SKILL.md
<!-- SKILL MAINTENANCE NOTES: - Dashboard 用 JSON.parse() 解析卡片内容 - 字段变更必须同步 dashboard/src/components/chat/ 中的渲染组件 - workspace file facts and supported raw search results are projected automatically - papercard is an Agent-deliberate highlight, not a raw-result dump - approvalcard 必须包含 approval_id(来自 exec.approval.requested) -->
Output Cards
Use fenced code blocks with the card type as the language tag only for Agent-authored card semantics. Content MUST be valid JSON — the dashboard parser uses JSON.parse(). Research-Claw automatically presents successful Workspace file facts and supported raw literature results; do not copy tool JSON merely to make those cards appear. Keep ordinary paths and safe links in the surrounding prose for non-Dashboard channels.
paper_card
ONLY for real academic publications deliberately highlighted by the Agent — from API queries, library_search, or user-identified papers. NEVER for concepts, tools, non-scholarly content, or every raw search hit. A raw result is only retrieved; the fence does not prove it was read, cited, saved, or verified.
Required: type, title, authors (string[]). Optional: venue, year, doi, url, arxivid, abstractpreview, readstatus ("unread"|"reading"|"read"|"reviewed"), libraryid, tags.
task_card
Required: type, title, tasktype ("human"|"agent"|"mixed"), status ("todo"|"inprogress"|"blocked"|"done"|"cancelled"), priority ("urgent"|"high"|"medium"|"low"). Optional: id, description, deadline (ISO 8601), relatedpapertitle, relatedfilepath.
progress_card
Required: type, period, papersread, papersadded, taskscompleted, taskscreated. Optional: writingwords, readingminutes, highlights (max 5).
approval_card
Required: type, action (string), context (string), risklevel ("low"|"medium"|"high"). Required (for exec approvals): approvalid from exec.approval.requested — without it, dashboard buttons are non-functional. Optional: details (must be a JSON object, not a string — e.g. {"paper_count": 7}).
file_card
Successful workspacesave, workspaceexport, workspaceappend, and workspacedownload results are presented automatically. State the ordinary workspace-relative path in prose; no fence is required. Legacy fences remain compatible, but NEVER fabricate a path.
monitor_digest
Required: type, monitorname, sourcetype (free-form), target, total_found, findings (array of {title, url?, summary?}, max 10). Optional: schedule.