This is a redesigned fork of fengshao1227/ccg-workflow. See What Changed below.

A competitive multi-model development system where every significant task is dispatched to ALL available models (Codex + Gemini) in parallel. Claude orchestrates, competes, compares outputs using benchmark-weighted evaluation criteria, and synthesizes the best solution.

Why competitive? Our benchmark analysis shows no single model dominates all tasks. Claude leads code quality (SWE-bench 80.8%), Codex leads autonomous execution (Terminal-Bench 77.3%), and Gemini leads visual design (WebDev Arena #1). Competitive dispatch ensures you always get the best output.

The original CCG uses static routing — frontend tasks always go to Gemini, backend tasks always go to Codex. This fork changes that completely.

Frontend task → Gemini only (assumed "frontend expert")
Backend task → Codex only (assumed "backend expert")
Claude → Orchestrator (doesn't compete)

The problem: these assumptions aren't backed by benchmarks. SWE-bench scores are nearly tied across all three models (~80%), and each model has different strengths that don't align with a simple frontend/backend split.

Any task → Send to ALL models in parallel → Compare outputs with weighted scores → Use the best one

Every task goes to Codex + Gemini + Claude at the same time. Claude then scores all three outputs using benchmark-weighted criteria and picks the best one (or combines the best parts from each).

We fact-checked the original claims against published benchmarks:

 ┌─→ Codex CLI ─→ Output A ─┐
 │ │
Task ─→ Claude ─────┼─→ Gemini CLI ─→ Output B ─┼─→ Weighted ─→ Best
 (also │ │ Compare Solution
 generates └─→ Claude Self ─→ Output C ─┘
 Output C)

External models have no write access — they only return patches/analysis, which Claude evaluates and applies.

A full /ccg:workflow run goes through 6 phases. At every phase that involves model output, all models compete in parallel:

flowchart TD
 Start["🚀 User runs /ccg:workflow"] --> P1
 subgraph P1["Phase 1 — Research"]
 R1["Analyze requirements"]
 R2["Search codebase via MCP"]
 R3["Score completeness (≥7 to proceed)"]
 R1 --> R2 --> R3
 end
 P1 --> P2
 subgraph P2["Phase 2 — Ideation (Competitive)"]
 direction LR
 C2["Codex analyzes"] & G2["Gemini analyzes"] & CL2["Claude analyzes"]
 end
 P2 --> W2["⚖️ Weighted scoring → Best analysis"]
 W2 --> P3
 subgraph P3["Phase 3 — Planning (Competitive)"]
 direction LR
 C3["Codex plans"] & G3["Gemini plans"] & CL3["Claude plans"]
 end
 P3 --> W3["⚖️ Weighted scoring → Best plan"]
 W3 --> Approve{"👤 User approves plan?"}
 Approve -- No --> P3
 Approve -- Yes --> P4
 subgraph P4["Phase 4 — Execution"]
 E1["Claude implements code"]
 E2["(sole code writer — external models have zero write access)"]
 E1 --> E2
 end
 P4 --> P5
 subgraph P5["Phase 5 — Review (Competitive)"]
 direction LR
 C5["Codex reviews"] & G5["Gemini reviews"] & CL5["Claude reviews"]
 end
 P5 --> W5["⚖️ Weighted consensus → Issues ranked by agreement (3/3, 2/3, 1/3)"]
 W5 --> Fix{"Critical issues?"}
 Fix -- Yes --> FixCode["Claude fixes"] --> P5
 Fix -- No --> P6
 subgraph P6["Phase 6 — Final Review"]
 F1["Verify against plan"]
 F2["Run tests"]
 F3["Report to user"]
 F1 --> F2 --> F3
 end
 P6 --> Done["✅ Done"]
 style P2 fill:#e8f4f8,stroke:#2196F3
 style P3 fill:#e8f4f8,stroke:#2196F3
 style P5 fill:#e8f4f8,stroke:#2196F3
 style W2 fill:#fff3e0,stroke:#FF9800
 style W3 fill:#fff3e0,stroke:#FF9800
 style W5 fill:#fff3e0,stroke:#FF9800

Each model's output is scored on multiple dimensions. The weights depend on the task type:

 Codex Gemini Claude
 ───── ────── ──────
Tech depth 0.30 0.20 0.35 ← Claude weighted highest
Risk detection 0.35 0.15 0.30 ← Codex weighted highest
UX / visual 0.10 0.40 0.20 ← Gemini weighted highest
Actionability 0.25 0.25 0.25 ← Equal
Score = Σ (dimension rating ×ばつ weight)
Winner = highest total score
Final output = winner + cherry-picked improvements from others

Different task types use different weight tables — see /ccg:routing-guide for all of them.

npx ccg-workflow

Requirements: Claude Code CLI, Node.js 20+

Important: This project depends on ora@9.x and string-width@8.x, which require Node.js >= 20.

Optional: Codex CLI (competitive dispatch), Gemini CLI (competitive dispatch)

Routing decisions are based on published benchmarks, not assumptions:

See /ccg:routing-guide for full weighted evaluation criteria per task type.

Every competitive dispatch uses task-type-specific weights:

Analysis: Tech depth (Claude .35) | Risk (Codex .35) | UX (Gemini .40) | Action (.25 each)
Planning: Architecture (Claude .40) | Scale (Claude .30) | UI (Gemini .40) | Detail (Codex .30)
Implementation: Correctness (Claude .35) | Visual (Gemini .45) | Edge cases (Codex .35) | Maintain (Claude .35)
Review: Security (Codex .35) | Visual (Gemini .40) | Logic (Claude .35) | Perf (Codex .30)

Integrates OPSX architecture to turn requirements into constraints:

/ccg:spec-init
/ccg:spec-research implement user authentication
/ccg:spec-plan
/ccg:spec-impl
/ccg:spec-review

/ccg:team-research implement real-time collaboration kanban API
/ccg:team-plan kanban-api
/ccg:team-exec
/ccg:team-review

Prerequisite: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in settings.json

~/.claude/
├── commands/ccg/ # Slash commands
├── agents/ccg/ # Sub-agents
├── skills/ # Quality gates
├── bin/codeagent-wrapper
└── .ccg/
 ├── config.toml
 └── prompts/{codex,gemini}/

CCG automatically installs a Hook to auto-authorize codeagent-wrapper commands.

Requirement: jq must be installed.

Code retrieval MCP (choose one):

• ace-tool (recommended)
• ContextWeaver (alternative)

Optional: Context7, Playwright, DeepWiki, Exa

npx ccg-workflow@latest # Update
npx ccg-workflow # Select "Uninstall"

• cexll/myclaude - codeagent-wrapper
• UfoMiao/zcf - Git tools
• GudaStudio/skills - Original routing design
• ace-tool - MCP tool

Star History Chart

MIT

v1.8.0 (Competitive Multi-Model Redesign) | Issues