Gemini CLI Integration

Always run through subagents for better iteration and summarization. Make the subagent always ask Gemini to provide evidence, and make the subagent verify the findings before summarizing for the main agent.

NEVER use --approval-mode yolo or --yolo. Gemini must not auto-approve destructive actions.

Quick Reference

# Basic one-shot query (non-interactive mode)
gemini "Explain this codebase architecture"

# Specify model (by alias or full ID)
gemini -m flash "Standard analysis task"
gemini -m pro "Deep reasoning task"
gemini -m gemini-2.5-flash "Use specific GA model"

# Include additional directories in context
gemini --include-directories ./libs,./shared "Analyze dependencies"

# Multi-line prompt using heredoc
gemini <<'__GEMINI_PROMPT__'
Analyze this codebase for security vulnerabilities.
Focus on: SQL injection, XSS, authentication issues
__GEMINI_PROMPT__

# Output formats
gemini -o json "query"         # structured JSON output
gemini -o text "query"         # default
gemini -o stream-json "query"  # real-time JSONL events

Available Models

Model Aliases (recommended for -m flag)

Alias	Resolves to	Use case
auto	gemini-3-pro-preview	Default, auto-routing
pro	gemini-3-pro-preview	Deep reasoning
flash	gemini-3-flash-preview	Fast agentic coding
flash-lite	gemini-2.5-flash-lite	Lowest latency/cost

Full Model IDs

Gemini 3 (Preview):

• gemini-3-pro-preview - Most powerful, deep reasoning
• gemini-3-flash-preview - Best for agentic coding (78% SWE-bench), 1/4 cost of Pro

Gemini 2.5 (GA):

• gemini-2.5-pro - Stable, balanced performance
• gemini-2.5-flash - Stable, fast
• gemini-2.5-flash-lite - Lowest latency and cost

Headless Mode Tool Restrictions

In non-interactive mode (positional query or -p flag), Gemini restricts certain tools by default:

Default	Excluded: shell, edit, write_file, web_fetch
--approval-mode auto_edit	Excluded: shell only

Override specific restrictions with --allowed-tools when needed:

gemini --allowed-tools ShellTool "Run the test suite and report results"

Subagent Delegation Pattern

CRITICAL: Always delegate Gemini tasks to a subagent using the Task tool for multiple retry attempts, better summarization, and error handling.

Pattern: Using Task Tool with Gemini

Task(
  subagent_type: "general-purpose",
  description: "Analyze codebase with Gemini",
  prompt: "Use the gemini-cli skill to analyze /path/to/project.
  Focus on security vulnerabilities and performance issues.
  See references/prompt_patterns.md for effective prompts.
  Ask Gemini to provide evidence for findings.
  Verify findings before summarizing key points in bullet form."
)

Pattern: Second Opinion

Task(
  subagent_type: "general-purpose",
  description: "Get Gemini second opinion",
  prompt: "Use gemini-cli skill to get Gemini's critique of the analysis.
  See references/prompt_patterns.md for second opinion prompts.
  Highlight disagreements or missing considerations."
)

When to Use Gemini vs Claude

Use Gemini for	Use Claude for
Massive codebase review (>100k tokens)	Quick code questions
Cross-validating critical analysis	Small file analysis
Multi-file architecture analysis	Interactive coding assistance
Tasks exceeding Claude's context	Tasks requiring file edits
	Tasks requiring follow-up questions

Additional Resources

• references/prompt_patterns.md - Effective prompt templates for codebase analysis, security audits, performance analysis, second opinions, and context maximization tips

Troubleshooting

Model not available: Gemini 3 models are still in preview. Fall back to Gemini 2.5 models for GA stability.

Slow responses: Large context takes time. Narrow scope with --include-directories or use flash-lite.

Tools not running in headless mode: Default headless excludes write tools. Use --allowed-tools to selectively enable specific tools.