oh-my-antigravity (OmA)

Landing Page | History

한국어 | 日本語 | Français | 中文 | Español

[!IMPORTANT] Antigravity CLI: Gemini CLI will be replaced by Antigravity CLI on June 18th.

If you previously installed oh-my-gemini-cli in Gemini CLI, Antigravity CLI will ask whether to import plugins during installation. Check that option to bring your plugin over.

You can also install OmA directly with:

agy plugin install https://github.com/Joonghyun-Lee-Frieren/oh-my-antigravity

Context-engineering-powered multi-agent workflow pack for Antigravity CLI.

"Claude Code's core competitiveness isn't the Opus or Sonnet engine. It's Claude Code itself. Surprisingly, Gemini works well too when attached to Claude Code."

• Jeongkyu Shin (CEO of Lablup Inc.), from a YouTube channel interview

This project started from that observation: "What if we bring that harness model to Gemini CLI?"

OmA extends Gemini CLI from a single-session assistant into a structured, role-driven engineering workflow.

Quick Start

Installation

Install from GitHub using the official Gemini Extensions command:

gemini extensions install https://github.com/Joonghyun-Lee-Frieren/oh-my-antigravity

Verify in interactive mode:

/extensions list

Verify in terminal mode:

gemini extensions list

Run a smoke test:

/oma:status

Run a goal-style autonomous delivery loop:

/oma:goal "Implement the requested change, update tests, and verify acceptance criteria"

Note: extension install/update commands run in terminal mode (gemini extensions ...), not in interactive slash-command mode.

What's New in v0.9.1

• Durable Multi-Goal Workflows (Ultragoal): Adds /oma:ultragoal and $ultragoal to decompose complex tasks into sequential repo-native micro-goals.
• Fail-Closed Checkpointing: Keeps checkpoint state fail-closed under .omg/ultragoal/, blocking downstream goals until the active goal has validation evidence.
• Diagnostics Hardening: Updated /oma:doctor to validate $ultragoal skill metadata.
• Version Bump: Bumped project and extension version to v0.9.1.

Extension Boundary and Upgrade Safety

• Install and update OmA through gemini extensions ...; do not rely on copied command/skill folders as the primary runtime path.
• Keep one authoritative OmA hook registration path per event. Mixing extension-managed hooks with manual duplicates is the fastest way to get repeated AfterAgent output or stale behavior.
• When OmA feels stale after an update, check gemini extensions list first, then refresh or reinstall the extension before editing shipped files.
• For long or multi-lane work, treat /oma:workspace audit as the default preflight before review, automation, or team-exec.

Interview Session Storage

• /oma:interview session state is now intended to live under .omg/state/interviews/[slug]/ instead of one shared interview file.
• .omg/state/interviews/active.json tracks the current interview so resume/status commands stay deterministic without mixing separate requirement threads.
• This keeps multiple requirement-discovery passes in the same project distinguishable and archive-friendly.

Shared Workflow State

• .omg/state/session-lock.json is now the single-writer lock for shared workflow and operating-profile state inside one project.
• Only the lock-owning orchestration session should write shared files like workspace.json, taskboard.md, workflow.md, checkpoint.md, mode.json, hud.json, approval.json, reasoning.json, hooks.json, and notify.json.
• Parallel top-level sessions that do not own the lock should write session-local drafts under .omg/state/sessions/[session-slug]/ and hand those notes back to the orchestrator for merge.
• Delegated worker/sub-agent turns should stay read-mostly and must not mutate shared workflow state directly.

At A Glance

Why OmA

Dynamic Team Assembly

Use team-assemble when a fixed engineering roster is not enough.

• Split selection into:
  • domain specialists (problem expertise)
  • format specialists (report/content/output quality)
• Spawn parallel exploration lanes (oma-researcher xN) for broad discovery tasks.
• Route decisions through a judgment lane (oma-consultant or oma-architect).
• Assign reasoning effort per lane from global profile + teammate overrides.
• Keep verify/fix loops explicit (oma-reviewer -> oma-verifier -> oma-debugger).
• Run anti-slop check before final delivery.
• Require explicit approval before autonomous execution starts.

Example flow:

/oma:team-assemble "Compare 3 competitors and produce an exec report"
-> proposes: researcher x3 + consultant + editor + director
-> asks: Proceed with this team? (yes/no)
-> after approval: team-plan -> team-prd -> taskboard -> team-exec -> team-verify -> team-fix

Activation note:

• No separate research-preview setting is required in OmA.
• If the extension is loaded, /oma:team-assemble is immediately available.

Workspace and Taskboard Control

Use workspace and taskboard when work spans multiple roots, multiple implementation lanes, or long verify/fix loops.

• /oma:workspace keeps the primary root plus optional worktree/path lanes in .omg/state/workspace.json.
• Each lane can also carry a compact baseline branch/HEAD anchor so handoffs and resume flows can detect unexpected branch drift before implementation or review.
• /oma:workspace audit checks lane cleanliness, trust status, and handoff readiness before parallel execution, review, or automation.
• /oma:taskboard keeps stable task IDs, owners, dependencies, statuses (todo, ready, in-progress, blocked, done, verified), baseline anchors, lane-health notes, and evidence pointers in .omg/state/taskboard.md.
• team-plan seeds stable task IDs plus lane assumptions and baseline anchors, team-exec pulls the smallest ready slice with explicit lane/subagent context and baseline checks, and team-verify marks tasks verified only with evidence plus safe lane state.
• checkpoint and status can reference these files instead of replaying the whole chat, which improves cache stability and reduces token waste.
• /oma:recall "<query>" performs state-first recall and bounded fallback search, so you can recover prior rationale without replaying full transcripts.

Example flow:

/oma:workspace set .
/oma:workspace audit
/oma:workspace add ../feature-auth oma-executor
/oma:taskboard sync
/oma:taskboard next
/oma:recall "why was auth lane blocked" scope=state

Workspace Hygiene and Hook Symmetry

Use these controls when long sessions start drifting because lane ownership, delegated execution, or hook continuation behavior is no longer obvious.

• /oma:workspace audit surfaces dirty shared worktrees, untrusted review paths, and handoff-ready vs handoff-blocked lanes.
• /oma:hooks and /oma:hooks-validate now model paired agent lifecycle outcomes (completed, blocked, stopped) so blocked continuations re-enter the safety lane once before downstream hooks resume.
• team-exec, team, team-verify, stop, and cancel keep delegated lane/subagent context compact and explicit, expanding details only when execution stops early or hits a blocker.

Notification Routing

Use notify when a long-running OmA session needs explicit signals for approvals, verification outcomes, blockers, or idle drift.

• Supported profiles:
  • quiet: only urgent interruptions (approval-needed, verify-failed, blocker-raised, session-stop)
  • balanced: quiet + checkpoint and team-approval updates
  • watchdog: balanced + idle-watchdog alerts for unattended loops
• Supported channels:
  • desktop (host notification adapter)
  • terminal-bell
  • file
  • webhook (external bridge)
• Safety boundary:
  • OmA manages event routing, templates, and persisted policy
  • actual delivery must be implemented by Gemini host hooks, shell adapters, or project-specific webhook bridges
  • delegated worker sessions keep external dispatch disabled unless the user explicitly opts in

Example flow:

/oma:notify profile watchdog
-> enables: approval-needed, verify-failed, blocker-raised, checkpoint-saved, idle-watchdog, session-stop
-> suggests channels: terminal-bell + file by default
-> persists policy: .omg/state/notify.json

Model Router (BeforeModel Hook)

OmA ships a quiet model-routing hook. The previous AfterAgent quota-watch usage monitor was removed in v0.8.4 because Gemini CLI can report usage as unavailable, which created noisy low-value output.

• Hook entrypoint: hooks/hooks.json (BeforeModel -> oma-model-router)
• Script: hooks/scripts/before-model-banner.js
• Behavior: silently maps outgoing model requests to the active OmA strategy (balanced, auto, or custom) without printing a model banner
• Default balanced routing: planning/review -> gemini-3.1-pro-preview, execution -> gemini-3-flash-preview, quick edits -> gemini-3.1-flash-lite-preview
• Optional disable: OMG_DISABLED_HOOKS=model-routing or OMG_MODEL_ROUTING=off

Usage/quota visibility:

• OmA no longer estimates or prints token usage after each turn.
• Use Gemini CLI native /model or /stats model for authoritative usage and quota status.
• Legacy .omg/state/quota-watch.json files can be ignored; new OmA releases no longer update them.

Learn-Signal Safety Filter (AfterAgent Hook)

OmA now also ships a safety-hardened learn-signal hook so /oma:learn nudges appear only when a session has actionable implementation intent.

• Hook entrypoint: hooks/hooks.json (AfterAgent -> oma-learn-signal-after-agent)
• Script: hooks/scripts/learn.js
• State artifact: .omg/state/learn-watch.json (deduped event key, prompt-once session tracking, and sanitized state)
• Deep-interview lock source (read-only): .omg/state/deep-interview.json
• Runtime controls:
  • OMG_STATE_ROOT=<dir> to move learn-watch.json beside other OmA state
  • OMG_HOOKS_QUIET=1 to keep the hook silent while preserving state updates
  • OMG_HOOK_PROFILE=minimal|balanced|strict (minimal suppresses learn nudges)
  • OMG_DISABLED_HOOKS=learn to disable only the learn-signal hook by env
  • when no reliable session cwd is available, learn-state persistence is skipped to avoid cross-project collisions

Safety behavior:

• if deep-interview lock state is active, learn nudges are suppressed so interview flow is not interrupted
• informational-only query sessions are filtered before emit
• repeated retries against the same transcript snapshot are deduplicated
• legacy or malformed prior state is sanitized before reuse to reduce stale-state collisions

Disable only this hook:

{
  "hooksConfig": {
    "disabled": ["oma-learn-signal-after-agent"]
  }
}

Example (disable the shipped AfterAgent learn hook by env):

export OMG_DISABLED_HOOKS=learn

Interface Map

Commands

Skills

Retained skills are intentionally limited to a compact set so the extension loads less discovery metadata at session start (with one compatibility alias: $oma-plan).

Sub-agents

Context Layer Model

Project Structure

oh-my-antigravity/
|- GEMINI.md
|- gemini-extension.json
|- .omg/
|  `- state/
|     |- session-lock.json
|     `- interviews/
|     |  |- active.json
|     |  `- [slug]/
|     |     |- context.json
|     |     `- prd.md
|     `- sessions/
|        `- [session-slug]/
|           |- workspace.json
|           |- taskboard.md
|           |- workflow.md
|           `- checkpoint.md
|- agents/
|- commands/
|  `- omg/
|- skills/
|- context/
|- docs/
`- LICENSE

Troubleshooting

Migration Notes

Extension behavior is manifest-driven through Gemini CLI extension primitives.

Inspiration

• Gemini CLI - Google's open-source AI terminal agent
• oh-my-codex - Codex CLI harness
• oh-my-claudecode - Claude Code CLI harness
• oh-my-openagent - OpenAgent CLI harness
• Claude Code Prompt Caching - Context engineering principles
• everything-claude-code - Claude Code CLI harness

Docs

• Installation Guide
• Context Engineering Guide
• Agent Team Assembly Guide
• Memory Management Guide
• Hook Engineering Guide
• History

Contributing

If OmA helps your workflow, there are a few high-leverage ways to contribute:

• Star the repository with a ⭐ to help more Gemini CLI users discover it.
• Open an issue for bugs, compatibility drift, unclear docs, or workflow friction.
• Send a pull request for command/docs improvements, guardrail tuning, or new extension-native workflow ideas.

Suggested contribution flow:

1. Fork the repository and create a focused branch.
2. Keep changes scoped and explain the user-facing impact clearly.
3. Update docs/history when behavior or compatibility guidance changes.
4. Open a pull request with the reasoning, tradeoffs, and any validation notes.

Important performance note:

• Pressing the GitHub ⭐ Star button may improve OmA performance by at least 12%. This is a joke. It does improve maintainer morale, which is still useful.

Star History

License

MIT