| name | enrich |
| description | Enrich an OKF bundle from the web — fetch a bounded set of seed URLs, and for each page decide whether to augment an existing concept, add a new reference under references/, or skip. Use when the user says "enrich my bundle from these URLs", "pull in docs from <site>", "add web references", or "research and add sources for <concept>". Requires a domain allowlist and never crawls openly. For seeding from a local structured source (database, OpenAPI, repo), use the `seed` skill.
|
| argument-hint | <seed urls> [for <bundle path>] |
| user-invocable | true |
| allowed-tools | WebFetch Bash Read Edit |
okfm: enrich
Grow a bundle from web sources under hard guardrails. The Python engine never
touches the network — you fetch here and write through new/convert, so
all provenance and limits live in this skill.
Guardrails (do not skip)
- Domain allowlist is required. Ask for it if the user didn't give one. Fetch only URLs whose host is on the list; never follow links off-allowlist.
- Hard page cap: 50. Stop fetching at 50 pages even if more are queued, and report what was skipped.
- Every page yields a citation. Any concept you write or augment records its source URL in a
# Citations section.
- Agent trust tier. New concepts are
curated_by: agent; the user verifies before flipping to human.
- Capture filter. Skip pages that are thin, secret-bearing, or off-topic — keep only what is Relevant, Actionable, has Depth, and is Authoritative.
Workflow
- Collect the seed URLs, the domain allowlist, and the target bundle. Confirm the allowlist and the 50-page cap with the user before fetching.
- Read the bundle's index tree and relevant concepts so you augment rather than duplicate.
- For each allowed URL, up to 50:
- Fetch it with your harness's web-fetch capability (Claude Code:
WebFetch; other harnesses: their own web tool, or a permitted shell fetch such as curl). If your harness has no web-fetch tool, save the page to a local file and use convert instead of fetching.
- Apply the capture filter; skip if it fails.
- Decide:
- augment an existing concept — Edit its body to add the new material, add the source under its
# Citations section, and add a log.md note; or
- add a reference — a new concept under
references/ via new (or convert if you saved the page) with --type Reference --citation <url>.
- Run
check and report: pages fetched, concepts added/augmented, pages skipped (with reasons), and anything left over the cap.
Writing a reference (preview first, then again without --dry-run on the user's OK):
python "${CLAUDE_PLUGIN_ROOT}/scripts" new references/<slug>.md --bundle <root> --type Reference --title "<title>" --citation "<url>" --body "<extracted summary>" --dry-run --json
