Skip to content

xchewtoyx/bundle-curator

v0.1.0

Harness-agnostic curation tooling for Zettelkasten-style OKF wiki bundles: skills to onboard sources, dispatch domain curators, review charters, and drain scoring/landscape backlogs, plus the scripts and shared agents they depend on. Domain-agnostic — reads the domain-slug list from the consuming repo's okf-core.toml. Domain-specific curator agents (<slug>-curator.md), okf-core.toml, and the fleeting/wiki content itself stay in the consuming repo; this plugin ships only the reusable mechanics.

GitHubOpen Website

bundle-curator

Harness-agnostic curation tooling extracted from rgh-pre and rgh-sme: the skills, shared agents, and scripts an OKF wiki repo uses to turn raw "fleeting" literature notes into curated, linked wiki bundles. This is not the consumer-facing wiki content plugin (see ../rgh-pre and ../rgh-sme in this marketplace) — it ships no wiki notes and answers no retrieval questions. It is the maintainer-side tooling that produces that content.

Authoring model: unlike ../rgh-pre and ../rgh-sme, which are vendored from Cloudsmith by scripts/sync_marketplace.py, this plugin is mastered directly in this repo — edit plugins/bundle-curator/ here and bump plugin.json's version. There is no .cloudsmith-source.json stamp, so scripts/sync_marketplace.py prints a warning and leaves this directory untouched every run (it only syncs plugins that carry that stamp). Once rgh-pre/rgh-sme (and any future domain-wiki repo) actually depend on this plugin instead of carrying their own copy of this tooling, mastering may move upstream to one of those repos, or to a dedicated repo of its own — this repo is the starting point, not necessarily the permanent home.

What a consuming repo must supply

This plugin is domain-agnostic — scripts/domains.py reads the current domain-slug list from the consuming repo's okf-core.toml, never a hardcoded table. Beyond that config, a repo that installs this plugin still needs, in its own tree (none of this ships here):

  • okf-core.toml — the domain-bundle declarations scripts/domains.py reads.
  • .claude/agents/<slug>-curator.md per domain — the curator charter (In scope / Boundaries) and curation persona. Domain-specific by nature; not something a shared plugin can supply.
  • fleeting/, curation/<slug>.txt, ratings/<slug>.jsonl, landscape/overlap.jsonl, watch/<slug>.jsonl, and the domain bundle directories themselves — all repo-local curation state these skills read and append to.
  • The okf CLI (install path: .venv/bin/okf, see scripts/setup.sh in rgh-pre/rgh-sme for the current install recipe) for validation and search during curation.
  • pymupdf (see requirements.txt here) if scripts/dump-source-chapters.py will be used for PDF/EPUB onboarding.

Layout

  • skills/ — onboard, curate, progress, charter-review, catch-up, score-catchup, source-watch. Each SKILL.md is the full harness-facing procedure.
  • agents/ — the domain-agnostic agents skills dispatch to: fleeting-researcher (charter-scoring and landscape-mapping modes), charter-researcher, chapter-writer, source-watch-researcher. Per-domain <slug>-curator agents are not here — they're consuming-repo content.
  • scripts/ — deterministic selection/reporting tooling the skills shell out to (candidate selection, progress reporting, chapter dumping, etc.). Pure Python 3, cwd-relative (they locate the repo root by walking up for okf-core.toml), so they work unmodified from wherever a harness installs this plugin.
  • docs/ — the detailed procedures (curation.md, landscape.md, charter-review.md, wiki-rules.md, source-watch.md) the skills point to rather than restate.
  • tests/ — pytest coverage for scripts/, ported alongside the scripts so this repo's own pytest tests/ -v (via requirements-dev.txt plus this directory's requirements.txt) still exercises them.

Provenance

Ported from rgh-pre (canonical source for everything except source-watch, which only rgh-sme has today) as of the repos' main at extraction time. Two real drifts between the two repos' copies were resolved in favour of rgh-pre's newer version: a fleeting/-scan bug fix (stray top-level files under fleeting/ are no longer treated as source notes) and consistent P>=0.7 (at-or-above) confidence-threshold semantics, matching what the selection scripts actually gate on. source-watch's implicit agentic-engineering default was replaced with a required --focused <domain-slug> argument, since a hardcoded default domain has no meaning across multiple consuming repos with different domain sets.