One task: explain clearly how a problem got to where it is today, by mining forward to accomplish this. Most reviews lay things out flat in a forward direction; this skill follows the lineage line backward, letting you see why each step happened.
First read references/sci-cli.md—CLI detection steps, usage, and iron rules.
Then read references/output-system.md—the three-layer output system and HTML shell specification.
Flags
- Default: trace back 3 layers
-d N: trace back N layers (1 to 5); by default 3 layers is about 4–5 retrievals, 5 layers is about 7–8
Steps (fixed order)
Step 0 — Confirm the CLI is available
Run sci --version. If not installed, give the installation command and stop; if there is an authentication error, prompt sci init and stop.
Step 1 — Anchor the current milestone
If the input is a topic, first sci search find the current landmark work as the starting point. If the input is a specific paper, use it directly as the anchor.
If the user inputs a specific paper, first use the title, DOI, or author-year to perform an exact search to confirm the anchor. When it cannot be confirmed, do not fill in paper information from memory; instead treat the input as a topic, and explain in the retrieval gaps that no seed paper was confirmed.
Step 2 — Trace back layer by layer
Starting from the anchor, perform a separate sci searchfor each layer: what method it depends on, what predecessor problem it solves, and which foundational conclusions it is built on. Each layer is an independent retrieval, not memory-based inference. The number of layers is determined by -d , default 3 layers.
Deep tracing means multiple retrievals. Tell the user "this run will perform about N retrievals" and then start directly, without waiting for confirmation.
Each lineage edge must be labeled with a relationship type:
- Method origin: the later work uses or adapts the method of the earlier work
- Problem origin: the later work continues the problem raised by the earlier work
- Evidence origin: the later work builds on the earlier work's conclusions or data
- Field turning point: the later work changes the research object, scale, or explanatory framework
Unless the retrieval results clearly show citation, method inheritance, or problem continuation, do not write it as direct mentorship. When evidence is insufficient, write "candidate predecessor" or "possibly related predecessor", and put it into the retrieval gaps.
Step 3 — Supplement the frontier Search in the other direction: the latest progress in the past 1–2 years, so that the tracing line extends to the present rather than only stopping in the past.
Step 4 — Assemble the lineage line String the retrieval results into a storyline from the earliest root to the current frontier, with each node:
- A actually retrieved paper (title, authors, year)
- One line saying "what it changed on this line"
The lineage line does not aim to fill in every year. Better to leave gaps than to insert transition nodes that cannot be anchored.
Output format
── 传承线 ──────────────────────────────────
2010 ● Smith et al. — "论文标题"
│ 改变了什么（一句话）
↓
2016 ● Jones et al. — "论文标题"
│ 改变了什么（一句话）
↓
2023 ● Lee et al. — "论文标题" ← 当前前沿
改变了什么（一句话）
────────────────────────────────────────────
贯穿问题：一直被重新问的那件事（一句话）
开放边缘：这条线下一步可能走向哪里（一句话）
检索缺口：有则列出并说明原因；无则写"无已知缺口"
The retrieval gaps line cannot be omitted—it lets the user know which nodes have empirical support and where blind spots exist.
Deliverables
After retrieval ends, create the {topic-slug}/ subfolder and output four things:
- Terminal — directly print in TUI format (vertical timeline)
report.md— generate according to the structured Markdown specification ofoutput-system.md: stable YAML front matter + at-a-glance conclusion + lineage chain + reading order + evidence table + detailed evidence + retrieval transparency + reusable JSONreport.html— self-contained HTML, generated according to theoutput-system.mdspecificationreferences.bib— merge the.bibfiles generated by all queries, deduplicate, and keep only the literature entries actually cited in the report
HTML signature experience:vertical lineage timeline, accent color #d97706(amber). Three summary cards at the end: overarching question / open edges / retrieval gaps. For HTML component specifications, see the bear-trace section of output-system.md .
For simple queries, give only 1–2 lightweight next steps. When entering a field or writing a review, give specific suggestions: reading order, which edge has the weakest evidence, and the segmentation logic of the review.
Honest boundaries
Database coverage is limited, and the lineage chain may have gaps. Mark gaps when you find them; do not fabricate a transitional node to fill them. A complete literature review or blank map is a product-level task; leave it to SciMaster.
Gotchas
- Do not use memory to fill nodes: when no suitable paper can be retrieved at a certain level, it is better to explain it in the "retrieval gaps" than to use training data to add a paper that "should exist".
- Do not force the timeline to be continuous: it is normal to have year gaps between two nodes; do not insert transitional papers without retrieval evidence to fill the timeline.
- Lineage edge relationships must be explicit: each edge must state the relationship type (method source/problem source/evidence source/field turning point); "related" or "influenced" is not a valid relationship description.