菜单
Archive AI deep-research dumps and articles into RKC source-assets. Hash, copy, write SourceDocument + ResearchTask shells. Optionally run the Phase 2 extractor. Idempotent on bytes plus prompt hash plus extractor version.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/SpillwaveSolutions/research-knowledge-capture --skill research-ingest命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Create missing Subject and ResearchArea nodes and write has_task / has_subject on an existing RKC tree. Idempotent repair for 0.2.0 archive-only ingest.
Bind a Cursor agent to RKC. Capture research into OKF. Do not project the graph.
Bind a Grok Bot agent to RKC. Isolation, identity, fail-closed writes.
正在显示 SKILL.md
| name | research-ingest |
| description | Archive AI deep-research dumps and articles into RKC source-assets. Hash, copy, write SourceDocument + ResearchTask shells. Optionally run the Phase 2 extractor. Idempotent on bytes plus prompt hash plus extractor version. |
Actor: grok-bot/research-knowledge-capture. Isolated session. Ready PR. No force-push. No write to main.
_inbox/research-dumps/ (not a type).knowledge/research/source-assets/<sha256>/original.*.SourceDocument + ResearchTask shells with ingested_from.reviewed | accepted | verified: true.python3 ${CLAUDE_PLUGIN_ROOT}/scripts/rkc_ingest.py _inbox/research-dumps \
--knowledge knowledge --vendor grok --subject <slug> --area <area-slug>
--subject is the slug. The Subject title is slug words, title-cased (ref-okf-plugin → Ref Okf Plugin), not the first source filename. Pass --subject-title "OKF plugin" for an explicit name.
Then extract:
python3 ${CLAUDE_PLUGIN_ROOT}/scripts/rkc_ingest.py _inbox/research-dumps \
--knowledge knowledge --vendor grok --subject <slug> --area <area-slug> --extract --subject-id <id>
or /research-extract against the archived asset.
vendor is free text. Conventional values: grok | gemini | claude | deepseek | chatgpt | article | perplexity | unknown.
--source-kind defaults to deep_research. Use reference_doc, published_medium, published_substack when that is what you have.
--extractor-version defaults to the current extractor (2). --extract does not change identity. Same bytes under another version are reported, not duplicated, unless --allow-reextract. Prompt hash: --prompt-hash or --prompt-file.
--dry-run prints the plan and writes nothing, including no research/catalogs/ directory.
Empty, whitespace-only, and frontmatter-only files are skipped (SKIP empty on stderr, listed in --errors-file). --min-source-bytes defaults to 1 and applies to the payload after stripping frontmatter and whitespace. One-line pointer documents are kept.
index.md and README.md titles use parent/filename.
Files over 200 KB skip heuristic extract unless --force-large. Progress goes to stderr. Failed files append to research/catalogs/ingest-errors.jsonl; the run continues.
Idempotency lookup is research/catalogs/ingest-index.json. --rebuild-index rebuilds it from sources + tasks.
Do not parallelize writers inside one knowledge tree. Shard by Subject (one tree or one process per slug). If you stop a run, kill workers by PID — pkill -f misses multiprocessing children.
Public samples are Northstar / Lumenfield fiction only.