/oma-read · /oma-probeCover the source, including the tail
Register versioned snapshots, page every byte window, and record real CLI help. Missing source coverage or a failed required probe prevents handoff.
v1.2.0 · Native Antigravity CLI plugin
Small skills.
Concrete limits.
Evidence first.
Read the full selected sources. Preserve dependency boundaries. Repair a few related failures at a time, then ask an independent reviewer for concrete counterexamples.
ANTIGRAVITY CLI / OmA
# Install the native plugin
agy plugin install https://github.com/Joonghyun-Lee-Frieren/oh-my-antigravity
TURN 01 / READ & DESIGN
/oma-team-plan/oma-read → /oma-probe → Scope Brief → Design Check. Stop.
TURN 02 / IMPLEMENT & VERIFY
/oma-team-execApply the plan. Run checks. Collect native review receipts.
≤20 lines per skill entrypointRead-only worker tools≤3 groups per repair batch≤120 lines per review
Make the next shortcut visible.
NEW / v1.2.0Built from the supplied Part 1–7 architecture: enforce source coverage, keep honest failure exits, and narrow both repair and review. Scope Brief, Design Check and Session Trail carry the decisions and evidence across turns.
Read-only boundariesKeep dependencies intact
Fingerprint declared dependency trees. Large-file writes need a planned rewrite; other edits target bounded existing chunks.
/oma-failureAn honest failure exit
Report DEPENDENCY_DEFECT with a real failed check, or name an external blocker. Preserve evidence and locks without inventing successful behavior.
AST + execution receiptsCheck structure. Record execution.
Parse scoped Python for suspicious success shortcuts. Bind real status, output and source/input hashes to immutable receipts. Missing Python fails honestly.
/oma-repair-contextFix related failures in order
Sealed dependency groups determine the next bounded batch. Real diagnostics are paged. A→B→A oscillation narrows focus to one group.
Blind section juryFind the counterexample first
Markdown shards follow headings; each uses a fresh reviewer conversation. A central judge requires criterion-level evidence and vetoes unrequested nonmaterial demands.
Read coverage proves delivery, not comprehension. AST patterns and tool guards complement behavioral tests and runtime sandboxing; they do not guarantee zero defects. Python scopes require Python 3.9+. Part 1–7 mapping, schemas and limits ↗
Care scales with the scope: routine, elevated, and critical work use review caps of 120, 90, and 60 lines and repair batches of 3, 2, and 1. Sensitive paths can only raise care. Every profile retains real checks and independent receipts. Read the operating guide ↗
A boundary at every handoff.
01 / WORKFLOWProcedures load when needed. Native capability limits and lifecycle hooks enforce the work that can happen in each phase.
Documentation first
Register and fully page source snapshots. A read-only reader returns constraints and unknowns. Required CLI probes use recorded execution.
Seal the contract
Resolve Scope Brief decisions, seal exact files and checks, and obtain an independent Design Check. End the user turn.
Stay inside scope
The owning conversation applies patches. PreToolUse rejects off-scope edits, shell bypasses, and evidence writes.
Prove each section
Real checks, AST results and blind reviews determine completion. Stop locks passed files and schedules dependency-ordered repairs.
Design → implementation is a user-turn boundary. Repair batches use separate execution slices scheduled by Stop, with verification between batches.
Constraints the harness can enforce.
02 / GUARDSHonest failure is a valid result. Missing evidence stays unknown, and a missing dependency can block the task.
PreToolUseA narrow write surface
Exact planned paths, one conversation owner, protected state and logs, no arbitrary shell calls. Tests and fixtures need explicit scope.
VerificationResults from execution
Sealed argv checks record real exit codes, output, and source hashes. There is no agent command to declare a failed check passed.
StopFix up to three
A batch contains at most three dependency groups or non-passing sections. Sensitive scopes and detected oscillation use fewer. Every batch returns to verification.
Hash locksKeep passed code stable
A file locks after its checks and every section pass. Further edits and stale evidence are rejected. Reopening requires operator recovery.
Section reviewA small, concrete assignment
The hook builds one source section of up to 120 lines per reviewer. Receipts come from the reviewer’s own conversation. The parent cannot write PASS or steer its verdict.
Fail / unknown / blockedNo manufactured success
Do not fake service responses, skip validation, or invent logs. Bounded retries and explicit blockers stop unproductive repair loops.
Tool constraints complement runtime sandboxing. Trusted project validators and relevant behavioral reviews remain essential; pattern checks cannot prove arbitrary code is truthful. A repair unit may contain multiple diagnostics.
Current models. Explicit selection.
03 / MODELSGemini 4 Argon for deeper design and review. Gemini 3.8 Flash for implementation and quick work. Choose a Flash-only profile when that fits your account.
| Lane | Balanced | Flash |
|---|---|---|
| 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 |
Verified 2026-10-16 against official model documentation. Run agy models for account availability, then select with agy --model <id>. Workers inherit that model. Profiles are recommendations; hooks do not switch models automatically.
Native from the entrypoint down.
04 / UPGRADEMoving from v1.0
Reinstall with agy plugin install. New plans require Scope Brief and native Design Check. Current clients must support child review messages and hooks.
Existing v1.0 sealed work can finish through its legacy path. New plans use native independent receipts; missing support remains unknown.
Migration details ↗A compact control plane
64 short skill entrypoints, 16 capability-limited workers, and three guard events. New workflows use the v2 reading and jury contract; existing sealed work retains its prior contract.
Use /oma-status for progress, /oma-doctor for diagnostics, and /oma-team-verify for remaining evidence.