⚡ BOOST SKILL (Universal /boost Protocol)

Multi-Agent Reasoning & Independent Physical Verification Pipeline for All AI Coding CLIs

---

⚡ 1-Line Quick Install

Install Boost Skill across all detected AI agent environments on your machine in one command:

curl -fsSL https://raw.githubusercontent.com/ervareza/boost-skill/main/install.sh | bash

Or install globally via npm / npx:

npx boost-skill install

Automatically discovers and installs adapters for Hermes Agent, Claude Code, OpenAI Codex CLI, OpenCode, and adds the universal boost CLI binary to your PATH.

---

📖 What is the /boost Protocol?

In September 2026, Google introduced the /boost slash command in the Google Antigravity CLI, moving AI development away from fragile single-turn generation into a distributed multi-agent reasoning and physical verification pipeline.

Instead of an AI assistant generating code and asking you to review unverified files, /boost:

1. Isolates all work in an ephemeral Git worktree — keeping your active branch and working directory 100% clean.
2. Decomposes the problem across specialized subagent streams (Investigator, Implementer, Test Crafter).
3. Validates all changes physically through actual build and test suite execution with a multi-round self-healing loop.
4. Delivers results only when all tests return 100% PASS (Exit code 0).

Boost Skill (boost-skill) brings this exact methodology and execution harness to every AI coding tool and CLI in the developer ecosystem.

---

🏛️ Architecture & Verification Flow

The 3 Core Execution Phases:

Phase	Responsibility	Scope & Mechanism
Phase 1: Planning & Worktree Isolation	Scans AST call-graphs, formulates hypothesis, and branches into an ephemeral worktree.	git worktree add -b boost-task .worktrees/boost-task (Zero uncommitted changes in active branch)
Phase 2: Distributed Subagent Topology	Spawns parallel specialist workers to investigate, patch code, and write regression tests.	• Investigator: Read-only AST & trace reproduction • Implementer: Minimal surgical patches • QA/Tester: TDD failing test cases (Red to Green)
Phase 3: Physical Verification & Reconcile	Compiles project, runs test runner, auto-heals failures (up to 5 rounds), and merges clean code.	bash scripts/boost-verify.sh (Delivers only with verified test execution logs)

---

📊 Comparison Matrix

Feature	Standard AI Coding Assistant	Traditional Loop / Agent	Boost Skill (/boost)
Working Tree Safety	❌ Edits live files directly	⚠️ Stashes or creates messy branches	✅ Ephemeral isolated Git worktree
Verification Method	❌ None / asks user to test	⚠️ Visual LLM code self-inspection	✅ Physical compiler & test runner execution
Architecture	❌ Single-turn linear prompt	⚠️ Monolithic multi-step loop	✅ 3-Phase Distributed Subagent Topology
Self-Healing	❌ Manual prompting required	⚠️ Prone to hallucinated infinite loops	✅ Bounded 5-round diagnostic feedback loop
Multi-CLI Portability	❌ Tied to specific vendor	❌ Single tool only	✅ Universal adapter for Top 10 AI CLIs

---

🚀 Supported AI Coding CLIs & Quickstarts

1. 🪽 Nous Research Hermes Agent

# Manual install
mkdir -p ~/.hermes/skills/autonomous-ai-agents/boost
cp adapters/hermes/SKILL.md ~/.hermes/skills/autonomous-ai-agents/boost/SKILL.md

Trigger inside Hermes:

/boost Fix intermittent race condition when WebSocket reconnects during token refresh

---

2. 🟣 Claude Code CLI (Anthropic)

# Add to project or global config
cp adapters/claude-code/CLAUDE.md ./CLAUDE.md

Trigger inside Claude Code:

claude "Run /boost on issue: optimize database N+1 query and add regression tests"

---

3. 🟢 OpenAI Codex CLI

# Run in non-interactive sandbox mode
codex exec --sandbox danger-full-access "Execute /boost protocol on task: refactor authentication middleware"

---

4. ⚡ OpenCode CLI

mkdir -p ~/.opencode/plugins
cp adapters/opencode/opencode.json ~/.opencode/plugins/boost.json

---

5. 🤖 Aider

cp adapters/aider/.aider.conf.yml ./.aider.conf.yml
aider --architect --test-cmd "bash scripts/boost-verify.sh"

---

6. 🖱️ Cursor & Composer

cp adapters/cursor/.cursorrules ./.cursorrules

---

7. 🌊 Windsurf (Cascade)

cp adapters/windsurf/.windsurfrules ./.windsurfrules

---

8. 🦘 Roo Code / Cline

cp adapters/roo-cline/.roomodes ./.roomodes

---

9. 🧠 Devin / Devin ACP

Include adapters/devin/devin-boost.md in your task playbook prompts.

---

10. 🐙 GitHub Copilot (Workspace & CLI)

mkdir -p .github && cp adapters/github-copilot/.github/copilot-instructions.md .github/copilot-instructions.md

---

🛠️ Standalone CLI Commands

When installed, the boost CLI binary allows you to drive the lifecycle manually from any terminal:

boost init <task_name>       # 1. Spawn isolated ephemeral Git worktree
boost test                   # 2. Run polyglot test verification in current repo
boost verify <task_name>     # 3. Run full verification inside the worktree
boost reconcile <task_name>  # 4. Merge verified commit into active branch & cleanup
boost abort <task_name>      # 5. Discard worktree and rollback changes

---

📜 Core Guarantees

1. Zero Pollution: Your uncommitted code is never touched or lost while agents experiment.
2. Real Test Proof: Completion reports always contain real test runner outputs, passing assertion counts, and execution exit codes.
3. Deterministic Cleanup: Ephemeral worktrees and temporary branches are automatically pruned upon successful reconciliation.

---

👤 Author & Attribution

• Engineered by: Ervareza Naurian (@ervareza)
• Inspiration: Google Antigravity CLI /boost Architecture (September 2026)
• Specification: Read the formal protocol in SPECIFICATION.md
• License: MIT