- name
- gemini
- description
- Interact with Google's Gemini model via CLI. Use when needing a second opinion from another LLM, cross-validation, or leveraging Gemini's Google Search grounding. Supports multi-turn conversations with session management.
Gemini CLI
Commands
Use the wrapper script for streamlined interaction:
# New conversation
~/.agents/skills/gemini/gemini-chat.sh "your prompt here"
# Continue conversation (session_id from previous output)
~/.agents/skills/gemini/gemini-chat.sh --resume <session_id> "follow-up prompt"The script:
- Streams output in real-time (no more waiting blindly)
- Shows retry/error info (429 등)
- Saves raw response to tmpfile for debugging
- Extracts session_id automatically
Examples
# Ask Gemini for code review
~/.agents/skills/gemini/gemini-chat.sh "Review this function for potential bugs: $(cat src/utils.ts)"
# Follow up (use session_id from previous output)
~/.agents/skills/gemini/gemini-chat.sh --resume abc-123 "How would you refactor it?"
# Leverage Google Search grounding
~/.agents/skills/gemini/gemini-chat.sh "What are the latest changes in TypeScript 5.8?"Raw CLI Usage (if needed)
tmpfile=$(mktemp)
echo "Response saved to: $tmpfile"
gemini --model auto-gemini-3 --approval-mode yolo "prompt" --output-format stream-json | tee "$tmpfile"
session_id=$(jq -rs 'last | .session_id' "$tmpfile")Model Selection
Available --model options:
| Value | Description |
|---|---|
auto-gemini-3 | 기본값. gemini-3-pro, gemini-3-flash 중 자동 선택 |
auto-gemini-2.5 | gemini-2.5-pro, gemini-2.5-flash 중 자동 선택 |
gemini-2.5-flash | 직접 지정 (가장 안정적) |
gemini-2.5-pro | 직접 지정 |
Quota Exhausted (429 에러) 대처
Gemini 3 quota 초과 시 (429, MODEL_CAPACITY_EXHAUSTED, RESOURCE_EXHAUSTED):
# auto-gemini-2.5로 fallback (2.5-pro 또는 2.5-flash 자동 선택)
~/.agents/skills/gemini/gemini-chat.sh --model auto-gemini-2.5 "your prompt"Fallback 순서:
auto-gemini-3(기본) → Gemini 3 계열 사용auto-gemini-2.5(fallback) → Gemini 2.5-pro/flash 중 자동 선택
Notes
- Gemini automatically uses
google_web_searchwhen needed - Image analysis: include file path in prompt, Gemini reads via
read_file - Raw response file remains after execution for debugging