tri-agent-orchestrator — скилл для Claude Code: настройка и установка
Скилл tri-agent-orchestrator для Claude Code распределяет инженерные задачи между Antigravity, Claude Code и Codex. Он подключается к запросам на исследование, разработку, аудит безопасности и тестирование.
MIT ★ 57 обновлён 2026-11-12
npx skills add abhigyan6/tri-agent-orchestrator --skill tri-agent-orchestrator -a claude-codeЧто делает tri-agent-orchestrator
Скилл координирует работу трёх агентов: Antigravity исследует проект и проектирует решение, Claude Code пишет и рефакторит код, а Codex генерирует тесты и проверяет сборку. Для передачи работы агенты обновляют HANDOFF.md в корне проекта.
Есть три сценария: последовательная разработка функции, сравнение архитектурных предложений и проверка кода на уязвимости. Для запуска из терминала используется CLI tri-agent; он поддерживает команды doctor, run и handoff. Для работы нужны CLI Antigravity (agy), Claude Code (claude), Codex (codex) и tri-agent.
Конвейер tri-agent run останавливается, если в HANDOFF.md неверно указаны Completed By и Next Agent или эти поля не изменились после этапа.
Какие задачи решает
- распределить исследование, реализацию и тестирование функции между агентами
- сравнить архитектурные варианты и записать решение в ADR
- найти уязвимости в логике аутентификации или загрузки файлов
- передать следующему агенту состояние работы через
HANDOFF.md
Примеры запросов агенту
- Организуй разработку функции: сначала исследуй требования, затем реализуй код и запусти тесты.
- Сравни два подхода к хранению сессий и запиши итоговое решение в ADR.
- Проведи аудит безопасности конечной точки загрузки файлов и проверь исправления тестами.
Как установить tri-agent-orchestrator
1Через Skills CLI (Vercel)
npx skills add abhigyan6/tri-agent-orchestrator --skill tri-agent-orchestrator -a claude-code2Вручную: положить файлы
~/.claude/skills/tri-agent-orchestrator/SKILL.mdПапка скилла целиком; для одного проекта — .claude/skills/tri-agent-orchestrator/
1Через Skills CLI (Vercel)
npx skills add abhigyan6/tri-agent-orchestrator --skill tri-agent-orchestrator -a codex2Вручную: положить файлы
.agents/skills/tri-agent-orchestrator/SKILL.mdДля всех проектов — ~/.agents/skills/
1Через Skills CLI (Vercel)
npx skills add abhigyan6/tri-agent-orchestrator --skill tri-agent-orchestrator -a cursor2Вручную: положить файлы
.cursor/skills/tri-agent-orchestrator/SKILL.mdCursor читает и .agents/skills/, и .claude/skills/
1GitHub CLI
gh skill install abhigyan6/tri-agent-orchestrator tri-agent-orchestrator2Вручную: положить файлы
.github/skills/tri-agent-orchestrator/SKILL.md1Через Skills CLI (Vercel)
npx skills add abhigyan6/tri-agent-orchestrator --skill tri-agent-orchestrator -a gemini-cli2Командой платформы
gemini skills install https://github.com/abhigyan6/tri-agent-orchestratorВручную: положить файлы
.devin/skills/tri-agent-orchestrator/SKILL.mdСтарый путь .windsurf/skills/ тоже читается
Claude.ai
Упакуйте папку в zip и загрузите в Settings → Capabilities → Skills
По шагам, с командами для каждой платформы: Как установить скилл в Claude Code: пять способов
Вопросы о tri-agent-orchestrator
Как установить tri-agent-orchestrator в Claude Code?
Выполни npx skills add abhigyan6/tri-agent-orchestrator --skill tri-agent-orchestrator -a claude-code. Файл скилла находится по пути ~/.claude/skills/tri-agent-orchestrator/SKILL.md.
Что делает tri-agent-orchestrator?
Распределяет этапы инженерной задачи между Antigravity, Claude Code и Codex. Агенты передают состояние работы через HANDOFF.md.
Что нужно для работы tri-agent-orchestrator?
Нужны CLI agy, claude, codex и tri-agent. Воркфлоу запускается через tri-agent run.
Исходный текст
SKILL.md · MIT
---
name: tri-agent-orchestrator
description: Coordinate complex engineering workflows across Google Antigravity (agy), Anthropic Claude Code (claude), and OpenAI Codex (codex). Orchestrates specialized research, architecture, implementation, adversarial auditing, testing, and deployment pipelines using standardized inter-agent handoffs.
---
# Tri-Agent Orchestrator: Antigravity × Claude Code × Codex
The **Tri-Agent Orchestrator** skill enables seamless coordination across the three industry-leading coding agent CLIs:
1. **Google Antigravity (`agy`)** – Architecture, multimodal analysis, deep research, and adversarial auditing.
2. **Anthropic Claude Code (`claude`)** – Domain modeling, high-taste implementation, nuanced refactoring, and clean diffs.
3. **OpenAI Codex (`codex`)** – Rapid scaffolding, test suite generation, terminal task execution, and deployment verification.
When this skill is active, any participating agent acts according to its core specializations and produces standardized handoff contracts (`HANDOFF.md`) so the next agent can immediately resume with zero loss of context.
---
## 1. Agent Specialization Matrix
| Agent | Primary Role | Core Strengths | Best Assigned Tasks |
| :--- | :--- | :--- | :--- |
| **Antigravity (`agy`)** | **Chief Architect & Auditor** | Multimodal/visual reasoning, broad codebase mapping, security review, complex system design | • Architectural specifications<br>• API design & RFC generation<br>• Adversarial edge-case discovery<br>• Multimodal UI/UX visual validation |
| **Claude Code (`claude`)** | **Lead Software Craftsman** | Contextual subtlety, clean idiomatic code, complex refactoring, high architectural taste | • Domain logic & core algorithms<br>• Clean component & API implementation<br>• Technical debt refactoring<br>• PR summary & documentation |
| **OpenAI Codex (`codex`)** | **Reliability & Execution Engineer** | Fast execution, test generation, bash/terminal automation, build & deploy pipelines | • Unit, integration, & e2e test harnesses<br>• Scaffolding boilerplate<br>• Git operations, CI/CD, Dockerfiles<br>• Build & lint verification |
---
## 2. Core Workflow Protocols
### Protocol A: The Relay Pipeline (Default)
Ideal for end-to-end features requiring deep planning, high-quality code, and exhaustive testing.
```
[Phase 1: Research & Spec] ────► [Phase 2: Core Build] ────► [Phase 3: Test & Deploy]
Antigravity Claude Code Codex
Outputs: SPEC.md Outputs: Source Code Outputs: Test Suite &
& HANDOFF.md & HANDOFF.md Deployment PR
```
1. **Step 1 (Antigravity)**: Research requirements, explore relevant repository files, evaluate trade-offs, and draft `docs/tri-agent/SPEC-<feature>.md`. Write a summary handoff in `HANDOFF.md`.
2. **Step 2 (Claude Code)**: Ingest `SPEC-<feature>.md` and `HANDOFF.md`. Implement production code cleanly, adhering strictly to the architectural constraints. Update `HANDOFF.md`.
3. **Step 3 (Codex)**: Ingest updated source files and `HANDOFF.md`. Generate comprehensive tests (unit, edge-case, integration), execute linting/test runners, fix any regressions, and stage the commit/PR.
---
### Protocol B: Architecture Debate & Consensus
Ideal for high-stakes technical decisions (database selection, state management, migration strategies).
1. **Antigravity** drafts Proposal A focusing on scalability, security, and ecosystem compatibility.
2. **Claude Code** drafts Proposal B focusing on developer ergonomic simplicity, maintainability, and clean abstractions.
3. **Codex** benchmarks or simulates both approaches via test scripts and generates a trade-off scorecard.
4. The winning consensus is logged into an Architecture Decision Record (`docs/adr/ADR-xxx.md`).
---
### Protocol C: Adversarial Red-Team Audit
Ideal for security-critical logic (authentication, payment processing, file uploads, permissions).
1. **Claude Code or Codex** writes the initial implementation.
2. **Antigravity** acts as the Adversarial Red-Team:
- Inspects attack vectors (IDOR, race conditions, path traversal, injection, unhandled promise rejections).
- Generates reproduction steps or failing test specs.
3. **Claude Code** fixes the vulnerabilities.
4. **Codex** verifies all security regression tests pass.
---
## 3. Standard Inter-Agent Handoff Contract (`HANDOFF.md`)
Whenever an agent finishes its phase in a multi-agent workflow, it must write or update `HANDOFF.md` in the workspace root with this exact schema. The `tri-agent run` pipeline checks `Completed By` and `Next Agent` after each phase and stops if they are wrong or unchanged, so a handoff that names the wrong successor halts the run:
```markdown
# Tri-Agent Handoff Contract
## Metadata
- **Workflow ID:** `<unique-task-id>`
- **Active Feature:** `<feature-name>`
- **Current Phase:** `[Phase 1: Spec | Phase 2: Build | Phase 3: Verify]`
- **Completed By:** `<agy | claude | codex>`
- **Next Agent:** `<agy | claude | codex>`
- **Timestamp:** `<ISO-8601>`
## What Was Done
- Brief bullet points of files created, modified, or researched.
## Key Files & Artifacts
- Specification: `path/to/spec.md`
- Source Files: `path/to/code.ts`
- Tests: `path/to/test.ts`
## Critical Constraints & Decisions
- Specific requirements or patterns the next agent must strictly preserve.
## Verification Checklist for Next Agent
- [ ] Task 1
- [ ] Task 2
- [ ] Run test command: `<command>`
```
---
## 4. CLI Helper Integration
When executing from terminal or scripts, use the bundled `tri-agent` CLI tool:
```bash
# Verify all agents are installed and healthy
tri-agent doctor
# Run an automated workflow: relay (default), debate, or audit
tri-agent run --workflow relay --prompt "Build modern auth middleware with JWT rotation"
tri-agent run --workflow debate --prompt "Choose a session store for the cart"
tri-agent run --workflow audit --feature uploads --prompt "Harden the file upload endpoint"
# Preview the exact commands without invoking any agent
tri-agent run --workflow relay --prompt "Build file upload component" --dry-run
# Generate a handoff document
tri-agent handoff create --completed agy --next claude --feature "auth-middleware"
# Summarize the current workspace handoff state (add --raw for the full file)
tri-agent handoff status
```
---
## 5. Agent Instructions When This Skill Is Triggered
When a user prompts you with a tri-agent task:
1. **Identify your current persona** (`agy`, `claude`, or `codex`).
2. **Check for existing `HANDOFF.md`** in the repository root:
- If `HANDOFF.md` exists and you are the `Next Agent`, resume execution according to the checklist.
- If no `HANDOFF.md` exists, you are the initiating agent. Begin Phase 1.
3. **Execute your scoped role** without overstepping into future phases unless explicitly instructed.
4. **Update `HANDOFF.md`** before concluding your turn and state clearly to the user what command to run next (e.g., `claude "Continue tri-agent workflow for auth middleware"` or `codex exec "..."`).
© abhigyan6 · MIT · abhigyan6/tri-agent-orchestrator