A plugin for engineering work with skills for product blueprints, scoped task execution, discovery interviews, context optimization, and pattern learning. It adds read-only documentation agents and gated edits with verification evidence.
Claude Code
$ npx -y skills add Joonghyun-Lee-Frieren/oh-my-antigravity --agent claude-code
Codex
$ npx skills add Joonghyun-Lee-Frieren/oh-my-antigravity -a codex
Cursor
$ npx skills add Joonghyun-Lee-Frieren/oh-my-antigravity -a cursor
Gemini CLI
$ npx skills add Joonghyun-Lee-Frieren/oh-my-antigravity -a gemini-cli
GitHub Copilot
$ npx skills add Joonghyun-Lee-Frieren/oh-my-antigravity -a github-copilot
OpenCode
$ npx skills add Joonghyun-Lee-Frieren/oh-my-antigravity -a opencode
Windsurf
$ npx skills add Joonghyun-Lee-Frieren/oh-my-antigravity -a windsurf
Website · History · 한국어 · 日本語 · Français · 中文 · Español
A native Antigravity CLI plugin for engineering work with separate design turns, read-only documentation agents, gated edits, real verification evidence, and small reviews.
Requires current Antigravity CLI and Node.js 20+. Python 3.9+ is additionally needed when the sealed scope contains Python files.
agy plugin install https://github.com/Joonghyun-Lee-Frieren/oh-my-antigravity
agy plugin list
Inside the CLI, inspect /plugin list and /hooks, then start with:
/oma-team-plan Describe the change and its acceptance criteria
Use /oma-read to register and fully page the source snapshots, then obtain the
documentation reader's concise summary. Use /oma-probe for required CLI help.
Record a Scope Brief, seal the
file scope and checks, and request an independent Design Check. After it passes,
end the design turn. In your next turn:
/oma-team-exec Implement the sealed plan
Run /oma-team-verify to collect remaining section reviews. Stop schedules repair batches
when checks or reviews fail. Missing services, permissions, or evidence can end in
blocked or unknown; success is never a required answer.
Read the guarded workflow guide for exact controller commands, JSON examples, recovery, and enforcement limits.
Validator scripts and declared check inputs now have recorded hashes. Changed
dependencies block start, verification, or acceptance of stale evidence. Session
Trail also reports validator drift. Every covering check result reaches the section
reviewer; up to two bounded log excerpts prioritize failures. See the
input schema and limits for legitimate scoped test
changes and dependency declarations. Existing sealed plans need resealing for the
new input protection.
The supplied Part 1–7 architecture now also drives enforced source-read coverage, read-only dependency fingerprints, large-file edit limits, typed failure exits, Python AST validation, dependency-ordered repair groups, and one-group focus on repair oscillation. Markdown reviews split at headings, use a fresh conversation per shard, and require disconfirmation evidence before a deterministic judge decides.
Start with /oma-read, /oma-probe, /oma-failure, and /oma-repair-context.
Read the Part 1–7 mapping and complete schemas,
including migration and enforcement limits. These safeguards do not guarantee
zero defects or prove that a model understood the delivered source.
| Surface | Behavior |
|---|---|
| Scope Brief | Record the objective, constraints, non-goals, decision owners, and source conflicts; unresolved items block sealing |
| Design Check | A read-only auditor checks the exact plan before the design turn can close; plan changes invalidate prior review |
| Evidence Window | Read integrity-checked log excerpts with a bounded UTF-8 byte cursor; full output remains on disk |
| Session Trail | Immutable JSON/Markdown checkpoints retain the current intent and next action; resume detects source/plan drift |
| Native review receipts | The reviewer submits its own result from another conversation; the parent cannot author PASS or steer the review |
| Care profiles | Routine/elevated/critical scopes use 120/90/60-line sections and 3/2/1-unit batches, while keeping all checks and locks |
Sensitive paths can only raise care. Compact status/context packs keep full logs and receipt tickets out of routine model context. Checks still run and missing proof stays unknown. Read schemas and examples.
Recommendations come from config/models.json, checked against official model availability on 2026-10-16.
| Lane | Balanced profile | Flash profile |
|---|---|---|
| Design/planning | gemini-4-argon-high | gemini-3.8-flash-high |
| Implementation | gemini-3.8-flash-medium | gemini-3.8-flash-medium |
| Quick work | gemini-3.8-flash-low | gemini-3.8-flash-low |
| Review/verification | gemini-4-argon-high | gemini-3.8-flash-high |
Check agy models for your account and choose with agy --model <id> or the runtime
selector. Subagents use model: inherit. /oma-model explains the profiles;
auto means keep the selected runtime model and custom requires all four lanes.
Profiles are recommendations, not automatic model switching: the native hook contract
has no model-override output. If Argon is unavailable, explicitly choose the Flash
profile. Older enterprise models may remain available; they are absent from shipped defaults.
The plugin exposes 64 short skills, including 53 /oma-<name> command entrypoints,
and 16 capability-limited agents. Use only the procedures relevant to the current task.
| Command | Purpose |
|---|---|
/oma-read, /oma-probe | Full source-window coverage and actual bounded CLI introspection |
/oma-failure, /oma-repair-context | Typed honest failure and active dependency-ordered repair diagnostics |
/oma-team-plan, /oma-team-prd | Documentation handoff, scope, acceptance criteria, and design |
/oma-brief | Objective, constraints, decision ownership, and source conflicts |
/oma-design-check | Independent check of the exact sealed plan |
/oma-evidence | Bounded, integrity-checked validator log windows |
/oma-checkpoint, /oma-resume | Immutable handoff and read-only continuity inspection |
/oma-team-exec | Apply the sealed scope in the next user turn |
/oma-team-verify | Real checks and one-section review handoffs |
/oma-team-fix | The Stop-selected batch of up to three repair units |
/oma-goal, /oma-ultragoal | Delivery tracking with the same turn and evidence gates |
/oma-status, /oma-doctor | Phase, ownership, evidence gaps, locks, and blockers |
/oma-model | Current model recommendations and runtime selection |
/oma-hooks, /oma-hooks-validate | Native hook inspection and safety diagnostics |
/oma-stop, /oma-cancel | Stop with evidence and locks retained |
oma-doc-reader reads specifications. oma-planner and oma-architect help design.
oma-executor proposes patches; only the owning main conversation applies them.
oma-design-auditor checks the sealed plan. oma-reviewer examines one bounded section; oma-verifier checks acceptance evidence.
Every shipped worker excludes write, shell, MCP, and delegation tools.
The authoritative workflow state is .omg/state/guard.json; execution logs live in
.omg/evidence/. Agents use .omg/drafts/ for requests and auxiliary notes. The controller
alone owns guard state, evidence, and checkpoints. Independent reviewer messages are
captured by the hook using the real sender identity. One workspace and one writer are allowed per workflow;
use separate workflows for isolated worktrees.
Reinstall the native plugin using agy plugin install. New plans require a settled
Scope Brief and native Design Check. Independent roles need native send_message
hooks with the actual child conversation ID; missing runtime support stays unknown.
Existing v1.0 sealed state can finish through its legacy receipt path, but new plans
always use the native path. Session Trail inspects current state without restoring
old ownership, approval, or locks.
Native commands use
/oma-team-exec, replacing /oma:team-exec. The old TOML protocol sources live in
protocols/oma/ and generate native skill wrappers; they are outside the runtime's
legacy commands/ discovery directory to avoid duplicate registrations.
gemini-extension.json remains a migration marker. Remove manually copied old hooks.
The Gemini BeforeModel router, router trace flags, and AfterAgent learn nudge are
removed. Use /oma-learn explicitly and select models through the CLI.
Passed files remain locked, including across subsequent workflows. Reopening a lock, changing sealed scope, or recovering blocked/foreign-owned state requires an explicit operator action outside the guarded agent. Preserve the prior evidence when recovering.
The harness constrains agent tools; it is not an OS sandbox. Validators are trusted project code. Behavioral checks and reviewers still need to catch semantic fabrication. A repair unit is one failed validator or one non-passing section, which may contain several diagnostics. See the full contract.
npm test
npm run test:skills
npm run test:generated
npm run test:version
agy plugin validate .
Edit command procedures in protocols/oma/, then run npm run build:skills.
Core skill procedures live beside their entrypoints in references/.
Tests cover turn boundaries, shell/path bypasses, ownership, failed/missing checks,
review coverage, stale evidence, locks, repair limits, cancellation, and retry bounds.
v1.1 also covers source conflicts, design receipt invalidation, actor spoofing, replay,
care escalation, UTF-8 log paging, checkpoint corruption, and source drift on resume.
Special thanks to @hyunuk.lim (threads). His threads and posts were a major help in shaping this refactor: smaller skills, physical tool constraints, honest failure, bounded repairs, and focused reviews.
Further reading: his Medium article on the ideas behind the v1.0 renewal. Thank you for sharing the concepts that helped shape OmA's engineering workflow.
The project also grew from Jeongkyu Shin's observation that the engineering harness around a model matters alongside the model itself.
Stars on GitHub
A development methodology for coding agents that makes the agent brainstorm before building, follow plans, run parallel work and finish branches in a disciplined way. You get fewer skipped steps and more consistent results from your coding assistant.
@obraHelps you take a change from a rough idea or bug report through to a tested, reviewed result, with each step handled by a focused skill. Its skills cover pinning down domain terms, designing module interfaces, diagnosing hard bugs, and reviewing changes against your repo's coding standards.
@mattpocockGives your coding agent a planning, testing and review routine to follow on every task. It plans before building, checks changes with tests, has a fresh reviewer look for regressions, and saves repeated wins as reusable skills.
@affaan-mDeveloper tools for working on the React codebase itself. Provides skills for extracting error messages, managing and debugging feature flags across release channels, running Flow type checks, and fixing lint and formatting issues before committing.
@facebookInstall it with the command on its page. It includes blueprint, context-optimize, deep-dive. Its skills do not fire on their own yet. Free and open source.
Is this plugin yours?
Claim it with GitHubSubmit a pluginPromote it