gemini-large-analysis

Large Codebase Analysis with Gemini

Invoke using /gemini-analyze or mcp__gemini__gemini_execute with include_directories set to target paths.

When to Use Gemini vs Claude

Use Gemini when the analysis scope exceeds ~200K tokens — roughly a full repository or multiple large modules. Claude can handle focused analysis of individual files or small groups, but whole-repo audits, dependency trees, and cross-cutting architectural questions benefit from Gemini's 1M-token context window.

Recommended Approach

1. Scope First, Then Send

Before calling gemini_execute, understand what you're sending:

• Use Glob to map the file tree: **/*.ts, **/*.py, etc.
• Use Grep to find specific patterns: entry points, exports, imports
• Estimate the token count (rough: 1 line ≈ 10 tokens)
• If scope exceeds even Gemini's context, plan multiple focused passes

2. Parameter Selection

• model: Always pro — codebase analysis requires deep reasoning
• timeout_ms: 2400000 (40 minutes) — full-repo analysis can take 20-40 minutes
• include_directories: Target specific directories rather than the entire repo when possible. This reduces noise and improves analysis quality.
• max_turns: Increase if you want Gemini to use tools (file reading, etc.)

3. Breaking Into Passes

For truly massive codebases, break into focused passes:

1. Architecture pass: "Map the high-level module structure and dependencies"
2. Pattern pass: "Identify design patterns and anti-patterns in src/"
3. Quality pass: "Find code quality issues, dead code, and tech debt in lib/"
4. Synthesis: Combine findings from each pass

4. Cross-Reference with Local Inspection

Gemini's findings are a starting point. For each finding:

• Verify with local file reads (the code may have changed)
• Check git blame for context on why code is the way it is
• Run relevant tests to confirm suspected issues

Output Format

Structure analysis output as:

1. Key findings — the most important observations
2. Architecture insights — module structure, dependency flow, coupling
3. Issues — with file:line references and severity
4. Recommendations — concrete, actionable next steps