Cascade Discover

Find new Claude Code tooling without re-recommending what's already in the stack.

When to invoke

• "What new Claude Code tools should I install?"
• "Discover new skills / agents / hooks"
• "Cascade discover"
• "What's new in the Claude ecosystem?"
• Weekly periodic check (recommended)

How it works

1. Inventory — scans ~/.claude/repos/, ~/.claude/plugins/marketplaces/, all skill/agent/command symlinks, and known_marketplaces.json. Resolves each to a GitHub owner/name via .git/config. Builds the installed-set.
2. Discovery — runs gh search repos against 9 default queries (claude-code skills/agents/hooks/plugin, mcp-server productivity, awesome lists, agentic personal os, etc.).
3. Crosscheck — filters out anything in the installed-set.
4. Rank — scores each by stars × exp(-days_since_push / 90). Recent + popular wins; old high-star repos decay.
5. Render — markdown table with rank, repo, stars, last push, score, description.

Invocation

python3 ~/.claude/skills/cascade-discover/discover.py

Flags

Schedule it

# weekly via the schedule skill
schedule "weekly cascade discover" "python3 ~/.claude/skills/cascade-discover/discover.py --save"

Freshness mode (--check-updates)

Discovery finds net-new tooling and throws away anything installed. Freshness asks the opposite question: of the repos I already have, which shipped upstream changes since my baseline? Answers "am I running stale skills/agents?"

python3 ~/.claude/skills/cascade-discover/discover.py --check-updates          # report drift
python3 ~/.claude/skills/cascade-discover/discover.py --check-updates --json   # for brain analyze
python3 ~/.claude/skills/cascade-discover/discover.py --check-updates --refresh-lock  # ack updates

How it works:

1. Inventory — same installed-set as discovery, minus repos Sam authors (owners swsounds42/*, swsounds/*, plus the own list in inventory-extras.json).
2. Probe — for each repo, gh api for latest default-branch commit (sha + date) and latest release.
3. Compare to baseline — freshness-lock.json stores the last-acknowledged sha/release per repo. New commits or a new release tag ⇒ STALE. Stale rows show commits-behind (gh compare) + a diff link.
4. Baseline lifecycle — first sight of a repo seeds its baseline (not flagged). Baselines only advance on --refresh-lock, so drift keeps nagging until you actually update + acknowledge. Uninstalled repos are pruned; --mute pins a repo you intentionally hold back.

Granularity caveat: this is repo-level, not skill-level. All 95 gws-* skills come from one repo (googleworkspace/cli), so a hit reads "googleworkspace/cli: 12 new commits" — a prompt to go look, not "gws-gmail-triage changed."

cascade-brain analyze calls this automatically (non-dry, so the weekly scan seeds/prunes the lock) and folds an "Updates available" section + a STALE editorial bucket into the brief.

Output

# Cascade Discovery Report — 2026-04-26

**Already installed:** 14 repos · **New candidates:** 47 · **Showing top:** 25

| # | Repo | ★ | Last push | Score | Description |
|---|---|---:|---|---:|---|
| 1 | [owner/repo](https://...) | 1,432 | 2026-04-22 | 1,348 | What it does |

Implementation notes

• Inventory dedup is by GitHub owner/name (lowercased). Symlinks pointing into a repo subdirectory still resolve to the parent repo via find_git_root.
• Score formula balances popularity vs freshness — a 1k-star repo last pushed 30 days ago beats a 5k-star repo from 2 years ago.
• Rate limits — gh search is rate-limited to ~30 req/min for authenticated users. The default 9 queries take ~10 seconds.
• Saved reports live at ~/.claude/discover-reports/<date>.md for trend tracking.