ctx-inspired improvements: routing feedback, session dedup, slug-token scoring, quality grades

[MaTriXy]

Summary

A review of stevesolun/ctx (MIT — graph-backed skill/agent recommendation engine for Claude Code, forked at MaTriXy/ctx) surfaced working solutions to several limitations Agent Registry's own README acknowledges: no relevance feedback, fixed scoring that doesn't learn, and no visibility into dead-weight agents.

Full plan (committed to this repo): docs/plans/ctx-inspired-improvements.md

We port ctx's feedback-loop principles only — no Python code, no graph artifacts, no new runtime dependencies. Everything stays zero-dep Bun + local JSONL.

Proposed changes (one PR per phase)

Phase	Change	Key files
1	Routing telemetry — local JSONL logging considered (hook injected) vs picked (get.js loaded); hit-rate with neutral 0.5 prior under 3 observations; 30-day truncation	new lib/routing-log.js, hooks/user_prompt_search.js, bin/get.js
2	Session dedup — don't re-inject the same agents on every prompt within one session_id; 24h pruning	hooks/user_prompt_search.js
3	Slug-token scoring — exact name-token matches ≫ substring, simple IDF over registry; folds into existing 0.6*BM25 + 0.4*keyword formula	lib/search.js
4	Hit-rate feedback in ranking — bounded multiplier final = combined * (0.85 + 0.3*hitRate) applied hook-side only; CLI search stays pure relevance	lib/search.js, hook
5	Quality grades — A–F per agent from usage (0.45) + routing hit-rate (0.30) + structure (0.25); hard floors: missing file ⇒ F, never-used ⇒ max D	new bin/quality.js, lib/quality.js
6	(stretch, parked) optional cosine/semantic ranker — deferred; breaks zero-dep promise, revisit after 1–5 ship	—

Dependency order: 1 → 4 → 5; phases 2 and 3 are independent and can ship anytime.

Note: this is distinct from the existing lib/telemetry.js (opt-in anonymous network analytics). All new state is local-only, lives under references/, and is git-ignored like registry.json.

Build & test

bun install                      # only dep: @clack/prompts (init UI)
bun test                         # full suite (101 tests today; each phase adds its own file)
bun bin/search.js "query"        # manual smoke
bun bin/quality.js               # after Phase 5
./install.sh                     # user-level install to ~/.claude/skills/agent-registry/

Constraints to preserve: Bun built-ins only for search/scoring, CommonJS in lib/, silent-fail hooks, user data stays git-ignored.

Attribution

Mechanisms adapted from ctx (MIT): routing-trace hit rate with minimum-observation prior, session suggest-shown flag, slug-token tiered scoring, multi-signal quality grades with hard floors, log truncation windows.