What It Does
AI coding agents skip diagnosis, forget reviews, and repeat mistakes. Summoner adds a process layer — checkpoints between phases, post-game reviews, and a memory chain that auto-recalls past fixes.
Quick Start
git clone https://github.com/johnson-xue/summoner.git ~/.claude/plugins/summoner/
cd ~/.claude/plugins/summoner/hooks && make build
cd your-project
~/.claude/plugins/summoner/scripts/summoner-init.sh
~/.claude/plugins/summoner/scripts/init-memory-db.sh $(grep -A2 '^project:' summoner.yaml | grep 'name:' | head -1 | awk '{print $2}')
How It Works
/summoner:fix "线上报错 SC_ErrInnerLogic..."
Phase 0 Memory — auto-recalls past bug patterns
Phase 1 Diagnosis — root cause (iron law: never skip)
⏸️ Checkpoint → continue / skip / recall (redo) / stop
Phase 2 Reproduce — Prove-It test (auto-skipped for config-only fixes)
Phase 3 Fix
Phase 4 Verify — run test suite
Phase 5 Review — code review
📋 Post-Game Review — 5 types, auto-journaled to SQLite
Commands
| Command | Pipeline |
|---|
/summoner:fix | diagnose → reproduce → fix → verify → review |
/summoner:new | define → plan → implement → test → review |
/summoner:ship | fan-out 1-3 reviewers → merge → go/no-go |
/summoner:debug | diagnosis only — no code changes |
/summoner:ops | server operations (delegated to project skill) |
/summoner:review | standalone code review |
Platform Support
| Platform | Commands | Memory | Hooks |
|---|
| Claude Code | ✅ | ✅ SQLite | ✅ Go |
| Gemini CLI | ✅ | ✅ bash | — |
| OpenCode | ✅ | ✅ bash | — |
| Cursor / Windsurf / Copilot / Aider | ✅ | ✅ bash | — |
Token Cost
| Workflow | Tokens | Overhead |
|---|
/summoner:fix (with memory) | ~9,300 | +86% |
/summoner:fix (no memory) | ~8,300 | +66% |
/summoner:debug | ~4,300 | +35% |
Multi-step workflows → Summoner. Single-step tasks → direct skill.
