research-setup

project-kit: Research a Setup Option

Turn "is there something better for X?" into a scored, recorded decision — and, when adopted, into concrete wiring in the kit. This keeps the setup evolving instead of frozen.

Stack & values to evaluate against

• Stack: uv-managed Python ≥3.11, beads for tracking, GitHub Actions CI, src layout.
• Profile / values (from github-map): applied ML + data engineering with own data and Peruvian regional domain as the differentiator; reproducibility; agentic dev (Claude Code + beads).
• Already installed (don't re-propose / weigh overlap): superpowers, frontend-design, github-map plugins; playwright + context7 MCP servers.

Workflow

1. Restate the need in one line, plus the constraint it must fit (e.g. "must work with uv", "stdio MCP").
2. Research with WebSearch/WebFetch: official MCP/plugin registries, GitHub repos, docs. Find 2–4 real candidates. Note maintenance signals (last release, stars, issues).
3. Score each candidate against github-map's rubric (schema.md in the github-map repo): weigh market_fit, real_impact_potential, code_quality, and agentic_dev_level as relevant — plus fit-to-profile and overlap with what's installed. Bias toward tools that raise the bar, not generic popularity.
4. Record a dated entry in catalog/setup-options.md (see its format) with a verdict: adopt / trial / reject, and a one-line rationale.
5. If adopt, offer wiring. Edit the source repo, never the installed plugin cache (~/.claude/plugins/cache/... is a copy that gets overwritten). Resolve the source via ${PROJECT_KIT_HOME:-$HOME/Code/project-kit}. If that path isn't writable/available, emit a unified diff the user can apply with git apply. Typical targets:
   • add an MCP server to templates/mcp.json.tmpl
   • add a permission to templates/claude-settings.local.json
   • add a dependency to templates/pyproject.toml.tmpl
   • note a global install (marketplace add / plugin install) in README.md

Notes

• Editing files in this repo requires a session restart before Claude Code picks up plugin changes.
• Prefer trial over adopt when a tool is promising but unproven in this stack — record why.