kb-ingest

name	kb-ingest
description	Save a URL, file path, pasted text, DOI, or PMID into the knowledgebase at $KB_ROOT/raw/manual/YYYY/MM/DD/. Enriches via the relevant MCP (pubmed/crossref) when input is a DOI or PMID. NON-DESTRUCTIVE — write-only into raw/, never into pages/.
version	0.1.0
author	dobry-ops
license	MIT
metadata	{"hermes":{"tags":["kb","ingest","raw","ms-research"],"category":"ms-knowledgebase","related_skills":["kb-research","kb-journal"]}}

kb-ingest

Save an artifact into the knowledgebase. Raw only — do not write to pages/ from this skill; use kb-research for that.

When to use

User says one of:

• "ingest this link: https://..."
• "save this PDF for later: /path/to/file.pdf"
• "remember this: "
• "look up DOI 10.1056/NEJMoa1601277"
• "look up PMID 28002679"

Env vars

• $KB_ROOT — kb tree root (/var/lib/ms-researcher/kb on engineer).

How

Use Europe/Sofia date partitioning: raw/manual/YYYY/MM/DD/. Slug the date as YYYY_MM_DD; slug the rest from input:

1. URL — curl -fsSL --max-time 15 -o $KB_ROOT/raw/manual/YYYY/MM/DD/<date>_<slug>.<ext>. Sniff content-type for .html, .pdf, .txt. If pubmed.ncbi.nlm.nih.gov, extract the PMID from the URL and ALSO call pubmed.pubmed_fetch(pmid) to save a structured <date>_<slug>.pubmed.json alongside.
2. File path — copy with install -m 0640 into raw/manual/YYYY/MM/DD/. Keep original extension. If .pdf and the title is not in the filename, leave it as is.
3. Pasted text — write to raw/manual/YYYY/MM/DD/<date>_<slug>.md with a single-line frontmatter (source: pasted, accessed: <date>).
4. DOI — call crossref.crossref_lookup(doi). Save the full JSON to raw/manual/YYYY/MM/DD/<date>_doi_<slug>.crossref.json. If crossref returns a PMID, also call pubmed.pubmed_fetch(pmid) and save its result.
5. PMID — call pubmed.pubmed_fetch(pmid). Save to raw/manual/YYYY/MM/DD/<date>_pmid_<pmid>.pubmed.json. If a DOI is in the response, ALSO call crossref.crossref_lookup(doi) and save that.

Then call kb-journal with event: ingested <slug> so the weekly digest sees it.

Slug rules

• Lowercase, ASCII, _ as separator.
• Strip articles and punctuation.
• Cap at 60 chars.
• Examples:
  • https://www.nejm.org/doi/full/10.1056/NEJMoa1601277 → nejm_ocrelizumab_opera_2017
  • PMID 28002679 → pmid_28002679 (the underlying paper is found via the JSON later)

Output format for the user

✓ Ingested: <kind> → kb/raw/manual/YYYY/MM/DD/<filename>
  Source:    <url|path|"pasted text">
  Enriched:  <yes/no — list the .json sidecar files if any>
  Next:      `research this` to extract findings into kb/content/<type>/YYYY/MM/<slug>.md

If the fetch fails, say so verbatim. Don't retry.

✗ Ingest failed (<reason>):
  <error body>

Rules

• Do NOT extract claims here. This is a save-raw step.
• Do NOT overwrite an existing <date>_<slug> file — append _2, _3 as a disambiguator.
• If legacy raw files exist at $KB_ROOT/raw/<date>_<slug>.*, leave them alone unless kb-maintain is explicitly running; this ingest skill writes only to raw/manual/YYYY/MM/DD/.
• Reference $KB_ROOT by name; never paste the value into chat.
• If the URL is behind a paywall and the fetch returns an HTML preview page, save it anyway and flag in the output: note: paywalled — only the preview was saved.
• If you can't reach the network (MCP error, curl timeout), do not fabricate the file. Report the failure.