Gemini CLI Runner

Invokes Google's Gemini CLI in non-interactive mode to get an independent response from a non-Claude model. The skill shells out via the Bash tool, captures stdout, and returns the response cleanly.

When to use

• The user explicitly asks to consult Gemini ("ask Gemini", "what does Gemini think", "second opinion from Gemini")
• A crossfire-style multi-model review needs an independent Google-side perspective
• The user wants an external check on a design, code change, or argument and Gemini is the chosen reviewer

Do NOT trigger when:

• The user is asking about Gemini CLI itself (use docs, not this skill)
• The task can be answered from your own knowledge — don't pay the latency for nothing
• Gemini CLI isn't installed (see "Setup check" below)

Setup the user needs

Before this skill works, the user must have:

1. Gemini CLI installed and on PATH. Install with npm install -g @google/gemini-cli (Node 20+).
2. Auth configured — pick one:
   • API key (headless-friendly): GEMINI_API_KEY env var, obtained from https://aistudio.google.com/app/apikey. This is the path of least friction for shell-out.
   • OAuth ("Sign in with Google"): First run gemini interactively, complete the browser flow, then cached credentials live in ~/.gemini/. Free-tier OAuth has stricter quotas than API key.
   • Workspace / Vertex: Set GOOGLE_APPLICATION_CREDENTIALS + GOOGLE_CLOUD_PROJECT (+ GOOGLE_CLOUD_LOCATION for Vertex).

If none is set, gemini -p will either drop into an OAuth flow (which won't work in non-interactive context) or error. Detect and tell the user.

How to invoke

Step 1 — Check that Gemini is available

command -v gemini && gemini --version

If unavailable, tell the user:

Gemini CLI doesn't appear to be installed or on PATH. Install with npm install -g @google/gemini-cli (Node 20+ required), then either set GEMINI_API_KEY or run gemini once interactively to complete OAuth, then retry.

Stop. Do not attempt the call.

Step 2 — Construct the prompt

Same as the Codex skill — assemble the prompt from the user's request, file contents, or a parent-flow-supplied template. For long prompts, use a file rather than inline shell-arg quoting.

Step 3 — Run Gemini from the project directory

gemini -p "$PROMPT" --output-format text -m flash 2>/dev/null

Run from the project's working directory so Gemini can read files referenced in the prompt. Crossfire reviewers read the project's docs and code via the CLI's filesystem tools — they don't just respond to inlined text.

Flag rationale:

• -p — mandatory for non-interactive use. Without it, Gemini drops into the REPL when stdin is a TTY.
• --output-format text — plain text on stdout. Use --output-format json for {response, stats, error} if the parent flow needs structured output.
• -m flash — pick the fast/cheap model. For higher-quality reviews use -m pro (or whatever the current top-tier model is). Defer to user preference if specified.
• 2>/dev/null — suppresses Gemini's progress/spinner output.

Note: Gemini auto-loads GEMINI.md from the cwd as additional context, similar to Claude's CLAUDE.md. For crossfire this is fine — it's project-scoped context the reviewer should see. If you specifically need a clean run with no project context, run from a temp directory (but you'll lose file-access to the very docs the prompt references — almost never what you want for crossfire).

For long prompts:

gemini -p "$(cat /path/to/prompt.md)" --output-format text -m flash 2>/dev/null

Or pipe stdin:

cat /path/to/prompt.md | gemini -p "Review the following:" --output-format text -m flash 2>/dev/null

Step 4 — Capture and return

Capture stdout. Present back with framing:

Gemini says:

<response>

Optional flags worth knowing

• --approval-mode yolo (or --yolo) — auto-approves any tool calls Gemini wants to make. Only relevant if the prompt encourages tool use; for pure text-in/text-out reviews, prompt phrasing keeps things clean and you don't need this.
• -m <model> — pin a specific model when the user requests one.
• --output-format stream-json — for streaming JSON events. Use only if the parent flow consumes streamed structured output.

Context-leak gotchas

Gemini auto-loads two things from the working directory that can contaminate output:

1. GEMINI.md — analogous to CLAUDE.md. If the cwd has one, it gets loaded as context. For deterministic crossfire, run from a clean directory or a temp dir.
2. .gemini/.env — environment overrides. Loaded from cwd up the directory tree, then ~/.gemini/.env. Variables can leak from project-specific configs.

If you want a guaranteed-clean Gemini call:

TEMP_DIR=$(mktemp -d)
cd "$TEMP_DIR"
gemini -p "$PROMPT" --output-format text -m flash 2>/dev/null
cd - >/dev/null
rm -rf "$TEMP_DIR"

Error handling

Windows / Cowork sandbox note

Same caveat as the Codex skill: the user's Gemini CLI is likely installed on Windows; Cowork's Bash runs in a Linux sandbox. If the sandbox can't reach the host binary, install Gemini CLI inside the sandbox (npm install -g @google/gemini-cli) and pass GEMINI_API_KEY as an env var so auth survives the ephemeral sandbox.