// the find
fengshao1227/ccg-workflow
多模型协作工作流引擎 — /ccg:go 一个命令，AI 自动分析意图、选择策略、编排 Codex + Gemini + Claude 协作执行
CCG is a workflow layer on top of Claude Code that orchestrates multiple LLMs — Claude as lead, with Codex, Gemini, and Grok available for parallel analysis and review. It installs as an npm package and injects slash commands, hooks, and strategy files into your `~/.claude/` config. The target user is someone who already lives in Claude Code and wants structured multi-agent workflows without writing the orchestration themselves.
The hook engine that re-injects task state every turn is the most practically useful idea here — context loss after Claude Code compaction is a real problem and this solves it with a simple `<ccg-state>` breadcrumb. The strategy selection (10 named strategies, auto-picked based on task complexity) keeps simple tasks fast and avoids unnecessary overhead. The Go binary bridge for calling Codex/Gemini in parallel is a clean solution to a real impedance mismatch — these models don't share a common subprocess interface. The persistent task directory under `.ccg/tasks/` with `task.json`, `plan.md`, and `review.md` gives you an audit trail and lets sub-agents share context without relying on in-memory state.
The repo is tagged as Go but is primarily TypeScript/Node.js installer code — the Go piece is a small subprocess wrapper. Anyone who picks this up expecting a Go codebase will be confused. The multi-model value proposition depends on having Codex CLI and Gemini CLI separately installed and authenticated, which is a non-trivial setup that the docs understate. The 100+ domain knowledge files auto-injected by keyword matching sounds useful but is essentially prompt stuffing — injecting a security knowledge file because your message mentions 'auth' adds tokens without any guarantee of relevance, and there's no mechanism to see what got injected or why. The quality gates (`/ccg:verify-security`, `/ccg:verify-quality`) are slash commands that call Claude again, so they only catch what Claude would catch anyway — calling it a 'gate' implies more enforcement than actually exists.