Excalibur — Research-Integrated Agent Harness with Self-Improvement
Full project orchestration powered by Claude Code
English | 한국어
xLoop is a Claude Code plugin that orchestrates planning, research, implementation, and self-improvement — all through a single keyword.
In Claude Code, run:
/plugin marketplace add HSUNEH/xLoop
/plugin install xloop@xloop
The plugin includes an optional HUD (status line) showing folder, git branch, context usage, tool/agent/skill counts, active plan, and 5h rate limit. The HUD requires a StatusLine hook (e.g., oh-my-claudecode). See hooks/scripts/xloop-hud.mjs.
Just type a keyword naturally in Claude Code:
excalibur "build a real-time chat app"
That's it. xLoop handles the rest:
Deep Interview → (Ralplan + Ralph + Eval) × N milestones → Done
| Keyword | What it does |
|---|---|
excalibur "..." |
Full project orchestration (office hours → interview → plan → implement → verify) |
ralph "..." |
PRD-driven implementation loop (TDD, code review, security audit) |
ralplan "..." |
6-agent consensus planning (design review, cross-model review, idea discovery) |
deep-interview "..." |
Socratic spec co-creation (used within excalibur or standalone) |
research "..." |
Multi-source investigation (web, arxiv, docs, NotebookLM) |
upgrade |
Self-improvement cycle (checksum → snapshot → implement → review → commit) |
rollback |
Restore from snapshot |
| Keyword | What it does |
|---|---|
slide-generator "..." |
Auto-generate presentations (research → outline → HTML design → PPTX) |
opportunity-scout "..." |
Trend-driven idea discovery (Google Trends + GitHub + YouTube + marketing frameworks) |
youtube-transcript <URL> |
Extract subtitles from YouTube videos (multi-language, 2-pass download) |
setup |
Installation wizard (CLAUDE.md, MCP server, config) |
Every task is automatically routed before orchestration:
| Score | Route | Example |
|---|---|---|
| 1 (simple) | Executor direct | "fix typo", "rename X" |
| 2 (medium) | Ralph only | multi-file task, clear scope |
| 3 (complex) | Ralplan + Ralph | architecture decisions, vague scope |
2-step gate: Structural heuristic (instant, 0 LLM cost) → Haiku micro-assessment (only when uncertain).
3-dimension scoring: Scope (40%) + Clarity (35%) + Decision (25%).
Big Loop #1 (M1: MVP)
├── Ralplan: Plan M1 (+ idea discovery, design review, cross-model review)
├── Ralph: Implement M1 stories (TDD, worktree, code review, security audit)
└── Eval: Auto-verify (5 metrics) + browser QA + sprint retro → learnings
│
Mode B: "M1 done (33%). Proceed / modify spec / upgrade?"
Mode C: 10-second checkpoint → auto-proceed
Big Loop #2 (M2 + M3 in parallel if independent)
├── Lane A: Ralplan(M2) → Ralph(M2)
├── Lane B: Ralplan(M3) → Ralph(M3) [concurrent]
└── Eval: Both complete → project done
Independent milestones run in parallel lanes. Conflict detected → sequential fallback.
Each loop generates .xloop/learnings/loop-{N}.json (technical, process, quality lessons) that feed into the next loop.
Ralph complete → Auto-verification (5 metrics) → Threshold check
│
Mode B: User decides → upgrade or accept
Mode C: Auto-upgrade if metrics fail (max 3 cycles)
│
▼
upgrade:
Checksum verify → Snapshot → Ralplan → Ralph → Review gate → Commit
Safety: SHA-256 checksum, git pre-commit hook, snapshot/rollback, multi-model review gate. Upgrades modify only xLoop files — never your project code.
| Agent | Model | Role |
|---|---|---|
| planner | Opus | Strategic planning, research needs identification |
| architect | Opus | Architecture review, steelman counterarguments |
| critic | Opus | Quality gate, principle-option consistency |
| researcher | Sonnet | Multi-source investigation |
| executor | Sonnet | Code implementation |
| verifier | Sonnet | Acceptance criteria verification |
| designer | Sonnet | UX/UI design review, component design |
| explorer | Haiku | Codebase search, quick lookups |
| Skill | Trigger | Purpose |
|---|---|---|
| excalibur | "excalibur" |
Full project orchestration (Interview → Ralplan → Ralph → Eval per milestone) |
| deep-interview | "deep-interview" |
Socratic spec co-creation with ambiguity gating |
| ralph | "ralph" |
PRD-driven implementation loop (TDD, code review, security audit) |
| ralplan | "ralplan" |
6-agent consensus planning (Planner → Researcher → Architect → Designer → Critic) |
| research | "research" |
Multi-source investigation (web, arxiv, docs, NotebookLM) |
| upgrade | "upgrade" |
Self-improvement cycle (checksum → snapshot → implement → review gate → commit) |
| rollback | "rollback" |
Snapshot restore |
| Skill | Trigger | Purpose |
|---|---|---|
| slide-generator | "slide-generator" / "PPT 만들어줘" |
Auto-generate presentations (research → outline → HTML design → PPTX) |
| opportunity-scout | "opportunity-scout" |
Trend-driven idea discovery with marketing framework analysis |
| youtube-transcript | "youtube-transcript" |
YouTube subtitle extraction (ko/en/ja, manual + auto-generated) |
| setup | "setup" |
Installation wizard (CLAUDE.md, MCP server, config) |
xLoop's Excalibur is a meta-orchestrator that absorbs the best ideas from 7 open-source projects:
| Source | What xLoop Takes | Excalibur Stage |
|---|---|---|
| OMC | Plugin/skill/hook architecture, Ralph PRD loop, Ralplan consensus | Foundation |
| Ouroboros | Self-improvement concept, constitutional governance, multi-model review | Upgrade |
| gstack | Office Hours reframing, Design Review, Code Review, Security Audit (CSO), Browser QA, Sprint Retro | Interview, Ralplan, Ralph, Eval |
| superpowers | TDD (Red-Green-Refactor), git worktree isolation, fresh subagent per task, 4-phase debugging | Ralph |
| ARIS | Cross-model adversarial review, idea discovery pipeline | Ralplan |
| autoresearch | Modify→eval→keep/discard experiment loop | Upgrade |
| notebooklm-py | NotebookLM API for source aggregation and synthesis | Research |
- Meta-Orchestrator: Absorbs ideas into SKILL.md — no runtime plugin dependency
- Complexity Gate: Bidirectional routing — scales DOWN for simple tasks, not just up
- Research-Integrated Planning: Research within the planning loop, not as a separate pre-step
- Milestone-Scoped Execution: Plan one chunk at a time, not the whole project
- Excalibur: Single keyword for entire project lifecycle
npm install # install dev dependencies
npm run typecheck # tsc --noEmit
npm test # vitest run (94 tests)MIT