AMAG: All Mighty AntiGravity

Agent orchestration for Antigravity. Curated rules, workflows, and skills that transform Antigravity's AI agent into a disciplined engineering partner.

Inspired by oh-my-opencode — rebuilt from scratch for Antigravity's native tooling.

Quick Start

cd /path/to/your/project
npx @totoroyyb/amag init

Or build from source:

git clone https://github.com/totoroyyb/amag.git
cd amag && npm install && npm run build && npm link

That's it. Your project now has a GEMINI.md root prompt, always-on rules, slash-command workflows, and on-demand skills.

What You Get

Core System (GEMINI.md)

The root prompt that orchestrates everything:

• Intent Gate — classifies every request before acting (trivial, explicit, exploratory, open-ended)
• True Intent Extraction — understands what you actually want, not just what you typed
• Auto-Ultrawork — maximum rigor is always on: full verification, zero scope reduction
• Codebase Assessment — evaluates project maturity and adapts behavior accordingly
• Verification Protocol — evidence-based completion with a 6-step checklist (spec compliance → build → tests → debug scan)

Rules (Always-On)

Loaded into every conversation automatically.

| Rule | What it does | |------|-------------| | code-quality | Think-first coding, surgical changes, no AI slop | | error-recovery | 3-failure escalation, blind retry prevention, hung command detection | | todo-enforcement | Task breakdown, progress tracking, never abandon work | | agentic-rules | Session start checks, active plan detection, auto-resume guidance |

Workflows (Slash Commands)

Type the command to activate.

| Command | What it does | |---------|-------------| | /plan | Planning interview with optional external consultant + critic review | | /start-work | Execute a plan task-by-task with category-aware delegation | | /resume | Cross-session resume from .amag/active-plan.md | | /debug | 6-phase systematic debugging with external agent escalation | | /debug-escalate | Immediately escalate to external CLI during an active /debug | | /explore | Read-only multi-phase codebase exploration and architecture synthesis | | /ultrawork | Maximum effort mode — 100% certainty before acting | | /init-deep | Generate hierarchical GEMINI.md context files across the codebase |

Skills (On-Demand)

Auto-loaded when relevant, or loaded explicitly by workflows.

| Skill | What it does | |-------|-------------| | deep-work | Autonomous exploration — read extensively, build mental model, act decisively | | git-master | Atomic commits, conventional format, rebasing, conflict resolution | | browser-testing | Visual testing via Antigravity's browser_subagent | | frontend-ui-ux | Design-first UI with bold aesthetics and responsive patterns | | writing | Anti-AI-slop technical writing — plain words, human tone | | architecture-advisor | Read-only design review with simplicity bias | | codebase-explorer | Structured parallel codebase research and cross-validation | | external-researcher | External library/API research — official docs and best practices | | external-cli-runner | Unified runner for Claude, Codex, and Gemini CLIs with retry logic | | plan-consultant | Pre-plan gap analysis — surface missing requirements before generation | | plan-critic | Post-plan adversarial review — verify references and executability |

Typical Workflow

You:   /plan add OAuth login with Google and GitHub
Agent: [explores codebase → asks clarifying questions → drafts plan]
       [runs plan through consultant + critic if configured]
       → implementation_plan.md ready for review

You:   /start-work
Agent: [executes plan task-by-task: build → type-check → test after each]
       → all tasks complete, verified with evidence

You:   /resume                      # pick up where you left off in a new session

You:   /debug login redirects fail on Safari
Agent: [reproduce → hypothesize → instrument → root-cause → fix → verify]
       [escalates to external CLI if stuck: /debug-escalate]

You:   /explore how does the auth middleware chain work?
Agent: [structural scan → module deep-dives → synthesizes architecture doc]
       → read-only, never modifies code

External Agent Integration

AMAG can delegate work to external CLI agents — Claude, Codex, or Gemini — to get a second opinion from a different model. This happens in two contexts:

Plan review (/plan workflow):

1. Consultant — before the plan is generated, an external agent analyzes your requirements for gaps, ambiguities, and missing edge cases
2. Critic — after the plan is generated, a different external agent stress-tests it for executability, missing references, and unrealistic assumptions

Debug consultation (/debug workflow):

• When debugging stalls after multiple failed hypotheses, AMAG sends the full debug context to an external agent for fresh analysis
• Trigger manually with /debug-escalate or let the workflow offer it at a natural checkpoint

Each external call includes full project context, retry logic (up to 3 attempts), and a configurable thinking level that controls how deeply the external model reasons.

Configuration

npx @totoroyyb/amag config show                              # View current config
npx @totoroyyb/amag config set review.consultant.cli claude   # Set plan consultant
npx @totoroyyb/amag config set review.critic.cli codex        # Set plan critic
npx @totoroyyb/amag config set debug.consultant.cli codex     # Set debug consultant
npx @totoroyyb/amag config set review.critic.thinking high    # Set thinking level
npx @totoroyyb/amag config reset                              # Reset to defaults

Thinking levels: max, high, medium, low, none.

[!NOTE] Configuration is stored in .amag/config.json in your project root. This file is created during init or update with sensible defaults. You can modify it via the CLI commands above or edit the JSON directly.

Prerequisites

External agent features require the corresponding CLI tools to be installed and authenticated on your machine:

| CLI | Install | Auth | |-----|---------|------| | Claude Code | curl -fsSL https://claude.ai/install.sh \| bash | Run claude and follow the OAuth login prompt | | Codex CLI | npm install -g @openai/codex | codex login | | Gemini CLI | npm install -g @google/gemini-cli | You might need to login with gemini CLI first |

[!TIP] If no external CLI is configured or available, AMAG falls back to native single-agent review — everything still works, just without the multi-model perspective.

CLI Reference

npx @totoroyyb/amag init                    # Install all components
npx @totoroyyb/amag update                  # Overwrite with latest templates
npx @totoroyyb/amag add <type> <name>       # Install one component (rule, workflow, skill)
npx @totoroyyb/amag remove <type> <name>    # Remove one component
npx @totoroyyb/amag uninstall               # Remove all AMAG files
npx @totoroyyb/amag list                    # Show available components
npx @totoroyyb/amag doctor                  # Check installation status
npx @totoroyyb/amag config show|set|reset   # Manage configuration

Design Philosophy

1. System prompts are the product — the CLI is just a delivery mechanism
2. Every template earns its place — no bloat, curated for real engineering workflows
3. Antigravity-native — built on AG's actual tools, not abstractions
4. Composable — add only what you need, remove what you don't

License

MIT