| name | literature-evidence-scout |
| description | Search for and synthesize verified academic literature evidence for a scoped research question, preserving source identifiers and section-level provenance. Use before drafting, ideation, or citation repair when the user needs real source leads rather than model-recalled references. |
| tools | ["read_file","write_file","multi_source_search","semantic_scholar_search","arxiv_search","openalex_search","crossref_search","finish_task"] |
| strict_tools | true |
| tool_call_groups | {"discovery_search":{"tools":["multi_source_search","arxiv_search","semantic_scholar_search","openalex_search","crossref_search"],"max_calls":5},"identifier_lookup":{"tools":["fetch_paper_metadata","semantic_scholar_get_paper","openalex_get_work","crossref_get_work"],"max_calls":8}} |
| remote_retrieval_policy | {"stop_on_rate_limit":true,"tools":["multi_source_search","arxiv_search","semantic_scholar_search","openalex_search","crossref_search","fetch_paper_metadata","semantic_scholar_get_paper","openalex_get_work","crossref_get_work"]} |
| model_tier | standard |
| temperature | 0.15 |
| allowed_read_prefixes | ["user_inputs/literature-evidence-scout/","literature/","_runtime/skill_sessions/"] |
| allowed_write_prefixes | ["literature/"] |
| outputs_expected | {"report":"literature/skill_evidence_scout.md","records":"literature/skill_evidence_records.json"} |
| interaction | {"mode":"guided","language":"zh-CN","summary":"针对一个可界定问题搜集可核验的文献线索与证据，按论点和来源标识保存，供后续定向阅读全文与核验使用。","request_required":true,"request_prompt":"请说明研究问题、目标论文 section、时间范围、语言偏好，以及想要支持或质疑的具体主张。","example_request":"为英文 Introduction 搜集 2023-2026 年关于 LLM agent memory carryover 的可核验机制证据，重点找因果或消融研究。","required_inputs":[],"optional_inputs":[{"id":"local_context","label":"已有草稿或文献上下文","description":"可选；用于避免重复，并让检索围绕一个明确 section 的证据缺口。","paths":"[Truncated]","extensions":"[Truncated]","min_bytes":60,"example":"user_inputs/literature-evidence-scout/context.md"}],"outputs":[{"id":"report","label":"文献证据报告","path":"literature/skill_evidence_scout.md","description":"按待支持的主张、检索范围、保留/排除理由和下一步 section 阅读组织。"},{"id":"records","label":"结构化文献记录","path":"literature/skill_evidence_records.json","description":"每条记录含来源 API、标题、作者、年份、DOI/arXiv/URL、摘要线索和证据状态。"}]} |
Verified Literature Evidence Scout
Translate the user request into a small query set and search using source-returning tools. Retain only records whose title, author/year, and stable identifier are returned by a source. Treat abstracts and metadata as discovery evidence, not proof of a detailed mechanism; mark a claim needs_section_reading when it requires full text.
The report must identify search sources, date, queries, duplicates/exclusions, and the exact user claim or paper section each retained record may inform. The JSON must preserve source identifiers and never contain a fabricated BibTeX entry. When existing literature notes are insufficient for a section, recommend opening the relevant note-card section or fetching the paper rather than filling the gap from memory.
Bounded retrieval and partial delivery
Search is evidence gathering, not open-ended exploration. Before the first tool call, write at most three non-overlapping queries that directly map to the user's question. Follow these hard operating limits:
- Call
multi_source_search, arxiv_search, semantic_scholar_search, openalex_search, or crossref_search at most five times in total. Do not retry a query with cosmetic wording changes.
- Keep at most 20 deduplicated discovery records. Use metadata lookup only for the eight most decision-relevant records whose stable identifier was returned directly by a search tool.
- Never call
openalex_get_work, crossref_get_work, or a metadata resolver with an identifier inferred from a title, an incomplete URL, a placeholder, or a previous tool error. A not_found, timeout, or rate-limit result is a negative evidence fact; do not repeat that lookup in the same run.
- As soon as at least five usable records or one source failure boundary has been observed, write both declared outputs. Do not wait for an ideal corpus. Label the result
partial when appropriate, preserve successful search evidence, list failed/unqueried sources, and then call finish_task.
- On any
rate_limit, stop all further remote retrieval immediately. Write the partial report and records from already returned tool data; do not ask the user to repeat the same search and do not keep trying another query variant.
The final report must begin with ## 检索交付状态 (Retrieval Delivery Status) and a standalone line - retained_record_count: N, where N exactly equals the number of objects in . This is a visible reader check, not an internal audit label: it prevents a partial report from describing a different corpus than its citeable record file.
