返回 Skill 列表
extension
分类: 开发与工程无需 API Key

dev_invoke_gemini-cli

通过markdown文件交接模式将QAQC和审查任务委托给Google Gemini CLI。在REVIEW.md中写入审查请求,Gemini分析后将发现结果输出到FINDINGS.md。可用于代码审查、安全审计、文档审查、大上下文分析等。触发词:gemini, gemini cli, delegate to gemini, gemini subagent, code review, QAQC, quality check, security audit, documentation review, large context, second opinion, architecture review, gemini-3-pro-preview, gemini-3-flash-preview 先决条件:Gemini CLI已认证(gemini登录或GEMINI_API_KEY)模型:gemini-3-pro-preview(默认),gemini-3-flash-preview(大上下文)

person作者: jakexiaohubgithub

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

| Model | Use Case | Status | |-------|----------|--------| | gemini-3-pro-preview | Default. Strong reasoning | CLI default | | gemini-3-flash-preview | Large context, fast analysis | For big codebases |

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

| Flag | Purpose | |------|---------| | -y | YOLO mode - auto-approve all actions | | -m <model> | Model: gemini-3-pro-preview, gemini-3-flash-preview | | -r <id> | Resume session by ID | | --output-format json | JSON output for parsing | | --list-sessions | List available sessions |

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

| Limit | Value | |-------|-------| | Requests/minute | 60 | | Requests/day | 1,000 |

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

| Task | Gemini | Codex | |------|:------:|:-----:| | Code review / QAQC | Best | OK | | Security audit | Best | OK | | Implementation | OK | Best | | Refactoring | OK | Best | | Large codebase analysis | Best | OK | | Documentation review | Best | OK | | Extended thinking (20-30 min) | OK | Best |

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