gemini（Gemini 助手）

Gemini CLI

Commands

New conversation (ask)

gemini --model auto "your prompt here" --output-format json 2>/dev/null

Response includes session_id for follow-up:

{
  "session_id": "uuid-here",
  "response": "...",
  "stats": {...}
}

Continue conversation (reply)

gemini --model auto --resume <session_id> -p "follow-up prompt" --output-format json 2>/dev/null

Context from previous turns is preserved.

Examples

# Ask Gemini for code review
result=$(gemini --model auto "Review this function for potential bugs: $(cat src/utils.ts)" --output-format json 2>/dev/null)
session_id=$(echo "$result" | jq -r '.session_id')
echo "$result" | jq -r '.response'

# Follow up on the same session (using session_id from above)
result=$(gemini --model auto --resume "$session_id" -p "How would you refactor it?" --output-format json 2>/dev/null)
session_id=$(echo "$result" | jq -r '.session_id')
echo "$result" | jq -r '.response'

# Leverage Google Search grounding
result=$(gemini --model auto "What are the latest changes in TypeScript 5.8?" --output-format json 2>/dev/null)
echo "$result" | jq -r '.response'

⚠️ CRITICAL: Session ID Handling

If you don't parse session_id, you CANNOT continue the conversation!

For every response, you MUST:

1. Extract session_id from the JSON response
2. Use --resume <session_id> for follow-up questions

❌ WRONG:

gemini "prompt" --output-format json 2>/dev/null | jq -r '.response'
# session_id lost → cannot continue conversation

✅ CORRECT:

result=$(gemini --model auto "prompt" --output-format json 2>/dev/null)
session_id=$(echo "$result" | jq -r '.session_id')
response=$(echo "$result" | jq -r '.response')
# session_id preserved for follow-up

Notes

• Gemini automatically uses google_web_search when needed
• Image analysis: include file path in prompt, Gemini reads via read_file