| name | sigma-scout |
| description | Discover skills relevant to sigma's domains on skillsmp.com and keep the bundle fresh, without ever auto-installing. Use to find trending or new community skills for classic-ml / deep-learning / nlp / rl / data / agents / mlops / llm work, vet them, and pull the keepers in. Triggers: "find new skills", "what skills should sigma add", "scout skillsmp", "keep the bundle current", "any trending skills for <domain>", or a periodic bundle refresh.
|
| origin | sigma |
sigma-scout
Keep sigma's skill/plugin bundle current against skillsmp.com
— a very large community catalog. The pure scoring/dedup/rank logic lives in
cli/scout.py; the network + install in cli/scout_run.py. This skill is the
curation rubric.
The one hard rule: surface, never auto-install
Scout proposes a relevance-ranked table; a human picks and checks each skill's
license before install. This is the same law as contradiction flagging and the
review gate — the tool never silently mutates the bundle.
Relevance bar (how candidates are judged)
- Relevance FLOOR — pure noise is dropped, not just out-ranked. A hit needs at
least one whole-token domain-keyword match to clear the floor; a popular-but-
off-topic skill scoring only on the capped star bump never surfaces. (Scout used
to surface whatever it found up to the cap — the floor closes that hole.)
- Token match, not substring. Domain terms match on whole tokens, so
rag
credits "rag retrieval" but NOT "storage"/"fragment", and data does not
match "metadata". Substring overlap inflated relevance on noise.
- Domain fit dominates; quality second. Keyword overlap drives the score; stars +
recency only break ties (the star bump is capped so popularity can't outrank fit).
- Per-author diversity cap. At most a few hits from one author surface, so a
prolific publisher can't flood the table and crowd out alternatives.
- Already-bundled → dropped. Anything whose repo or name already lives under
skills/ or skills/vendor/ is deduped out — scout only ever surfaces new skills.
Before adding a skill (vetting checklist)
- License — confirm it permits redistribution if going into the sigma bundle
(
--vendor). Surface it; a human decides.
- Maintenance — recent commits, non-trivial stars, a real README.
- Overlap — does it duplicate an existing sigma skill or context-engine? If so,
prefer improving ours over vendoring a near-duplicate (YAGNI).
- Scope — does it serve a domain sigma actually covers? Don't widen scope just
because a skill is popular.
Targets
- project (default) — clone into
.claude/skills/ to use in the current repo.
--vendor (maintainer) — clone into sigma's skills/vendor/; then review +
commit to grow the shipped bundle. Treat skills/vendor/ as unmodified upstream
copies (the existing vendor law) — don't hand-edit in place; re-vendor to update.
Rate + keys
Anonymous skillsmp access is rate-limited (≈50/day). A free SKILLSMP_API_KEY in
~/.sigma/.env (never the committed config) raises it (~500/day). Optional.
Fail-safe
API down / rate-limited / malformed payload → empty result + a banner, never a
crash. A partial sweep still ranks what returned.
Compose, don't duplicate
Scout grows the bundle; sigma-prune trims what's unused; sigma-cost sizes
the token cost of carrying them. Three distinct layers of bundle hygiene.