feat: add context7-cache skill — local documentation cache

[m2015agg]

Summary

Adds context7-cache — a skill that caches Context7 library documentation locally in SQLite+FTS5, serving docs via CLI instead of MCP round-trips.

• npm package: @m2015agg/context7-skill (npm | GitHub)
• Auto-detects project dependencies from package.json, requirements.txt, pyproject.toml, etc.
• Pre-caches documentation weighted by import frequency
• FTS5 full-text search across all cached libraries
• Version-aware caching from pinned dependency versions
• Offline resilient — serves from cache when API is unreachable

Anthropic SkillsBench Results

Benchmarked using the skill-creator framework with LLM grader (Sonnet 4.6). 3 evals × 3 runs per config.

Metric	CLI Cache (with_skill)	MCP Server (without_skill)	Delta
Pass Rate	93% ± 10%	87% ± 12%	+6%
Speed	17.9s	31.9s	1.8x faster
Hallucinations	None	Fabricated FastAPI scope param (3/3 runs)	—

Per-eval breakdown

• FastAPI Depends(): CLI 100% vs MCP 75% (MCP hallucinated non-existent scope parameter)
• Celery+Redis config: CLI 80% vs MCP 87% (both missed task decorator sometimes)
• Supabase RLS policies: CLI 100% vs MCP 100% (tied)

Key finding

The MCP server returned Context7 docs that included a fabricated scope parameter on FastAPI's Depends() — caught by the LLM grader in all 3 runs. The CLI cache served pre-verified cached docs without this hallucination.

Files added

• skills/context7-cache/SKILL.md — 131 lines, follows spec
• skills/context7-cache/references/benchmark.json — full benchmark data
• skills/context7-cache/references/benchmark.md — human-readable summary
• skills/context7-cache/scripts/check-setup.sh — setup verification

Test plan

• SKILL.md follows Agent Skills spec (name format, description < 1024 chars, < 500 lines)
• Directory name matches name field
• Benchmarked with skill-creator grader agent
• Fair comparison: MCP given 10 turns (enough for resolve + fetch + respond)
• npm package published and installable

🤖 Generated with Claude Code