你想做什么）"
If user picks citation trace
Route to paper_citation_trace(paper_id). Carry the paper's method_family and topics as context for the citation exploration.
Rules
- 2 checkpoints (analysis angle + after analysis). Both workspace note-add and file save are automatic.
- If user explicitly says "不要记录", skip the note_add call and the save_report call.
- FORK suggestions must be based on the actual analysis result (citation count, code availability, method novelty). Never give the same generic options.
