Omega Gemini CLI

Omega Gemini CLI (headless)

This skill follows the Agent Skills open standard. Codex discovers it from REPO scope at $REPO_ROOT/.agents/skills/. It runs the Gemini CLI in headless mode so Codex can call Gemini from the project—no MCP server or MCP configuration required.

Response times

Gemini CLI runs as a subprocess and includes model startup time. Typical wall-clock times observed on this setup:

• Simple Q&A / news query: ~2 minutes
• Codebase or large-file review: ~5–10 minutes

Set expectations with the user before running long tasks.

Overview

From the project root (the directory that contains .claude and .agents), run the headless script. It invokes gemini -p "..." with optional model, sandbox, and JSON flags and returns the response. Requires Node.js and Google Gemini CLI to be installed.

How to run Gemini

Input: A user prompt (and optional @ file refs). Output: The script’s stdout (Gemini’s response).

From the project root, run:

node .claude/skills/omega-gemini-cli/scripts/ask-gemini.mjs "USER_PROMPT"

Optional flags (append to the command):

• --model MODEL / -m MODEL — e.g. gemini-3.7-flash, gemini-3.1-flash-lite. Reduce quota issues by using Flash-lite.
• --sandbox / -s — sandbox mode for running code.
• --json — output a JSON object {"response":"..."} (consistent envelope on success and error).

Gemini CLI also supports --output-format stream-json, --approval-mode, --resume, and more. Full reference: .claude/skills/omega-gemini-cli/references/headless.md.

Examples:

• Analysis: node .claude/skills/omega-gemini-cli/scripts/ask-gemini.mjs "Review @src/main.js for bugs"
• With model: node .claude/skills/omega-gemini-cli/scripts/ask-gemini.mjs "Summarize this" --model gemini-3.7-flash
• Stdin: echo "Explain recursion" | node .claude/skills/omega-gemini-cli/scripts/ask-gemini.mjs

1. Build the prompt from the user’s request and any @ file refs. 2. Run the command above from the project root. 3. Return the script’s stdout to the user as Gemini’s response.

When to use this skill

• Large files or codebases — Use Gemini's context for analysis.
• Second opinion / feedback — User wants Gemini's take on code or docs.
• Sandbox — Run or test code via Gemini (add --sandbox when supported).
• Brainstorming — Build a prompt that includes the challenge and optional method (e.g. SCAMPER) and run the script.

If Gemini CLI is not set up

If the script fails (e.g. "gemini not found"), direct the user to install Node.js and Gemini CLI and complete one-time auth. Setup can be verified by running from project root:

node .claude/skills/omega-gemini-cli/scripts/verify-setup.mjs

Full installation and auth details live in the Claude skill at .claude/skills/omega-gemini-cli/references/installation.md and references/auth.md (same repo).

Models (2026)

Gemini CLI model IDs: gemini-3.7-flash (latest, GA), gemini-3.6-flash, gemini-3.5-flash, gemini-3.5-flash-lite, gemini-3.1-pro-preview, gemini-3.1-flash-lite, gemini-3-flash-preview, gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite. Use --model gemini-3.7-flash (or gemini-3.1-flash-lite) for most tasks. Note: gemini-3-pro-preview shut down March 9, 2026 — use gemini-3.1-pro-preview instead. See .claude/skills/omega-gemini-cli/references/models-2026.md.

Shared scripts

Scripts are shared with the Claude skill: .claude/skills/omega-gemini-cli/scripts/. No duplicate copies are required for Codex.