Invoking Gemini CLI

Invoking Gemini CLI

Delegate QAQC, code review, and analysis tasks to Gemini CLI using markdown files for input and output.

Pattern: Markdown File Handoff

Claude Code                         Gemini CLI
    |                                   |
    +-- Write REVIEW.md ----------------+
    |   (code + review criteria)        |
    |                                   |
    +-- Execute: gemini -y "Read        |
    |   REVIEW.md, analyze, write       |
    |   findings to FINDINGS.md"        |
    |                                   |
    |                                   +-- Reads REVIEW.md
    |                                   +-- Analyzes content
    |                                   +-- Writes FINDINGS.md
    |                                   |
    +-- Read FINDINGS.md <--------------+
    |   (analysis + recommendations)    |
    v                                   v

Benefits:

• No shell escaping issues
• Full context in structured format
• Explicit output structure
• Session resume available

When to Use

• QAQC / Code review - Second opinion on implementation
• Architecture review - Validate design decisions
• Documentation review - Check clarity and completeness
• Security audit - Identify vulnerabilities
• Large context analysis - 1M+ token context window

Model Selection

Recommendation: Use default gemini-3-pro-preview for most tasks. Use gemini-3-flash-preview for very large context (>100K tokens) via -m flag.

Note: If you encounter 429 capacity errors with gemini-3-pro-preview, retry or wait briefly.

Invocation

Standard Pattern (Recommended)

cd /​path/​to/​project && gemini -y "Read REVIEW.md in the current directory. Follow the review criteria. Write all findings to FINDINGS.md."

With Explicit Model

cd /​path/​to/​project && gemini -y -m gemini-3-flash-preview "Read REVIEW.md, analyze per criteria, write findings to FINDINGS.md"

Resume Session

gemini -r <session_id> "Read REVIEW.md for updated criteria, append to FINDINGS.md"

Core Flags Reference

Review Request Template (REVIEW.md)

# Review Request: [Brief Title]

## Objective
[What kind of review/​analysis is needed]

## Review Criteria
Rate each finding: CRITICAL | HIGH | MEDIUM | LOW

### Security
- SQL injection vulnerabilities
- XSS risks
- Authentication/​authorization issues
- Hardcoded secrets

### Reliability
- Error handling gaps
- Edge cases not covered
- Race conditions
- Resource leaks

### Performance
- N+1 query patterns
- Unnecessary allocations
- Missing caching opportunities

### Maintainability
- Code clarity
- Naming conventions
- Documentation gaps

## Code to Review

### File: src/​api/​users.ts
```typescript
// Paste code here or reference file

File: src/​services/​auth.ts

// Paste code here

Context

[Any relevant background, constraints, or requirements]

Output Format

Write to FINDINGS.md with:

• Summary of review
• Findings table with severity, location, description
• Specific recommendations
• Overall assessment
• Session ID for follow-up

## Findings Template (FINDINGS.md)

Gemini should produce:

```markdown
# Review Findings: [Title]

## Summary
[Brief overview of review results]

## Findings

| Severity | Location | Issue | Recommendation |
|----------|----------|-------|----------------|
| CRITICAL | users.ts:45 | SQL injection via unsanitized input | Use parameterized queries |
| HIGH | auth.ts:23 | Missing rate limiting on login | Add rate limiter middleware |
| MEDIUM | users.ts:78 | No input validation | Add schema validation |
| LOW | auth.ts:12 | Magic number | Extract to named constant |

## Detailed Analysis

### CRITICAL: SQL Injection (users.ts:45)
[Detailed explanation with code example]

### HIGH: Missing Rate Limiting (auth.ts:23)
[Detailed explanation with recommendation]

## Recommendations
1. [Priority action items]
2. [Secondary improvements]

## Overall Assessment
[Summary judgment: PASS | NEEDS WORK | FAIL]

## Session
Session ID: `<id>` (for follow-up questions)

Workflow Example (ras-commander)

1. Write REVIEW.md

# Review Request: Remote Execution Security

## Objective
Security review of the PsExec-based remote execution module.

## Review Criteria
Rate findings: CRITICAL | HIGH | MEDIUM | LOW

Focus on:
- Command injection vulnerabilities
- Credential exposure
- Path traversal risks
- Session hijacking

## Code to Review

### File: ras_commander/​remote/​PsexecWorker.py
```python
[paste relevant code]

File: ras_commander/​remote/​Execution.py

[paste relevant code]

Context

• This module executes HEC-RAS remotely via PsExec
• Runs with elevated privileges on remote Windows machines
• Handles user credentials for authentication

Output Format

Write to FINDINGS.md with severity-rated findings and specific mitigations.

### 2. Execute Gemini

```bash
cd "C:/​GH/​ras-commander" && gemini -y "Read REVIEW.md, perform security review per criteria, write findings to FINDINGS.md"

3. Read FINDINGS.md

Review findings, address issues, continue development.

Environment Variables

GEMINI_API_KEY=xxx         # Required (or use OAuth login)
GOOGLE_API_KEY=xxx         # Alternative (takes precedence)

Rate Limits

Session Management

• Session ID appears in Gemini output
• Use gemini --list-sessions to see available sessions
• Resume with gemini -r <id> "follow-up instruction"

Tips

1. Paste actual code - Don't describe, include the real code in REVIEW.md
2. Define severity levels - CRITICAL/​HIGH/​MEDIUM/​LOW helps prioritization
3. Specify output structure - Tell Gemini exactly what FINDINGS.md should contain
4. Provide context - Domain, compliance requirements, constraints
5. Ask focused questions - One review type per request gets better results

Decision Matrix: Gemini vs Codex

When to Escalate

Use Gemini for:

• Quick code reviews
• Documentation checks
• Large context analysis
• Security audits (initial scan)

Use Codex for:

• Complex implementation
• Extended thinking tasks
• Multi-file refactoring

Use specialized ras-commander agents for:

• HDF analysis -> hdf-analyst
• Geometry parsing -> geometry-parser
• USGS integration -> usgs-integrator

---

See Also:

• code-oracle-gemini agent - Full orchestration capabilities
• .claude/​rules/​subagent-output-pattern.md - Output format standards