ccg-workflow

CCG - Claude + Codex + Gemini Multi-Model Collaboration

[GitHub stars] [NPM Downloads] [npm version] [License: MIT] [CI] [codecov] [Claude Code] [Node] [Follow on X] [star] [Docs] [Ask DeepWiki]

简体中文 | English | Documentation

♥️ Sponsor

[PackyCode]

Thanks to PackyCode for sponsoring this project! PackyCode is a stable, high-throughput API gateway for leading models — one domain, one key, automatic failover, 97% availability. Dedicated high-speed routes for Codex and Claude Code. Pay in RMB 1:1 with no FX markup or top-up fees; new users get a first-charge discount plus $1 in free credits, and eligible groups start from 20% of list price. Sign up here to get started.

💡 PackyCode also exposes an Anthropic-compatible endpoint, so it can serve as your API provider for Claude Code itself — run npx ccg-workflow, pick PackyCode at Step 1, and paste your key.

---

[APIMart]

Thanks to APIMart for sponsoring this project! APIMart is a low-cost API platform for AI image & video generation — GPT-Image-2 from $0.006/image, 160+ images per dollar. One async API covers both image and video: submit a task, get an ID, fetch results via polling or callback. Batch tens of thousands of images without timeouts, switch models without changing code. Pay-as-you-go with no monthly fee — sign up here to get started.

💡 APIMart also exposes an Anthropic-compatible endpoint, so it can serve as your API provider for Claude Code itself — run npx ccg-workflow, pick APIMart at Step 1, and paste your key.

---

[x6.ai]

x6.ai: One API key for every major model. Official-key and proxy channels, up to 90% off, free credits on sign-up. Pay with Alipay, Google Pay, Apple Pay, Visa or Mastercard.

💡 x6.ai also exposes an Anthropic-compatible endpoint, so it can serve as your API provider for Claude Code itself — run npx ccg-workflow, pick x6.ai at Step 1, and paste your key.

---

Choose the skills you use (v3.7.0)

Install native skills independently, or select them during the CCG setup wizard. Each selected skill gets a ccg- entry; its scripts and references live in a versioned source cache.

npx ccg-workflow skills catalog
npx ccg-workflow skills install verify-quality seo-page-builder
npx ccg-workflow skills list --installed --json
npx ccg-workflow skills update
npx ccg-workflow skills remove seo-page-builder
# Codex native skills
npx ccg-workflow skills install verify-quality --target codex
# Configure the framework with a saved selection
npx ccg-workflow init --skills verify-quality,verify-security
npx ccg-workflow doctor --json

Claude Code uses /ccg-verify-quality; Codex uses $ccg-verify-quality. The manager exposes only selected entries, keeps supporting resources together, refuses ownership conflicts and preserves edited caches. The menu has S. Manage Skills and V. Doctor. Reinstalling retains the latest selection; --skills all selects the catalog and --skills none selects none. Existing full installations remain available as compatibility mode.

doctor --json returns a versioned report with check IDs, status, remediation, and failure/warning counts. It detects the actual Claude/Codex/Skills-only mode, validates the selected backends (including review routes), Hook registrations and managed skills. Failures exit with code 1. Version and filesystem checks establish configuration health; they do not call a model or prove vendor API access.

Standalone Codex mode checks its own CLI, configuration and installed files. Missing shared Claude routing is optional; missing external delegation tools are reported as warnings with setup instructions.

Model-action Hooks now follow the configured routes and concrete model IDs, avoid duplicate calls to one backend, and use native Agents for pure Claude setups. Native mode uses on-demand skill discovery; legacy knowledge excerpts have a fixed size limit.

Use an OpenCode worker, including DeepSeek

Configure your provider in OpenCode with /connect, then query opencode models (or opencode models deepseek) for its actual model IDs. In CCG's model menu choose OpenCode and enter that provider/model. Keep Claude Code as the coordinator; the worker runs through the existing wrapper. CCG preserves the chosen model when changing routing, reinstalling or updating, and does not manage the provider credentials.

🐳 CCG on DeepSeek Harness — dsh-ccg

The same role matrix, running natively inside DeepSeek Harness. It ships inside this package — no second install, no second version to track.

npx ccg-workflow dsh install    # every profile found; --profile <name> for one
npx ccg-workflow dsh list       # which profiles have it

Or pick D. DeepSeek Harness from npx ccg-workflow.

Requires dsh 0.2.0-rc.2 or 0.2.1-alpha.2. After updating CCG, run npx ccg-workflow@latest dsh install again and restart dsh. Configuration now belongs to each profile's cordis.patch.yml; the CCG card is in Settings › Built-in plugins.

Seven role-pinned delegation tools, each on its own model with its own expert persona — and two things the Claude Code side cannot do as cleanly:

• Model panels. Give one role several models and they all answer the same brief independently, rendered side by side in the conversation. Nothing votes, nothing is averaged — disagreement is the finding.
• Live teammates. ccg_team hires a role as a colleague that stays alive across turns, owns its own files (a colliding hire is refused, not warned about), and reports back on its own. Every hire asks you to approve it first.

No external CLI, no binary bridge, no cold-start tax — every hop is a provider API request. Source lives in dsh-ccg/; full README →

🧩 Also by the author

DSH Marketplace — a directory of DeepSeek Harness plugins. 2,500+ indexed across 14 categories, bilingual EN / 中文.

Not a sponsor — same author as CCG, built out of a plain annoyance: DSH plugins are multiplying fast, but figuring out what one does, whether it still installs, and what the command actually is meant opening one GitHub README after another. So the install commands here are put through a throwaway container first. A plugin only gets a check mark if it really installs — and if it can't, the listing says why (buried in a monorepo subdirectory, never published to npm, and so on). A verified index, not another awesome list.

dsh plugin --profile web add dshmarketplace-plugin   # browse the store from inside DSH
npx dshmarketplace-cli add owner/repo                # install any plugin from your shell

There is also a Python SDK and a public API. Early days — bug reports and plugin submissions are very welcome on GitHub.

---

What is CCG?

CCG is a workflow engine for Claude Code. It turns Claude into a multi-model orchestrator — Claude stays in control while dispatching specialized work to Codex (OpenAI), Grok (xAI), Kimi Code (Moonshot), and Antigravity through a Go binary bridge.

One command. Describe what you want. The engine handles the rest.

npx ccg-workflow    # Install in 60 seconds

Architecture

Claude Code is the lead orchestrator. It analyzes your intent, selects a strategy, and manages the entire workflow. The Hook Engine injects state every turn so Claude never loses context — even after compaction. The codeagent-wrapper (a compiled Go binary) bridges Claude to external models for parallel analysis and review.

How It Works

You: /ccg:go add JWT authentication to this API

CCG Engine:
  1. Reads project context (git status, tech stack, file structure)
  2. Classifies: feature / L complexity / backend / high risk
  3. Selects strategy: full-collaborate
  4. Creates .ccg/tasks/add-jwt-auth/task.json
  5. Launches dual-model analysis (Codex + Gemini in parallel)
  6. Produces plan → HARD STOP for your approval
  7. Spawns Agent Teams Builders for parallel implementation
  8. Runs quality gates + dual-model cross-review
  9. Reports results

Every turn, a hook injects:
  <ccg-state>
  Task: add-jwt-auth (in_progress)
  Strategy: full-collaborate
  Phase: 4-implementation
  </ccg-state>

10 Built-in Strategies

The engine auto-selects the right strategy based on task type and complexity:

Strategy	When	External Models	Agent Teams
direct-fix	Simple bug, single file	—	—
quick-implement	Small feature, clear scope	—	—
guided-develop	Medium feature, needs planning	Single	—
full-collaborate	Complex feature, multi-module	Dual parallel	✓
debug-investigate	Complex bug, unknown cause	Dual diagnosis	—
refactor-safely	Code restructuring	Dual review	—
deep-research	Technical research	Dual exploration	—
optimize-measure	Performance optimization	Optional	—
review-audit	Code review	Dual cross-review	—
git-action	commit, rollback, branches	—	—

Simple tasks run fast with zero overhead. Complex tasks get the full engine.

Core Features

Hook Engine — Never Lose Context

4 JavaScript hooks inject state into every Claude Code session:

Hook	Event	What it does
workflow-state.js	Every turn	Injects current task state as breadcrumb
session-start.js	Session start/compact	Re-injects full project context
subagent-context.js	Agent/Bash spawn	Injects spec directly into subagent prompts
skill-router.js	Every turn	Auto-injects domain knowledge by keyword

Context survives compaction. Sub-agents born with spec in their prompt. Zero state loss.

Task System — Persistent Lifecycle

Medium+ complexity tasks get a persistent directory:

.ccg/tasks/add-jwt-auth/
├── task.json         # Status, strategy, phase, gate
├── requirements.md   # Enhanced requirements
├── plan.md           # Approved implementation plan
├── context.jsonl     # Spec files for sub-agent injection
├── review.md         # Review results
└── research/         # Persisted research findings

Quality Gates — Built-in Security & Quality

Gate	Trigger
/ccg:verify-security	New modules, security changes
/ccg:verify-quality	Changes > 30 lines
/ccg:verify-change	Doc sync check
/ccg:verify-module	Module structure check
/ccg:gen-docs	Auto-generate README + DESIGN

100+ Domain Knowledge Files

When your message mentions security, caching, RAG, Kubernetes, etc., the relevant knowledge file is auto-injected. 10 domains, 61 files:

Security · Architecture · DevOps · AI/MLOps · Development · Frontend Design · Infrastructure · Mobile · Data Engineering · Orchestration

Standalone Skills — battle-tested website workflows (v3.5.1)

Self-contained skills, each usable on its own or auto-triggered by intent. Drawn from real site-building work and shipped with their scripts:

Skill	What it does
/ccg:bt-panel	Drive a BaoTa / aaPanel server over its HTTP API — deploy a build, update a live site, read/write files, run shell, run MySQL. No SSH or rsync; just a panel URL + API key. Ships bt_client.py + one-shot bt_deploy.py.
/ccg:seo-page-builder	Build, audit, or optimize SEO tool pages (AI generator / remover / enhancer / converter / editor). SERP-intent driven, with a runnable onpage-audit.py keyword-density meter.
/ccg:adsense-site-auditor	Audit a site for Google AdSense readiness against the full official checklist — eligibility, ownership, content quality, ads.txt, privacy, Publisher Policies — before you apply or after a rejection.

bt-panel reads credentials only from env vars / a git-ignored sites.json; nothing is hardcoded. seo-page-builder is adapted from yuzeiki's skill of the same name (see its SKILL.md).

Commands

Core (v3.0 default: 13 commands)

Command	Description
/ccg:go	Smart entry — describe what you want, engine handles the rest
/ccg:commit	Smart conventional commit
/ccg:rollback	Interactive rollback
/ccg:clean-branches	Clean merged branches
/ccg:worktree	Worktree management
/ccg:init	Initialize project CLAUDE.md
/ccg:context	Project context management

OpenSpec Integration

Command	Description
/ccg:spec-init	Initialize OPSX environment
/ccg:spec-research	Requirements → constraints
/ccg:spec-plan	Constraints → zero-decision plan
/ccg:spec-impl	Execute plan + archive
/ccg:spec-review	Dual-model cross-review

Legacy Mode (18 additional commands)

Includes /ccg:workflow, /ccg:plan, /ccg:execute, /ccg:frontend, /ccg:backend, /ccg:analyze, /ccg:debug, /ccg:optimize, /ccg:test, /ccg:review, /ccg:team, and more.

Quick Start

Option A — full install (multi-model orchestration + commands + hooks + binary, recommended)

# Install (interactive 4-step wizard)
npx ccg-workflow

# Or non-interactive with defaults
npx ccg-workflow init --skip-prompt

Requires Node.js 20+ and Claude Code CLI. Codex CLI, Grok CLI, Kimi Code CLI, and Antigravity are optional (enable multi-model features).

Option B — native plugin (skills only, zero dependencies, v3.5.1+)

Want just the skills (web-ops toolkit, quality gates, frontend-design, domain knowledge) without the multi-model orchestration?

claude plugin marketplace add fengshao1227/ccg-workflow
claude plugin install ccg@ccg

Skills are invoked as /ccg:<skill>. No npx, no binary. The multi-model commands ship only through Option A (their templates carry install-time placeholders that a native plugin can't resolve).

CLI Commands

npx ccg-workflow                          # Interactive menu
npx ccg-workflow init                     # 4-step install wizard
npx ccg-workflow doctor                   # Environment health check
npx ccg-workflow doctor --json            # Machine-readable health report
npx ccg-workflow skills list              # Catalog and native selection
npx ccg-workflow status                   # Installation overview
npx ccg-workflow codex-mode install       # Install Codex-Led mode
npx ccg-workflow codex-mode uninstall     # Uninstall Codex-Led mode
npx ccg-workflow dsh install              # Install CCG into DeepSeek Harness
npx ccg-workflow dsh list                 # Which dsh profiles have it
npx ccg-workflow dsh uninstall            # Remove it from every profile
npx ccg-workflow uninstall                # Uninstall CCG
npx ccg-workflow config mcp               # Configure MCP tokens
npx ccg-workflow diagnose-mcp             # Diagnose MCP issues

Configuration

~/.claude/
├── commands/ccg/          # Slash commands
├── hooks/ccg/             # Hook scripts (5 files)
├── skills/ccg/            # Legacy bundled skills
├── skills/ccg-<slug>      # Selected native skill entries
├── .ccg-skills/           # Native selection manifest and resource cache
├── rules/                 # Auto-trigger rules
├── .ccg/
│   ├── config.toml        # Model routing, MCP, performance
│   ├── engine/            # 10 strategy files + model router
│   └── prompts/           # Expert prompts (codex/gemini/claude)
└── bin/codeagent-wrapper  # Multi-model bridge (Go binary)

Environment Variables

Set in ~/.claude/settings.json under "env":

Variable	Default	Description
CODEX_TIMEOUT	7200	Wrapper timeout (seconds)
CODEAGENT_POST_MESSAGE_DELAY	5	Post-completion delay
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS	unset	Set 1 for parallel Agent Teams

Update / Uninstall

npx ccg-workflow@latest     # Update to latest
npx ccg-workflow doctor     # Check health after update
npx ccg-workflow uninstall  # Clean uninstall

Credits

• stellarlinkco/myclaude — codeagent-wrapper inspiration
• UfoMiao/zcf — Git tools reference
• mindfold-ai/Trellis — Hook-based workflow state patterns
• ace-tool — MCP code retrieval

Contributors

Contact

• X (Twitter): @CCG_Workflow
• Email: fengshao1227@gmail.com
• Issues: GitHub Issues
• Community: Linux.do

Star History

[Star History Chart]

License

MIT

---

v3.7.3 | Issues | Contributing