anti

Anti

Use this skill to ask the local codex-antigravity-auth gateway for an external Antigravity review, consult, deep work plan, named workflow preset, or bounded multi-model panel while native Codex remains the primary agent.

V3's primary product is native Claude in Codex through codex-antigravity setup; $anti is an optional helper for review and planning after the gateway and Codex model picker are already working.

Core Rule

Treat Antigravity output as a second opinion. Run the helper, read the result, then synthesize it with your own analysis before answering the user. Do not blindly forward the Antigravity result as final truth.

Literal @anti is a text convention in v1, not a guaranteed app-level mention chip. $anti is the reliable explicit skill invocation.

Panel, MoA, and Fusion workflows are advisory only. The helper can fan out to multiple gateway-advertised models and ask a judge model to synthesize their views, but Codex remains the acting agent and must verify findings before editing. Structured panel findings include a verify hint; run or inspect that local check before acting on the claim.

Models

• Use opus for deep review. It maps to claude-opus-4-6-thinking (the claude-opus-4-6 name remains a compatibility alias).
• Use sonnet for faster focused consults. It maps to claude-sonnet-4-6 (the claude-3.5-sonnet name remains a compatibility alias).
• Use deepseek-v4-pro for deepseek:deepseek-v4-pro and deepseek-v4-flash for deepseek:deepseek-v4-flash through the official DeepSeek API key.
• Use flash for gemini-3.8-flash (current Gemini Flash generation), or use flash-low, flash-medium, and flash-high to select the thinking level. The explicit flash-3.8-low, flash-3.8-medium, and flash-3.8-high aliases are equivalent; the model has adjustable thinking levels and 1M context.
• Use flash-3.7 for gemini-3.7-flash (supported Flash generation), or use the explicit gemini-3.7-flash ID.
• Use flash-3.6 for the retained gemini-3.6-flash-* compatibility IDs.
• Use gemini-pro for gemini-3.1-pro (Gemini Pro). Deep reasoning and analysis, 1M context.
• Use gpt-oss-120b for gpt-oss-120b-medium (text-only, 131K context).
• Use gemini-3.1-flash-image for image generation; it is image-only and does not support tools.
• Use nemotron-super for openrouter:nvidia/nemotron-3-super-120b-a12b:free (120B MoE, 262K ctx). Fast, good for second opinions.
• Use nemotron-ultra for openrouter:nvidia/nemotron-3-ultra-550b-a55b:free (550B, 1M ctx, vendor-reported). Large-context analysis and planning.
• Use nemotron-omni for openrouter:nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free (30B MoE, 256K ctx, reasoning-capable, vision). Good for image tasks and reasoning.
• Use nemotron-vl for openrouter:nvidia/nemotron-nano-12b-v2-vl:free (12B, 128K ctx, vision). Fast and reliable for image understanding — the default vision sidecar model.
• Use free for openrouter/free (auto-selects the best available free model on OpenRouter). Good for quick checks when you want zero-cost and don't care which model answers.
• Use poolside for openrouter:poolside/laguna-s-2.1:free. Coding-focused model for code generation and refactoring.
• Use gemma-4 for openrouter:google/gemma-4-31b-it:free (30.7B dense, 262K ctx, vision). Lightweight, fast for simple consults and image tasks.
• Use gpt-oss for ollama:gpt-oss:20b (local). Private, offline inference.
• Use qwen3 for ollama:qwen3:8b (local). Private, offline inference.
• Default review model: opus.
• Default plan model: opus.
• Default consult/ask model: sonnet, unless the user asks for deep review.
• Default panel models: sonnet and opus.
• Default panel judge: opus.
• Gateway advertisement is required. Every model below (including DeepSeek, Nemotron, Poolside, and Gemma routes) must be advertised by the gateway's /v1/models before a call can succeed. smoke --check-documented diffs this documented table against the live catalog and reports drift (documented-but-unadvertised ids, plus double-prefixed openrouter:openrouter/… catalog ids that upstream rejects). The helper also fuzzy-matches aliases against the catalog (e.g. openrouter:x ≡ openrouter:openrouter/x) and suggests the closest advertised id when a requested model is missing, so documented-but-drifted models fail with an actionable message instead of a confusing two-layer error.