HN 표시: VibeHQ 여러 CLI 에이전트를 실제 회사 팀으로 조율

🌐 Language: English | 繁體中文 | 日本語 Running 5 AI agents in parallel is easy. Making them not break each other's code is the hard part. VibeHQ adds contracts, task tracking, and idle-aware messaging to Claude Code, Codex & Gemini CLI — so they work like an actual engineering team, not 5 interns editing the same file. Every "multi-agent" tool lets you run multiple CLI agents in parallel. But parallel ≠ collaboration. Here's what actually happens when 5 agents build the same app: | What Goes Wrong | Real Example from Our Logs | |---|---| | Schema conflicts — each agent invents its own JSON format | Frontend expects { data: [] } , backend writes { results: [] } , third agent creates its own copy | | Orchestrator role drift — the PM starts writing code | PM spent 6 manual JS patches fixing integration bugs instead of coordinating | | Ghost files — agents publish 43-byte stubs instead of real content | Agent writes full file via share_file , then puts "See local file..." in publish_artifact . Loop repeats for 68 minutes | | Premature execution — agents start before dependencies are ready | Agent sees QUEUED task description, ignores the status, starts coding with hardcoded data | | Silent failures — crashed agents produce no signal | Orchestrator waits 18 minutes for a response from a dead process | These aren't edge cases. They're LLM-native behavioral patterns that reliably appear across model families. We documented 7 of them with full session logs. VibeHQ is a teamwork protocol layer that sits on top of real CLI agents. Each agent stays a full Claude Code / Codex / Gemini process with all native features — VibeHQ adds the coordination they're missing: | Problem | VibeHQ's Fix | |---|---| | Schema conflicts | Contract system — agents must sign API specs before coding begins | | Role drift | Structured task lifecycle — create → accept → in_progress → done with required artifacts | | Ghost files | Hub-side validation — rejects publish_artifact calls with stub content (/SKILL.md | ~/.claude/skills/ | | Codex CLI | .agents/skills//SKILL.md | ~/.codex/skills/ | The SKILL.md format is an emerging cross-platform standard — same frontmatter (name , description ), same markdown body. A skill created for one platform works on the other. Claude Code — skills are already included in .claude/skills/ . Just use them: # In Claude Code, type: /run-teamwork "Build an AI investment analysis platform" /benchmark-loop "Build a todo app" --grade A /optimize-protocol v1 Codex CLI — copy the skills to Codex's directory: # Project-level (committed to repo) mkdir -p .agents/skills cp -r .claude/skills/run-teamwork .agents/skills/ cp -r .claude/skills/optimize-protocol .agents/skills/ cp -r .claude/skills/benchmark-loop .agents/skills/ # Or user-level (available in all projects) cp -r .claude/skills/run-teamwork ~/.codex/skills/ cp -r .claude/skills/optimize-protocol ~/.codex/skills/ cp -r .claude/skills/benchmark-loop ~/.codex/skills/ Then in Codex CLI, invoke with /skills or type $ to mention a skill. Give it a project description — it designs the team, spawns agents, and builds it. No analysis, no loop. /run-teamwork "Build an e-commerce site with payments and admin panel" - Analyzes the prompt to determine required domains and team size - Generates PM system prompt with research-first workflow (research before implementation) - Spawns agents in tmux (macOS/Linux) or Windows Terminal - Waits for all tasks to complete - Reports the output directory and file count Reads analysis data and writes real code fixes (not parameter tuning): /optimize-protocol v1 # Read analysis for run v1, implement fixes - Loads current run + all previous optimization reports - Builds cross-run trend table (what's improving, what regressed, what's a side-effect) - Classifies each problem as NEW, RECURRING, or SIDE-EFFECT of a previous fix - Implements real TypeScript changes to the framework - Verifies build passes - Saves a detailed changelog to ~/.vibehq/analytics/optimizations/ Runs the full self-improving cycle automatically: /benchmark-loop "Build a Todo app with REST API, React frontend, and WebSocket real-time updates" - Spawns a fresh team with a standardized project - Waits for the team to finish (heartbeat monitoring) - Analyzes session logs (13 rules + LLM grading) - Triggers /optimize-protocol to write code fixes - Rebuilds the framework ( npx tsup ) - Repeats with a new team — zero human intervention The underlying tools are regular CLI commands — no skills required: # 1. Run a benchmark vibehq start --team your-team # 2. Analyze vibehq-analyze --team your-team --with-llm --save --run-id v1 # 3. Auto-optimize (Claude Code / Codex skill) /optimize-protocol v1 # 4. Run again, compare vibehq start --team your-team vibehq-analyze --team your-team --with-llm --save --run-id v2 vibehq-analyze compare v1 v2 All optimization reports are saved to ~/.vibehq/analytics/optimizations/ for tracking and auditing. Supports both Claude Code and Codex CLI native JS